Power system load forecasting is a key aspect of grid scheduling and operation, which is affected by factors such as national policies, population growth, seasonal changes, and weather changes. In this paper, a prediction model based on improved AlexNet-GRU is proposed to address the problem of short-term load prediction accuracy in distribution networks. Firstly, the basic principles and characteristics of power system load forecasting are analyzed, and the load data processing methods, including abnormal data correction and missing data completion, are studied. Then the AlexNet network in the field of image recognition is improved into a one-dimensional convolutional structure, and combined with the GRU network to construct the prediction model, fully utilizing AlexNet’s ability to extract complex features and GRU’s advantage of processing time-series data. The analyses of the algorithms show that the model reduces the average absolute percentage error MAPE by 1.082%, 1.314%, 1.939%, and 2.323%, and improves the average prediction accuracy by 1.085%, 1.236%, 1.876%, and 2.223%, respectively, when compared with the CNN-GRU, GRU, LSTM, and RNN models, during the consecutive six-month test in a province of Southwest China, 2.223%. In the validation of a regional dataset in Australia, the mean absolute error MAE is reduced by 22.77 MW and the root mean square error RMSE is reduced by 18.48 MW compared with the CNN-GRU model.The experimental results show that the proposed model can effectively improve the accuracy and stability of short-term load forecasting of the distribution network, and provide decision-making support for the safe and economic operation of the power system.
