Aiming at the low accuracy of short-term load forecasting in power market, a combined forecasting model based on LSTM artificial neural network is proposed. This paper analyzes the unique advantages of LSTM neural network and its variant GRU neural network in learning time series features in load forecasting. Convolution neural network is used as the feature extraction layer of load data, combined with GRU network, a combined model is constructed, and the residual prediction model is established to correct the results. The simulation results show that the prediction effect of neural network with memory function is better than Ann and SVM models, and the average relative error of residual prediction model proposed in this paper is about 1.79%, and its accuracy is higher than that of single algorithm load prediction model.