Horizon-Dependent Performance of Tree-Based Machine Learning Models for Short-Term Solar PV Forecasting

Abstract

Short-term photovoltaic (PV) generation forecasting across multiple horizons, from sub-hourly dispatch to hourly storage scheduling, is critical for modern grid operations, yet systematic multi-horizon comparisons within a unified framework are limited. This paper presents such a framework, benchmarking six tree-based machine learning models against reference baselines across five forecast horizons from 15 minutes to 3 hours ahead. Using 15-minute resolution data from an approximately 1.2 MWp plant in Kayseri, Turkey, we construct a comprehensive feature engineering pipeline integrating solar geometry, lag-based persistence signals, meteorological inputs, and trigonometric time encodings. Models are trained exclusively on daytime observations to better capture the dynamics of PV output, with hyperparameters tuned independently per horizon using Bayesian optimization. Ridge meta-learning stacking achieved the lowest RMSE at four of the five horizons; however, Diebold-Mariano tests with Holm correction showed a statistically significant advantage over the best individual model only at the 45-minute horizon. SHAP-based interpretability analysis reveals that lag and persistence features dominate at short horizons, while irradiance-derived variables gain importance as the forecast window extends.