Enhancing solar power prediction using machine learning: A comparative study of ensemble approaches with explainable AI insights

Solar power prediction is important for effective energy planning and grid stability. However, it is a complex process owing to changing atmospheric conditions and the nonlinear characteristics of photovoltaic (PV) systems. This paper compares the efficacy of various machine learning methods based on their ability to predict solar power by assessing the performance of 6 ensemble models: Extra Trees, Random Forest (RF), Gradient Boosting Machine (GBM), LightGBM, Extreme Gradient Boosting (XGBoost), and Categorical Boosting. The study uses a high-resolution dataset collected from a PV installation at UTP Solar Research Park, Malaysia, located at a latitude of 4.361° and a longitude of 100.98°. The dataset covers the period from 2 June 2025 to 2 July 2025 and consists of 51,714 samples incorporating time-series, meteorological, and electrical variables, including irradiance, ambient temperature, humidity, PV surface temperature, current, voltage, and power, measured using the site’s PV monitoring and data acquisition system. Coefficient of determination (R 2 ), root mean square error (RMSE), mean absolute error (MAE), and Akaike information criterion (AIC) are used to evaluate the models’ performance. All models showed high predictive power. Nonetheless, RF proved more generalized, with the best test R 2 of 0.9849, RMSE of 0.7442 W, MAE of 0.5374 W, and the lowest AIC of 34760.0520. XGBoost yielded the best overall R 2 of 0.9949 and a near-perfect training R 2 of 0.9999. However, there were considerable differences between the training and testing performances, which implied overfitting. Permutation importance and Shapley Additive Explanations values were consistent in identifying Hour and Humidity as the most significant predictors, while Surface Temperature played a moderate role. The results of the feature importance analysis are consistent with PV and atmospheric principles, suggesting the validity of the models.