WBML-PV: Window-based machine learning for ultra-short-term photovoltaic power forecasting

Abstract

Accurate ultra-short-term photovoltaic (PV) power forecasting is essential for grid management and the integration of renewable energy. However, the stochastic and volatile nature of PV power, along with inherent uncertainty, challenges stable grid operation as PV penetration grows. Currently, deep learning (DL) and reinforcement learning (RL) models often struggle to generalize under new conditions, manage computational demands, and address the uncertainty in PV forecasting. To address these issues, a window-based machine learning (WBML) approach is proposed, utilizing light gradient boosting machine (WB-LGBM) and extreme gradient boosting (WB-XGBoost) models. These proposed models outperform attention-based and non-attention-based RL and DL baselines in deterministic metrics like mean absolute error (MAE) and $R^2$, while significantly reducing training time. Optimized via Optuna and evaluated using fuzzy C-means clustering, their performance is validated by the Diebold–Mariano test. Uncertainty is assessed using non-parametric kernel density estimation (NPKDE) and confidence intervals (CIs) at 99%, 95%, 90%, and 80% confidence levels within the WBML framework, demonstrating robust and conservative forecast uncertainty quantification. Amplitude and phase errors are analyzed with standard deviation error, bias, dispersion, skewness, and kurtosis. The models demonstrate reduced imbalance penalties and enhanced revenue through improved forecasting accuracy.