A scalable forecasting framework for PV systems using hyper-tuned regressors and environmental data

Abstract

Forecasting photovoltaic (PV) power output is essential for reliable grid integration, operational planning, and supporting the global transition toward renewable energy. This paper proposes an integrated machine learning framework that improves prediction accuracy through systematically designed preprocessing, model selection, and advanced hyperparameter optimization. Using a high-resolution dataset from the Sharda University PV system, 13 regression models, including ensemble methods and neural networks, are tested and compared with the aim of maximizing generalizability and predictive performance. Performance gains are achieved through structured hyperparameter optimization using Randomized Search Cross-Validation (RSCV) and Grid Search Cross-Validation (GSCV), where the Random Forest Regressor achieved an R 2 of 0.9561 before tuning and 0.9893 after tuning, representing the highest improvement. Gradient Boosting Regressor and K-Nearest Neighbours also benefited from hyperparameter optimization. A comparative study with benchmark approaches shows that the optimized models in this work are superior in both predictive accuracy and computational efficiency. The proposed framework is scalable, as it can be adapted to different PV datasets while requiring fewer computational resources than deep learning methods, thereby bridging the gap between traditional machine learning approaches and practical energy management systems. • Proposed a novel "Hyper Tune" optimization framework for enhancing PV power forecasting accuracy through systematic hyperparameter tuning. • Evaluated and compared 13 state-of-the-art machine learning regression models, including ensemble and neural network-based methods, using high-resolution real-world data from a PV system. • Achieved exceptional prediction performance with the Random Forest Regressor, reaching an R 2 score of 0.9893 after hyperparameter optimization using RSCV and GSCV. • Integrated rigorous data preprocessing, feature engineering, and correlation-based feature selection to improve model generalizability and robustness. • Demonstrated superior accuracy and computational efficiency compared to existing state-of-the-art models, establishing a scalable and practical framework for real-time solar energy forecasting.