Developing surrogate models for the early-stage design of residential blocks using graph neural networks

Abstract Building simulation based on physical modeling is commonly adopted for performance prediction, but the high time costs hinder its application in the early design stage of buildings. Data-driven surrogate models have been proposed as a means to replicate computationally expensive simulation models. However, existing surrogate models for sustainable residential block design are limited in scope, focusing either on individual buildings or on specific cases within multi-block projects. This study leverages graph neural networks to develop optimal surrogate models incorporating inter-building effects to predict multiple indicators of sustainable performance for residential blocks at a region level. A graph schema is proposed to represent the general geometric features and relations among buildings in residential block design. A regional dataset is generated for model training and testing, using real residential zones in Hong Kong. The surrogate models are developed and evaluated, using two kinds of architectures (individual architectures for specific indicators and an integrative architecture) and three different neural networks (graph attention network (GAT), graph convolutional network, and artificial neural network). The results showed that the surrogate models using the individual architectures and GAT outperform the models using other architectures and neural networks. These surrogate models achieve a high prediction accuracy with CV(RMSE)s of 11.79%, 7.63%, and 8.00% in terms of energy consumption, indoor thermal comfort, and daylighting, respectively, on the regional test set. Moreover, they enable a significant acceleration of the performance evaluation, reducing the calculation time from 6.346 min to 1.565 ms (243,297 times) per case compared to physics-based simulations.