Short-term power load forecasting based on spatial-temporal dynamic graph and multi-scale Transformer

Abstract Short-term power load forecasting is essential for ensuring power system stability and facilitating market planning. However, the multi-periodic nature of load data and its complex correlations with external factors pose significant challenges to accurate predictions. To address these issues, we propose a novel spatial-temporal dynamic graph Transformer (SDGT), which integrates a multi-scale Transformer module with a patch-based multi-scale encoder to capture multiple periodic patterns and extract temporal dependencies. Additionally, a spatial-temporal correlation graph (STCG) is constructed based on shape similarity and semantic relevance, and further enhanced using a graph convolution module to model dynamic spatial correlations between load data and external influencing factors. Experimental results on two public benchmark datasets demonstrate that SDGT surpasses state-of-the-art forecasting models, exhibiting superior predictive accuracy and robustness. The findings validate the effectiveness of SDGT in capturing multi-periodic patterns and spatial-temporal dependencies, making it a promising approach for improving short-term power load forecasting and supporting real-world power system operations and energy market planning.