A generalized machine learning approach for multi-site solar power forecasting using histogram-based gradient boosting

Abstract

The integration of distributed photovoltaic (PV) systems into modern smart grids requires highly accurate, low-latency power forecasting. Usually, forecasting models are trained on a per-site basis, which scales worse across large grid networks. This paper proposes a Generalized Capacity Factor machine learning framework capable of predicting the power output of numerous distinct solar arrays using a single unified model. By normalizing the target variable against the physical capacity (kWp) of each site and combining it with localized meteorological data, the model effectively learns the underlying hardware efficiency regardless of the installation scale. To determine the optimal architecture, an empirical evaluation of four State-of-the-Art (SOTA) tree-based algorithms – Histogram-based Gradient Boosting Regressor (HGBR), XGBoost, LightGBM, and Random Forest – was conducted using a dataset of 42 PV sites. The results demonstrate that the proposed HGBR model yields the highest predictive accuracy, achieving a Root Mean Square Error (RMSE) of 6.12 kW and an R² score of 0.67, significantly outperforming traditional bagging ensembles. This generalized approach offers grid operators a scalable, high-accuracy, and computationally efficient tool for managing decentralized renewable energy resources.