A Rapid Wind Velocity Prediction Method in 

Built Environment Based on CycleGAN Model



| Keywords |




Deep learning, Wind velocity prediction, Pix2pix, CycleGAN



Abstract |



Although the wind microclimate and wind environment play important roles in urban prediction, the time-consuming and complicated setup and process of wind simulation are widely regarded as challenges. There are several methods to use deep learning (DL) models for wind speed prediction by labeling pairs of wind simulation dataset samples. However, many wind simulation experiments are needed to obtain paired datasets, which is still time-consuming and cumbersome. Compared with previous studies, we propose a method to train a DL model without labelling paired data, which is based on Cycle Generative Adversarial Network (cycleGAN). To verify our hypothesis, we evaluate the results and process of the pix2pix model (requires paired datasets) and cycleGAN (does not requires paired datasets), and explore the difference of results between these two DL models and professional CFD software. The result shows that cycleGAN can perform as well as pix2pix in accuracy, indicating that some random city plans image samples and random wind simulation samples can train surrogate models as accurate as labelled DL methods. Although the DL method has similar results to the professional CFD method, the details of the wind flow results still need improvement. This study can help designers and policymakers to make informed decisions to choose Dl methods for real-time wind speed prediction for early-stage design exploration.




Ximing Zhong




Chuheng Tan

Conference Video 



| Figures |



Fig. 1 The whole workflow of comparing two DL models for wind prediction in this study

Fig. 2. Datasets samples for the pix2pix model and cycleGAN model separately

Fig. 5 Training loss diagram and FID of pix2pix and cycleGAN model.




| Acknowledgment |