A hydrid LSTM-XGBoost model for short-term photovoltaic power forecasting under complex weather conditions

With the growing penetration of photovoltaic (PV) systems in modern power grids, accurate short-term PV power forecasting has become essential for maintaining grid stability and enabling reliable scheduling of distributed renewable energy. However, current forecasting methods predominantly rely on patterns observed in historical power generation data, failing to respond promptly to complex weather conditions. To address this challenge, this study proposes an LSTM-XGBoost PV power forecasting model, which integrates Long Short-Term Memory (LSTM) networks with Extreme Gradient Boosting (XGBoost). LSTM captures long-term dependencies and dynamic patterns in time series data, while XGBoost performs feature selection and structured data processing to extract latent features from multidimensional variables. This approach resolves the conflict between complex environmental variables and time series characteristics in photovoltaic power forecasting. The new model inherits the respective strengths of LSTM and XGBoost while avoiding drawbacks like gradient explosion, significantly enhancing prediction accuracy. Results show that, compared with the traditional ResNet-LSTM model, prediction errors decrease by more than 50% under normal weather conditions and by 20–35% under extreme conditions. In addition, the model responds to weather fluctuations 0.7 s faster, response speed is twice as fast, enabling earlier adaptation to sudden changes than existing approaches. This hybrid learning model construction method offers new insights for exploring precise PV power generation forecasting.