Abstract: With the continuous development of China’s national economy and the acceleration of urbanization, the construction of subway is developing rapidly, and the power supply system of subway is also growing correspondingly. The load of the power supply system of subway has become an important part of the load of the urban power system. Due to the characteristics of mobility, time-varying and non-linearity of the load of the metro power supply system, the research on the load prediction technology of the metro power supply system is becoming more and more important. Firstly, the load prediction of the subway power supply system is studied, and the multi-dimensional factors such as subway historical load, subway transfer station, subway over-ground/underground form, passenger flow, weather, temperature are considered. Then, the feature learning of the multi-dimensional factors is carried out based on the stacked denoising autoencoders. Based on the deep recurrent neural network which is suitable for dealing with nonlinear problems of sequence properties, a load prediction method of the subway power supply system is proposed. Finally, the actual operation data of Nanjing metro proves the effectiveness and superiority of the prediction method is proposed, and the method has good effect on short term and medium-long term load forecast of metro power supply system. The medium and long term load prediction is carried out for the subway station to be built in Nanjing, which provides reference for the capacity regulation of the main station.