Abstract: With the increasing coupling and diversity of energy use on the customer side, accurate multivariate load forecasting is crucial for fine-grained regional dispatch management. A predictive solution is proposed based on the MMoE-CNN-Informer model for longer-term multivariate load prediction while ensuring the accuracy of short-term multivariate load prediction. Firstly, a convolutional neural network is used for supervised feature extraction of multivariate load sequences and their feature sequences. The resulting features are then input into the multi-gate mixture-of-experts(MMoE) multi-task model to learn the coupling strength between multivariate load sequences. The results of this learning process are fed into each load Informer forecasting model to achieve the multivariate load longer-term forecasting task. Experiments are conducted using the multivariate load dataset and compared with six other related prediction methods. The results demonstrate that the proposed improved model has certain advantages in the long and short-term time series prediction of multivariate loads, and improves the ability of long-term prediction of multivariate loads while ensuring the accuracy of short-term prediction of multivariate loads, which reflects the effectiveness and feasibility of the scheme.