Abstract: In the background of increasing coupling degree among different loads in the regional integrated energy system（IES） and increasing demand for more accurate and reliable energy consumption forecasting, a short-term forecasting method of cooling,heating and electrical loads in IES based on coupling feature construction and multi-task learning is proposed. First, from the perspective of feature engineering, the coupling feature mining algorithm is used to construct the coupling feature variables of cooling, heating and electrical loads in IES, and the coupling features among different energy load demands are extracted. Then,taking the exogenous variables such as load history data, coupling feature variables and the temperature as the model inputs, the load forecasting model of IES is established by using the sharing mechanism of multi-task learning, which makes the highdimensional features and model parameters among energy forecasting subtasks be used for mutual reference through a shared learning layer built by long short-term memory neural network, so that the full exploitation and utilization of load coupling features can be realized. Taking the IES of Tempe Campus of Arizona State University as an example, this paper proves that the proposed forecasting method can effectively improve the short-term forecasting accuracy of cooling, heating and electrical loads in the regional IES, through the accuracy comparison of forecasting results and the interpretability research of the deep learning model.