Load forecasting is crucial to the economic operation of the power grid. In order to improve the accuracy of short-term load forecasting and reduce the training time of the hybrid neural network, a short-term load forecasting method based on basic network with multilayer perceptron (MLP), simple recurrent units(SRU) and principal component analysis (PCA) is proposed. Firstly, the method considers various power load influencing factors to establish input feature sets of the load forecasting task, and uses PCA to transform and reduce the historical load and temperature characteristics which are the original inputs of the network. Then, the method uses new data information obtained after PCA as the inputs of the hybrid neural network model, and trains the network with Adam gradient descent algorithm. Finally, the outputs of the proposed model are load forecasting results. The results of the experiments show that the MAPE of the hybrid model including SRU on all test set samples is 2.126%, which is much lower than that of the single model with only basicnet and the hybrid model including DNN, and compared with the hybrid model including LSTM, the training time is reduced by 22.74%, and the application of PCA also accelerates the convergence of the model, which greatly reduces the number of training epochs.