An Evolutionary Stacked Ensemble for Improving Accuracy and Computational Efficiency in Daily Photovoltaic Energy Prediction

This study presents a computationally efficient stacked ensemble model for daily photovoltaic energy prediction. The ensemble combines five base regressors, including extreme gradient boosting, decision trees, random forests, neural networks, and k-nearest neighbors. A weighted linear combination merges these outputs, with weights optimized through differential evolution. The proposed method is evaluated against Bayesian optimization and equal-weighted ensembles, focusing on prediction accuracy and specific optimizer computational costs. While Bayesian optimization is widely used in solar forecasting, results show that differential evolution outperforms it with a root mean square error of 77.019 kWh and a coefficient of determination of 96.215% while converging in as few as 30 iterations. In contrast, Bayesian optimization plateaued without improvement after 100 iterations and incurred 31 to 121 times higher optimizer overhead than differential evolution. This highlights the superior scalability of differential evolution, particularly as the computational cost of Gaussian process-based updates increases with sample size. Seasonal evaluation shows highest accuracy in winter, and validation using external data confirms the generalizability of the method. Overall, the results demonstrate that differential evolution remains a robust and scalable solution for big-data solar energy prediction applications.