Explainable optimized voting ensemble for photovoltaic power forecasting

Forecasting photovoltaic (PV) power accurately is essential for enhancing grid reliability and optimizing energy management in renewable power systems. This study proposes an optimized ensemble-based regression framework for improving PV power predictions. Three machine learning models (LSTM, CNN-LSTM, and SVR) are utilized to generate baseline forecasts. To enhance predictive performance, particle swarm optimization (PSO) is utilized for hyperparameter tuning, ensuring optimal model configurations. Furthermore, a weighted ensemble strategy is introduced, where simple voting and grid search-based ensemble voting are compared to refine final predictions. Experimental results demonstrate that the optimized grid search ensemble model achieves superior forecasting accuracy, with MSE of 163.02 kW 2 , RMSE of 12.77 kW, nRMSE of 1.55%, rRMSE of 6.31%, MAE of 5.26 kW, and R 2 of 0.9976. The scalability and robustness of the model are tested by utilizing data from different regions and a variable cloud analysis respectively, yielding superior results. These findings highlight the critical role of hyperparameter optimization and ensemble weighting in enhancing solar PV power forecasting, offering a robust framework for grid operators and energy planners to improve decision-making in solar-integrated power systems. • This work employs LSTM, CNN-LSTM, and SVR models for initial PV power forecast. • The PSO algorithm is employed to optimize hyperparameters of the forecasting models. • A simple voting regressor approach is implemented to enhance the initial forecasts. • Grid search is implemented to improve the simple voting regressor forecasts. • SHapley Additive exPlanations is adopted to explain model predictions.