Abstract: Existing methods for mining the electricity consumption characteristics of high-voltage users often use data clustering or linear dimensionality reduction to extract features. However, these methods often focus on single dimensional information and rely on linear correlation assumptions, resulting in poor fit between the mined features and real electricity consumption behavior, and limited scene adaptability. A multi-scale feature mining method for high-voltage user electricity consumption patterns based on spatiotemporal graph neural network was studied to address this issue. Firstly, a spatiotemporal graph model is constructed that integrates the static attributes of high-voltage users, dynamic electricity consumption timing, and grid topology correlation. The spatiotemporal dependency relationship of electricity consumption data is integrated using spatial adjacency matrix and temporal adjacency matrix. By using spatiotemporal graph neural networks for hierarchical extraction, three types of features can be obtained: local fine-grained fluctuations, mesoscale cluster collaboration, and global macro trends, thus fully covering the multidimensional core attributes of electricity consumption patterns. Finally, multi-scale feature dimensions are aligned through linear mapping, and cross scale feature aggregation is achieved by combining spatiotemporal adaptive weight learning with the "weighted fusion+feature interaction" strategy to achieve deep coupling of multi-scale information. The experimental results show that the mean square error between the electricity consumption pattern features mined by this method and the real electricity consumption behavior is reduced by 18.2% compared to method 1 and 25.7% compared to method 2. This method can more accurately and stably capture the spatiotemporal characteristics of high-voltage users" electricity consumption patterns.