A hybrid graph attention mechanism for load forecasting based on efficient spatiotemporal feature extraction

Load forecasting is fundamental to optimising and dispatching power systems. Despite the efficiency of existing load forecasting methods, they fall short in extracting more comprehensive feature representations. To solve this problem, this paper presents a spatiotemporal load forecasting method that integrates key factors, including historical load patterns, footfall, and meteorological conditions. Our method leverages residual graph convolutional networks (ResGCN) and long short-term memory (LSTM) as the primary models for forecasting. Firstly, we pinpoint the most significant variables by correlation analysis. Subsequently, we extract spatiotemporal features from load graphs and input these features into our forecasting model. The model integrates a local-global graph attention (LGGA) mechanism to incorporate local and global information, enhancing the understanding of load data. Additionally, we employ a convolutional block attention module (CBAM) to fine-tune the feature representations, thereby improving model sensitivity. Experimental results demonstrate the superiority of our method.