Developing GNN-Based Surrogate Models for Multi-Objective Sustainable Performance Predictions of Residential Blocks

Although building performance simulation using physical models is frequently utilized for performance prediction, its significant computational demands pose challenges to its implementation in the early design stage. Surrogate models have been proposed to replicate computationally expensive physics-based simulation models, but existing surrogate models for sustainable residential block design are limited in scope, focusing on specific cases. Graph neural network (GNN) could be a solution to enhance the generality of the surrogate models for residential block design. However, the optimal architectures of the surrogate model and the time costs compared with physics-based simulation models have not been discussed yet. To fill these gaps, this study explores the development of GNN-based surrogate models for multi-objective sustainable performance predictions of residential blocks. Firstly, we introduce a graph schema to represent the general geometric features and relations, and a regional dataset for training and testing of the surrogate models. Secondly, we propose two kinds of architectures (individual architectures for specific indicators and an integrative architecture) for the surrogate models. Thirdly, we train and optimize the models utilizing the graph schema, regional dataset and architectures. Finally, the optimized surrogate models are evaluated in two aspects: 1) the optimized models using the individual architectures for specific indicators and the ones using the integrative architecture are compared in terms of prediction accuracy and time costs; and 2) the time costs of the optimized model are analyzed by comparing with physics-based simulations. The results showed that surrogate models based on individual architectures outperform the model using the integrative architecture in terms of prediction accuracy and time costs for all sustainable performance indicators. Although the model preparation time of the surrogate models exceeds that of the physics-based simulations, the surrogate models reduce the calculation time from 6.346 min to 1.565 ms per case compared with the physics-based simulations.