Heatwave-Gated Adaptive Spatiotemporal Learning for Short-Term Load Forecasting of Light-Industrial Multiusers

Abstract Heatwaves are becoming more frequent under global warming, leading to concurrent load surges and intensified fluctuations in light-industrial electricity consumption. This heatwave-induced nonstationarity weakens the accuracy of conventional short-term load forecasting approaches. To capture both heatwave-driven interuser coordination and dynamic spatiotemporal dependencies, we propose a heatwave-gated adaptive spatiotemporal graph neural network (HWG-STGNN). Specifically, gated temporal convolutional network blocks with causal dilated convolutions model long-range temporal dependencies; a learnable adjacency matrix adaptively characterizes spatial correlations among multiple users; and a heatwave gating module adjusts feature contributions according to meteorological intensity, enabling the model to respond to varying heat-stress levels. We evaluate the proposed method using 15-min multiuser load data from a southern Chinese city (2018–2019). Results show that HWG-STGNN achieves competitive overall performance across different forecasting settings, with more pronounced and consistent gains during heatwave intervals. On heatwave samples, the mean absolute error and mean absolute percentage error are reduced by approximately 18% and 14%, respectively, with root mean square error also improved. These findings indicate that incorporating extreme meteorological information into spatiotemporal forecasting frameworks is an effective way to enhance accuracy and robustness in high-temperature, strongly nonstationary conditions.