A Study on Novel Solar Power Forecasting Using an XGB-LiGBM-RF Hybrid Model and the L-BFGS-B Optimization Algorithm

Abstract

Accurate forecasting of solar power is essential for enhancing the stability and efficiency of power systems with high Photovoltaic (PV) penetration. This paper proposes a novel hybrid model based on a Stacking Ensemble (SE) of XGBoost, LightGBM, and Random Forest (RF), with optimal weights determined using the Limited Memory Broyden–Fletcher Goldfarb Shanno with Box constraints (L-BFGS-B) algorithm. The model is trained and tested on real-world data from a 49.5 MW solar power plant in Vietnam. The experimental results show that the proposed SE-XGB-LGBM-RF-OW model outperforms individual learners and deep learning baselines in both accuracy and training time. It consistently achieves a Normalized Mean Absolute Percentage Error (NMAPE) below 1.2% across all seasons. Compared to LSTM and GRU models, SE reduces Root Mean Square Error (RMSE) by more than 90% and shortens training time by over 20 times. Additionally, it significantly lowers the MAPE and NMAPE values, with improvements exceeding 90% in most seasonal test cases, highlighting the model’s superior accuracy and generalization capability.