Physics-informed stacking ensemble with random forest for short-term photovoltaic power forecasting

High penetration of photovoltaic (PV) generation introduces strong intermittency and uncertainty into power systems, making accurate short-term PV power forecasting essential for secure operation and economic dispatch. Conventional statistical models and single machine-learning predictors often struggle to capture the nonlinear relationships among historical power, meteorological variables, and numerical weather prediction (NWP) outputs, especially under rapidly changing cloud conditions. This paper proposes a physics-informed stacking ensemble framework for short-term PV power forecasting at a grid-connected plant. A multisource feature set integrates historical PV power, on-site meteorological measurements, NWP-derived variables, and time-solar-geometry descriptors, with normalization by installed capacity or clear-sky irradiance. Several heterogeneous base regressors, including Random Forest, gradient boosting, support vector regression, and a lightweight neural network, are trained using a sliding-window scheme, and their out-of-fold predictions are combined by a meta-learner to improve accuracy. A physics-informed post-processing layer enforces non-negativity, capacity limits, and night-time behavior. Experiments on real PV power and weather data demonstrate that the proposed method consistently outperforms persistence, individual base models, and simple averaging ensembles in terms of root-mean-square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and R2. Improved tracking of ramps and reduced forecast errors highlight the value of the approach for reserve scheduling, intra-day dispatch, and storage management in PV-integrated power systems.