Decoupling visual semantics via explicit representation learning for urban building energy modeling

Deep learning (DL) offers a computational paradigm for Urban Building Energy Modeling (UBEM) that surpasses traditional physical modeling. However, existing end-to-end black-box models face a severe semantic gap. Furthermore, prediction bias caused by the lack of explicit physical semantics remains unquantified. Although recent hybrid physics-ML approaches have made progress in UBEM by integrating physical constraints at the loss function or model architecture level, little attention has been paid within UBEM to enforcing physical semantic consistency at the intermediate representation layer. To address this gap, this study proposes a knowledge-guided explicit representation learning framework. Rather than relying on implicit mapping, we construct a hybrid architecture integrating a physical regularization decoder, ResNet-18, and LSTM. This design aligns visual features with an explicit parameter ontology at the intermediate layers. Based on an empirical study of 1,729 buildings in Changsha, China, we developed a high-precision prediction tool and revealed two key findings in engineering modeling. First, we identified a dimensionality paradox in geometric decoding. Increasing the bottleneck width admits visual redundancy and lowers decoding accuracy, an effect confirmed across five random seeds. Second, we quantified the decoupling benefits of knowledge embedding. Explicit constraints restored the model’s sensitivity weight for geometric morphology from less than 2% in implicit models to over 50%, correcting the model’s neglect of geometric form. Ultimately, this framework replaces an opaque statistical mapping with an interpretable one, in which prediction is mediated by explicitly defined physical quantities. It provides a robust, low-cost solution for UBEM based on diagrammatic images derived from urban vector data. Moreover, it offers a general methodological reference for addressing multi-source data fusion and interpretability challenges in engineering.