Rapid prediction of pollutant dispersion in residential blocks using generative adversarial networks

Abstract

The morphological layout of urban residential blocks has a significant impact on pollutant dispersion. However, the optimization of pollutant dispersion-oriented residential blocks studies used to rely on computational fluid dynamics (CFD) simulations with abstract models, which not only disconnects from the actual design situation, but also limits the efficiency of scenario iteration. Therefore, in order to realise real-time prediction of pollutant concentration fields for arbitrary layouts at the early morphological design stage of urban settlements, this study clusters 1997 residential blocks in Shenzhen, China, and constructs a CFD dataset representative of Shenzhen's residential block. Then, based on this dataset, the prediction performance of three GAN models (Pix2Pix, CycleGAN and Pix2PixHD) on pollutant dispersion under two scenarios (design optimization and real-time monitoring) is trained and compared, and the impacts of two optimization methods (training stability and training data optimization) on the performance of the models are explored. The results show that the clustered dataset can reflect the real Shenzhen residential blocks characteristics. On the test set, the MAE of Pix2PixHD reaches 0.132 and the inference time is less than 1 s. This study is the first to realise second-scale prediction of pollutant dispersion in urban residential blocks by coupling a clustered residential block dataset with generative adversarial networks, thereby establishing the methodological and data foundation for future cross-city generalization and real-time morphology optimisation.