Dynamic Prediction of Carbon Emissions from Building Operation Stages by Time-Series Convolutional Networks (TCNs)

In this paper, a dynamic prediction model of carbon emission in the operation phase of buildings based on time-sequence convolutional network (TCN) is proposed, which utilizes dilated convolution, causal convolution and residual linkage to construct a long-series modeling framework, and forms a high-quality time-sequence sample based on the preprocessing of multi-source energy consumption and meteorological data. The model performs well in the comparison experiments of LSTM, GRU, and SVR, with a 1-h prediction RMSE of 2.318 kg CO₂, which is 11.3% lower than that of LSTM, and a 24-h prediction R² of 0.957, which is higher than that of LSTM (0.946) and SVR (0.918), and a single-epoch training time of 12.3 s, with a delay of only 0.87 ms/sample, combining high accuracy and high precision. sample, both high precision and high efficiency, suitable for dynamic management and real-time decision-making of building carbon emissions.