Multi-point forecasting of photovoltaic power generation by light gradient boosting machine

Abstract

Machine learning methods have been developed for forecasting multi-point photovoltaic generation (PV). Multi-point forecasting is expected to improve forecast accuracy by learning Spatio-temporal cloud variability from the relationship of power generation among neighboring PVs. But multi-point forecastings may also complicate the calculation process and increase the number of unique models to forecast every PV. This study proposes a multi-Light Gradient Boosting Machine (LGBM) stacking model to predict PV generation 30 minutes ahead. The proposed multi-LGBM stacking can forecast multiple PV units by a single model, which doesn’t require multiple models for multiple PVs. Also, the proposed multi-LGBM stacking improved RMSE by about 1.29% compared to the existing LGBM, which trains each PV separately.