Abstract: Fast and accurate prediction of hot spot temperature is an important precondition for condition detection, fault prediction and dynamic overload of transformers. The key is to realize dynamic prediction of transformer hot spot temperature and improve anti-noise performance of the hot spot temperature prediction model. In this paper, the hot spot temperature training samples of different ambient temperature and load change conditions are obtained through flow-thermal field coupling simulation calculation. The long short-term memory network (LSTM) is used to construct a deep learning model so as to realize the dynamic prediction of hot spot temperature. The ensemble empirical mode decomposition (EEMD) is used to reduce the noise interference in the input data, improving the anti-noise performance of the deep learning model. A 20 MVA/110 kV oil-immersed transformer is taken as example for analysis, and an experimental platform of transformer hot spot temperature rise is built to verify the validity of the model. The average error of the hot spot temperature predicted by the EEMD-LSTM network compared with the experimental results is only 1.35 ℃. After introducing random noise with amplitude of 5 ℃, the maximum error only increases by 0.47 ℃. The results show that the deep learning model based on EEMD-LSTM network can realize the dynamic prediction of transformer hot spot temperature, and it has good anti-noise performance, which is of great significance to the study of dynamic evaluation of transformer load capacity and dynamic capacity increase.