Photovoltaic power generation forecasting with Bayesian optimization and stacked ensemble learning
Mohamed A. Atiea, Abdelrhman A. Abdelghaffar, Houssem Ben Aribia, Ferchichi Noureddine, Abdullah M. Shaheen
Abstract
• Introduced Bayesian Optimization through the Optuna framework in PV energy forecasting. • Developed preprocessing techniques for missing values, outliers, and feature engineering. • Designed a Stacked Model with tuned RFR, GBR, and KNN with LR as a meta-learner. • Validated findings on real-world PV datasets of Sharda University, India and DKASC, Australia. • Achieved R² scores of 0.9777 (Sharda) and 0.9967 (DKASC), surpassing previous approaches. Accurate photovoltaic (PV) power forecasting is essential for optimizing energy management and ensuring grid stability in renewable energy systems. This study leverages advanced machine learning regression models and hyper-parameter optimization techniques to enhance predictive accuracy for PV power generation. Using two distinct datasets, the Sharda University PV dataset (2022 Edition) and the DKASC Hanwha Solar dataset, this research evaluates model performance across diverse geographic and climatic conditions. Extensive data preprocessing addressed missing values, outliers, and feature engineering. Among the evaluated models, Random Forest Regressor (RFR) and Gradient Boosting Regressor (GBR) demonstrated superior baseline performance, achieving accuracies of 95.61 % (Sharda) and 99.61 % (DKASC). Bayesian Optimization via the Optuna framework was employed to fine-tune hyperparameters, resulting in significant improvements. The optimized RFR achieved R² scores of 0.9777 (Sharda) and 0.9967 (DKASC), while a Stacked Model combining tuned RFR, GBR, and KNN further enhanced accuracy, achieving R² scores of 0.9788 (Sharda) and 0.9970 (DKASC). These findings highlight the importance of robust preprocessing, effective feature selection, and advanced optimization in improving machine learning model performance for renewable energy forecasting. The proposed methodology offers a scalable framework for integrating machine learning into PV systems, supporting the transition to sustainable energy solutions.