Machine Learning Ensembles for Solar Power Forecasting

Solar power forecasting is crucial for better integration of renewable energy, grid stability and energy resources planning and allocation. In this paper, we propose a 24-hour ahead solar power forecasting system by harnessing the power of blending ensemble techniques. Our ensemble framework combines the predictive capabilities of Long ShortTerm Memory (LSTM) and XGBoost models as base learners, further refined by Ridge Regression meta model. We conducted extensive experiments on real-world solar power generation data, assessing the effectiveness of our ensemble methodology. The results demonstrate an improvement in forecasting results compared to individual models based on the Root Mean Square Error (RMSE) score. The blending of LSTM and XGBoost models allows us to capture both the temporal dependencies and nonlinear relationships inherent in solar power generation data, while the Ridge Regression meta model was used for regularization and predictive stability. Our findings highlight the potential of ensemble techniques in advancing solar power forecasting, contributing to the reliable integration of renewable energy sources into the grid.