The spacecraft power system is one of the key subsystems, and its operational status directly affects the lifespan and performance of the entire spacecraft system. Therefore, employing advanced technology for fault diagnosis of the power system to improve the reliability and safety of spacecraft in orbit has become a research focus in the field of fault diagnosis. Methods based on deep learning have advantages such as strong fitting ability and rich feature extraction, making them the mainstream approach in the field of fault diagnosis. However, in the field of fault diagnosis of spacecraft power systems, mainstream fault diagnosis methods cannot capture the long-term dependency of sequences and are limited to modeling in the time dimension, severely affecting the performance of fault diagnosis methods. Therefore, this paper proposes a method based on spatio-temporal self-attention mechanism for efficient and accurate fault diagnosis of spacecraft power system. The method adopts a Transformer-based encoder structure to extract high-dimensional features from spacecraft telemetry data, optimizes the self-attention mechanism therein, uses temporal convolution to extract temporal feature information, and employs both temporal and spatio bidirectional self-attention mechanisms to extract spatio-temporal features from the data. Finally, the features extracted by the model are mapped to obtain the fault diagnosis results of spacecraft power system. Relevant experiments are conducted on the spacecraft power system dataset. The experimental results show that compared with commonly used methods in the field of fault diagnosis, the proposed method has stronger fault representation extraction ability, which can effectively improve the fault diagnosis capability of spacecraft power systems.