Performance Analysis of Photovoltaic Power Forecasting Based on Input Feature Configuration and Selection

This study investigated the impact of different input feature configurations on photovoltaic (PV) power forecasting performance.Four feature groups were defined: a basic feature group comprising solar position and temporal information, a meteorological feature group, an autoregressive feature group, and a dynamic and lagged feature group.These feature groups were subsequently combined to construct eight input feature configurations.The resulting configurations were evaluated using a tree-based model (LightGBM) and a time-series deep learning model (LSTM) to compare their forecasting performance.The experimental results showed that incorporating temporal features reduced prediction errors in terms of MAE, NMAE, RMSE, and MSE.Overall, LightGBM consistently achieved lower prediction errors than LSTM across all feature configurations.These findings highlight the importance of incorporating temporal dependencies when designing input feature structures for PV power forecasting.