The Regional Integrated Energy Systems (RIES) amalgamate traditional fossil fuels with a variety of renewable energy sources such as solar and wind power. By utilizing energy in a cascading manner, these systems can efficiently meet diverse l oad demands including electricity, heat, and cooling, playing a crucial role in energy conservation and carbon reduction. Accurate load forecasting is essential for the flexible, reliable, and efficient operation of these systems. This study proposes a multi-task learning (MTL) model based on Residual Neural Networks (ResNet) and Gated Recurrent Units (GRU) for multi-load forecasting. Initially, a combination of Local Outlier Factor and K-Means++ algorithms were employed for data anomaly cleaning. Subsequently, establishing the prediction model involves using multi-layer ResNet for data feature extraction and GRU network for further exploration and prediction of the temporal features of multi-load data. Finally, employing MTL to achieve joint forecasting of multi-loads. Applying the proposed method with multi-load data from Arizona State University (ASU) Tempe campus case conducting multi-load forecasting experiments in summer, winter, and transitional seasons cases, and comparing the results with other forecasting models. The results indicate that the model in this study has the highest prediction accuracy with MAPE for electricity load at 1.67%, 1.37%, 1.44%, heat load at 2.43%, 1.22%, 1.35%, and cooling load at 0.79%, 0.99%, 1.12%. Additionally, comparing the running times of different models reveals that the model in this study has the shortest runtime, demonstrating the highest computational efficiency.