A GNN-Based Surrogate Model for Rapid Energy Consumption Prediction of Residential Floor Plans for the Early Design Stage

Under the global sustainability goal of reducing carbon emissions, it is crucial to consider building energy consumption at the early design stage. However, previous studies mainly focused on physical features like building form and materials, overlooking the impact of spatial layout on building energy performance. This study proposes a graph-based spatial representation method and develops a Graph Fusion Network (GFN) model for rapid room-level energy consumption prediction. Experiment results demonstrate that the proposed model effectively captures the relationship between graph-based spatial features and energy consumption, significantly outperforming traditional machine learning methods in predictive accuracy and generalization capability. This research validates the influence of spatial layout on building energy performance. It provides an efficient and accurate energy assessment tool for early design stages, supporting carbon-conscious spatial decision-making and design optimization.