Abstract

In recent years, rapid urban population growth has increased the focus on urban microclimate studies due to their significant impact on human comfort and building energy performance. Computational Fluid Dynamics (CFD) plays a vital role in accurately simulating fluid flows, essential for evaluating the effects of urban microclimate parameters. However, balancing computational cost with precision simulation accuracy is challenging when using traditional numerical methods. Recent advances in deep learning offer promising solutions to accelerate CFD simulations and reduce computational costs. However, one of the major limitations of deep learning models is their tendency to accumulate errors over time, resulting in insufficient reliability in long-term predictions. In response, this study proposes the implementation of Denoising Diffusion Probabilistic Models (DDPM) as a post-processing approach for reducing error accumulation in deep learning models.
The research utilizes CityFFD, an accurate Fast Fluid Dynamics (FFD) model, to simulate complex urban airflow dynamics. The initial phase of the study includes a comprehensive validation of CityFFD against established CFD benchmarks, confirming its effectiveness in generating high-fidelity airflow patterns around buildings.
Subsequently, Convolutional Autoencoders (CAE) and U-Net, two established deep learning models, are employed to predict airflow around a single building. DDPM is then applied as a post-processing technique to refine the reconstructed flow fields from these models to enhance the fidelity of the generated flow fields and reduce error accumulation in sequential timestep predictions.
Lastly, to extend the application of DDPM in reducing error accumulation in urban microclimate simulations, the Fourier Neural Operator (FNO) is employed to simulate the dynamics of urban-scale wind flow simulations. This approach demonstrates the potential of DDPM as a post-processing technique to improve the accuracy and reliability of advanced deep learning methods such as FNO in urban airflow simulations.