Abstract: Transformer oil temperature is an index that directly reflects the heat dissipation performance of transformer. Accurate prediction about top oil temperature of the transformer is instrumental in monitoring its operation. By analyzing the traditional mathematical model of top oil temperature of the transformer, considering the influence of load rate and ambient temperature on oil temperature, and determining the active power, reactive power and ambient temperature of load data peak and valley as the characteristic quantity, a prediction model of transformer top oil temperature based on long short term memory（LSTM） network is proposed. Taking the real data of substation as an example, the proposed LSTM prediction model is trained, and 5 random samples is selected for prediction; the BP neural network（BPNN） and the recurrent neural network（RNN） prediction models are built respectively to predict the same samples and intercept the prediction data of the first 30 minutes, and compared with the LSTM model. The simulation results show that the temperature prediction model based on LSTM has the highest calculation accuracy, the error rate is controlled within 5%, the predictive value is basically consistent with the actual value, and the temperature prediction can be effectively realized.