Abstract: A multi-load short-term joint forecasting model of integrated energy system is proposed to improve the accuracy of multi-load forecasting of integrated energy system based on multi-scale feature extraction by comprehensively considering the interaction mechanism of multi-energy, the coupling characteristics of multi-load and the correlation of meteorological factors. Firstly, the coupling characteristics of multi-load and the correlation of influencing factors are studied by the maximum information coefficient, and the prediction characteristics are selected. Secondly, the input features are decomposed by variational modal decomposition technology to enhance features purity. Finally, the CNN-BiLSTM multi-task learning model is used for feature fusion, and the Attention mechanism is used to select important features differently to realize multi-scale feature extraction. In addition, hyperparameter optimization of the VMD and CNN-BiLSTM multi-task learning model is achieved by the snow ablation optimizer to realize joint forecasting of IES multivariate loads. Experiments were conducted using real-world data from Arizona, USA. The results indicate that the proposed joint forecasting method possesses lower root mean square error and higher accuracy compared with single forecasting method or other models and greater robustness in IES multivariate load forecasting.