At present, there are generally few feature dimensions, unclear relationship among different features, and small effective data volume to the data collected in load forecasting research. In order to improve the accuracy of short-term load forecasting, a new load forecasting model based on XGBoost algorithm is proposed. The load prediction model based on XGBoost algorithm uses CART tree as the basic learner, inputs the preprocessed historical load and characteristic data, then builds several weak learner, layer by layer trains models to get the model, and finally inputs the features of test set into model to get the final predicted results. The load forecasting model established in this paper has the advantages of avoiding the standardization of data features, processing data missing fields, not caring whether the features are interdependent or not, and good learning effect. According to the experimental results of real power grid data, the mean absolute percent error of load prediction based on XGBoost algorithm reaches 3.46%, which is more accurate than the predicted value of load model based on BP, GRNN and DBN neural network, indicating the superiority of the proposed model.