Abstract: With the continuous development of integrated energy utilization technologies and the diversification of users ’ energy consumption needs, the existing method for single type load prediction is difficult to reflect the coupling characteristics of multiple loads. Accurate multiple load prediction will become the primary prerequisite for optimal dispatch and economic operation of integrated energy systems. Based on this, a multi-task learning load prediction method is proposed by using a long short-term memory（LSTM） neural network as a shared layer. In this method, the coupling characteristics of multiple loads are simulated through the shared layer to improve the prediction accuracy. Firstly, a multi-task learning load prediction model is developed by means of "hard sharing mechanism + LSTM sharing layer", and the sharing mechanism is used to learn the coupling information provided by different subtasks. Secondly, the neural network interpretability technology is used to visually explain the offline training results, which demonstrates that the proposed model can use the coupling information provided by the subtasks to improve the prediction accuracy. Finally, compared with the traditional model, the results show that the proposed model has better effect on prediction accuracy and time.