A graph-based multi-modal deep learning framework for automated energy performance assessment and retrofit decision support in Chinese residential buildings

Over recent decades, the increase in global energy consumption has led to negative environmental impacts and has become a major concern that hinders social development. Since 2007, the Chinese government has implemented large-scale energy-efficiency retrofit programmes for existing buildings. However, in practical projects, the evaluation and decision-making process for retrofit options is still constrained by incomplete building information and the limited efficiency of current assessment workflows. To support retrofit decision-making, this thesis proposes a data-driven approach for household-level energy prediction and retrofit decision support. The research is organised into four interconnected stages. First, based on existing floor plans of ageing residential buildings, an automated framework for architectural element recognition and semantic retrieval is developed, enabling the automatic generation of baseline Building Energy Models (BEMs). Second, these baseline models are enriched with thermal material properties and occupant behaviour parameters, defined in accordance with the current literature. Large-scale parametric energy simulations are then carried out to produce a comprehensive residential building energy consumption dataset tailored to the Chinese housing context. Third, a multi-modal deep learning model, FusionGNN, is proposed to predict household annual Energy Use Intensity (EUI). The model is first pre-trained on the simulation dataset and then fine-tuned using measured data collected from 120 households. Finally, the calibrated FusionGNN is applied to a case study of four typical residential typologies to rapidly evaluate energy performance under various retrofit scenarios. By integrating predicted energy outcomes with retrofit cost assessments, a multi-objective optimisation is conducted to identify the optimal retrofit strategies. Overall, this research provides a scalable data-driven decision-support framework for residential building retrofits in China, offering evidence to help stakeholders achieve a more appropriate balance between energy savings and economic feasibility.