Prediction of dissolved gas concentration in oil is very important for transformer early fault detection. a prediction model based on singular spectrum analysis (SSA) combined with long and short-term memory network (LSTM) is proposed to improve the prediction accuracy. First, To solve the problem of data leakage in traditional sequence decomposition, a sampling strategy based on SSA stepwise decomposition is proposed. Then based on this strategy, the concentration sequence of dissolved gas in original oil with complex characteristics is decomposed into relatively single trend component and fluctuation component. Finally using LSTM network for each component for single step and multi-step prediction respectively. The predicted values of each component are accumulated to obtain the prediction result of original gas concentration. The example shows that compared with the single LSTM model, the overall prediction accuracy of the proposed method is higher in the experimental days.