Power grid is diagnosed after a failure to prevent the fault occurred by inferring the information that the fault generated.The method of model prediction (MP) and abductive reasoning network(ARN) is proposed to forecast the power system fault. MP predicted the troublefree operation data of the power grid by using historical data, and compared with the actual grid runtime data, the difference was calculated and used as the input of fault diagnosis system. ARN was used to bulid the fault diagnosis system and solve the complicated relationships between data processing and the corresponding candidate fault section. The fault location can be found before protection device and circuit breaker by combining the method of MP and ARN. The test results showed that the model prediction method can quickly and accurately diagnose the fault compared with BP neural network method.