A dual-wing hybrid model with stacking and short-term optimization for day-ahead photovoltaic power forecasting

Photovoltaic (PV) power generation offers significant potential due to its clean and renewable nature. To address the challenges posed by the volatility of PV generation on grid integration and scheduling, this paper proposes a hybrid forecasting method for 24-h-ahead PV power forecasting. We first use the Spearman correlation coefficient and feature selection experiments to improve data input quality. Subsequently, we introduce the Dual-Wing Collaborative Model (DWCM), incorporating a stacking algorithm, GRU-CNN-SE-LR, in the left-wing module. This module leverages the strengths of the Gated Recurrent Unit, Convolutional Neural Network, and the Squeeze-and-Excitation attention mechanism for initial forecasts. In the right-wing module, we integrate an Extreme Gradient Boosting (XGB) model optimized through random grid search (RGS), referred to as RGS-XGB, specifically tailored for short-term predictions. This model enhances accuracy by optimizing left-wing forecasts. The left wing focuses on longer-term forecasting (24 h), while the right wing addresses shorter-term predictions (12 h). By partially substituting left-wing predictions with those from the right wing, the DWCM demonstrates exceptional day-ahead forecasting performance. Validation experiments using historical data of varying resolutions confirm the DWCM's superiority over five advanced models in terms of mean absolute error, mean squared error, and relative absolute error, validating the advantages of the foundational model and short-term optimization strategy.