Integration of weather environment similarity set output features with LSTM-iTransformer for short-term photovoltaic power forecasting

To address the challenges of insufficient feature representation and the difficulty in jointly capturing local and global temporal dependencies in existing photovoltaic power forecasting models, this paper proposes a short-term forecasting model that integrates weather-environment similarity set output features with LSTM-iTransformer. First, based on one year of measured data from a 35 MW photovoltaic power station in China, similar time periods are selected and output features are extracted using normalization, Spearman correlation coefficient weighting, and weighted Euclidean distance. Second, an LSTM-iTransformer hybrid architecture is constructed, where LSTM captures local temporal dependencies and iTransformer uncovers global feature relationships, enabling multi-scale information collaborative modeling. Experimental results show that the model achieves a minimum MAE of 0.5498 MW across four quarters, with a reduction in MAE by an average of 11.57% to 26.36% compared to traditional models. Ablation experiments confirm that the inclusion of similarity set output features reduces the average MAE by 16.0%, significantly improving prediction accuracy and generalization capability, providing a novel approach for accurate photovoltaic power forecasting.