Accurate prediction of multi-energy load is crucial for the optimal scheduling and economic operation of integrated energy systems（IES）. Aiming at the strong randomness of regional IES and the coupling relationship between multi-energy sources, a multi-task short-term load prediction model based on SEResNet-BiLSTM network and attention mechanism is proposed. Firstly, the model of squeezeand-excitation networks-residual network（SEResNet）is used as the high-dimensional feature extraction unit to mine the coupling relationship between multiple energy sources. The high-dimensional feature extraction of multi-energy load data is realized. Then, bidirectional long short-term memory（BiLSTM）network is used to capture the time series characteristics between data to realize the prediction of load data. Multi-task load learning is realized by hard weight sharing to realize multivariate load forecasting. Finally, the effectiveness of proposed method is verified by simulation experiments, and the accuracy of the proposed method is significantly improved compared with other models.