Neural Network Models for Solar Irradiance Forecasting in Polluted Areas: A Comparative Study

ABSTRACT Increasing global energy demand and renewable energy expansion have heightened the importance of accurate solar irradiance forecasting for effective grid management and capacity planning. Atmospheric pollution significantly affects solar irradiance measurements, requiring air quality integration for precise forecasting in polluted urban environments. This study develops a comprehensive multi‐city data set spanning eight geographically diverse locations with systematically categorized pollution levels, from pristine environments (Copenhagen, Sydney) to heavily polluted urban centers (Beijing, New Delhi, Lahore). A pollution‐aware neural network training methodology is introduced, representing the first systematic investigation of ensemble model performance across explicitly categorized atmospheric quality levels. The study presents a novel ensemble architecture integrating multi‐layer perceptrons, recurrent neural networks, and nonlinear autoregressive with exogenous inputs, specifically designed for forecasting under varying atmospheric pollution conditions. The proposed ensemble model achieves superior performance with R² of 0.8702, RMSE of 1.0809, and MAE of 0.8137, consistently outperforming individual models across all pollution categories and geographical locations. Validation using the HI‐SEAS data set confirms superiority over three contemporary state‐of‐the‐art methodologies. The framework incorporates SHapley Additive exPlanations (SHAP) analysis for model interpretability and comprehensive cross‐validation procedures. This study establishes a foundational framework for pollution‐aware solar forecasting, addressing critical gaps regarding atmospheric variability's impact on prediction accuracy.