Modeling Atmospheric Radiative Transfer for Solar Energy Harvesting: A Chemical-Informed Deep Learning Approach with CNN-LSTM-Attention Networks

Abstract

Accurate prediction of solar irradiance is fundamentally a problem of modeling radiative transfer through a dynamic, multi-component chemical system—the atmosphere. Variability in atmospheric composition and physical state, dictated by parameters like humidity, pressure, and aerosol content, poses a significant challenge for optimizing photovoltaic (PV) energy conversion, a key technology in green chemistry initiatives. This study presents a chemically informed deep learning model that interprets meteorological time-series data as proxies for atmospheric state to forecast sky radiance. We develop a hybrid architecture combining Convolutional and Bidirectional Long Short-Term Memory networks (CNN-BiLSTM) with an attention mechanism to extract spatiotemporal features and identify critical historical time steps governing irradiance. The model, trained on high-resolution monitoring data, achieves a Mean Absolute Error (MAE) of 30 W/m² and an R² score of 0.78 for single-step prediction. Interpretation via LIME (Local Interpretable Model-agnostic Explanations) reveals the dominant influence of recent radiance and temperature features, validating the model's alignment with physical expectations. This work demonstrates how deep learning can decode complex atmospheric chemistry interactions to provide actionable intelligence for solar panel angle optimization, enhancing the efficiency of solar-to-electrical energy conversion.