A Multimodal Deep Learning Framework for Rapid Prediction of Operational Carbon Emissions in Early-Stage Residential Building Design

This study introduces a domain-specific multimodal deep learning framework, centered on a Vision Transformer (ViT), to accelerate the prediction of operational carbon emissions in residential buildings. Our approach uniquely fuses two data modalities: the geometric information captured in floorplan images and the quantitative data from vector-based building parameters. By training and testing on a comprehensive dataset of 17,000 residential samples derived from a large-scale open-source Chinese database, the proposed model demonstrates exceptional predictive capability. On the test set, it achieved a mean bias error of 1.75%, a mean absolute percentage error of 2.14%, and a coefficient of determination (R2) of 0.95. Further validation through comparative analysis shows that our framework significantly outperforms established deep learning architectures, including ResNet-18, Inception-V4, and VGG-19, in both accuracy and robustness. The developed tool provides architects with a reliable and rapid method for assessing the carbon footprint of design options, thereby offering crucial scientific support for sustainable building design.