Integrated Energy Systems (IES), which jointly manage electricity, heating, and cooling within a unified framework, are critical for achieving energy sustainability and operational flexibility. Accurate load forecasting in IES is essential for optimal scheduling, demand-side management, and resilience enhancement. However, the heterogeneous temporal dynamics and complex cross-modal interactions among energy carriers pose significant modeling challenges. In this paper, we propose a novel deep learning architecture named Multi-Scale Temporal Convolutional Network with Gated Cross-modal Fusion (MTCN-GCF) for multi-energy load forecasting. The model incorporates a multi-scale TCN encoder to capture fine-to-coarse temporal patterns, a gated fusion mechanism to selectively integrate complementary information across energy forms, and a multi-head channel attention module to enhance feature representation. Experiments conducted on real-world campus-level data from Arizona State University’s Tempe campus demonstrate that MTCN-GCF consistently outperforms baseline models such as standard TCN and LSTM across electricity, cold, and heat load forecasting tasks, achieving significant improvements in MAE, RMSE, MAPE, and R² metrics. These results confirm the effectiveness of multi-scale and cross-modal modeling strategies in enhancing the accuracy and robustness of IES load forecasting.

Integrated energy system; Deep learning; Multi-Scale temporal convolutional network.