Improving the performance of a CAE-based reduced-order model for predicting turbulent airflow field around an isolated high-rise building

Abstract

Computational fluid dynamics (CFD) has been considered as a promising numerical approach in fluid dynamics problems, such as urban airflow prediction. However, airflow field prediction using CFD models is time-consuming. Thus, they cannot be used for (near) real-time and long-term simulations. Reduced-order models (ROMs) are emerged to obviate this limitation. Deep learning (DL) algorithms have been used for developing non-intrusive ROMs (NIROMs) in fluid dynamics applications. In the present study, three different approaches, namely, convolutional autoencoder (CAE), multi-scale CAE (MS-CAE), and self-attention CAE (SA-CAE), are developed for dimensionality reduction, which is considered the first step of the development of a NIROM. The developed models are then used to find a low-dimensional representation of the original data. Afterward, a parallel long short-term memory (LSTM) network is employed for computing the temporal dynamics of the obtained low-dimensional space. The models are trained to reconstruct a turbulent airflow field in the wake region of an isolated high-rise building, located in an unstable thermal stratification condition, using validated CFD data. The models show promising performance in reconstructing the airflow field. However, discrepancies can be observed in the regions with intense gradients. Also, power spectral density functions (PSD) obtained from the reconstructed data are in good agreement with those obtained from the CFD results. On the whole, SA-CAE performs better in reconstructing the airflow field than the other models, followed by MS-CAE and CAE.