Tripping is a common fault in power transmission and distribution systems. In recent years, in order to deal with this kind of fault, academic circles have proposed protection methods based on relay protection action and electrical element action. However, these methods for electrical protection have hysteresis in handling tripping faults. Therefore, the prediction of tripping faults in advance plays a vital role in dealing with hidden problems and power recovery. In this paper, a method of power system trip fault prediction based on multi-source time series data is proposed. LSTM is used to extract the time characteristics of multivariate data, which alleviates the problem of RNN gradient disappearance on long time series. The model adds a peephole connection structure on the three-layer grid to enable a single unit to view the LSTM unit status in the previous stage, thereby strengthening the network timing memory capability. Secondly, we use L2 regularization measures such as parameter normalization to mitigate the impact of over fitting in fault prediction on the results. Finally, support vector machine classifier is introduced to improve the generalization ability and robustness of the overall model. The experimental data were obtained from relevant institutions of the State Grid of China. Experiments show that the proposed method has the advantage of high classification accuracy compared with existing data mining methods. In the last part of this paper, the practical application is discussed to prove its feasibility in the actual scene.