Efficient and accurate shortterm power load forecasting plays an important role in the reliable operation and sustainable development of the power grid. In view of the complexity and timing characteristics of the factors that affect the changes in power grid load, and the shortcomings of the existing machine learning prediction methods in selecting key parameters based on experience, the convolutional neural network (CNN) is used to extract the multidimensional feature vector representing the load change, which is constructed as a time series and input to the gated recurrent unit (GRU), and the improved sparrow search algorithm (ISSA) is used to iteratively optimize the hyperparameters in the GRU network. The prediction experimental sample comes from the load data of a certain area in Yunnan, and the prediction accuracy of the proposed method reaches 98.624%. Compared with the neural network prediction methods of RNN, GRU and LSTM, the calculation example shows that the proposed method overcomes the difficulty of selecting key parameters based on experience and has higher prediction accuracy.