A Metrological System Architecture for AI-Driven Digital Twins: Uncertainty Propagation in Smart Grid Load Forecasting

Load profile forecasting and aggregation are essential for power system planning and operation, yet traditional deterministic AI models often function as black boxes, neglecting the rigorous quantification of input uncertainties. This study proposes a Software-in-the-Loop (SIL) Digital Twin architecture that integrates the Guide to the Expression of Uncertainty in Measurement (GUM) directly into the computational pipeline. Utilizing a Long Short-Term Memory (LSTM) forecasting core and Monte Carlo simulations, the system propagates uncertainties originating from physical measurement noise and SCADA data imputation. To establish a traceable metrological baseline, initial validation is conducted using a highly controlled synthetic load profile at a 15-minute granularity. Our results reveal that degraded input data quality can account for up to 40 % of the total prediction variance during high-volatility periods, exposing the "false confidence" inherent in deterministic point predictions. By outputting a probabilistic mean enveloped by a 95% coverage uncertainty band, this Digital Twin framework establishes a human-mediated closed loop, empowering "human-on-the-loop" operators to execute risk-informed decisions and safeguard grid stability. Given the importance of effective uncertainty propagation for reliable power system operation, informed decision-making, and risk mitigation, this study aims to develop artificial intelligence and/or machine learning (AI/ML) based load profile forecasting and aggregation models. This initial investigation assesses the models’ potential as digital representations to assist operators, specifically considering how uncertainty propagation can be modeled and assessed within them