Forecasting Solar Photovoltaic Power Generation: A Machine Learning Time Series Model Approach

This article presents a novel hybrid machine learning time series model (MLTSM) for predicting the electrical output of solar photovoltaic (PV) systems, integrating a physics‐based theoretical model with an ensemble of data‐driven regressors. The study addresses the challenge of solar energy’s variability by enhancing predictability for grid integration. Using a 34‐day dataset from two solar power plants in India, we engineer critical features—including irradiation and ambient temperature, transformed via a third‐degree polynomial derived from PV system physics—to improve forecasting accuracy. We conduct a comprehensive evaluation of multiple machine learning (ML) models, including linear regression, ridge regression, decision trees (DTree), random forest (RForest), and K‐nearest neighbors, and propose a weighted hybrid ensemble that combines the top performers. Among the individual models, linear and ridge regression demonstrated superior performance. The proposed hybrid model achieved a notable R 2 value of 98% for Plant 1 and 91% for Plant 2, with root mean squared errors (RMSEs) of 36–66 and 42–127, respectively. This study contributes a publicly available dataset, a novel physics‐informed feature engineering methodology, and a scalable hybrid forecasting framework that offers a practical balance of accuracy, computational efficiency, and interpretability for real‐world solar energy forecasting.