Abstract

Different thermal stratification conditions, namely, stable, isothermal (or neutral), and unstable, can locally occur in urban areas. Alteration in the thermal condition of an urban area may significantly change the airflow pattern and pollutant dispersion process by affecting both the mean and fluctuating components of the variables. The unstable effects can increase the vertical flow movement, while the stable ones can suppress it. Furthermore, unstable conditions increase turbulence kinetic energy (TKE), which increases the fluctuations in concentration. On the other hand, stable conditions lead to buoyant destruction. Due to frequent changes in the boundary conditions, a model is required for monitoring these situations, which can be used as a fast-response (near real-time) model. This thesis aims to propose a systematic approach for analysis and reduced-order modeling of the airflow and concentration fields under non-isothermal conditions.
The present study uses a high-fidelity computational fluid dynamics approach, i.e., embedded large eddy simulation (ELES), to simulate the impact of the aforementioned thermal conditions on the airflow and concentration fields. The model considers the pros of both the Reynolds-averaged Navier-Stokes, RANS, (i.e., high speed), and large eddy simulation, LES, (i.e., high accuracy) approaches. After thoroughly analyzing the results, the proper orthogonal decomposition (POD) and frequency analyses are performed to investigate the impact of thermal conditions on the turbulence structure of the flow field. Considering the most energetic POD modes can lead to a good approximation of the whole airflow field, which is an important finding in developing a reduced-order model (ROM). Due to the limitations arising from the linear nature of POD, convolutional autoencoder (CAE)-based methods are used for model order reduction, using the unstable dataset generated by ELES. In addition to the conventional CAE, multiscale CAE (MS-CAE) and self-attention CAE (SA-CAE) are developed to capture multiscale and long-range dependencies among the datapoints, respectively. Afterwards, a parallel long short-term memory (LSTM) network is used to compute the temporal dynamics of the low-dimensional subspaces. ROMs maintain prediction accuracy at an acceptable level compared to ELES, while reducing the data reconstruction time to the order of seconds.