To address the issues of low efficiency in urban building energy management and insufficient accuracy in carbon emission source identification – critical challenges amid global dual-carbon goals – this paper proposes the TRD-Net model, which integrates Temporal Convolutional Network (TCN), ResNet, and Deep Reinforcement Learning (DRL). Urban buildings account for over 30% of global energy consumption and 20% of carbon emissions, but traditional methods rely on manual experience or simple statistical models, failing to capture multi-scale temporal fluctuations of energy data and achieve precise emission source tracing. TRD-Net leverages TCN for multi-scale temporal feature extraction, ResNet for deep nonlinear relationship mining, and DRL for dynamic decision optimization, forming an end-to-end framework. Experiments on four datasets (including 500,000 building energy consumption records and 400,000 time-series carbon emission data) show that TRD-Net significantly outperforms baseline models such as TCN-LSTM and Transformer in metrics like Accuracy and AUC. For instance, on the City Carbon Emission Sources Dataset, TRD-Net achieves an F1 score of 98.16%, with a superior parameter scale (318–339M) and inference time (5.35–5.62 ms). Ablation experiments validate the necessity of component synergy, and the DRL optimization strategy enhances efficiency. In the future, the distributed architecture will be optimized to enhance scalability, providing a reliable technical solution for intelligent low-carbon building management.
Wang et al. (Thu,) studied this question.