Accurate multi-energy load forecasting plays a vital role in the optimal scheduling of integrated energy systems (IES). However

existing forecasting methods face two major challenges: insufficient extraction of complex features within load data

which limits prediction accuracy

and ever-increasing model complexity

which leads to a sharp rise in computational resource demands. To tackle the above-mentioned challenges

this research introduces a load forecasting scheme derived from multiscale lightweight deep learning. First

the dynamic coupling characteristics of load data are analysed using correlation coefficients to determine model-specific input configurations. Then

an improved variant channel multiscale bottleneck residual network (VC-MBResNet) is proposed to obtain high-dimensional load feature data. Finally

by leveraging shared underlying parameters and an enhanced adaptive loss function for each subtask-using both soft and hard weight-sharing strategies—the extracted features are integrated into a multi-task learning (MTL) framework. A bidirectional long short-term memory (BiLSTM) network is employed as the shared layer

with an attention mechanism embedded in its hidden layers to enhance the focus on critical temporal features. Experimental results demonstrate that the proposed approach surpasses existing models in terms of both prediction accuracy and computational efficiency.