Abstract: The prediction of the dissolved gas concentration in oil can provide important data basis for the condition assessment and early fault diagnosis of power transformers. Therefore, a hybrid prediction model combining feature selection of extreme gradient boosting (XGBoost) and time series modeling of long short-term memory (LSTM) network is proposed, which is used for predicting the dissolved gas concentration in transformer oil. This model screens key features through mutual information entropy analysis, and then uses a two-layer LSTM network to capture the time series dependence of the gas concentration changes. Finally, through experimental verification, compared with the traditional autoregressive integrated moving average model(ARIMA)and single LSTM models, the average absolute error of prediction of this hybrid model is reduced by 32.7%, and the coefficient of determination R² is increased to 0.921. It shows that this hybrid model has high practical potential and can be used as the core prediction module of the online monitoring system for transformers.