In order to solve the problem of low precision of long time series load forecasting, a load forecasting model LR-Informer based on improved Informer is proposed in this paper. In this paper, the LSTM layer is added between Informer's self-attention block and “distillation” operation to further enhance Informer's local dependence and long dependence on time series. At the same time, residual connection is added to the encoder, so that the information in the middle and low layers of the network can be transmitted to the upper layers, reducing the information loss in the network stack. The prediction accuracy of the model is improved. The experimental results show that compared with the mainstream model Informer, Informerstack, MSE decreases by 5.8% on average, showing better prediction performance and stability.