Predicting latent faults of transformers is essential to evaluate their health status. This paper proposes a new transformer fault prediction method. First of all, a prediction framework of temporal attention mechanism is built based on LSTM network, and the IALO algorithm is used to optimize the hyperparameters of LSTM. Afterwards, use the optimized model to predict the dissolved gas in transformer oil. Then, the SVM model optimized by the MPA algorithm is used to diagnose the gas prediction results. Finally, the fault diagnosis results are counted, and the model is verified by comparing with the actual operation state. The experimental results show that the abnormal operation status is up to 29 times form the 42th to 58th day, and the abnormal operation probability is 86.89% in the next two months, among which the proportion of medium temperature overheating fault is highest, 88.67%, and the errors from the actual situations are only 2.46% and 1.29%. The predicted results are in good agreement with the actual operating situations of transformers, which proves the feasibility of the proposed method in accurately predicting the time point and fault type of abnormal operation states of transformers.