Dynamic stacking ensemble hybrid model for enhanced short-term photovoltaic power forecasting with self-organizing maps and advanced deep learning

To address the challenges posed by the inherent volatility of photovoltaic (PV) power generation on grid stability, this paper introduces the Dynamic Stacking Ensemble Hybrid Model (DSEHM). By integrating hybrid deep neural networks, attention-based mechanisms, tree-based models, and dynamic model selection, DSEHM enhances the prediction accuracy of non-stationary PV power series. Key components include advanced models such as Informer, Attention-Enhanced Gated Recurrent Unit (AttnGRU), and Temporal Convolutional Network (TCN). Dimensionality reduction is performed using a Self-Organizing Map (SOM), preserving topological relationships, while Gaussian Mixture Models (GMM) clustering aids in selecting optimal models for stacking. Experimental results demonstrate that DSEHM outperforms standalone models, achieving significant improvements in prediction accuracy. For instance, in the Austria dataset’s 1-step forecast, Mean Absolute Error (MAE), Relative Squared Error (RSE), and Root Mean Squared Error (RMSE) decreased by 22.55 %, 36.73 %, and 19.01 %, respectively. Furthermore, a comparative study with previous approaches further validates the effectiveness of DSEHM. These findings highlight DSEHM’s potential as a robust tool for improving PV power forecasting, with broader implications for renewable energy grid integration.