Abstract: In order to solve the problems of randomness and uncertainty of short-term power load with time, we propose a fusion model based on weighted grey relational projection algorithm Bagging-Blending. First, we use the weighted grey relational projection algorithm to analyze the influencing factors of power load (such as weather, temperature, humidity, date type, etc.) in order to select the characteristics of historical load. On this basis, we embed each single model in the Bagging integration algorithm, including support vector regression (SVR), K-nearest neighbor (KNN), gate recurrent unit (GRU), eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM) and Categorical features gradient Boosting (CatBoost), to improve the stability and generalization ability of the model. At the same time, we use the Pearson correlation coefficient to analyse the correlation of each single model. Second, according to the different angles of the data observation space of the model, we use the Blending model to fuse the models with low correlation. Finally, we verify the model by the power load data of in New England, ISO New England. Compared with the traditional single model (SVR, GRU) and other fusion models (Bagging-XGBoost, optimal weighted GRU-XGBoost), the proposed fusion model has stronger generalization ability, higher stability and prediction accuracy.