Abstract: To address the complex multi-energy coupling relationships, as well as the non-stationarity and dependency of multivariate time series in integrated energy systems, a multi-task learning model based on multi-channel feature extraction and deep interactive fusion was proposed. First, the maximal information coefficient (MIC) and autocorrelation function (ACF) were combined to analyze load characteristics and select key input features. Second, a multi-channel gated temporal convolutional network (MGTCN) was employed to independently model each input feature and dynamically extract effective information. The outputs of each channel achieved deep interaction through a cross multi-head attention module, explicitly mining the correlation information among multi-source features. Then, an adaptive routing fusion (ARF) module was introduced to conduct global semantic modeling and efficient fusion for each load and its interactive information, followed by a bidirectional long short-term memory (BiLSTM) network to further capture the long- and short-term temporal dependencies of the fused features. Finally, under the multi-task learning framework, the coupling information among loads was shared, and balanced training across tasks was achieved by optimizing the loss function, thereby enhancing the overall performance of the prediction model. Case analysis demonstrates that the proposed method achieves higher prediction accuracy in short-term multi-load forecasting.