In order to improve the accuracy and generalization ability of load forecasting, a GRU-BiLSTM-Self-attention model based on Bagging integrated algorithm is proposed. In order to fully extract multiple features of high-dimensional input data, the model uses BiLSTM- Self- attention model to extract local features and GRU model to extract temporal features, independently extract sample subsets from the same training set and train them, integrate the output results and obtain the final prediction results. The real data of a power supply company in Nanjing are selected for the experiment and compared with the prediction models such as LSTM neural network, GRU neural network and BiSTM neural network.The experimental data show that the root mean square error of the model is 50.770 3, and the accuracy is 97.36% . Compared with other models used for comparison, the results show that this model has certain advantages in prediction effect, which shows that the proposed model has better generalization ability and prediction accuracy.