Weather-guided decomposition and hard-routing expert modeling for short-term photovoltaic power forecasting

Abstract

Solar energy has become a pivotal renewable resource in modern power systems; however, the stochastic and weather-dependent variability of photovoltaic (PV) generation continues to challenge accurate power forecasting and grid operation. This study proposes a weather-adaptive forecasting framework that integrates meteorological classification, signal decomposition, and hybrid deep learning for short-term PV power prediction. Specifically, a Gaussian Mixture Model (GMM) partitions historical samples into homogeneous weather subsets, after which Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) is applied to each subset to extract intrinsic mode functions (IMFs). A weather-guided hard-routing expert framework is then employed. Three long short-term memory network-Transformer(LSTM–Transformer) experts with identical architectures but independent parameters are trained separately using the sunny, cloudy, and rainy subsets. For each forecasting day, a rule-based weather router activates the expert associated with the dominant weather category. The activated expert jointly models meteorological variables and IMF components to represent weather-induced temporal variations and cross-time dependencies. Experimental results on a two-year PV dataset demonstrate that, compared with the baseline models, the proposed method reduces the average mean absolute error (MAE) and root mean square error (RMSE) by 33.7% and 30.2%, respectively, while achieving an average coefficient of determination (R²) of 98.69%.These results demonstrate the effectiveness of the proposed framework in the evaluated PV power forecasting task.