Photovoltaic Short-Term Power Prediction Using LGBM

This paper presents a LightGBM-based model for short-term photovoltaic (PV) power forecasting, developed to address the unique challenges associated with large-scale renewable energy datasets, including data complexity, nonlinearity, and seasonality. The proposed model leverages the efficiency and robustness of LightGBM, which optimizes Gradient Boosting Decision Tree (GBDT) with histogram-based decision tree learning and gradient-based one-side sampling (GOSS) techniques. Through experimental validation across three PV sites with varying data characteristics, we compared the LightGBM model with the LSTM model, a conventional approach in PV forecasting. Results show that LightGBM achieves higher prediction accuracy, greater computational efficiency, and increased stability across sites, outperforming LSTM, particularly in data quality-sensitive scenarios. Further analysis revealed that LightGBM's robust handling of sparse and heterogeneous data makes it especially suited for PV forecasting applications. This study highlights the potential for further improvement in prediction accuracy by extending the model to a multivariate framework, which could capture a broader range of influencing factors in PV power output.