An Approach to Predicting Buildings Energy Consumption in Urban Block Based on Multimodal Deep Learning Model

Traditional methods of predicting building energy consumption often fail to comprehensively capture the complex environmental features of urban blocks. This study proposes a multimodal deep learning framework that integrates microclimatic environments, and building information modelling with demographic attributes for the complete construction of underlying scenarios for building energy prediction in urban blocks. Based on this, multi-model comparison experiments are carried out by adjusting the sample size and dataset weighting ratio, resulting in the multi-modal dataset with the best overall performance and corresponding prediction model. Finally, the interactions between multi-modal factors and their effects on block-scale building energy consumption levels are analysed using ablation experiments, SHAP and Grad-CAM methods. The proposed multimodal model-based approach for modelling building energy consumption outperforms the single modal dataset model in terms of multifactor interpretability and prediction efficiency.