A Study on LSTM-Based PM2.5 Forecasting with Increased Training Data Volume in Seoul, Korea

PM2.5 air pollution has become a critical environmental concern in Korea, requiring accurate forecasting systems for effective air quality management and public health protection. Numerical models face limitations including high computational costs, complex processes, and parameterization uncertainties. This study developed an LSTM-based PM2.5 forecasting model for the Seoul area using integrating observational data and numerical model outputs to overcome the limitations. To address input data gaps in future time steps where observational data are unavailable, WRF-CMAQ model outputs are incorporated as supplementary inputs. Three LSTM models with different training periods are developed: T3V19(3-year), T5V21(5-year), and T6V22(6-year training). Performance evaluation during January-March 2023 demonstrated significant improvements over the CMAQ model. The T6V22 model achieves a 96% improvement in NMB (1.3 vs. 32% for CMAQ), meeting “Goal” benchmark criteria. The correlation coefficient increased from 0.79 to 0.85, while NME decrees from 43.3% to 22.6%. LSTM models consistently outperformed conventional numerical models across all forecast lead times (D+0, D+1, D+2). The results suggest that as input data volume increases, model performance becomes more superior and enables more stable air quality predictions, providing a promising framework for operational forecasting systems