In order to effectively excavate and utilize the daily and weekly periodicity of power load, accurately implement short-term load forecasting, further deepen the coordinated control of source-grid-load-storage, and promote the integration of energy digitization and low-carbon development, a short-term prediction method of power load considering daily classification attention is proposed; Firstly, the results of single-day linear regression is compared and analyzed; Then, the whole historical data is classified by classification methods, the attention mechanism is introduced to add weights to historical data of different days; Trained by neural network, three different models are obtained; Finally, the prediction results considering daily classification are compared with multiple linear regression and Bidirectional Long Short-Term Memory Network (BiLSTM); The results show good prediction accuracy on the test dataset. The errors of the three models are lower than linear regression and BiLSTM prediction algorithm under evaluation of Mean Absolute Percentage Error (MAPE), Normalized Root Mean Square Error (NRMSE) and Root Mean Square Percentage Error (RMSPE).