Intelligent Hybrid Forecasting of Photovoltaic Power using LSTM and XGBoost: Seasonal and Comparative Insights

Abstract

Accurate forecasting of photovoltaic (PV) power generation is essential for the reliability and cost-effectiveness of renewable-based power systems to be maintained. This study provides a comprehensive comparison of several statistical and machine learning approaches, including SARIMA, SVR, k-NN, RF, XGBoost, and LSTM networks. To enhance the consistency of predictions, a hybrid ensemble method is proposed, which integrates LSTM and XGBoost through a weighted averaging approach. The goal of this configuration is to take advantage of the temporal learning ability of LSTM and the nonlinear feature modeling strength of XGBoost. The hourly seasonal datasets were evaluated using common performance indicators, including RMSE, MAE, MAPE, and R2. The analysis shows that each individual model performs better under certain seasonal or meteorological conditions. The LSTM–XGBoost hybrid generally yields the lowest prediction errors. It has particular effectiveness in the capture of both short-term variations and broader seasonal patterns. The results highlight the importance of hybrid intelligent systems in enhancing PV forecasting accuracy and show their potential to support more stable renewable energy operations.