Interpretable short-term electricity load forecasting considering small sample heatwaves

Heatwaves have become more frequent, causing significant fluctuations in electricity load and making it difficult for traditional forecasting methods to capture the load variations accurately. Simultaneously, as the power system undergoes digital transformation, the growing complexity of forecasting algorithms makes the internal reasoning processes of models harder to interpret. To address the heatwave and interpretability induced challenges, this study proposes a BiLSTM-based electricity load forecasting method incorporating TimeGAN and a dual attention mechanism. The method first uses TimeGAN to alleviate the imbalance in historical data's distribution of heatwaves samples. Subsequently, a dual attention mechanism, which accounts for both features and time, is integrated into the BiLSTM, enabling the model to more effectively capture the temporal characteristics of the load data and the impact of meteorological factors. Case experiment reveals that the proposed method outperforms traditional models in short-term electricity load forecasting during summer heatwaves. The proposed method achieves superior performance in three key aspects: higher forecasting accuracy with the mean absolute percentage error reductions of 18.03 % and 33.62 % on residential and public service datasets respectively; better generalization ability; and enhanced interpretability. Overall, the proposed approach offers a practical and effective solution for interpretable load forecasting under extreme weather conditions.