A Leakage-Controlled Meteorology-Aware Framework for Multi-Horizon Photovoltaic Power Forecasting

Abstract

Short-term photovoltaic (PV) power forecasting sup-ports reserve scheduling, storage control, and renewable-energy dispatch. This paper presents a leakage-controlled meteorology-aware framework for multi-horizon PV power forecasting. This study provides a detailed PV site and dataset description, clarifies unavailable sensor and inverter metadata, explains the physical meaning of meteorological variables, and reports validation-set-based model selection. A two-year 5-min Sydney PV dataset from 2022–2023 is evaluated using chronological splits, train-set-only normalization, and split-internal sliding-window construction. Persistence, linear, SVR, random forest (RF), histogram gradient boosting (HGB), XGBoost (XGB), compact LSTM, and tuned LSTM baselines are compared. For 5-min hold-out testing, HGB, XGB, RF, and tuned LSTM obtain 11.836, 11.847, 11.910, and 12.108 kW RMSE, respectively. Additional 2023 seasonal and GHI-level evaluations show that errors vary substantially across seasons and irradiance regimes. As an additional external-site reproducibility check, an external Stanford 30-kW rooftop PV dataset from 2017–2019 is collected and evaluated using history-only inputs. These results show that recurrent models are competitive, but strong tree-based baselines and leakage-controlled protocols remain necessary for credible PV forecasting evaluation.