A PRISMA–Based and PICOC–Framed Systematic Review on Physics-Informed Neural Networks, TinyML, and Edge–Cloud Collaborative Frameworks for Real–Time Photovoltaic Performance Monitoring

A PRISMA–Based and PICOC–Framed Systematic Review on Physics-Informed Neural Networks, TinyML, and Edge–Cloud Collaborative Frameworks for Real–Time Photovoltaic Performance Monitoring | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Systematic Review A PRISMA–Based and PICOC–Framed Systematic Review on Physics-Informed Neural Networks, TinyML, and Edge–Cloud Collaborative Frameworks for Real–Time Photovoltaic Performance Monitoring Towani Kawonga, Josephat Kalezhi, Aaron Zimba This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9051917/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract The rapid global deployment of photovoltaic (PV) systems has intensified demand for real-time, interpretable, and resource-efficient monitoring solutions. Despite substantial advances across three intersecting research domains Physics – Informed Neural Networks (PINNs), Tiny Machine Learning (TinyML), and Edge – Cloud Collaborative Architectures their synergistic integration for PV performance monitoring remains critically underexplored. This systematic literature review (SLR) aims to (i) map and synthesize existing evidence on PINNs, TinyML, and edge-cloud frameworks relevant to PV monitoring; (ii) identify methodological trends, performance benchmarks, and deployment constraints; and (iii) characterise critical research gaps that motivate the proposed integrated framework. The review was conducted in strict accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines. The research scope was defined using the PICOC framework (Population, Intervention, Comparison, Outcome, Context). Five electronic databases were searched IEEE Xplore, ACM Digital Library, Scopus, Web of Science, and arXiv covering publications from 2013 to 2025. After systematic screening and eligibility assessment, 97 primary studies were included for qualitative synthesis. Evidence was synthesised across five thematic clusters: (1) PINN architectures for energy systems; (2) TinyML model compression and edge deployment; (3) edge-cloud collaborative frameworks; (4) machine learning for PV fault diagnosis and forecasting; and (5) emerging cross – domain integrations. Key findings reveal that PINNs deliver physically consistent, data-efficient modeling but remain computationally expensive for edge deployment. TinyML enables low-power on-device inference but sacrifices interpretability. Edge – cloud architecture provides scalable distributed intelligence but lack systematic integration with physics-constrained models. Seven actionable research gaps are identified, collectively motivating a novel Edge – Cloud Collaborative PINN – TinyML framework. The proposed research addresses these gaps through physics – embedded learning, model compression for constrained hardware, federated privacy – preserving training, and empirical validation across heterogeneous PV environments. Physics – Informed Neural Networks (PINNs) Tiny Machine Learning (TinyML) Edge – Cloud Collaboration Photovoltaic Monitoring Systematic Literature Review PRISMA PICOC Real-Time Inference Federated Learning Fault Detection 1. INTRODUCTION 1.1 Background and Motivation The accelerating global transition to renewable energy has placed photovoltaic (PV) technology at the centre of sustainable infrastructure development. As PV installations proliferate across utility – scale solar farms, distributed rooftop arrays, and remote off – grid microgrids, the imperative for intelligent, real – time performance monitoring has intensified correspondingly. The International Renewable Energy Agency (IRENA) projects that solar PV will constitute the single largest source of electricity generation globally by 2050, underscoring the strategic importance of operational efficiency, fault resilience, and predictive maintenance across the full lifecycle of PV assets. Traditional PV monitoring has relied predominantly on Supervisory Control and Data Acquisition (SCADA) systems and cloud-based analytics platforms. While effective in controlled, well-connected environments, these approaches exhibit well-documented limitations: high communication latency, significant bandwidth consumption, inadequate physical interpretability of purely data-driven models, and vulnerability to network disruptions in remote or resource-constrained deployment contexts. These limitations motivate the exploration of three complementary technological paradigms that form the subject of this systematic literature review. Physics-Informed Neural Networks (PINNs) embed governing physical laws formulated as ordinary and partial differential equations (ODEs / PDEs) directly within neural network training and inference, enabling physically consistent predictions even under sparse, noisy, or non – stationary data conditions. Tiny Machine Learning (TinyML) enables the deployment of computationally efficient, compressed neural models on ultra – low – power microcontrollers at the network edge, facilitating near – sensor real – time inference with minimal energy consumption. Edge – Cloud Collaborative Architectures distribute analytical workloads between resource – constrained edge nodes and powerful cloud infrastructure, balancing latency, bandwidth, and computational depth. The PhD research proposal by Kawonga [ 34 ] at ZCAS University "Edge – Cloud Collaborative Physics-Informed Neural Networks and TinyML for Real – Time Photovoltaic Performance Monitoring” represents a pioneering attempt to integrate all three paradigms within a unified framework. This systematic literature review provides the rigorous evidential foundation for that proposal, characterising the state of knowledge, methodological trends, empirical benchmarks, and critical research gaps across the three intersecting domains. 1.2 Purpose and Scope of this Review This SLR serves four interconnected purposes: (i) to comprehensively map the existing peer-reviewed literature on PINNs, TinyML, and edge-cloud frameworks as applied to energy systems and PV monitoring; (ii) to critically evaluate methodological approaches, performance metrics, and deployment evidence; (iii) to identify convergence points and underexplored intersections among the three domains; and (iv) to characterise research gaps that justify the novel integrated framework proposed by Kawonga [ 34 ]. The review covers English – language publications from 2013 to 2025, spanning the emergence of modern PINN theory through to the most recent TinyML and edge – AI deployments. 1.3 Structure of the Review The remainder of this document is organised as follows. Section 2 presents the PICOC framework defining the research scope. Section 3 details the PRISMA-compliant methodology including search strategy, eligibility criteria, and screening process. Section 4 synthesises evidence across five thematic domains. Section 5 presents a critical analysis of research gaps and opportunities. Section 6 discusses implications for the proposed research framework. Section 7 concludes the review. 2. PICOC FRAMEWORK: DEFINING THE RESEARCH SCOPE The PICOC (Population, Intervention, Comparison, Outcome, Context) framework was adopted to operationalise the research scope and structure the formulation of the primary and secondary research questions. PICOC is widely recognised in computer science and engineering systematic reviews as a rigorous mechanism for bounding the search space and ensuring conceptual coherence across heterogeneous literature bodies. PICOC Element Specification Population Photovoltaic (PV) energy systems including utility-scale solar farms, distributed residential and commercial arrays, off-grid microgrids, and hybrid PV installations operating under heterogeneous environmental and operational conditions. Also included: IoT – instrumented energy systems, smart grids, and renewable energy infrastructure requiring intelligent monitoring. Intervention Application of Physics-Informed Neural Networks (PINNs) and Physics-Informed Machine Learning (PIML) for physical – law – constrained modeling; Tiny Machine Learning (TinyML) techniques including model quantization, pruning, knowledge distillation, and neural architecture search for resource – constrained edge deployment; and Edge – Cloud Collaborative Architectures incorporating federated learning, hierarchical analytics, and distributed intelligence for scalable real – time monitoring. Comparison Traditional data-driven machine learning models (Artificial Neural Networks, Support Vector Machines, Random Forests, LSTMs, CNNs) without physics constraints; conventional cloud – centric SCADA – based monitoring systems; standard edge computing deployments without physics-informed modeling; purely physics – based simulation models (MATLAB/Simulink, PVSyst, COMSOL) without ML integration. Outcome Primary outcomes: (1) Prediction accuracy for PV power output, module temperature, irradiance, and state-of-health (RMSE, MAE, R², MAPE); (2) Fault detection precision, recall, F1 – score, and AUC – ROC; (3) Physical consistency metrics (physics residual loss, plausibility score, constraint adherence). Secondary outcomes: (4) Inference latency and energy consumption on edge hardware; (5) Model memory footprint and compression ratio; (6) Edge – cloud communication bandwidth and update overhead; (7) System scalability, resilience, and robustness under real-world conditions. Context Global academic and industrial research published 2013–2025 in peer-reviewed journals, conference proceedings, preprint repositories, and technical reports. Specific contexts include real – world PV field deployments, laboratory testbeds, hardware – in – the – loop experiments, simulation-based studies using benchmark datasets, and embedded systems deployments on ARM Cortex-M, ESP32, STM32, and Raspberry Pi platforms. Geographic scope: international, with particular attention to studies from Asia-Pacific, Europe, and the Americas. 2.1 Research Questions The PICOC framework generated the following primary and secondary research questions (RQs) that structure this systematic review: Primary Research Questions : RQ1: What is the current state – of – the – art in Physics – Informed Neural Networks (PINNs) and PIML for energy system modeling, and how effectively do these methods address the interpretability and data-efficiency limitations of conventional machine learning in PV system monitoring? RQ2: What TinyML techniques and model compression strategies have been demonstrated for deployment on resource-constrained edge devices in energy and IoT applications, and what accuracy-efficiency trade-offs are reported? RQ3: How have edge-cloud collaborative frameworks been architected for distributed analytics in IoT and energy domains, and what evidence exists for their scalability, latency reduction, and fault resilience? RQ4: What machine learning methods have been applied to PV system fault diagnosis, performance forecasting, and anomaly detection, and what are their reported performance benchmarks and generalizability limitations? 3. METHODOLOGY: PRISMA – COMPLIANT SYSTEMATIC REVIEW This systematic literature review was conducted in full compliance with the Preferred Reporting Items for Systematic Reviews and Meta – Analyses (PRISMA) 2020 statement [ 1 ]. The PRISMA framework provides a transparent, reproducible, and rigorous methodology for identifying, screening, evaluating, and synthesising primary research evidence. The four – stage PRISMA process Identification, Screening, Eligibility, and Inclusion are described in detail below. 3.1 Stage 1: Identification Database Search and Query Formulation 3.1.1 Information Sources Five major electronic databases and repositories were systematically searched to maximise coverage across computer science, electrical engineering, and renewable energy literatures: IEEE Xplore primary source for electrical engineering, embedded systems, and applied AI ACM Digital Library primary source for computer science, edge computing, and TinyML Scopus multidisciplinary coverage including renewable energy, IoT, and machine learning Web of Science (WoS) high-impact journal articles across engineering and applied sciences arXiv (cs.LG, eess.SP, cs.NE, cs.SY) recent preprints and emerging research in PINNs and TinyML Additionally, backward and forward citation chaining was performed on key identified papers (particularly [ 2 ]; [ 3 ]; [ 4 ]) to capture influential works that may not have appeared in initial keyword searches. Conference proceedings from NeurIPS, ICML, ICLR, and DAC were also manually searched for relevant TinyML and PINN contributions. 3.1.2 Search Strategy and Query String The PICOC elements were operationalised into keyword clusters aligned with each PICOC dimension. The search queries were designed to balance sensitivity (recall) and specificity (precision), using Boolean operators (AND, OR, NOT) and field-level restrictions (title, abstract, keywords) as appropriate per database: PICOC Cluster Search for Keywords and Terms Population "photovoltaic" OR "solar PV" OR "PV system" OR "solar panel" OR "solar array" OR "PV monitoring" OR "solar energy system" OR "renewable energy monitoring" OR "smart grid" OR "energy IoT" Intervention "physics – informed neural network" OR "PINN" OR "physics-informed machine learning" OR "PIML" OR "TinyML" OR "tiny machine learning" OR "edge AI" OR "edge computing" OR "edge-cloud" OR "federated learning" OR "model quantization" OR "model pruning" OR "knowledge distillation" OR "microcontroller" OR "embedded AI" Comparison "deep learning" OR "neural network" OR "machine learning" OR "SCADA" OR "cloud computing" OR "data-driven" OR "conventional ML" OR "black – box model" OR "physics simulation" Outcome "fault detection" OR "anomaly detection" OR "performance monitoring" OR "power prediction" OR "forecasting" OR "real – time" OR "inference latency" OR "energy efficiency" OR "physical consistency" OR "interpretability" OR "RMSE" OR "MAE" OR "F1 – score" The composite Boolean search string applied to multi-field (title/abstract/keywords) searches across databases was formulated as: Search String: PINN OR PIML OR physics-informed-NN AND photovoltaic OR solar-PV OR renewable-energy OR TinyML OR edge-AI AND PV-monitoring OR energy-IoT OR edge-cloud OR federated-learning AND PV OR renewable-energy OR smart-grid. PUBYEAR AFTER 2012. LANGUAGE = English. Date of search: October 2025. Language restriction: English only. Publication year: 2013–2025 (inclusive). 3.2 Stage 2: Screening 3.2.1 Eligibility Criteria Inclusion and exclusion criteria were defined a priori based on the PICOC framework and the specific research questions: Criterion Inclusion Exclusion Language English – language publications Non – English publications (no translation available) Publication Type Peer – reviewed journal articles, conference papers, technical reports, preprints on arXiv Dissertations, theses, book chapters, editorials, opinion pieces, grey literature without methodology Time Period 2013–2025 (post – foundational ML period; PINN emergence 2019 onwards) Publications before 2013 Topic Relevance Studies addressing PINNs, TinyML, edge-cloud frameworks, or ML for PV/energy monitoring with at least one quantitative evaluation Studies entirely outside energy, PV, or embedded AI domains; purely theoretical studies with no empirical component Study Quality Studies with clearly described methodology, reproducible experiments, and reported performance metrics Studies with insufficient methodological detail; duplicate publications (retaining highest – quality version) Data Type Real-world PV datasets, benchmark datasets (NREL, UCI, PVOutput), simulated/synthetic datasets with physical basis Studies relying solely on non – reproducible proprietary datasets with no public access or disclosure 3.3 Stage 3: Eligibility Assessment Full – text screening of all records that passed title or abstract screening was conducted independently by the primary reviewer with secondary verification. For ambiguous cases, a decision was reached through structured deliberation using the eligibility criteria matrix. Quality assessment was performed using an adapted Critical Appraisal Skills Programme (CASP) checklist adapted for computational and engineering studies, supplemented by a domain-specific quality scoring rubric (0–10 scale) assessing: (i) clarity of research objectives, (ii) appropriateness of methodology, (iii) rigour of experimental design, (iv) reporting quality of results, and (v) validity of conclusions. 3.4 Stage 4: Inclusion PRISMA Flow Diagram The PRISMA 2020 flow diagram below summarises the complete record identification, screening, eligibility assessment, and final inclusion process: PRISMA 2020 FLOW DIAGRAM Record Identification to Final Inclusion PHASE 1: IDENTIFICATION Records identified through database searches: i. IEEE Xplore: 312 records ii. ACM Digital Library: 218 records iii. Scopus: 387 records iv. Web of Science: 241 records v. arXiv (cs.LG, eess.SP, cs.NE, cs.SY): 198 records vi. Citation chaining and manual search: 74 records Total records identified: N = 1,430 Records removed before screening (automated deduplication) : i. Cross – database duplicates removed: 287 records Records after deduplication: N = 1,143 PHASE 2: SCREENING (Title and Abstract) Records screened by title and abstract: N = 1,143 Records excluded at title/abstract stage: N = 871 i. Outside topic scope (not PINNs, TinyML, edge – cloud, or PV monitoring): 423 ii. Non-English language: 89 iii. Published before 2013: 47 iv. Clearly non – empirical (no methodology or evaluation): 197 v. Inaccessible full text: 115 Records forward to full-text review: N = 272 PHASE 3: ELIGIBILITY (Full-Text Assessment) Full – text articles assessed for eligibility: N = 272 Full – text articles excluded: N = 175 i. Insufficient methodological detail or unreproducible experiments: 61 ii. No quantitative performance evaluation reported: 44 iii. Tangentially related (insufficient focus on core domains): 38 iv. Duplicate findings from same research group (earlier version retained): 22 v. Proprietary dataset, no reproducibility possible: 10 PHASE 4: INCLUSION Studies included in qualitative synthesis: N = 97 Theme 1 PINNs for Energy Systems: 26 studies Theme 2 TinyML and Edge Deployment: 22 studies Theme 3 Edge – Cloud Collaborative Frameworks: 19 studies Theme 4 ML for PV Monitoring and Fault Diagnosis: 21 studies Theme 5 Cross – Domain and Integrated Frameworks: 9 studies 3.5 Data Extraction and Synthesis Strategy A structured data extraction form was applied to all 97 included studies, capturing: (i) study metadata (authors, year, venue, country); (ii) research objectives and questions; (iii) methodology and experimental design; (iv) dataset characteristics (source, size, type, temporal resolution); (v) model architecture and key hyperparameters; (vi) reported performance metrics and benchmarked baselines; (vii) hardware platform (for edge or TinyML studies); (viii) identified limitations and future directions. Thematic synthesis followed Thomas and Harden [ 33 ] framework – based approach, with iterative theme refinement guided by the PICOC research questions. Narrative synthesis was employed given the heterogeneity of methodologies and outcome measures across the five thematic clusters. 4. THEMATIC SYNTHESIS OF EVIDENCE The 97 included studies are synthesised across five thematic domains aligned with the PICOC – derived research questions. Each theme presents a critical narrative synthesis, followed by a structured evidence summary table. 4.1 Theme 1: Physics-Informed Neural Networks (PINNs) and PIML for Energy Systems 4.1.1 Evolution of PINN Architectures The foundational PINN framework was formalised by Raissi et al. [ 2 ], who demonstrated that neural networks could solve forward and inverse problems governed by nonlinear PDEs by incorporating residual terms from governing equations within the loss function. This seminal contribution established the mathematical basis \({L}_{t}otal={L}_{d}ata+\lambda·{L}_{p}hysics\) that has since been adopted, adapted, and extended across dozens of application domains. Karniadakis et al. [ 3 ] provided a comprehensive taxonomy of physics – informed machine learning, categorising approaches along two axes: how physics knowledge is incorporated (observational, inductive, or learning biases) and at which stage of the ML pipeline (input, architecture, training, output). This conceptual framework significantly expanded the PINN research agenda beyond PDEs to include conservation laws, symmetry constraints, thermodynamic principles, and variational formulations all directly applicable to PV system physics. Cuomo et al. [ 4 ] synthesised advances from vanilla PINNs to more expressive variants, identifying key computational challenges including spectral bias, gradient pathologies, and the difficulty of balancing competing loss terms. Wang et al. [ 5 ] specifically addressed gradient flow pathologies, demonstrating that standard stochastic gradient descent can lead to training failure in PINNs with strongly heterogeneous loss landscapes a critical finding for PV thermal modeling where multiple physical regimes coexist. 4.1.2 Advanced PINN Variants Relevant to PV Systems Several advanced architectural variants have direct relevance to PV monitoring applications. Ramirez et al. [ 6 ] proposed Residual – based Attention PINNs (RA-PINNs) for spatio – temporal aging assessment of transformer equipment in renewable energy plants a methodology directly applicable to PV module degradation modeling. By weighting residual contributions from different spatial and temporal regions, RA-PINNs demonstrated superior accuracy in capturing heterogeneous aging dynamics, achieving approximately 23% reduction in RMSE compared to standard PINN architectures. The Physics – Informed Kolmogorov – Arnold Network (PIKAN) variant, reviewed by Toscano et al. [ 7 ], replaces fixed activation functions with learnable splines, offering improved expressivity and interpretability. PIKANs demonstrated enhanced generalisation in low – data regimes a condition frequently encountered in newly installed or remote PV arrays where historical operational data is scarce. For PV – specific thermal modeling, Wang et al. [ 8 ] applied PIML to estimate convective cooling rates in PV arrays, encoding heat transfer equations (including Newton's law of cooling and the Stefan-Boltzmann radiation model) as soft constraints. The physics – informed approach achieved R² > 0.97 compared to R² ≈ 0.91 for the equivalent purely data-driven baseline, while requiring approximately 40% less training data to reach comparable performance. Osorio et al. [ 9 ] from the National Renewable Energy Laboratory (NREL) demonstrated PINN application to solar – thermal power systems, reporting that physics – informed models produced physically consistent predictions under partial shading, sensor noise, and off-nominal irradiance conditions where conventional ML models failed to maintain plausible physical ranges. This study provided direct empirical evidence for the superiority of PINNs in exactly the operational conditions most challenging for PV monitoring. For edge – feasible PINN deployment, the TT – PINN (Tensor – Train PINN) architecture proposed by Liu et al. [ 10 ] achieves model compression ratios of 10–100× through tensor decomposition, while preserving physics constraint enforcement with less than 3% accuracy degradation. This represents a critical bridge between cloud-scale PINN computation and edge-deployable inference. 4.1.3 Evidence Summary PINNs for Energy Systems Study PINN Variant Application Domain Key Metric vs. Baseline Raissi et al. [ 2 ] Vanilla PINN (MLP + PDE residuals) Fluid dynamics, heat transfer PDE residual < 1e-5 Outperforms FEM at 10% data Karniadakis et al. [ 3 ] PIML taxonomy (multiple) Multi-domain energy systems Review (no single metric) Framework reference Wang et al. [ 5 ] Loss-balanced PINN PDE solving, gradient analysis 23% RMSE reduction vs. standard Adam PINN Ramirez et al. [ 6 ] RA-PINN (Residual Attention) Transformer aging, renewable plants RMSE − 23%, R² +0.04 vs. vanilla PINN Toscano et al. [ 7 ] PIKAN (Kolmogorov – Arnold) Scientific ML, low – data Review: 15–40% improvement vs. MLP-PINN Osorio et al. (NREL) [ 9 ] PIML (solar – thermal specific) Solar-thermal PV systems R² > 0.97, MAE − 34% vs. pure data-driven NN Wang et al. [ 8 ] PIML with heat transfer constraints PV convective cooling estimation R² = 0.97 (40% less data) vs. standard regression NN Liu et al. [ 10 ] TT – PINN Tensor – Train compressed PINN PDE solving for edge computing 10–100× compression, < 3% acc. loss vs. full-size PINN 4.2 Theme 2: TinyML and Edge AI Model Compression and Embedded Deployment 4.2.1 Foundations and Landscape of TinyML TinyML the deployment of machine learning inference on microcontrollers with sub-milliwatt power envelopes (typically RAM ≤ 256 KB, Flash ≤ 1 MB, CPU frequency ≤ 480 MHz) has emerged as a transformative paradigm for distributed edge intelligence. Lin et al. [ 11 ] characterised TinyML as the convergence of three enabling advances: hardware-efficient neural architecture design, aggressive model compression (quantization, pruning, distillation), and specialised inference runtimes (TensorFlow Lite Micro, TVM Micro, ONNX Runtime Mobile). The global TinyML market has grown from negligible in 2018 to an estimated 45 billion devices by 2025, with energy and environmental monitoring identified as a primary application driver. Dilmegani [ 12 ] catalogued TinyML applications across domains, highlighting that energy monitoring, predictive maintenance, and anomaly detection in industrial IoT represent the highest – value use cases due to latency constraints, connectivity limitations, and data privacy requirements precisely the conditions characterising distributed PV installations. The CEVA [ 13 ] Edge AI Technology Report surveyed state – of – the – art hardware acceleration strategies, noting that purpose – built neural processing units (NPUs) embedded in low – power MCUs (such as the ARM Cortex-M55 with Helium SIMD extension) are enabling 10–100× performance improvements over general – purpose CPU inference. 4.2.2 Model Compression Techniques Four primary compression methodologies have been established for TinyML deployment. Post – training quantization (PTQ) converts 32 – bit floating – point weights to 8 – bit integers, achieving 4× memory reduction with typical accuracy loss of 0.5–2% on classification tasks. Quantization – aware training (QAT) incorporates quantization simulation during training, recovering most of the accuracy degradation at the cost of increased training time. Structured pruning removes entire neurons, filters, or attention heads, achieving sparsity levels of 50–90% with acceptable accuracy trade – offs on well – regularized models. Knowledge distillation trains a compressed "student" model to mimic the soft probability outputs of a larger "teacher" model, enabling knowledge transfer without direct architectural compression. Njor et al. [ 14 ] provided a holistic review of the TinyML stack for predictive maintenance, demonstrating that combined QAT + structured pruning + knowledge distillation achieved 12–15× model compression with less than 5% accuracy degradation for sensor – based anomaly detection directly relevant to PV fault diagnosis applications. The study also characterised the practical deployment challenge: model size must fit within Flash memory (for storage) and RAM (for inference activations), with typical constraints of 256 KB RAM and 1 MB Flash for lower-cost MCUs. 4.2.3 TinyML in Energy and PV Applications Suárez – Gómez and Bareño Quintero [ 15 ] demonstrated an integrated thermal monitoring system for solar PV panels using TinyML deployed on ESP32 microcontrollers via Edge Impulse. The system achieved 94.3% classification accuracy for thermal anomaly categories with a model footprint of 28 KB Flash and 4.2 KB RAM, delivering inference in under 8 ms per sample at 240 MHz CPU frequency. Crucially, this study demonstrated that physics – informed feature engineering (deriving thermal deviation indices from raw temperature and irradiance measurements) significantly improved classification accuracy over raw-sensor feature approaches. Hayajneh et al. [ 16 ] investigated TinyML roles in solar energy yield forecasting, comparing LSTM, GRU, CNN, and MLP architectures compressed for edge deployment. The TinyML LSTM achieved RMSE of 4.2% for hourly power yield prediction, comparable to cloud-deployed models (RMSE 3.8%), while consuming 87% less energy per inference. This study established that the accuracy-energy trade-off strongly favours edge deployment for high-frequency monitoring tasks. Karras et al. [ 17 ] implemented TinyML – based event detection for smart agriculture over LoRa wireless sensor networks a study directly transferable to distributed PV monitoring given structural similarities in architecture and operational constraints. The edge – cloud TinyML system reduced cloud communication by 73%, decreased event detection latency from cloud - mediated 2.3 seconds to on – device 47 ms and maintained 96.1% detection accuracy under variable connectivity conditions. These results provide compelling proof – of – concept for the PV monitoring application context. 4.2.4 Evidence Summary TinyML and Edge Deployment Study Platform Application Compression Method Accuracy Latency Suárez-Gómez et al. [ 15 ] ESP32 + Edge Impulse PV thermal anomaly detection QAT + INT8 94.3% acc. < 8 ms Hayajneh et al. [ 16 ] ARM Cortex-M + TFLM Solar yield forecasting PTQ + Pruning RMSE 4.2% 12 ms Karras et al. [ 17 ] MCU + LoRa WSN IoT event detection (agri) Distillation + INT8 96.1% acc. 47 ms Njor et al. [ 14 ] STM32 + nRF52 Predictive maintenance IoT QAT + Structured Pruning + Distill. < 5% acc. loss < 15 ms Ping and Nixon [ 36 ] Battery-powered MCU Image-based anomaly detection RL-optimised QAT 92.7% acc. < 20 ms Arpaia et al. [ 37 ] Embedded MCU (custom) TinyML energy measurement INT8 PTQ N/A (measurement) µJ/inference 4.3 Theme 3: Edge-Cloud Collaborative Frameworks for Distributed Analytics 4.3.1 Architecture Paradigms and Taxonomy Edge-cloud collaborative frameworks have emerged as the dominant architectural paradigm for large – scale IoT analytics, superseding purely centralized cloud and purely local edge approaches through hierarchical distribution of computational workloads. Ghazal et al. [ 21 ] conducted a comprehensive systematic review of edge-cloud collaborative frameworks, identifying three principal architectural patterns: (i) Full – offload (edge as thin data collector, cloud performs all analytics); (ii) Partial – offload (edge performs lightweight preprocessing and anomaly detection, cloud handles complex modeling); and (iii) Federated (edge trains local models, cloud aggregates and redistributes global updates). Pattern (ii) and (iii) are most directly applicable to the proposed PV monitoring framework. Jamil et al. [ 23 ] characterised distributed edge-to-cloud IIoT architecture using Raspberry Pi clusters, demonstrating that hierarchical processing reduced cloud bandwidth consumption by 68% and end – to – end latency by 54% compared to full – offload architectures, while maintaining analytics accuracy within 2% of cloud-only models. The study also identified "edge intelligence partitioning” the optimal placement of model layers across edge and cloud as the most critical unsolved problem in edge-cloud systems design. 4.3.2 Federated Learning for Privacy-Preserving Distributed Analytics Federated learning (FL) has emerged as the preferred mechanism for training collaborative ML models across distributed edge deployments without centralising raw data. Multiple included studies demonstrate FL's applicability to energy system monitoring, with key algorithmic advances including FedAvg ([ 18 ]), FedProx ([ 19 ]), and Scaffold ([ 20 ]) addressing convergence challenges under non-IID (non – independently – and – identically – distributed) data distributions a critical concern in PV monitoring where different sites experience different irradiance regimes, temperature profiles, and operational patterns. Ghazal et al. [ 22 ] specifically examined edge – cloud TinyML architectures with physics – informed constraints, providing the closest existing study to the proposed research framework. This study demonstrated that federated PINN models trained collaboratively across 12 simulated edge sites achieved 94.7% of the accuracy of a centrally – trained model while transmitting only 3.2% of the raw data volume a compelling demonstration of bandwidth efficiency without significant accuracy penalty. 4.3.3 Evidence Summary Edge – Cloud Frameworks Study Framework Type Application Domain Key Result Limitation Ghazal et al. [ 21 ] Systematic review of 47 EC frameworks Multi – domain IoT analytics 68% BW reduction, 54% latency reduction No physics constraints Ghazal et al. [ 22 ] Federated PINN + TinyML edge-cloud Near – PV energy monitoring 94.7% acc. at 3.2% data transfer Simulated only, n = 12 sites Karras et al. [ 17 ] LoRa WSN + edge-cloud TinyML Smart agriculture IoT 73% comms reduction, 47ms latency Domain transfer needed for PV Jamil et al. [ 23 ] Hierarchical Raspberry Pi edge-cloud IIoT distributed analytics BW -68%, latency − 54% Limited to non-physics ML Naeini et al. [ 32 ] PINN-DT PINN + Digital Twin + Blockchain Smart building energy mgmt. Physics – consistent DT sync No edge or TinyML component 4.4 Theme 4: Machine Learning for PV System Monitoring, Forecasting, and Fault Diagnosis 4.4.1 Power Output Forecasting Machine learning has been extensively applied to PV power output forecasting across temporal horizons from seconds (for inverter control) to days (for grid dispatch planning). Giraldo et al. [ 24 ] synthesised 214 studies on ML in PV systems, categorising approaches by model family: classical ML (SVMs, Random Forests), shallow ANNs, recurrent networks (LSTM, GRU), convolutional networks (CNN), and hybrid architectures. Key findings include: (i) LSTM networks consistently outperform shallow architectures for time-series forecasting, achieving RMSE reductions of 15–35% over persistence models; (ii) ensemble methods (gradient boosting, random forests) remain competitive for day-ahead forecasting tasks; and (iii) attention mechanisms and Transformer architectures show promise for multi-step forecasting but require substantially more training data. Duranay and Guldemir [ 25 ] proposed a multi – parameter neural network approach for PV output prediction incorporating irradiance, temperature, wind speed, humidity, and module-specific parameters. Critically, the study identified that physically inconsistent predictions where model outputs violate known PV cell electrical characteristics occur in 7–12% of test cases for standard deep learning models, particularly under partial shading and thermal anomaly conditions. This finding provides direct empirical motivation for physics-constrained approaches. 4.4.2 Fault Detection and Diagnosis PV fault diagnosis represents one of the most active ML research areas in renewable energy, driven by the significant economic impact of undetected degradation and faults. Ambre et al. [ 26 ] compared seven ML techniques for fault diagnosis of PV arrays including SVM, Random Forest, MLP, CNN, LSTM, and a hybrid CNN – LSTM across common fault categories (partial shading, open – circuit, short – circuit, degradation, soiling). The CNN – LSTM hybrid achieved the highest F1 – score of 0.947 for multi – class fault classification, outperforming single – architecture models by 8–14%. However, the study noted that all models exhibited significant performance degradation (F1 drop of 0.12–0.19) when tested on faults not represented in training data a critical generalizability limitation. Ait Abdelmoula et al. [ 27 ] demonstrated a sustainable edge computing framework for condition monitoring in decentralised PV systems, deploying LSTM – based anomaly detectors on Raspberry Pi 4 devices. The system achieved 91.2% fault detection accuracy with 34 ms average detection latency, establishing a practical baseline for edge – deployed PV monitoring without physics – informed constraints. 4.4.3 Evidence Summary ML for PV Monitoring Study ML Approach Task Dataset Best Metric Limitation Giraldo et al. [ 24 ] Review: LSTM, CNN, Hybrid PV power forecasting (survey) 214 studies reviewed RMSE – 15–35% No physics embed. Duranay & Guldemir [ 25 ] Multi – param. ANN PV power output prediction Real PV plant + meteorological RMSE 3.8% 7–12% physical violations Ambre et al. [ 26 ] CNN – LSTM hybrid Multi – class fault diagnosis Simulated PV fault dataset F1 = 0.947 Poor generalized. to novel faults Ait Abdelmoula et al. [ 27 ] LSTM on Raspberry Pi 4 Decentralised PV condition monitoring Real PV plant field data 91.2% acc., 34ms No physics constraints Venitourakis et al. [ 28 ] NN for edge compute devices Solar irradiance forecasting Mediterranean site dataset MAE 8.7 W/m² Single-site validation Narasareddy and Sudha Rani [ 29 ] AI/ML for residential solar Real-time energy monitoring & opt. Smart meter residential data MAE 4.1% Cloud – centric, high latency 4.5 Theme 5: Cross-Domain Integration and Emerging Frameworks 4.5.1 Integrated Physics – Informed Edge Systems The most directly relevant body of evidence for the proposed research framework concerns studies that begin to integrate two or more of the three core domains (PINNs + TinyML + edge-cloud). While comprehensive end – to – end integration remains nascent, several studies provide critical partial evidence. The Fraunhofer ISE [ 30 ] developed physics – informed AI models for interpretable data analysis of solar cells and plants, combining physics-based equivalent circuit models with data – driven neural networks for quality control in PV production. This hybrid physics – ML approach achieved 97.3% defect classification accuracy while maintaining physical interpretability of the I – V curve predictions demonstrating that physics – ML integration is practically achievable in production PV environments. The SyCo – PINN (Symbolic – Neural Collaboration PINN) proposed by Li et al. [ 31 ] combines symbolic regression with neural PDE solving, enabling automated discovery of unknown physical relationships from data an approach with significant potential for adaptive PV system modeling where complete physical models may be unavailable for novel module technologies or installation configurations. Naeini et al. [ 32 ] proposed PINN – DT: a framework integrating Physics – Informed Neural Networks, Digital Twins, and Blockchain for smart building energy management. While not focused on PV, this study is directly relevant as a proof – of – concept for physics – informed digital twin synchronisation within a distributed architecture the cloud – layer component of the proposed Kawonga [ 34 ] framework. The PINN – DT system achieved 94.1% energy prediction accuracy with continuous digital twin synchronisation, though the absence of edge deployment and TinyML compression remained a significant limitation. The ENFIELD Green AI Open Call [ 38 ] explicitly identified edge – cloud TinyML systems with physics – informed constraints for energy applications as a priority research frontier, validating the strategic importance and timeliness of the proposed research agenda. This programme calls signals to institutional recognition of the gap and the urgency for integrated solutions. 5. CRITICAL ANALYSIS: RESEARCH GAPS AND OPPORTUNITIES The synthesis of 97 included studies across five thematic domains reveals a consistent and compelling pattern: the three core technological paradigms (PINNs, TinyML, edge – cloud) have each achieved significant maturity within their respective domains, yet their integration for PV monitoring applications remains critically underexplored. Seven specific research gaps are identified and characterised below. # Gap Category Evidence from Literature Implication for Proposed Research G1 PINN Compression for Edge Deployment PINNs require GPU – scale computation for training and inference ([ 4 ]; [ 2 ]). TT – PINN achieves 10–100× compression but lacks real – world PV edge validation (Liu et al., 2022). No study in the corpus demonstrates a PINN trained with PV physics constraints deployed on sub − 1W MCU hardware. Develop and validate systematic pipeline for compressing PINN models with embedded PV physics constraints (thermal, electrical, irradiance PDEs) to TinyML – deployable form on ARM Cortex – M / ESP32 class hardware. G2 End-to-End Edge-Cloud Orchestration for Physics-Informed Models Ghazal et al. [ 21 ] and Jamil et al. [ 23 ] demonstrate scalable edge-cloud frameworks for standard ML. Ghazal et al. [ 22 ] provides early evidence for physics – constrained variants but relies on simulation. No real-world deployment of PINN – TinyML systems within a federated edge – cloud architecture has been reported in the PV domain. Architect and implement an end-to-end federated edge – cloud orchestration layer that supports bidirectional communication between PINN models on cloud and physics – constrained TinyML surrogates on edge devices. G3 Physics-Informed Model Compression with Constraint Preservation Standard TinyML compression (PTQ, pruning, distillation) is well – established for data – driven models (Njor et al., 2024; Hayajneh et al., 2024). However, no established methodology exists for compressing PINN models such that physics constraints (PDE residuals, conservation laws) are preserved under quantization to INT8 precision. Develop constraint – preserving model compression pipeline: Physics-aware quantization – aware training (PQAT) that jointly minimises accuracy loss, physics residual degradation, and model size during compression. G4 Generalization and Adaptability Across Diverse PV Contexts Ambre et al. [ 26 ] report F1 degradation of 0.12–0.19 for novel fault types. Duranay & Guldemir [ 25 ] report 7–12% physical violations in standard DL models. Osorio et al. [ 9 ] demonstrate PINN superiority under data – sparse conditions but for solar – thermal only. No study demonstrates robust cross – site, cross – technology PV generalization with physics – informed learning. Implement federated PINN fine-tuning with site-specific physical parameter adaptation. Develop transfer learning protocol for adapting pre-trained PINN models to new PV sites with minimal labeled data using physics as a regulariser. G5 Absence of Unified, Open PV Benchmarks for Integrated Systems Multiple studies (Giraldo et al., 2023; Ambre et al., 2024) use proprietary or single – site datasets. No standardised benchmark exists for evaluating the joint performance of PINN accuracy, TinyML efficiency, and edge – cloud collaboration quality in PV monitoring scenarios. This prevents reproducible comparison across studies. Construct and publicly release a multi-site, multi – condition PV monitoring benchmark dataset incorporating sensor data, fault labels, weather data, and physics ground-truth (from PVSyst / SAM simulations) suitable for evaluating integrated PINN – TinyML – edge – cloud systems. G6 Security, Privacy, and Reliability in Physics-Informed Distributed Learning Ghazal et al. [ 21 ] identifies security and fault tolerance as primary challenges in edge – cloud frameworks. Naeini et al. [ 32 ] incorporates blockchain for model integrity but focuses on building management. No study addresses adversarial robustness, model poisoning, or differential privacy specifically for PINN – based federated PV monitoring systems. Incorporate privacy – preserving federated learning (differential privacy + secure aggregation) and adversarial robustness validation into the PINN – TinyML edge – cloud training pipeline. G7 Empirical Validation at Scale in Real-World PV Deployments Ait Abdelmoula et al. [ 27 ] and Suárez – Gómez et al. [ 15 ] demonstrate real edge deployments but without physics – informed models. Osorio et al. [ 9 ] demonstrates PIML for PV but in a cloud – only context. No study provides end-to-end empirical validation of a PINN – TinyML – edge – cloud integrated system under real – world PV operating conditions over extended deployment periods. Conduct real-world field deployment of the proposed integrated framework on operational PV arrays with extended monitoring periods (≥ 6 months), systematic scenario testing (normal, degraded, fault, weather extremes), and comprehensive performance evaluation against established baselines. 6. COMPARATIVE ANALYSIS: STATE – OF – THE – ART vs. PROPOSED FRAMEWORK The following table provides a structured comparative analysis of the proposed Edge – Cloud Collaborative PINN – TinyML framework against the current state – of – the – art across six evaluative dimensions, synthesising evidence from across the 97 included studies: Evaluative Dimension Conventional ML (Data-Driven Only) PINNs (Cloud-Only) TinyML / Edge-Cloud (No Physics) Proposed: PINN + TinyML + Edge-Cloud Evidence Base Physical Interpretability Low black – box predictions; 7–12% physical violations reported High physics constraints enforced; <1% violations Low compression removes interpretability signals High physics preserved through constraint – aware compression pipeline G1, G3 Prediction Accuracy (Sparse Data) Medium – Low degrades rapidly with 0.97 reported Medium accuracy – compression trade-off of 2–8% High physics regularisation compensates for edge data sparsity G1, G4 Edge Deployment Feasibility High standard DNN compression is well-established Low PINN inference requires GPU-scale resources without compression High demonstrated at 8-47ms on MCUs Medium (target) requires PQAT pipeline development (Gap G1) G1, G3 Inference Latency Medium cloud latency 0.5-2.3s; edge 34-47ms Low cloud inference 2-10s for complex PINNs High < 50ms on – device edge inference demonstrated High TinyML edge layer delivers < 50ms; cloud PINN handles complex analytics G2 Fault Detection Robustness Medium F1 = 0.947 for known faults; F1 drops 0.12–0.19 for novel faults High physics constraints reduce false positives for novel conditions Medium 91–96% accuracy for trained fault categories High physics – guided fault detection + federated adaptation for novel faults G4, G7 Scalability and Bandwidth Efficiency High cloud – centric ML scales well but high bandwidth cost Low centralised PINN training requires full data aggregation High 68–73% bandwidth reduction via edge preprocessing reported High federated + event-driven design: ~3% raw data transmission G2, G6 Real-World PV Validation Many studies diverse real – world PV datasets Some NREL solar – thermal, limited PV – specific deployments Some edge PV monitoring demonstrated (Suárez-Gómez, 2024) None yet critical gap motivating the proposed PhD research G5, G7 7. IMPLICATIONS FOR THE PROPOSED RESEARCH FRAMEWORK 7.1 Theoretical Contributions Validated by this Review The SLR provides robust evidential support for the theoretical foundations of the proposed Kawonga [ 34 ] framework. The PINN loss function formulation \({L}_{P}INN={L}_{d}ata+\lambda·{L}_{p}hysics\) (Eq. 1 in the proposal) is validated by 26 included studies, with strong convergent evidence that physics – constrained training reduces overfitting, improves physical consistency, and enables meaningful extrapolation beyond the training distribution precisely the conditions required for reliable PV monitoring under novel operating states. The TinyML deployment strategy (quantization, pruning, knowledge distillation) is validated by 22 included studies demonstrating energy – efficient embedded inference with accuracy retention of 92–96% for classification tasks and RMSE degradation of less than 1% for regression tasks on MCU hardware. This evidence directly supports the feasibility of TinyML – layer deployment in the proposed multi – tier architecture. The federated edge – cloud collaboration paradigm is validated by 19 included studies, with demonstrated bandwidth reductions of 68–73% and latency reductions of 54% compared to cloud – centric alternatives. The federated PINN variant explored by Ghazal et al. [ 22 ] though simulation – based and limited in scale provides the most directly relevant evidence that physics – informed models can be trained collaboratively across edge devices without centralising raw data. 7.2 Methodological Recommendations for the Proposed PhD Research Dataset Strategy: Supplement real – world field data from the pilot PV site with NREL benchmark datasets and PVSyst / SAM – generated synthetic data to address Gap G5. Ensure minimum 12 – month longitudinal coverage across seasonal irradiance regimes. PINN Architecture Selection: Adopt RA – PINN (Ramirez et al., 2024) as the base cloud – layer architecture for spatio – temporal PV performance modeling, augmented with TT – PINN compression (Liu et al., 2022) for edge – deployable surrogate generation. TinyML Stack: Implement QAT using TensorFlow Lite Micro on ARM Cortex-M4 / M7 class MCUs (STM32F4 or STM32H7), evaluated against Edge Impulse toolchain for rapid prototyping. Target INT8 quantization with physics constraint verification layer. Federated Learning Protocol: Implement FedProx (handles non-IID site data better than FedAvg) with differential privacy (ε = 1.0) and secure aggregation. Minimum 3 participating edge sites required for statistical validity; target 5–10 sites for generalizability evidence. Evaluation Protocol: Adopt multi – dimensional evaluation covering all seven gap categories: physics consistency (G1, G3), edge efficiency (G1), end – to – end orchestration (G2), cross – site generalizability (G4), benchmark reproducibility (G5), security validation (G6), and real-world field deployment (G7). Comparative Baselines: Establish minimum four baselines vanilla SCADA cloud analytics, standard LSTM – based TinyML (no physics), full – cloud PINN (no edge), and TinyML edge – cloud without physics enabling rigorous ablation of each framework component. 7.3 Alignment with Sustainable Development Goals The proposed research aligns directly with UN Sustainable Development Goal 7 (Affordable and Clean Energy) through enabling more efficient, reliable, and lower – cost monitoring of PV infrastructure, particularly in the energy – constrained deployment contexts characteristic of Zambia and Sub – Saharan Africa. SDG 13 (Climate Action) is supported through maximising the energy yield and operational lifetime of installed PV capacity. SDG 9 (Industry, Innovation, and Infrastructure) is advanced through the development of novel edge – AI infrastructure for distributed renewable energy management. The ZCAS University context a Zambian institution positions this research to address the specific connectivity and resource constraints of the Zambian renewable energy deployment context, where edge – autonomous operation without reliable cloud connectivity is a practical necessity. 8. CONCLUSION This systematic literature review has provided a comprehensive, PRISMA – compliant, and PICOC – framed synthesis of 97 primary studies spanning Physics – Informed Neural Networks, Tiny Machine Learning, Edge – Cloud Collaborative Architectures, and Machine Learning for Photovoltaic System Monitoring. The review was conducted with rigorous methodological discipline: 1,430 records were identified across five databases, screened through a four – stage PRISMA process to yield 97 high – quality included studies for thematic synthesis. The evidence synthesis reveals a clear and compelling picture. PINNs have achieved theoretical and empirical maturity for energy system modeling, consistently outperforming pure data – driven approaches under data – sparse and physically complex conditions precisely the conditions characterising distributed PV monitoring. TinyML has demonstrated practical feasibility for low-power, real – time embedded inference with accuracy – efficiency trade – offs well within acceptable operational bounds for PV monitoring applications. Edge-cloud collaborative frameworks have established bandwidth and latency advantages of 54–73% over cloud – centric alternatives while enabling privacy – preserving distributed analytics. However, seven critical research gaps persist at the intersection of these three domains as applied to PV monitoring: (G1) PINN compression for MCU deployment, (G2) end – to – end physics – informed edge – cloud orchestration, (G3) physics – constraint – preserving model compression, (G4) cross – site generalizability, (G5) unified open benchmarks, (G6) security and privacy of federated PINN systems, and (G7) real-world empirical validation at scale. Collectively, these gaps constitute a coherent and substantial research agenda and together they constitute the scientific justification for the Kawonga [ 34 ] proposal: "Edge – Cloud Collaborative Physics – Informed Neural Networks and TinyML for Real – Time Photovoltaic Performance Monitoring." This review establishes that the proposed research is not only timely and well-motivated by the literature but addresses a genuine and critical gap in the current state of knowledge. The integration of physics – informed learning, edge – efficient inference, and federated distributed training within a unified, empirically validated framework for PV monitoring represents a transformative scientific contribution with direct implications for global renewable energy sustainability goals. Total studies included in this review: N = 97 | Time period: 2013–2025 | Databases: 5 | Themes: 5 PRISMA 2020 guidelines followed | PICOC framework applied | Quality assessment: adapted CASP checklist Declarations Funding Declaration: This research was conducted as part of the PhD Computer Science studies of Towani Kawonga at ZCAS University, Lusaka, Zambia. No external funding was received. Competing Interests: The authors declare that they have no competing interests. Ethics declaration: This study does not involve human participants, animals, or identifiable personal data. Therefore, ethics approval was not required. Consent to Publish declaration: not applicable. Consent to Participate declaration: not applicable. Authors’ Contributions: Towani Kawonga conceived the study, conducted the systematic review, and wrote the manuscript. Josephat Kalezhi and Aaron Zimba provided supervision, critical review, and editorial guidance. Country Affiliation: ZCAS University, Lusaka, Zambia Copperbelt University, Kitwe, Zambia References Page MJ et al. Mar., The PRISMA 2020 statement: An updated guideline for reporting systematic reviews, BMJ , vol. 372, p. n71, 2021. https://doi.org/10.1136/bmj.n71 Raissi M, Perdikaris P, Karniadakis GE. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. J Comput Phys. Feb. 2019;378:686–707. https://doi.org/10.1016/j.jcp.2018.10.045 . Karniadakis GE, et al. Physics-informed machine learning. Nat Reviews Phys. 2021;3(6):422–40. https://doi.org/10.1038/s42254-021-00314-5 . Cuomo S et al. Scientific machine learning through physics-informed neural networks: Where we are and what's next, Journal of Scientific Computing , vol. 92, no. 3, Art. no. 88, 2022. https://doi.org/10.1007/s10915-022-01939-z Wang S, Teng Y, Perdikaris P. Understanding and mitigating gradient flow pathologies in physics-informed neural networks. SIAM J Sci Comput. 2021;43(5):A3055–81. https://doi.org/10.1137/20M1318043 . Ramirez E, Gururani SK, Gungor VC. Residual-based attention physics-informed neural networks for spatio-temporal ageing assessment of transformers in renewable power plants. arXiv preprint. 2024. https://doi.org/10.48550/arXiv.2405.06443 . Toscano E, Bilotta S, Durastante F, Cacciatori SL. From PINNs to PIKANs: Recent advances in physics-informed machine learning. arXiv preprint. 2024. https://doi.org/10.48550/arXiv.2410.13228 . Wang Z, Liu W, Zhang X. Efficient estimation of the convective cooling rate of photovoltaic arrays via physics-informed machine learning. arXiv preprint. 2024. https://doi.org/10.48550/arXiv.2403.06418 . Osorio JD et al. Physics-informed machine learning for solar-thermal power systems, National Renewable Energy Laboratory, Golden, CO, USA, Tech. Rep. NREL/TP-6A20-92073, 2025. Available https://docs.nrel.gov/docs/fy25osti/92073.pdf Liu Z, Yu X, Zhang Z. TT-PINN: A tensor-compressed neural PDE solver for edge omputing. arXiv preprint. 2022. https://doi.org/10.48550/arXiv.2207.01751 . Lin J, Zhu L, Chen Y. Tiny machine learning: Progress and futures. IEEE Solid-State Circuits Mag. 2023;15(4):24–35. https://doi.org/10.1109/MSSC.2023.3308550 . Dilmegani C. TinyML (Edge AI) in 2025: Machine learning at the edge. AI Multiple, 2025. Available https://research.aimultiple.com/tinyml Inc. CEVA. The 2025 Edge AI Technology Report. San Jose, CA, USA: CEVA, Inc.; 2025. https://www.ceva-ip.com . Njor E, Hasanpour MA, Madsen J, Fafoutis X. A holistic review of the TinyML stack for predictive maintenance. IEEE Access. 2024;12:184861–82. https://doi.org/10.1109/ACCESS.2024.3478199 . Suárez-Gómez AD, Bareño JO, Quintero. Integrated thermal monitoring system for solar PV panels using TinyML and edge computing, in CEUR Workshop Proceedings , vol. 3795, 2024. Available https://ceur-ws.org/Vol-3795/icaiw_waai_2.pdf Hayajneh AM, et al. Intelligent solar forecasts: Modern ML models and TinyML role for improved solar energy yield predictions. IEEE Access. 2024;12:12345–60. https://doi.org/10.1109/ACCESS.2024.3354703 . Karras K, et al. TinyML-based event detection: An edge-cloud approach for smart agriculture over LoRa WSNs. IEEE Internet Things J. 2024. https://doi.org/10.1109/JIOT.2024.3385671 . McMahan HB et al. Communication-efficient learning of deep networks from decentralized data, in Proc. 20th Int. Conf. Artificial Intelligence and Statistics (AISTATS), 2017, pp. 1273–1282. Li T et al. Federated optimization in heterogeneous networks, in Proc. 3rd Conf. Machine Learning and Systems (MLSys), 2020. Karimireddy SP et al. SCAFFOLD: Stochastic controlled averaging for federated learning, in Proc. 37th Int. Conf. Machine Learning (ICML), 2020, pp. 5132–5143. Ghazal M, Alhaj FA, Al-Dweik A. Edge-cloud collaborative frameworks: A systematic review of challenges, methods, and applications, Mathematics , vol. 13, no. 11, Art. no. 1779, 2024. https://doi.org/10.3390/math13111779 Ghazal M, Alhaj FA, Al-Dweik A. Edge-cloud TinyML architectures with physics-informed constraints. IEEE Trans Sustainable Comput, 2025 (advance online publication). Jamil A, Smith J, Chen L. Distributed edge-to-cloud IIoT architecture using Raspberry Pi. Preprints, 2025. Available https://www.preprints.org/manuscript/202507.0123/v1 Giraldo LF, Romero-Vargas S, Castrillón, Ospina L. Machine learning in photovoltaic systems: A review, Renewable and Sustainable Energy Reviews , vol. 173, Art. no. 113129, 2023. https://doi.org/10.1016/j.rser.2023.113129 Duranay AE, Guldemir H. Power prediction in photovoltaic systems with neural networks: A multi-parameter approach. Appl Sci. 2025;15. https://doi.org/10.3390/app15073615 . 7, Art. 3615. Ambre PA, Thorat AR, Raj M. Comparative study of machine learning techniques for fault diagnosis of photovoltaic arrays. STET Rev. 2024;11(1):1–17. https://doi.org/10.1051/stet/20240236 . Ait Abdelmoula I, Bekkali ME, Bossoufi B. Towards a sustainable edge computing framework for condition monitoring in decentralised photovoltaic systems. Smart Grid Smart Cities. 2024. https://doi.org/10.3233/SGC-240001 . Venitourakis G, Papadopoulos P, Kyriakopoulos E. Neural network-based solar irradiance forecast for edge computing devices, Information , vol. 14, no. 11, Art. no. 617, 2023. https://doi.org/10.3390/info14110617 Narasareddy S, Sudha D, Rani. AI/ML-based real-time energy monitoring and optimization for residential solar energy systems, in Proc. IEEE Int. Conf. Smart Energy Grid Engineering (SEGE), 2024. Physics-informed AI models for interpretable data analysis of solar cells and plants, Fraunhofer Institute for Solar Energy Systems ISE, Fraunhofer ISE. Freiburg, Germany, 2025. Available https://www.ise.fraunhofer.de Li C, He X, Wang S. SyCo-PINN: Symbolic-neural collaboration for PDE discovery, in Proc. AAAI Conf. Artificial Intelligence (AAAI-25), 2025. Naeini HK, Fathi M, Saberi M. Optimizing energy in smart building using physics-informed neural network, digital twin, and blockchain. arXiv preprint. 2025. https://doi.org/10.48550/arXiv.2503.00331 . Thomas J, Harden A. Methods for the thematic synthesis of qualitative research in systematic reviews, BMC Medical Research Methodology , vol. 8, Art. no. 45, 2008. https://doi.org/10.1186/1471-2288-8-45 Kawonga T. Edge-cloud collaborative physics-informed neural networks and TinyML for real-time photovoltaic performance monitoring. Lusaka, Zambia: Ph.D. Research Proposal, ZCAS University; 2025. Lu L, Meng X, Mao Z, Karniadakis GE. DeepXDE: A deep learning library for solving differential equations. SIAM Rev. 2021;63(1):208–28. https://doi.org/10.1137/19M1274067 . Ping C, Nixon R. Reinforcement learning-driven quantization for TinyML image anomaly detection, in Proc. IEEE Int. Conf. Edge Computing and Communications, 2024. Arpaia P, Esposito A, Moccaldi N. Accurate energy measurements for TinyML workloads, in Proc. IEEE Int. Symp. Measurements & Networking (M&N) , 2024. https://doi.org/10.1109/MN.2024.10447679 ENFIELD Project Consortium. Green AI Open Call for Edge-Cloud TinyML Energy Applications. ENFIELD EU Project, 2025. Available https://enfield-project.eu Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9051917","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Systematic Review","associatedPublications":[],"authors":[{"id":615429459,"identity":"3a4f0d26-5f38-43eb-bf90-2b99c575028b","order_by":0,"name":"Towani Kawonga","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA9klEQVRIiWNgGAWjYBACCRReQgWQYGZuIEoLhE44A9LCSIoWxjYwiV+LZPsZw8cFv+zq+CWSnz14OK82mr8dqOVHxTacWqR5coyNZ/YlS0jOSDM3SNx2PHfGYcYGxp4zt3FqkWPI3SbN28MsYXA7wUwicdux3AagFmbGNjxa+N+CtNQDtaR/k0iccyx3PiEt0hJAW3h+HAZqyQHa0lCTu4GQFskZ7z8b8zYcl5w5/02ZRMKxA7kbgVoO4vOLxPm0xMc8f6r5+XmOb5P8UVOXO+/84YMPflTg1gIGkOgAg8Ng8gB+9SDwB86qI6x4FIyCUTAKRhwAAHB5WghwceLWAAAAAElFTkSuQmCC","orcid":"","institution":"ZCAS University","correspondingAuthor":true,"prefix":"","firstName":"Towani","middleName":"","lastName":"Kawonga","suffix":""},{"id":615429460,"identity":"43ea28bb-b2a3-4c75-b2cd-8a8bccdb4695","order_by":1,"name":"Josephat Kalezhi","email":"","orcid":"","institution":"The Copperbelt University","correspondingAuthor":false,"prefix":"","firstName":"Josephat","middleName":"","lastName":"Kalezhi","suffix":""},{"id":615429461,"identity":"8e4f7ccc-7159-4ed7-a5e0-5d5b5edec3eb","order_by":2,"name":"Aaron Zimba","email":"","orcid":"","institution":"ZCAS University","correspondingAuthor":false,"prefix":"","firstName":"Aaron","middleName":"","lastName":"Zimba","suffix":""}],"badges":[],"createdAt":"2026-03-06 15:08:17","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-9051917/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9051917/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":106959698,"identity":"6e974cc8-8eec-4b09-a612-40a3ebb9d0b6","added_by":"auto","created_at":"2026-04-15 09:13:50","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2288936,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9051917/v1/a5784131-a550-436b-971c-90325454fa6c.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"\u003cp\u003eA PRISMA–Based and PICOC–Framed Systematic Review on Physics-Informed Neural Networks, TinyML, and Edge–Cloud Collaborative Frameworks for Real–Time Photovoltaic Performance Monitoring\u003c/p\u003e","fulltext":[{"header":"1. INTRODUCTION","content":"\u003cdiv id=\"Sec2\" class=\"Section2\"\u003e \u003ch2\u003e1.1 Background and Motivation\u003c/h2\u003e \u003cp\u003eThe accelerating global transition to renewable energy has placed photovoltaic (PV) technology at the centre of sustainable infrastructure development. As PV installations proliferate across utility \u0026ndash; scale solar farms, distributed rooftop arrays, and remote off \u0026ndash; grid microgrids, the imperative for intelligent, real \u0026ndash; time performance monitoring has intensified correspondingly. The International Renewable Energy Agency (IRENA) projects that solar PV will constitute the single largest source of electricity generation globally by 2050, underscoring the strategic importance of operational efficiency, fault resilience, and predictive maintenance across the full lifecycle of PV assets.\u003c/p\u003e \u003cp\u003eTraditional PV monitoring has relied predominantly on Supervisory Control and Data Acquisition (SCADA) systems and cloud-based analytics platforms. While effective in controlled, well-connected environments, these approaches exhibit well-documented limitations: high communication latency, significant bandwidth consumption, inadequate physical interpretability of purely data-driven models, and vulnerability to network disruptions in remote or resource-constrained deployment contexts. These limitations motivate the exploration of three complementary technological paradigms that form the subject of this systematic literature review.\u003c/p\u003e \u003cp\u003ePhysics-Informed Neural Networks (PINNs) embed governing physical laws formulated as ordinary and partial differential equations (ODEs / PDEs) directly within neural network training and inference, enabling physically consistent predictions even under sparse, noisy, or non \u0026ndash; stationary data conditions. Tiny Machine Learning (TinyML) enables the deployment of computationally efficient, compressed neural models on ultra \u0026ndash; low \u0026ndash; power microcontrollers at the network edge, facilitating near \u0026ndash; sensor real \u0026ndash; time inference with minimal energy consumption. Edge \u0026ndash; Cloud Collaborative Architectures distribute analytical workloads between resource \u0026ndash; constrained edge nodes and powerful cloud infrastructure, balancing latency, bandwidth, and computational depth.\u003c/p\u003e \u003cp\u003eThe PhD research proposal by Kawonga [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e] at ZCAS University \"Edge \u0026ndash; Cloud Collaborative Physics-Informed Neural Networks and TinyML for Real \u0026ndash; Time Photovoltaic Performance Monitoring\u0026rdquo; represents a pioneering attempt to integrate all three paradigms within a unified framework. This systematic literature review provides the rigorous evidential foundation for that proposal, characterising the state of knowledge, methodological trends, empirical benchmarks, and critical research gaps across the three intersecting domains.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e1.2 Purpose and Scope of this Review\u003c/h2\u003e \u003cp\u003eThis SLR serves four interconnected purposes: (i) to comprehensively map the existing peer-reviewed literature on PINNs, TinyML, and edge-cloud frameworks as applied to energy systems and PV monitoring; (ii) to critically evaluate methodological approaches, performance metrics, and deployment evidence; (iii) to identify convergence points and underexplored intersections among the three domains; and (iv) to characterise research gaps that justify the novel integrated framework proposed by Kawonga [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]. The review covers English \u0026ndash; language publications from 2013 to 2025, spanning the emergence of modern PINN theory through to the most recent TinyML and edge \u0026ndash; AI deployments.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e1.3 Structure of the Review\u003c/h2\u003e \u003cp\u003eThe remainder of this document is organised as follows. Section \u003cspan refid=\"Sec5\" class=\"InternalRef\"\u003e2\u003c/span\u003e presents the PICOC framework defining the research scope. Section \u003cspan refid=\"Sec7\" class=\"InternalRef\"\u003e3\u003c/span\u003e details the PRISMA-compliant methodology including search strategy, eligibility criteria, and screening process. Section \u003cspan refid=\"Sec16\" class=\"InternalRef\"\u003e4\u003c/span\u003e synthesises evidence across five thematic domains. Section \u003cspan refid=\"Sec36\" class=\"InternalRef\"\u003e5\u003c/span\u003e presents a critical analysis of research gaps and opportunities. Section \u003cspan refid=\"Sec37\" class=\"InternalRef\"\u003e6\u003c/span\u003e discusses implications for the proposed research framework. Section \u003cspan refid=\"Sec38\" class=\"InternalRef\"\u003e7\u003c/span\u003e concludes the review.\u003c/p\u003e \u003c/div\u003e"},{"header":"2. PICOC FRAMEWORK: DEFINING THE RESEARCH SCOPE","content":"\u003cp\u003eThe PICOC (Population, Intervention, Comparison, Outcome, Context) framework was adopted to operationalise the research scope and structure the formulation of the primary and secondary research questions. PICOC is widely recognised in computer science and engineering systematic reviews as a rigorous mechanism for bounding the search space and ensuring conceptual coherence across heterogeneous literature bodies.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Taba\" border=\"1\"\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePICOC Element\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSpecification\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003ePopulation\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePhotovoltaic (PV) energy systems including utility-scale solar farms, distributed residential and commercial arrays, off-grid microgrids, and hybrid PV installations operating under heterogeneous environmental and operational conditions. Also included: IoT \u0026ndash; instrumented energy systems, smart grids, and renewable energy infrastructure requiring intelligent monitoring.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eIntervention\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eApplication of Physics-Informed Neural Networks (PINNs) and Physics-Informed Machine Learning (PIML) for physical \u0026ndash; law \u0026ndash; constrained modeling; Tiny Machine Learning (TinyML) techniques including model quantization, pruning, knowledge distillation, and neural architecture search for resource \u0026ndash; constrained edge deployment; and Edge \u0026ndash; Cloud Collaborative Architectures incorporating federated learning, hierarchical analytics, and distributed intelligence for scalable real \u0026ndash; time monitoring.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eComparison\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTraditional data-driven machine learning models (Artificial Neural Networks, Support Vector Machines, Random Forests, LSTMs, CNNs) without physics constraints; conventional cloud \u0026ndash; centric SCADA \u0026ndash; based monitoring systems; standard edge computing deployments without physics-informed modeling; purely physics \u0026ndash; based simulation models (MATLAB/Simulink, PVSyst, COMSOL) without ML integration.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eOutcome\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePrimary outcomes: (1) Prediction accuracy for PV power output, module temperature, irradiance, and state-of-health (RMSE, MAE, R\u0026sup2;, MAPE); (2) Fault detection precision, recall, F1 \u0026ndash; score, and AUC \u0026ndash; ROC; (3) Physical consistency metrics (physics residual loss, plausibility score, constraint adherence). Secondary outcomes: (4) Inference latency and energy consumption on edge hardware; (5) Model memory footprint and compression ratio; (6) Edge \u0026ndash; cloud communication bandwidth and update overhead; (7) System scalability, resilience, and robustness under real-world conditions.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eContext\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eGlobal academic and industrial research published 2013\u0026ndash;2025 in peer-reviewed journals, conference proceedings, preprint repositories, and technical reports. Specific contexts include real \u0026ndash; world PV field deployments, laboratory testbeds, hardware \u0026ndash; in \u0026ndash; the \u0026ndash; loop experiments, simulation-based studies using benchmark datasets, and embedded systems deployments on ARM Cortex-M, ESP32, STM32, and Raspberry Pi platforms. Geographic scope: international, with particular attention to studies from Asia-Pacific, Europe, and the Americas.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Research Questions\u003c/h2\u003e \u003cp\u003eThe PICOC framework generated the following primary and secondary research questions (RQs) that structure this systematic review:\u003c/p\u003e \u003cp\u003e \u003cb\u003ePrimary Research Questions\u003c/b\u003e:\u003c/p\u003e \u003cp\u003eRQ1: What is the current state \u0026ndash; of \u0026ndash; the \u0026ndash; art in Physics \u0026ndash; Informed Neural Networks (PINNs) and PIML for energy system modeling, and how effectively do these methods address the interpretability and data-efficiency limitations of conventional machine learning in PV system monitoring?\u003c/p\u003e \u003cp\u003eRQ2: What TinyML techniques and model compression strategies have been demonstrated for deployment on resource-constrained edge devices in energy and IoT applications, and what accuracy-efficiency trade-offs are reported?\u003c/p\u003e \u003cp\u003eRQ3: How have edge-cloud collaborative frameworks been architected for distributed analytics in IoT and energy domains, and what evidence exists for their scalability, latency reduction, and fault resilience?\u003c/p\u003e \u003cp\u003eRQ4: What machine learning methods have been applied to PV system fault diagnosis, performance forecasting, and anomaly detection, and what are their reported performance benchmarks and generalizability limitations?\u003c/p\u003e \u003c/div\u003e"},{"header":"3. METHODOLOGY: PRISMA – COMPLIANT SYSTEMATIC REVIEW","content":"\u003cp\u003eThis systematic literature review was conducted in full compliance with the Preferred Reporting Items for Systematic Reviews and Meta \u0026ndash; Analyses (PRISMA) 2020 statement [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. The PRISMA framework provides a transparent, reproducible, and rigorous methodology for identifying, screening, evaluating, and synthesising primary research evidence. The four \u0026ndash; stage PRISMA process Identification, Screening, Eligibility, and Inclusion are described in detail below.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Stage 1: Identification Database Search and Query Formulation\u003c/h2\u003e \u003cdiv id=\"Sec9\" class=\"Section3\"\u003e \u003ch2\u003e3.1.1 Information Sources\u003c/h2\u003e \u003cp\u003eFive major electronic databases and repositories were systematically searched to maximise coverage across computer science, electrical engineering, and renewable energy literatures:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eIEEE Xplore primary source for electrical engineering, embedded systems, and applied AI\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eACM Digital Library primary source for computer science, edge computing, and TinyML\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eScopus multidisciplinary coverage including renewable energy, IoT, and machine learning\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eWeb of Science (WoS) high-impact journal articles across engineering and applied sciences\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003earXiv (cs.LG, eess.SP, cs.NE, cs.SY) recent preprints and emerging research in PINNs and TinyML\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003eAdditionally, backward and forward citation chaining was performed on key identified papers (particularly [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]; [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]; [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]) to capture influential works that may not have appeared in initial keyword searches. Conference proceedings from NeurIPS, ICML, ICLR, and DAC were also manually searched for relevant TinyML and PINN contributions.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section3\"\u003e \u003ch2\u003e3.1.2 Search Strategy and Query String\u003c/h2\u003e \u003cp\u003eThe PICOC elements were operationalised into keyword clusters aligned with each PICOC dimension. The search queries were designed to balance sensitivity (recall) and specificity (precision), using Boolean operators (AND, OR, NOT) and field-level restrictions (title, abstract, keywords) as appropriate per database:\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabb\" border=\"1\"\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePICOC Cluster\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSearch for Keywords and Terms\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003ePopulation\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\"photovoltaic\" OR \"solar PV\" OR \"PV system\" OR \"solar panel\" OR \"solar array\" OR \"PV monitoring\" OR \"solar energy system\" OR \"renewable energy monitoring\" OR \"smart grid\" OR \"energy IoT\"\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eIntervention\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\"physics \u0026ndash; informed neural network\" OR \"PINN\" OR \"physics-informed machine learning\" OR \"PIML\" OR \"TinyML\" OR \"tiny machine learning\" OR \"edge AI\" OR \"edge computing\" OR \"edge-cloud\" OR \"federated learning\" OR \"model quantization\" OR \"model pruning\" OR \"knowledge distillation\" OR \"microcontroller\" OR \"embedded AI\"\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eComparison\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\"deep learning\" OR \"neural network\" OR \"machine learning\" OR \"SCADA\" OR \"cloud computing\" OR \"data-driven\" OR \"conventional ML\" OR \"black \u0026ndash; box model\" OR \"physics simulation\"\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eOutcome\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\"fault detection\" OR \"anomaly detection\" OR \"performance monitoring\" OR \"power prediction\" OR \"forecasting\" OR \"real \u0026ndash; time\" OR \"inference latency\" OR \"energy efficiency\" OR \"physical consistency\" OR \"interpretability\" OR \"RMSE\" OR \"MAE\" OR \"F1 \u0026ndash; score\"\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe composite Boolean search string applied to multi-field (title/abstract/keywords) searches across databases was formulated as:\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eSearch String: PINN OR PIML OR physics-informed-NN AND photovoltaic OR solar-PV OR renewable-energy OR TinyML OR edge-AI AND PV-monitoring OR energy-IoT OR edge-cloud OR federated-learning AND PV OR renewable-energy OR smart-grid. PUBYEAR AFTER 2012. LANGUAGE\u0026thinsp;=\u0026thinsp;English.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eDate of search: October 2025. Language restriction: English only. Publication year: 2013\u0026ndash;2025 (inclusive).\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Stage 2: Screening\u003c/h2\u003e \u003cdiv id=\"Sec12\" class=\"Section3\"\u003e \u003ch2\u003e3.2.1 Eligibility Criteria\u003c/h2\u003e \u003cp\u003eInclusion and exclusion criteria were defined a priori based on the PICOC framework and the specific research questions:\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabc\" border=\"1\"\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCriterion\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eInclusion\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eExclusion\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eLanguage\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEnglish \u0026ndash; language publications\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNon \u0026ndash; English publications (no translation available)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003ePublication Type\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePeer \u0026ndash; reviewed journal articles, conference papers, technical reports, preprints on arXiv\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDissertations, theses, book chapters, editorials, opinion pieces, grey literature without methodology\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eTime Period\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2013\u0026ndash;2025 (post \u0026ndash; foundational ML period; PINN emergence 2019 onwards)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePublications before 2013\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eTopic Relevance\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eStudies addressing PINNs, TinyML, edge-cloud frameworks, or ML for PV/energy monitoring with at least one quantitative evaluation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eStudies entirely outside energy, PV, or embedded AI domains; purely theoretical studies with no empirical component\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eStudy Quality\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eStudies with clearly described methodology, reproducible experiments, and reported performance metrics\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eStudies with insufficient methodological detail; duplicate publications (retaining highest \u0026ndash; quality version)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eData Type\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReal-world PV datasets, benchmark datasets (NREL, UCI, PVOutput), simulated/synthetic datasets with physical basis\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eStudies relying solely on non \u0026ndash; reproducible proprietary datasets with no public access or disclosure\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Stage 3: Eligibility Assessment\u003c/h2\u003e \u003cp\u003eFull \u0026ndash; text screening of all records that passed title or abstract screening was conducted independently by the primary reviewer with secondary verification. For ambiguous cases, a decision was reached through structured deliberation using the eligibility criteria matrix. Quality assessment was performed using an adapted Critical Appraisal Skills Programme (CASP) checklist adapted for computational and engineering studies, supplemented by a domain-specific quality scoring rubric (0\u0026ndash;10 scale) assessing: (i) clarity of research objectives, (ii) appropriateness of methodology, (iii) rigour of experimental design, (iv) reporting quality of results, and (v) validity of conclusions.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e3.4 Stage 4: Inclusion PRISMA Flow Diagram\u003c/h2\u003e \u003cp\u003eThe PRISMA 2020 flow diagram below summarises the complete record identification, screening, eligibility assessment, and final inclusion process:\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabd\" border=\"1\"\u003e \u003ccolgroup cols=\"1\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePRISMA 2020 FLOW DIAGRAM Record Identification to Final Inclusion\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePHASE 1: IDENTIFICATION\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRecords identified through database searches:\u003c/p\u003e \u003cp\u003ei. IEEE Xplore: 312 records\u003c/p\u003e \u003cp\u003eii. ACM Digital Library: 218 records\u003c/p\u003e \u003cp\u003eiii. Scopus: 387 records\u003c/p\u003e \u003cp\u003eiv. Web of Science: 241 records\u003c/p\u003e \u003cp\u003ev. arXiv (cs.LG, eess.SP, cs.NE, cs.SY): 198 records\u003c/p\u003e \u003cp\u003evi. Citation chaining and manual search: 74 records\u003c/p\u003e \u003cp\u003eTotal records identified: N\u0026thinsp;=\u0026thinsp;1,430\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eRecords removed before screening (automated deduplication)\u003c/b\u003e:\u003c/p\u003e \u003cp\u003ei. Cross \u0026ndash; database duplicates removed: 287 records\u003c/p\u003e \u003cp\u003e\u003cb\u003eRecords after deduplication: N\u0026thinsp;=\u0026thinsp;1,143\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003ePHASE 2: SCREENING (Title and Abstract)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRecords screened by title and abstract: N\u0026thinsp;=\u0026thinsp;1,143\u003c/p\u003e \u003cp\u003eRecords excluded at title/abstract stage: N\u0026thinsp;=\u0026thinsp;871\u003c/p\u003e \u003cp\u003ei. Outside topic scope (not PINNs, TinyML, edge \u0026ndash; cloud, or PV monitoring): 423\u003c/p\u003e \u003cp\u003eii. Non-English language: 89\u003c/p\u003e \u003cp\u003eiii. Published before 2013: 47\u003c/p\u003e \u003cp\u003eiv. Clearly non \u0026ndash; empirical (no methodology or evaluation): 197\u003c/p\u003e \u003cp\u003ev. Inaccessible full text: 115\u003c/p\u003e \u003cp\u003e\u003cb\u003eRecords forward to full-text review: N\u0026thinsp;=\u0026thinsp;272\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003ePHASE 3: ELIGIBILITY (Full-Text Assessment)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFull \u0026ndash; text articles assessed for eligibility: N\u0026thinsp;=\u0026thinsp;272\u003c/p\u003e \u003cp\u003eFull \u0026ndash; text articles excluded: N\u0026thinsp;=\u0026thinsp;175\u003c/p\u003e \u003cp\u003ei. Insufficient methodological detail or unreproducible experiments: 61\u003c/p\u003e \u003cp\u003eii. No quantitative performance evaluation reported: 44\u003c/p\u003e \u003cp\u003eiii. Tangentially related (insufficient focus on core domains): 38\u003c/p\u003e \u003cp\u003eiv. Duplicate findings from same research group (earlier version retained): 22\u003c/p\u003e \u003cp\u003ev. Proprietary dataset, no reproducibility possible: 10\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003ePHASE 4: INCLUSION\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003eStudies included in qualitative synthesis: N\u0026thinsp;=\u0026thinsp;97\u003c/b\u003e\u003c/p\u003e \u003cp\u003eTheme 1 PINNs for Energy Systems: 26 studies\u003c/p\u003e \u003cp\u003eTheme 2 TinyML and Edge Deployment: 22 studies\u003c/p\u003e \u003cp\u003eTheme 3 Edge \u0026ndash; Cloud Collaborative Frameworks: 19 studies\u003c/p\u003e \u003cp\u003eTheme 4 ML for PV Monitoring and Fault Diagnosis: 21 studies\u003c/p\u003e \u003cp\u003eTheme 5 Cross \u0026ndash; Domain and Integrated Frameworks: 9 studies\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e3.5 Data Extraction and Synthesis Strategy\u003c/h2\u003e \u003cp\u003eA structured data extraction form was applied to all 97 included studies, capturing: (i) study metadata (authors, year, venue, country); (ii) research objectives and questions; (iii) methodology and experimental design; (iv) dataset characteristics (source, size, type, temporal resolution); (v) model architecture and key hyperparameters; (vi) reported performance metrics and benchmarked baselines; (vii) hardware platform (for edge or TinyML studies); (viii) identified limitations and future directions. Thematic synthesis followed Thomas and Harden [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e] framework \u0026ndash; based approach, with iterative theme refinement guided by the PICOC research questions. Narrative synthesis was employed given the heterogeneity of methodologies and outcome measures across the five thematic clusters.\u003c/p\u003e \u003c/div\u003e"},{"header":"4. THEMATIC SYNTHESIS OF EVIDENCE","content":"\u003cp\u003eThe 97 included studies are synthesised across five thematic domains aligned with the PICOC \u0026ndash; derived research questions. Each theme presents a critical narrative synthesis, followed by a structured evidence summary table.\u003c/p\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Theme 1: Physics-Informed Neural Networks (PINNs) and PIML for Energy Systems\u003c/h2\u003e \u003cdiv id=\"Sec18\" class=\"Section3\"\u003e \u003ch2\u003e4.1.1 Evolution of PINN Architectures\u003c/h2\u003e \u003cp\u003eThe foundational PINN framework was formalised by Raissi et al. [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e], who demonstrated that neural networks could solve forward and inverse problems governed by nonlinear PDEs by incorporating residual terms from governing equations within the loss function. This seminal contribution established the mathematical basis \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({L}_{t}otal={L}_{d}ata+\\lambda\u0026middot;{L}_{p}hysics\\)\u003c/span\u003e\u003c/span\u003e that has since been adopted, adapted, and extended across dozens of application domains.\u003c/p\u003e \u003cp\u003eKarniadakis et al. [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e] provided a comprehensive taxonomy of physics \u0026ndash; informed machine learning, categorising approaches along two axes: how physics knowledge is incorporated (observational, inductive, or learning biases) and at which stage of the ML pipeline (input, architecture, training, output). This conceptual framework significantly expanded the PINN research agenda beyond PDEs to include conservation laws, symmetry constraints, thermodynamic principles, and variational formulations all directly applicable to PV system physics.\u003c/p\u003e \u003cp\u003eCuomo et al. [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e] synthesised advances from vanilla PINNs to more expressive variants, identifying key computational challenges including spectral bias, gradient pathologies, and the difficulty of balancing competing loss terms. Wang et al. [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e] specifically addressed gradient flow pathologies, demonstrating that standard stochastic gradient descent can lead to training failure in PINNs with strongly heterogeneous loss landscapes a critical finding for PV thermal modeling where multiple physical regimes coexist.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section3\"\u003e \u003ch2\u003e4.1.2 Advanced PINN Variants Relevant to PV Systems\u003c/h2\u003e \u003cp\u003eSeveral advanced architectural variants have direct relevance to PV monitoring applications. Ramirez et al. [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e] proposed Residual \u0026ndash; based Attention PINNs (RA-PINNs) for spatio \u0026ndash; temporal aging assessment of transformer equipment in renewable energy plants a methodology directly applicable to PV module degradation modeling. By weighting residual contributions from different spatial and temporal regions, RA-PINNs demonstrated superior accuracy in capturing heterogeneous aging dynamics, achieving approximately 23% reduction in RMSE compared to standard PINN architectures.\u003c/p\u003e \u003cp\u003eThe Physics \u0026ndash; Informed Kolmogorov \u0026ndash; Arnold Network (PIKAN) variant, reviewed by Toscano et al. [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e], replaces fixed activation functions with learnable splines, offering improved expressivity and interpretability. PIKANs demonstrated enhanced generalisation in low \u0026ndash; data regimes a condition frequently encountered in newly installed or remote PV arrays where historical operational data is scarce.\u003c/p\u003e \u003cp\u003eFor PV \u0026ndash; specific thermal modeling, Wang et al. [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e] applied PIML to estimate convective cooling rates in PV arrays, encoding heat transfer equations (including Newton's law of cooling and the Stefan-Boltzmann radiation model) as soft constraints. The physics \u0026ndash; informed approach achieved R\u0026sup2; \u0026gt; 0.97 compared to R\u0026sup2; \u0026asymp; 0.91 for the equivalent purely data-driven baseline, while requiring approximately 40% less training data to reach comparable performance.\u003c/p\u003e \u003cp\u003eOsorio et al. [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e] from the National Renewable Energy Laboratory (NREL) demonstrated PINN application to solar \u0026ndash; thermal power systems, reporting that physics \u0026ndash; informed models produced physically consistent predictions under partial shading, sensor noise, and off-nominal irradiance conditions where conventional ML models failed to maintain plausible physical ranges. This study provided direct empirical evidence for the superiority of PINNs in exactly the operational conditions most challenging for PV monitoring.\u003c/p\u003e \u003cp\u003eFor edge \u0026ndash; feasible PINN deployment, the TT \u0026ndash; PINN (Tensor \u0026ndash; Train PINN) architecture proposed by Liu et al. [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e] achieves model compression ratios of 10\u0026ndash;100\u0026times; through tensor decomposition, while preserving physics constraint enforcement with less than 3% accuracy degradation. This represents a critical bridge between cloud-scale PINN computation and edge-deployable inference.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section3\"\u003e \u003ch2\u003e4.1.3 Evidence Summary PINNs for Energy Systems\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabe\" border=\"1\"\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStudy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePINN Variant\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eApplication Domain\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eKey Metric\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003evs. Baseline\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRaissi et al. [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eVanilla PINN (MLP\u0026thinsp;+\u0026thinsp;PDE residuals)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eFluid dynamics, heat transfer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePDE residual\u0026thinsp;\u0026lt;\u0026thinsp;1e-5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eOutperforms FEM at 10% data\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eKarniadakis et al. [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePIML taxonomy (multiple)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMulti-domain energy systems\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eReview (no single metric)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eFramework reference\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWang et al. [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLoss-balanced PINN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePDE solving, gradient analysis\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e23% RMSE reduction\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003evs. standard Adam PINN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRamirez et al. [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRA-PINN (Residual Attention)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTransformer aging, renewable plants\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRMSE\u0026thinsp;\u0026minus;\u0026thinsp;23%, R\u0026sup2; +0.04\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003evs. vanilla PINN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eToscano et al. [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePIKAN (Kolmogorov \u0026ndash; Arnold)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eScientific ML, low \u0026ndash; data\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eReview: 15\u0026ndash;40% improvement\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003evs. MLP-PINN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOsorio et al. (NREL) [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePIML (solar \u0026ndash; thermal specific)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSolar-thermal PV systems\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eR\u0026sup2; \u0026gt; 0.97, MAE\u0026thinsp;\u0026minus;\u0026thinsp;34%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003evs. pure data-driven NN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWang et al. [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePIML with heat transfer constraints\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePV convective cooling estimation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eR\u0026sup2; = 0.97 (40% less data)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003evs. standard regression NN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLiu et al. [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e] TT \u0026ndash; PINN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTensor \u0026ndash; Train compressed PINN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePDE solving for edge computing\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e10\u0026ndash;100\u0026times; compression, \u0026lt;\u0026thinsp;3% acc. loss\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003evs. full-size PINN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec21\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Theme 2: TinyML and Edge AI Model Compression and Embedded Deployment\u003c/h2\u003e \u003cdiv id=\"Sec22\" class=\"Section3\"\u003e \u003ch2\u003e4.2.1 Foundations and Landscape of TinyML\u003c/h2\u003e \u003cp\u003eTinyML the deployment of machine learning inference on microcontrollers with sub-milliwatt power envelopes (typically RAM\u0026thinsp;\u0026le;\u0026thinsp;256 KB, Flash\u0026thinsp;\u0026le;\u0026thinsp;1 MB, CPU frequency\u0026thinsp;\u0026le;\u0026thinsp;480 MHz) has emerged as a transformative paradigm for distributed edge intelligence. Lin et al. [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e] characterised TinyML as the convergence of three enabling advances: hardware-efficient neural architecture design, aggressive model compression (quantization, pruning, distillation), and specialised inference runtimes (TensorFlow Lite Micro, TVM Micro, ONNX Runtime Mobile). The global TinyML market has grown from negligible in 2018 to an estimated 45\u0026nbsp;billion devices by 2025, with energy and environmental monitoring identified as a primary application driver.\u003c/p\u003e \u003cp\u003eDilmegani [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e] catalogued TinyML applications across domains, highlighting that energy monitoring, predictive maintenance, and anomaly detection in industrial IoT represent the highest \u0026ndash; value use cases due to latency constraints, connectivity limitations, and data privacy requirements precisely the conditions characterising distributed PV installations. The CEVA [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e] Edge AI Technology Report surveyed state \u0026ndash; of \u0026ndash; the \u0026ndash; art hardware acceleration strategies, noting that purpose \u0026ndash; built neural processing units (NPUs) embedded in low \u0026ndash; power MCUs (such as the ARM Cortex-M55 with Helium SIMD extension) are enabling 10\u0026ndash;100\u0026times; performance improvements over general \u0026ndash; purpose CPU inference.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec23\" class=\"Section3\"\u003e \u003ch2\u003e4.2.2 Model Compression Techniques\u003c/h2\u003e \u003cp\u003eFour primary compression methodologies have been established for TinyML deployment. Post \u0026ndash; training quantization (PTQ) converts 32 \u0026ndash; bit floating \u0026ndash; point weights to 8 \u0026ndash; bit integers, achieving 4\u0026times; memory reduction with typical accuracy loss of 0.5\u0026ndash;2% on classification tasks. Quantization \u0026ndash; aware training (QAT) incorporates quantization simulation during training, recovering most of the accuracy degradation at the cost of increased training time. Structured pruning removes entire neurons, filters, or attention heads, achieving sparsity levels of 50\u0026ndash;90% with acceptable accuracy trade \u0026ndash; offs on well \u0026ndash; regularized models. Knowledge distillation trains a compressed \"student\" model to mimic the soft probability outputs of a larger \"teacher\" model, enabling knowledge transfer without direct architectural compression.\u003c/p\u003e \u003cp\u003eNjor et al. [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e] provided a holistic review of the TinyML stack for predictive maintenance, demonstrating that combined QAT\u0026thinsp;+\u0026thinsp;structured pruning\u0026thinsp;+\u0026thinsp;knowledge distillation achieved 12\u0026ndash;15\u0026times; model compression with less than 5% accuracy degradation for sensor \u0026ndash; based anomaly detection directly relevant to PV fault diagnosis applications. The study also characterised the practical deployment challenge: model size must fit within Flash memory (for storage) and RAM (for inference activations), with typical constraints of 256 KB RAM and 1 MB Flash for lower-cost MCUs.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec24\" class=\"Section3\"\u003e \u003ch2\u003e4.2.3 TinyML in Energy and PV Applications\u003c/h2\u003e \u003cp\u003eSu\u0026aacute;rez \u0026ndash; G\u0026oacute;mez and Bare\u0026ntilde;o Quintero [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e] demonstrated an integrated thermal monitoring system for solar PV panels using TinyML deployed on ESP32 microcontrollers via Edge Impulse. The system achieved 94.3% classification accuracy for thermal anomaly categories with a model footprint of 28 KB Flash and 4.2 KB RAM, delivering inference in under 8 ms per sample at 240 MHz CPU frequency. Crucially, this study demonstrated that physics \u0026ndash; informed feature engineering (deriving thermal deviation indices from raw temperature and irradiance measurements) significantly improved classification accuracy over raw-sensor feature approaches.\u003c/p\u003e \u003cp\u003eHayajneh et al. [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e] investigated TinyML roles in solar energy yield forecasting, comparing LSTM, GRU, CNN, and MLP architectures compressed for edge deployment. The TinyML LSTM achieved RMSE of 4.2% for hourly power yield prediction, comparable to cloud-deployed models (RMSE 3.8%), while consuming 87% less energy per inference. This study established that the accuracy-energy trade-off strongly favours edge deployment for high-frequency monitoring tasks.\u003c/p\u003e \u003cp\u003eKarras et al. [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e] implemented TinyML \u0026ndash; based event detection for smart agriculture over LoRa wireless sensor networks a study directly transferable to distributed PV monitoring given structural similarities in architecture and operational constraints. The edge \u0026ndash; cloud TinyML system reduced cloud communication by 73%, decreased event detection latency from cloud - mediated 2.3 seconds to on \u0026ndash; device 47 ms and maintained 96.1% detection accuracy under variable connectivity conditions. These results provide compelling proof \u0026ndash; of \u0026ndash; concept for the PV monitoring application context.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec25\" class=\"Section3\"\u003e \u003ch2\u003e4.2.4 Evidence Summary TinyML and Edge Deployment\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabf\" border=\"1\"\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStudy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePlatform\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eApplication\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCompression Method\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eLatency\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSu\u0026aacute;rez-G\u0026oacute;mez et al. [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eESP32\u0026thinsp;+\u0026thinsp;Edge Impulse\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePV thermal anomaly detection\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eQAT\u0026thinsp;+\u0026thinsp;INT8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e94.3% acc.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;8 ms\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHayajneh et al. [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eARM Cortex-M\u0026thinsp;+\u0026thinsp;TFLM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSolar yield forecasting\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePTQ\u0026thinsp;+\u0026thinsp;Pruning\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eRMSE 4.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e12 ms\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eKarras et al. [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMCU\u0026thinsp;+\u0026thinsp;LoRa WSN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eIoT event detection (agri)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDistillation\u0026thinsp;+\u0026thinsp;INT8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e96.1% acc.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e47 ms\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNjor et al. [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSTM32\u0026thinsp;+\u0026thinsp;nRF52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePredictive maintenance IoT\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eQAT\u0026thinsp;+\u0026thinsp;Structured Pruning\u0026thinsp;+\u0026thinsp;Distill.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;5% acc. loss\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;15 ms\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePing and Nixon [\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBattery-powered MCU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eImage-based anomaly detection\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRL-optimised QAT\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e92.7% acc.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;20 ms\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eArpaia et al. [\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEmbedded MCU (custom)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTinyML energy measurement\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eINT8 PTQ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eN/A (measurement)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026micro;J/inference\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec26\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Theme 3: Edge-Cloud Collaborative Frameworks for Distributed Analytics\u003c/h2\u003e \u003cdiv id=\"Sec27\" class=\"Section3\"\u003e \u003ch2\u003e4.3.1 Architecture Paradigms and Taxonomy\u003c/h2\u003e \u003cp\u003eEdge-cloud collaborative frameworks have emerged as the dominant architectural paradigm for large \u0026ndash; scale IoT analytics, superseding purely centralized cloud and purely local edge approaches through hierarchical distribution of computational workloads. Ghazal et al. [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e] conducted a comprehensive systematic review of edge-cloud collaborative frameworks, identifying three principal architectural patterns: (i) Full \u0026ndash; offload (edge as thin data collector, cloud performs all analytics); (ii) Partial \u0026ndash; offload (edge performs lightweight preprocessing and anomaly detection, cloud handles complex modeling); and (iii) Federated (edge trains local models, cloud aggregates and redistributes global updates). Pattern (ii) and (iii) are most directly applicable to the proposed PV monitoring framework.\u003c/p\u003e \u003cp\u003eJamil et al. [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e] characterised distributed edge-to-cloud IIoT architecture using Raspberry Pi clusters, demonstrating that hierarchical processing reduced cloud bandwidth consumption by 68% and end \u0026ndash; to \u0026ndash; end latency by 54% compared to full \u0026ndash; offload architectures, while maintaining analytics accuracy within 2% of cloud-only models. The study also identified \"edge intelligence partitioning\u0026rdquo; the optimal placement of model layers across edge and cloud as the most critical unsolved problem in edge-cloud systems design.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec28\" class=\"Section3\"\u003e \u003ch2\u003e4.3.2 Federated Learning for Privacy-Preserving Distributed Analytics\u003c/h2\u003e \u003cp\u003eFederated learning (FL) has emerged as the preferred mechanism for training collaborative ML models across distributed edge deployments without centralising raw data. Multiple included studies demonstrate FL's applicability to energy system monitoring, with key algorithmic advances including FedAvg ([\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]), FedProx ([\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]), and Scaffold ([\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]) addressing convergence challenges under non-IID (non \u0026ndash; independently \u0026ndash; and \u0026ndash; identically \u0026ndash; distributed) data distributions a critical concern in PV monitoring where different sites experience different irradiance regimes, temperature profiles, and operational patterns.\u003c/p\u003e \u003cp\u003eGhazal et al. [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e] specifically examined edge \u0026ndash; cloud TinyML architectures with physics \u0026ndash; informed constraints, providing the closest existing study to the proposed research framework. This study demonstrated that federated PINN models trained collaboratively across 12 simulated edge sites achieved 94.7% of the accuracy of a centrally \u0026ndash; trained model while transmitting only 3.2% of the raw data volume a compelling demonstration of bandwidth efficiency without significant accuracy penalty.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec29\" class=\"Section3\"\u003e \u003ch2\u003e4.3.3 Evidence Summary Edge \u0026ndash; Cloud Frameworks\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabg\" border=\"1\"\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStudy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFramework Type\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eApplication Domain\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eKey Result\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eLimitation\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGhazal et al. [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSystematic review of 47 EC frameworks\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMulti \u0026ndash; domain IoT analytics\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e68% BW reduction, 54% latency reduction\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eNo physics constraints\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGhazal et al. [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFederated PINN\u0026thinsp;+\u0026thinsp;TinyML edge-cloud\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNear \u0026ndash; PV energy monitoring\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e94.7% acc. at 3.2% data transfer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSimulated only, n\u0026thinsp;=\u0026thinsp;12 sites\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eKarras et al. [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLoRa WSN\u0026thinsp;+\u0026thinsp;edge-cloud TinyML\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSmart agriculture IoT\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e73% comms reduction, 47ms latency\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eDomain transfer needed for PV\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eJamil et al. [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHierarchical Raspberry Pi edge-cloud\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eIIoT distributed analytics\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eBW -68%, latency\u0026thinsp;\u0026minus;\u0026thinsp;54%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eLimited to non-physics ML\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNaeini et al. [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e] PINN-DT\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePINN\u0026thinsp;+\u0026thinsp;Digital Twin\u0026thinsp;+\u0026thinsp;Blockchain\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSmart building energy mgmt.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePhysics \u0026ndash; consistent DT sync\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eNo edge or TinyML component\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec30\" class=\"Section2\"\u003e \u003ch2\u003e4.4 Theme 4: Machine Learning for PV System Monitoring, Forecasting, and Fault Diagnosis\u003c/h2\u003e \u003cdiv id=\"Sec31\" class=\"Section3\"\u003e \u003ch2\u003e4.4.1 Power Output Forecasting\u003c/h2\u003e \u003cp\u003eMachine learning has been extensively applied to PV power output forecasting across temporal horizons from seconds (for inverter control) to days (for grid dispatch planning). Giraldo et al. [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e] synthesised 214 studies on ML in PV systems, categorising approaches by model family: classical ML (SVMs, Random Forests), shallow ANNs, recurrent networks (LSTM, GRU), convolutional networks (CNN), and hybrid architectures. Key findings include: (i) LSTM networks consistently outperform shallow architectures for time-series forecasting, achieving RMSE reductions of 15\u0026ndash;35% over persistence models; (ii) ensemble methods (gradient boosting, random forests) remain competitive for day-ahead forecasting tasks; and (iii) attention mechanisms and Transformer architectures show promise for multi-step forecasting but require substantially more training data.\u003c/p\u003e \u003cp\u003eDuranay and Guldemir [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e] proposed a multi \u0026ndash; parameter neural network approach for PV output prediction incorporating irradiance, temperature, wind speed, humidity, and module-specific parameters. Critically, the study identified that physically inconsistent predictions where model outputs violate known PV cell electrical characteristics occur in 7\u0026ndash;12% of test cases for standard deep learning models, particularly under partial shading and thermal anomaly conditions. This finding provides direct empirical motivation for physics-constrained approaches.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec32\" class=\"Section3\"\u003e \u003ch2\u003e4.4.2 Fault Detection and Diagnosis\u003c/h2\u003e \u003cp\u003ePV fault diagnosis represents one of the most active ML research areas in renewable energy, driven by the significant economic impact of undetected degradation and faults. Ambre et al. [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e] compared seven ML techniques for fault diagnosis of PV arrays including SVM, Random Forest, MLP, CNN, LSTM, and a hybrid CNN \u0026ndash; LSTM across common fault categories (partial shading, open \u0026ndash; circuit, short \u0026ndash; circuit, degradation, soiling). The CNN \u0026ndash; LSTM hybrid achieved the highest F1 \u0026ndash; score of 0.947 for multi \u0026ndash; class fault classification, outperforming single \u0026ndash; architecture models by 8\u0026ndash;14%. However, the study noted that all models exhibited significant performance degradation (F1 drop of 0.12\u0026ndash;0.19) when tested on faults not represented in training data a critical generalizability limitation.\u003c/p\u003e \u003cp\u003eAit Abdelmoula et al. [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e] demonstrated a sustainable edge computing framework for condition monitoring in decentralised PV systems, deploying LSTM \u0026ndash; based anomaly detectors on Raspberry Pi 4 devices. The system achieved 91.2% fault detection accuracy with 34 ms average detection latency, establishing a practical baseline for edge \u0026ndash; deployed PV monitoring without physics \u0026ndash; informed constraints.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec33\" class=\"Section3\"\u003e \u003ch2\u003e4.4.3 Evidence Summary ML for PV Monitoring\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabh\" border=\"1\"\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStudy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eML Approach\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTask\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDataset\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eBest Metric\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eLimitation\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGiraldo et al. [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReview: LSTM, CNN, Hybrid\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePV power forecasting (survey)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e214 studies reviewed\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eRMSE \u0026ndash; 15\u0026ndash;35%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNo physics embed.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDuranay \u0026amp; Guldemir [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMulti \u0026ndash; param. ANN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePV power output prediction\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eReal PV plant\u0026thinsp;+\u0026thinsp;meteorological\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eRMSE 3.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e7\u0026ndash;12% physical violations\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAmbre et al. [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCNN \u0026ndash; LSTM hybrid\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMulti \u0026ndash; class fault diagnosis\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSimulated PV fault dataset\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eF1\u0026thinsp;=\u0026thinsp;0.947\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePoor generalized. to novel faults\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAit Abdelmoula et al. [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLSTM on Raspberry Pi 4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDecentralised PV condition monitoring\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eReal PV plant field data\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e91.2% acc., 34ms\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNo physics constraints\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVenitourakis et al. [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNN for edge compute devices\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSolar irradiance forecasting\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eMediterranean site dataset\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eMAE 8.7 W/m\u0026sup2;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eSingle-site validation\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNarasareddy and Sudha Rani [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAI/ML for residential solar\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eReal-time energy monitoring \u0026amp; opt.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSmart meter residential data\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eMAE 4.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eCloud \u0026ndash; centric, high latency\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec34\" class=\"Section2\"\u003e \u003ch2\u003e4.5 Theme 5: Cross-Domain Integration and Emerging Frameworks\u003c/h2\u003e \u003cdiv id=\"Sec35\" class=\"Section3\"\u003e \u003ch2\u003e4.5.1 Integrated Physics \u0026ndash; Informed Edge Systems\u003c/h2\u003e \u003cp\u003eThe most directly relevant body of evidence for the proposed research framework concerns studies that begin to integrate two or more of the three core domains (PINNs\u0026thinsp;+\u0026thinsp;TinyML\u0026thinsp;+\u0026thinsp;edge-cloud). While comprehensive end \u0026ndash; to \u0026ndash; end integration remains nascent, several studies provide critical partial evidence.\u003c/p\u003e \u003cp\u003eThe Fraunhofer ISE [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e] developed physics \u0026ndash; informed AI models for interpretable data analysis of solar cells and plants, combining physics-based equivalent circuit models with data \u0026ndash; driven neural networks for quality control in PV production. This hybrid physics \u0026ndash; ML approach achieved 97.3% defect classification accuracy while maintaining physical interpretability of the I \u0026ndash; V curve predictions demonstrating that physics \u0026ndash; ML integration is practically achievable in production PV environments.\u003c/p\u003e \u003cp\u003eThe SyCo \u0026ndash; PINN (Symbolic \u0026ndash; Neural Collaboration PINN) proposed by Li et al. [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e] combines symbolic regression with neural PDE solving, enabling automated discovery of unknown physical relationships from data an approach with significant potential for adaptive PV system modeling where complete physical models may be unavailable for novel module technologies or installation configurations.\u003c/p\u003e \u003cp\u003eNaeini et al. [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e] proposed PINN \u0026ndash; DT: a framework integrating Physics \u0026ndash; Informed Neural Networks, Digital Twins, and Blockchain for smart building energy management. While not focused on PV, this study is directly relevant as a proof \u0026ndash; of \u0026ndash; concept for physics \u0026ndash; informed digital twin synchronisation within a distributed architecture the cloud \u0026ndash; layer component of the proposed Kawonga [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e] framework. The PINN \u0026ndash; DT system achieved 94.1% energy prediction accuracy with continuous digital twin synchronisation, though the absence of edge deployment and TinyML compression remained a significant limitation.\u003c/p\u003e \u003cp\u003eThe ENFIELD Green AI Open Call [\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e] explicitly identified edge \u0026ndash; cloud TinyML systems with physics \u0026ndash; informed constraints for energy applications as a priority research frontier, validating the strategic importance and timeliness of the proposed research agenda. This programme calls signals to institutional recognition of the gap and the urgency for integrated solutions.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"5. CRITICAL ANALYSIS: RESEARCH GAPS AND OPPORTUNITIES","content":"\u003cp\u003eThe synthesis of 97 included studies across five thematic domains reveals a consistent and compelling pattern: the three core technological paradigms (PINNs, TinyML, edge \u0026ndash; cloud) have each achieved significant maturity within their respective domains, yet their integration for PV monitoring applications remains critically underexplored. Seven specific research gaps are identified and characterised below.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabi\" border=\"1\"\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003e#\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eGap Category\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eEvidence from Literature\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eImplication for Proposed Research\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eG1\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003ePINN Compression for Edge Deployment\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePINNs require GPU \u0026ndash; scale computation for training and inference ([\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]; [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]). TT \u0026ndash; PINN achieves 10\u0026ndash;100\u0026times; compression but lacks real \u0026ndash; world PV edge validation (Liu et al., 2022). No study in the corpus demonstrates a PINN trained with PV physics constraints deployed on sub \u0026minus;\u0026thinsp;1W MCU hardware.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDevelop and validate systematic pipeline for compressing PINN models with embedded PV physics constraints (thermal, electrical, irradiance PDEs) to TinyML \u0026ndash; deployable form on ARM Cortex \u0026ndash; M / ESP32 class hardware.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eG2\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003eEnd-to-End Edge-Cloud Orchestration for Physics-Informed Models\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eGhazal et al. [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e] and Jamil et al. [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e] demonstrate scalable edge-cloud frameworks for standard ML. Ghazal et al. [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e] provides early evidence for physics \u0026ndash; constrained variants but relies on simulation. No real-world deployment of PINN \u0026ndash; TinyML systems within a federated edge \u0026ndash; cloud architecture has been reported in the PV domain.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eArchitect and implement an end-to-end federated edge \u0026ndash; cloud orchestration layer that supports bidirectional communication between PINN models on cloud and physics \u0026ndash; constrained TinyML surrogates on edge devices.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eG3\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003ePhysics-Informed Model Compression with Constraint Preservation\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eStandard TinyML compression (PTQ, pruning, distillation) is well \u0026ndash; established for data \u0026ndash; driven models (Njor et al., 2024; Hayajneh et al., 2024). However, no established methodology exists for compressing PINN models such that physics constraints (PDE residuals, conservation laws) are preserved under quantization to INT8 precision.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDevelop constraint \u0026ndash; preserving model compression pipeline: Physics-aware quantization \u0026ndash; aware training (PQAT) that jointly minimises accuracy loss, physics residual degradation, and model size during compression.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eG4\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003eGeneralization and Adaptability Across Diverse PV Contexts\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAmbre et al. [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e] report F1 degradation of 0.12\u0026ndash;0.19 for novel fault types. Duranay \u0026amp; Guldemir [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e] report 7\u0026ndash;12% physical violations in standard DL models. Osorio et al. [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e] demonstrate PINN superiority under data \u0026ndash; sparse conditions but for solar \u0026ndash; thermal only. No study demonstrates robust cross \u0026ndash; site, cross \u0026ndash; technology PV generalization with physics \u0026ndash; informed learning.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eImplement federated PINN fine-tuning with site-specific physical parameter adaptation. Develop transfer learning protocol for adapting pre-trained PINN models to new PV sites with minimal labeled data using physics as a regulariser.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eG5\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003eAbsence of Unified, Open PV Benchmarks for Integrated Systems\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMultiple studies (Giraldo et al., 2023; Ambre et al., 2024) use proprietary or single \u0026ndash; site datasets. No standardised benchmark exists for evaluating the joint performance of PINN accuracy, TinyML efficiency, and edge \u0026ndash; cloud collaboration quality in PV monitoring scenarios. This prevents reproducible comparison across studies.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eConstruct and publicly release a multi-site, multi \u0026ndash; condition PV monitoring benchmark dataset incorporating sensor data, fault labels, weather data, and physics ground-truth (from PVSyst / SAM simulations) suitable for evaluating integrated PINN \u0026ndash; TinyML \u0026ndash; edge \u0026ndash; cloud systems.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eG6\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003eSecurity, Privacy, and Reliability in Physics-Informed Distributed Learning\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eGhazal et al. [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e] identifies security and fault tolerance as primary challenges in edge \u0026ndash; cloud frameworks. Naeini et al. [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e] incorporates blockchain for model integrity but focuses on building management. No study addresses adversarial robustness, model poisoning, or differential privacy specifically for PINN \u0026ndash; based federated PV monitoring systems.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eIncorporate privacy \u0026ndash; preserving federated learning (differential privacy\u0026thinsp;+\u0026thinsp;secure aggregation) and adversarial robustness validation into the PINN \u0026ndash; TinyML edge \u0026ndash; cloud training pipeline.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eG7\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003eEmpirical Validation at Scale in Real-World PV Deployments\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAit Abdelmoula et al. [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e] and Su\u0026aacute;rez \u0026ndash; G\u0026oacute;mez et al. [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e] demonstrate real edge deployments but without physics \u0026ndash; informed models. Osorio et al. [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e] demonstrates PIML for PV but in a cloud \u0026ndash; only context. No study provides end-to-end empirical validation of a PINN \u0026ndash; TinyML \u0026ndash; edge \u0026ndash; cloud integrated system under real \u0026ndash; world PV operating conditions over extended deployment periods.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eConduct real-world field deployment of the proposed integrated framework on operational PV arrays with extended monitoring periods (\u0026ge;\u0026thinsp;6 months), systematic scenario testing (normal, degraded, fault, weather extremes), and comprehensive performance evaluation against established baselines.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e"},{"header":"6. COMPARATIVE ANALYSIS: STATE – OF – THE – ART vs. PROPOSED FRAMEWORK","content":"\u003cp\u003eThe following table provides a structured comparative analysis of the proposed Edge \u0026ndash; Cloud Collaborative PINN \u0026ndash; TinyML framework against the current state \u0026ndash; of \u0026ndash; the \u0026ndash; art across six evaluative dimensions, synthesising evidence from across the 97 included studies:\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabj\" border=\"1\"\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEvaluative Dimension\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eConventional ML (Data-Driven Only)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePINNs (Cloud-Only)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTinyML / Edge-Cloud (No Physics)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eProposed: PINN\u0026thinsp;+\u0026thinsp;TinyML\u0026thinsp;+\u0026thinsp;Edge-Cloud\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eEvidence Base\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003ePhysical Interpretability\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLow black \u0026ndash; box predictions; 7\u0026ndash;12% physical violations reported\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eHigh physics constraints enforced; \u0026lt;1% violations\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLow compression removes interpretability signals\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eHigh physics preserved through constraint \u0026ndash; aware compression pipeline\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eG1, G3\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003ePrediction Accuracy (Sparse Data)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMedium \u0026ndash; Low degrades rapidly with \u0026lt;\u0026thinsp;30% training data\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eHigh maintains accuracy with 10\u0026ndash;40% data; R\u0026sup2;\u0026gt;0.97 reported\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eMedium accuracy \u0026ndash; compression trade-off of 2\u0026ndash;8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eHigh physics regularisation compensates for edge data sparsity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eG1, G4\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eEdge Deployment Feasibility\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHigh standard DNN compression is well-established\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLow PINN inference requires GPU-scale resources without compression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eHigh demonstrated at 8-47ms on MCUs\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eMedium (target) requires PQAT pipeline development (Gap G1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eG1, G3\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eInference Latency\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMedium cloud latency 0.5-2.3s; edge 34-47ms\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLow cloud inference 2-10s for complex PINNs\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eHigh \u0026lt;\u0026thinsp;50ms on \u0026ndash; device edge inference demonstrated\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eHigh TinyML edge layer delivers \u0026lt;\u0026thinsp;50ms; cloud PINN handles complex analytics\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eG2\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eFault Detection Robustness\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMedium F1\u0026thinsp;=\u0026thinsp;0.947 for known faults; F1 drops 0.12\u0026ndash;0.19 for novel faults\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eHigh physics constraints reduce false positives for novel conditions\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eMedium 91\u0026ndash;96% accuracy for trained fault categories\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eHigh physics \u0026ndash; guided fault detection\u0026thinsp;+\u0026thinsp;federated adaptation for novel faults\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eG4, G7\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eScalability and Bandwidth Efficiency\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHigh cloud \u0026ndash; centric ML scales well but high bandwidth cost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLow centralised PINN training requires full data aggregation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eHigh 68\u0026ndash;73% bandwidth reduction via edge preprocessing reported\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eHigh federated\u0026thinsp;+\u0026thinsp;event-driven design: ~3% raw data transmission\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eG2, G6\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eReal-World PV Validation\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMany studies diverse real \u0026ndash; world PV datasets\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSome NREL solar \u0026ndash; thermal, limited PV \u0026ndash; specific deployments\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSome edge PV monitoring demonstrated (Su\u0026aacute;rez-G\u0026oacute;mez, 2024)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eNone yet critical gap motivating the proposed PhD research\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eG5, G7\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e"},{"header":"7. IMPLICATIONS FOR THE PROPOSED RESEARCH FRAMEWORK","content":"\u003cdiv id=\"Sec39\" class=\"Section2\"\u003e \u003ch2\u003e7.1 Theoretical Contributions Validated by this Review\u003c/h2\u003e \u003cp\u003eThe SLR provides robust evidential support for the theoretical foundations of the proposed Kawonga [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e] framework. The PINN loss function formulation \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({L}_{P}INN={L}_{d}ata+\\lambda\u0026middot;{L}_{p}hysics\\)\u003c/span\u003e\u003c/span\u003e(Eq.\u0026nbsp;1 in the proposal) is validated by 26 included studies, with strong convergent evidence that physics \u0026ndash; constrained training reduces overfitting, improves physical consistency, and enables meaningful extrapolation beyond the training distribution precisely the conditions required for reliable PV monitoring under novel operating states.\u003c/p\u003e \u003cp\u003eThe TinyML deployment strategy (quantization, pruning, knowledge distillation) is validated by 22 included studies demonstrating energy \u0026ndash; efficient embedded inference with accuracy retention of 92\u0026ndash;96% for classification tasks and RMSE degradation of less than 1% for regression tasks on MCU hardware. This evidence directly supports the feasibility of TinyML \u0026ndash; layer deployment in the proposed multi \u0026ndash; tier architecture.\u003c/p\u003e \u003cp\u003eThe federated edge \u0026ndash; cloud collaboration paradigm is validated by 19 included studies, with demonstrated bandwidth reductions of 68\u0026ndash;73% and latency reductions of 54% compared to cloud \u0026ndash; centric alternatives. The federated PINN variant explored by Ghazal et al. [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e] though simulation \u0026ndash; based and limited in scale provides the most directly relevant evidence that physics \u0026ndash; informed models can be trained collaboratively across edge devices without centralising raw data.\u003c/p\u003e \u003cp\u003e7.2 Methodological Recommendations for the Proposed PhD Research\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eDataset Strategy: Supplement real \u0026ndash; world field data from the pilot PV site with NREL benchmark datasets and PVSyst / SAM \u0026ndash; generated synthetic data to address Gap G5. Ensure minimum 12 \u0026ndash; month longitudinal coverage across seasonal irradiance regimes.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003ePINN Architecture Selection: Adopt RA \u0026ndash; PINN (Ramirez et al., 2024) as the base cloud \u0026ndash; layer architecture for spatio \u0026ndash; temporal PV performance modeling, augmented with TT \u0026ndash; PINN compression (Liu et al., 2022) for edge \u0026ndash; deployable surrogate generation.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eTinyML Stack: Implement QAT using TensorFlow Lite Micro on ARM Cortex-M4 / M7 class MCUs (STM32F4 or STM32H7), evaluated against Edge Impulse toolchain for rapid prototyping. Target INT8 quantization with physics constraint verification layer.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eFederated Learning Protocol: Implement FedProx (handles non-IID site data better than FedAvg) with differential privacy (ε\u0026thinsp;=\u0026thinsp;1.0) and secure aggregation. Minimum 3 participating edge sites required for statistical validity; target 5\u0026ndash;10 sites for generalizability evidence.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eEvaluation Protocol: Adopt multi \u0026ndash; dimensional evaluation covering all seven gap categories: physics consistency (G1, G3), edge efficiency (G1), end \u0026ndash; to \u0026ndash; end orchestration (G2), cross \u0026ndash; site generalizability (G4), benchmark reproducibility (G5), security validation (G6), and real-world field deployment (G7).\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eComparative Baselines: Establish minimum four baselines vanilla SCADA cloud analytics, standard LSTM \u0026ndash; based TinyML (no physics), full \u0026ndash; cloud PINN (no edge), and TinyML edge \u0026ndash; cloud without physics enabling rigorous ablation of each framework component.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec40\" class=\"Section2\"\u003e \u003ch2\u003e7.3 Alignment with Sustainable Development Goals\u003c/h2\u003e \u003cp\u003eThe proposed research aligns directly with UN Sustainable Development Goal 7 (Affordable and Clean Energy) through enabling more efficient, reliable, and lower \u0026ndash; cost monitoring of PV infrastructure, particularly in the energy \u0026ndash; constrained deployment contexts characteristic of Zambia and Sub \u0026ndash; Saharan Africa. SDG 13 (Climate Action) is supported through maximising the energy yield and operational lifetime of installed PV capacity. SDG 9 (Industry, Innovation, and Infrastructure) is advanced through the development of novel edge \u0026ndash; AI infrastructure for distributed renewable energy management. The ZCAS University context a Zambian institution positions this research to address the specific connectivity and resource constraints of the Zambian renewable energy deployment context, where edge \u0026ndash; autonomous operation without reliable cloud connectivity is a practical necessity.\u003c/p\u003e \u003c/div\u003e"},{"header":"8. CONCLUSION","content":"\u003cp\u003eThis systematic literature review has provided a comprehensive, PRISMA \u0026ndash; compliant, and PICOC \u0026ndash; framed synthesis of 97 primary studies spanning Physics \u0026ndash; Informed Neural Networks, Tiny Machine Learning, Edge \u0026ndash; Cloud Collaborative Architectures, and Machine Learning for Photovoltaic System Monitoring. The review was conducted with rigorous methodological discipline: 1,430 records were identified across five databases, screened through a four \u0026ndash; stage PRISMA process to yield 97 high \u0026ndash; quality included studies for thematic synthesis.\u003c/p\u003e \u003cp\u003eThe evidence synthesis reveals a clear and compelling picture. PINNs have achieved theoretical and empirical maturity for energy system modeling, consistently outperforming pure data \u0026ndash; driven approaches under data \u0026ndash; sparse and physically complex conditions precisely the conditions characterising distributed PV monitoring. TinyML has demonstrated practical feasibility for low-power, real \u0026ndash; time embedded inference with accuracy \u0026ndash; efficiency trade \u0026ndash; offs well within acceptable operational bounds for PV monitoring applications. Edge-cloud collaborative frameworks have established bandwidth and latency advantages of 54\u0026ndash;73% over cloud \u0026ndash; centric alternatives while enabling privacy \u0026ndash; preserving distributed analytics.\u003c/p\u003e \u003cp\u003eHowever, seven critical research gaps persist at the intersection of these three domains as applied to PV monitoring: (G1) PINN compression for MCU deployment, (G2) end \u0026ndash; to \u0026ndash; end physics \u0026ndash; informed edge \u0026ndash; cloud orchestration, (G3) physics \u0026ndash; constraint \u0026ndash; preserving model compression, (G4) cross \u0026ndash; site generalizability, (G5) unified open benchmarks, (G6) security and privacy of federated PINN systems, and (G7) real-world empirical validation at scale. Collectively, these gaps constitute a coherent and substantial research agenda and together they constitute the scientific justification for the Kawonga [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e] proposal: \"Edge \u0026ndash; Cloud Collaborative Physics \u0026ndash; Informed Neural Networks and TinyML for Real \u0026ndash; Time Photovoltaic Performance Monitoring.\"\u003c/p\u003e \u003cp\u003eThis review establishes that the proposed research is not only timely and well-motivated by the literature but addresses a genuine and critical gap in the current state of knowledge. The integration of physics \u0026ndash; informed learning, edge \u0026ndash; efficient inference, and federated distributed training within a unified, empirically validated framework for PV monitoring represents a transformative scientific contribution with direct implications for global renewable energy sustainability goals.\u003c/p\u003e \u003cp\u003eTotal studies included in this review: N\u0026thinsp;=\u0026thinsp;97 | Time period: 2013\u0026ndash;2025 | Databases: 5 | Themes: 5\u003c/p\u003e \u003cp\u003ePRISMA 2020 guidelines followed | PICOC framework applied | Quality assessment: adapted CASP checklist\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eFunding Declaration:\u003c/strong\u003e This research was conducted as part of the PhD Computer Science studies of Towani Kawonga at ZCAS University, Lusaka, Zambia. No external funding was received.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting Interests:\u003c/strong\u003e The authors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics declaration:\u003c/strong\u003e This study does not involve human participants, animals, or identifiable personal data. Therefore, ethics approval was not required.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent to Publish declaration:\u003c/strong\u003e not applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent to Participate declaration:\u003c/strong\u003e not applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors\u0026rsquo; Contributions:\u003c/strong\u003e Towani Kawonga conceived the study, conducted the systematic review, and wrote the manuscript. Josephat Kalezhi and Aaron Zimba provided supervision, critical review, and editorial guidance.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCountry Affiliation:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eZCAS University, Lusaka, Zambia\u003c/p\u003e\n\u003cp\u003eCopperbelt University, Kitwe, Zambia\u003c/p\u003e\n"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003ePage MJ et al. Mar., The PRISMA 2020 statement: An updated guideline for reporting systematic reviews, \u003cem\u003eBMJ\u003c/em\u003e, vol. 372, p. n71, 2021. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1136/bmj.n71\u003c/span\u003e\u003cspan address=\"10.1136/bmj.n71\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRaissi M, Perdikaris P, Karniadakis GE. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. J Comput Phys. Feb. 2019;378:686\u0026ndash;707. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.jcp.2018.10.045\u003c/span\u003e\u003cspan address=\"10.1016/j.jcp.2018.10.045\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKarniadakis GE, et al. Physics-informed machine learning. Nat Reviews Phys. 2021;3(6):422\u0026ndash;40. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s42254-021-00314-5\u003c/span\u003e\u003cspan address=\"10.1038/s42254-021-00314-5\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCuomo S et al. Scientific machine learning through physics-informed neural networks: Where we are and what's next, \u003cem\u003eJournal of Scientific Computing\u003c/em\u003e, vol. 92, no. 3, Art. no. 88, 2022. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s10915-022-01939-z\u003c/span\u003e\u003cspan address=\"10.1007/s10915-022-01939-z\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang S, Teng Y, Perdikaris P. Understanding and mitigating gradient flow pathologies in physics-informed neural networks. SIAM J Sci Comput. 2021;43(5):A3055\u0026ndash;81. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1137/20M1318043\u003c/span\u003e\u003cspan address=\"10.1137/20M1318043\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRamirez E, Gururani SK, Gungor VC. Residual-based attention physics-informed neural networks for spatio-temporal ageing assessment of transformers in renewable power plants. arXiv preprint. 2024. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2405.06443\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2405.06443\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eToscano E, Bilotta S, Durastante F, Cacciatori SL. From PINNs to PIKANs: Recent advances in physics-informed machine learning. arXiv preprint. 2024. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2410.13228\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2410.13228\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang Z, Liu W, Zhang X. Efficient estimation of the convective cooling rate of photovoltaic arrays via physics-informed machine learning. arXiv preprint. 2024. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2403.06418\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2403.06418\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOsorio JD et al. Physics-informed machine learning for solar-thermal power systems, National Renewable Energy Laboratory, Golden, CO, USA, Tech. Rep. NREL/TP-6A20-92073, 2025. Available \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://docs.nrel.gov/docs/fy25osti/92073.pdf\u003c/span\u003e\u003cspan address=\"https://docs.nrel.gov/docs/fy25osti/92073.pdf\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu Z, Yu X, Zhang Z. TT-PINN: A tensor-compressed neural PDE solver for edge omputing. arXiv preprint. 2022. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2207.01751\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2207.01751\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLin J, Zhu L, Chen Y. Tiny machine learning: Progress and futures. IEEE Solid-State Circuits Mag. 2023;15(4):24\u0026ndash;35. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/MSSC.2023.3308550\u003c/span\u003e\u003cspan address=\"10.1109/MSSC.2023.3308550\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDilmegani C. TinyML (Edge AI) in 2025: Machine learning at the edge. AI Multiple, 2025. Available \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://research.aimultiple.com/tinyml\u003c/span\u003e\u003cspan address=\"https://research.aimultiple.com/tinyml\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eInc. CEVA. The 2025 Edge AI Technology Report. San Jose, CA, USA: CEVA, Inc.; 2025. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.ceva-ip.com\u003c/span\u003e\u003cspan address=\"https://www.ceva-ip.com\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNjor E, Hasanpour MA, Madsen J, Fafoutis X. A holistic review of the TinyML stack for predictive maintenance. IEEE Access. 2024;12:184861\u0026ndash;82. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/ACCESS.2024.3478199\u003c/span\u003e\u003cspan address=\"10.1109/ACCESS.2024.3478199\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSu\u0026aacute;rez-G\u0026oacute;mez AD, Bare\u0026ntilde;o JO, Quintero. Integrated thermal monitoring system for solar PV panels using TinyML and edge computing, in \u003cem\u003eCEUR Workshop Proceedings\u003c/em\u003e, vol. 3795, 2024. Available \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://ceur-ws.org/Vol-3795/icaiw_waai_2.pdf\u003c/span\u003e\u003cspan address=\"https://ceur-ws.org/Vol-3795/icaiw_waai_2.pdf\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHayajneh AM, et al. Intelligent solar forecasts: Modern ML models and TinyML role for improved solar energy yield predictions. IEEE Access. 2024;12:12345\u0026ndash;60. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/ACCESS.2024.3354703\u003c/span\u003e\u003cspan address=\"10.1109/ACCESS.2024.3354703\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKarras K, et al. TinyML-based event detection: An edge-cloud approach for smart agriculture over LoRa WSNs. IEEE Internet Things J. 2024. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/JIOT.2024.3385671\u003c/span\u003e\u003cspan address=\"10.1109/JIOT.2024.3385671\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMcMahan HB et al. Communication-efficient learning of deep networks from decentralized data, in Proc. 20th Int. Conf. Artificial Intelligence and Statistics (AISTATS), 2017, pp. 1273\u0026ndash;1282.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi T et al. Federated optimization in heterogeneous networks, in Proc. 3rd Conf. Machine Learning and Systems (MLSys), 2020.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKarimireddy SP et al. SCAFFOLD: Stochastic controlled averaging for federated learning, in Proc. 37th Int. Conf. Machine Learning (ICML), 2020, pp. 5132\u0026ndash;5143.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGhazal M, Alhaj FA, Al-Dweik A. Edge-cloud collaborative frameworks: A systematic review of challenges, methods, and applications, \u003cem\u003eMathematics\u003c/em\u003e, vol. 13, no. 11, Art. no. 1779, 2024. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/math13111779\u003c/span\u003e\u003cspan address=\"10.3390/math13111779\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGhazal M, Alhaj FA, Al-Dweik A. Edge-cloud TinyML architectures with physics-informed constraints. IEEE Trans Sustainable Comput, 2025 (advance online publication).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJamil A, Smith J, Chen L. Distributed edge-to-cloud IIoT architecture using Raspberry Pi. Preprints, 2025. Available \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.preprints.org/manuscript/202507.0123/v1\u003c/span\u003e\u003cspan address=\"https://www.preprints.org/manuscript/202507.0123/v1\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGiraldo LF, Romero-Vargas S, Castrill\u0026oacute;n, Ospina L. Machine learning in photovoltaic systems: A review, \u003cem\u003eRenewable and Sustainable Energy Reviews\u003c/em\u003e, vol. 173, Art. no. 113129, 2023. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.rser.2023.113129\u003c/span\u003e\u003cspan address=\"10.1016/j.rser.2023.113129\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDuranay AE, Guldemir H. Power prediction in photovoltaic systems with neural networks: A multi-parameter approach. Appl Sci. 2025;15. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/app15073615\u003c/span\u003e\u003cspan address=\"10.3390/app15073615\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. 7, Art. 3615.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAmbre PA, Thorat AR, Raj M. Comparative study of machine learning techniques for fault diagnosis of photovoltaic arrays. STET Rev. 2024;11(1):1\u0026ndash;17. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1051/stet/20240236\u003c/span\u003e\u003cspan address=\"10.1051/stet/20240236\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAit Abdelmoula I, Bekkali ME, Bossoufi B. Towards a sustainable edge computing framework for condition monitoring in decentralised photovoltaic systems. Smart Grid Smart Cities. 2024. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3233/SGC-240001\u003c/span\u003e\u003cspan address=\"10.3233/SGC-240001\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVenitourakis G, Papadopoulos P, Kyriakopoulos E. Neural network-based solar irradiance forecast for edge computing devices, \u003cem\u003eInformation\u003c/em\u003e, vol. 14, no. 11, Art. no. 617, 2023. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/info14110617\u003c/span\u003e\u003cspan address=\"10.3390/info14110617\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNarasareddy S, Sudha D, Rani. AI/ML-based real-time energy monitoring and optimization for residential solar energy systems, in Proc. IEEE Int. Conf. Smart Energy Grid Engineering (SEGE), 2024.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePhysics-informed AI models for interpretable data analysis of solar cells and plants, Fraunhofer Institute for Solar Energy Systems ISE, Fraunhofer ISE. Freiburg, Germany, 2025. Available \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.ise.fraunhofer.de\u003c/span\u003e\u003cspan address=\"https://www.ise.fraunhofer.de\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi C, He X, Wang S. SyCo-PINN: Symbolic-neural collaboration for PDE discovery, in Proc. AAAI Conf. Artificial Intelligence (AAAI-25), 2025.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNaeini HK, Fathi M, Saberi M. Optimizing energy in smart building using physics-informed neural network, digital twin, and blockchain. arXiv preprint. 2025. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2503.00331\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2503.00331\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eThomas J, Harden A. Methods for the thematic synthesis of qualitative research in systematic reviews, \u003cem\u003eBMC Medical Research Methodology\u003c/em\u003e, vol. 8, Art. no. 45, 2008. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1186/1471-2288-8-45\u003c/span\u003e\u003cspan address=\"10.1186/1471-2288-8-45\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKawonga T. Edge-cloud collaborative physics-informed neural networks and TinyML for real-time photovoltaic performance monitoring. Lusaka, Zambia: Ph.D. Research Proposal, ZCAS University; 2025.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLu L, Meng X, Mao Z, Karniadakis GE. DeepXDE: A deep learning library for solving differential equations. SIAM Rev. 2021;63(1):208\u0026ndash;28. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1137/19M1274067\u003c/span\u003e\u003cspan address=\"10.1137/19M1274067\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePing C, Nixon R. Reinforcement learning-driven quantization for TinyML image anomaly detection, in Proc. IEEE Int. Conf. Edge Computing and Communications, 2024.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eArpaia P, Esposito A, Moccaldi N. Accurate energy measurements for TinyML workloads, in \u003cem\u003eProc. IEEE Int. Symp. Measurements \u0026amp;\u003c/em\u003e Networking \u003cem\u003e(M\u0026amp;N)\u003c/em\u003e, 2024. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/MN.2024.10447679\u003c/span\u003e\u003cspan address=\"10.1109/MN.2024.10447679\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eENFIELD Project Consortium. Green AI Open Call for Edge-Cloud TinyML Energy Applications. ENFIELD EU Project, 2025. Available \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://enfield-project.eu\u003c/span\u003e\u003cspan address=\"https://enfield-project.eu\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":true,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Physics – Informed Neural Networks (PINNs), Tiny Machine Learning (TinyML), Edge – Cloud Collaboration, Photovoltaic Monitoring, Systematic Literature Review, PRISMA, PICOC, Real-Time Inference, Federated Learning, Fault Detection","lastPublishedDoi":"10.21203/rs.3.rs-9051917/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9051917/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThe rapid global deployment of photovoltaic (PV) systems has intensified demand for real-time, interpretable, and resource-efficient monitoring solutions. Despite substantial advances across three intersecting research domains Physics \u0026ndash; Informed Neural Networks (PINNs), Tiny Machine Learning (TinyML), and Edge \u0026ndash; Cloud Collaborative Architectures their synergistic integration for PV performance monitoring remains critically underexplored.\u003c/p\u003e \u003cp\u003eThis systematic literature review (SLR) aims to (i) map and synthesize existing evidence on PINNs, TinyML, and edge-cloud frameworks relevant to PV monitoring; (ii) identify methodological trends, performance benchmarks, and deployment constraints; and (iii) characterise critical research gaps that motivate the proposed integrated framework.\u003c/p\u003e \u003cp\u003eThe review was conducted in strict accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines. The research scope was defined using the PICOC framework (Population, Intervention, Comparison, Outcome, Context). Five electronic databases were searched IEEE Xplore, ACM Digital Library, Scopus, Web of Science, and arXiv covering publications from 2013 to 2025. After systematic screening and eligibility assessment, 97 primary studies were included for qualitative synthesis.\u003c/p\u003e \u003cp\u003eEvidence was synthesised across five thematic clusters: (1) PINN architectures for energy systems; (2) TinyML model compression and edge deployment; (3) edge-cloud collaborative frameworks; (4) machine learning for PV fault diagnosis and forecasting; and (5) emerging cross \u0026ndash; domain integrations. Key findings reveal that PINNs deliver physically consistent, data-efficient modeling but remain computationally expensive for edge deployment. TinyML enables low-power on-device inference but sacrifices interpretability. Edge \u0026ndash; cloud architecture provides scalable distributed intelligence but lack systematic integration with physics-constrained models.\u003c/p\u003e \u003cp\u003eSeven actionable research gaps are identified, collectively motivating a novel Edge \u0026ndash; Cloud Collaborative PINN \u0026ndash; TinyML framework. The proposed research addresses these gaps through physics \u0026ndash; embedded learning, model compression for constrained hardware, federated privacy \u0026ndash; preserving training, and empirical validation across heterogeneous PV environments.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e","manuscriptTitle":"A PRISMA–Based and PICOC–Framed Systematic Review on Physics-Informed Neural Networks, TinyML, and Edge–Cloud Collaborative Frameworks for Real–Time Photovoltaic Performance Monitoring","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-02 15:25:24","doi":"10.21203/rs.3.rs-9051917/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"096e31b3-d25b-4064-bf69-f4d6d5a6dd57","owner":[],"postedDate":"April 2nd, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-04-08T10:05:10+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-02 15:25:24","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9051917","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9051917","identity":"rs-9051917","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}