Subway load forecasting can facilitate the stable and efficient operation of subway power systems. Most existing methods for forecasting subway power load utilize statistical or machine learning models, such as linear regression or support vector machines. However, due to the difficulty in effectively capturing the spatial-temporal characteristics of subway power systems, particularly time-varying nature and non-linear complexities of the load, these methods are limited in the precision of prediction. To further enhance the precision of subway load forecasting, a subway power load forecasting method based on spatial-temporal graph neural networks (STGNN) is proposed to predict the power traction load of each station during subway operations. STGNN extracts spatial-temporal relationships from multiple perspectives of subway stations by constructing multiple-perspective spatial-temporal graphs that integrate a geographical distance graph, a load similarity graph, and a dynamic learning graph. It comprehensively captures the spatial-temporal dynamic changes of the subway power system, where the dynamic learning graph mechanism adaptively adjusts the adjacency matrix, thereby improving the ability of the model to perceive non-linearity and the evolutionary temporal characteristics. Experiments are conducted on historical data of power load fromsubway stationsin some city. Results show that STGNN achieves a high prediction precision of 89.37%, which is 3.16%, 3.90%, 11.38% and 2.10% higher than those of XGBoost, LightGBM, LSTM and MTGNN models respectively, indicating that STGNN has broad application prospects in subway power load forecasting.