The power router (PR) has been widely used to improve the utilization of renewable energy sources (RES). The intermittent and random nature of RES inevitably increases the fault probability of PR. However, none of the existing methods can realize the fault diagnosis of PR. To address this issue, a novel fault diagnosis method based on transfer learning and gated recurrent unit fully convolutional network (GRU-FCN) is proposed in this paper for internal-diagnosis and external-diagnosis of PR, which includes fault data extraction, fault localization, and fault type classification. First, the current data is extracted from the bus to facilitate fault localization. Subsequently, sensor data from the localized port is utilized for specific fault type classification. Moreover, various fault scenarios of a four-port PR are simulated in MATLAB/Simulink to generate datasets for evaluating the effectiveness of the proposed method. Finally, a comparison between the proposed GRU-FCN model and other neural network models is carried out, which further proves the superiority of the proposed method in the field of fault diagnosis for PR.