Improving Photovoltaic Power Forecasting Accuracy by Integrating Aerosol Optical Features: A Dual-Channel Deep Learning Approach

This paper proposes a short-term photovoltaic (PV) power prediction method that integrates aerosol optical feature mining with a dual-channel attention mechanism to address the complex non-linear attenuation effects of atmospheric aerosols and the limitations of existing models in handling sudden meteorological changes and aerosol evolution. Using the optical properties of aerosols and clouds (OPAC) database, a high-dimensional aerosol optical feature set is constructed, which is subsequently optimized using the minimum redundancy maximum relevance (mRMR) algorithm. The prediction scenarios are categorized into polluted and clean regimes through K-means clustering. A dual-channel encoder–decoder network, combining bidirectional long short-term memory (BiLSTM) and iTransformer, is developed to capture high-frequency meteorological volatility and low-frequency aerosol evolution. A bidirectional cross-attention mechanism enables deep feature interaction between the optical and meteorological channels. The method is validated using in situ measurements from a PV station in Hebei, China, along with aerosol data from the Copernicus Atmosphere Monitoring Service (CAMS) and meteorological data from the ECMWF Reanalysis v5 (ERA5). Experimental results demonstrate an average reduction of approximately 29.83% in mean absolute error (MAE) on polluted days and 15.22% on clean days. Interpretability analysis reveals distinct physical mechanisms driving the predictions, emphasizing the role of extinction on polluted days and scattering on clean days.