Short-Term Load Forecasting Under Extreme Weather via Sample Screening and Fine-Tuning

In recent years, intensifying global climate change has led to more frequent extreme weather events, causing abrupt fluctuations in electricity demand and renewable generation that threaten power system stability, especially under high renewable penetration. However, extreme weather data are often sparse, incomplete and discontinuous, posing significant challenges to load forecasting. To address these issues, this article proposes a general forecasting framework for extreme weather. The framework first applies a weighted dynamic time warping (WDTW)-based sample selection method to identify historical samples similar to target extreme weather conditions, thereby enriching the training dataset. It then develops a deep learning-based fine-tuning strategy to enhance adaptability to extreme conditions while retaining the general feature extraction capability of the pre-trained model. During forecasting, the fine-tuning model is used for extreme weather samples, whereas the baseline model is retained for normal conditions, improving forecasting accuracy in extreme scenarios without degrading the overall performance. Two case studies are conducted to evaluate the proposed framework. Under extreme temperature conditions, the proposed framework reduces the mean absolute percentage error (MAPE) by 3.25 percentage points in the load forecasting case and by 5.59 percentage points in the net load forecasting case. These results demonstrate the effectiveness and applicability of the proposed framework.