The invention discloses a short-term electric power load forecasting method based on a multi-label neural network. Firstly, the historical data is processed by segmentation coding and standardization.Secondly, k-Means clustering algorithm divides the original data into k clusters; Again, the K- NN-based multi-labeling algorithm learns the similarity between the load to be forecasted and k clusters. Finally, the BPNN model is trained with each clustering data, and the load forecasting results are obtained. This method has been proved to be greatly improved in prediction accuracy and running time by experiments.