Non-intrusive reduced order model of urban airflow with dynamic boundary conditions

Abstract To address the problem that urban wind field simulations are limited by high computational requirements, a supervised machine learning framework for the non-intrusive model order reduction of urban airflow is proposed to provide fast predictions with dynamic boundary conditions. An integrated atmospheric modeling system is established by coupling the Parallelized LES Model (PALM) with the Urban Canopy Model (UCM) which embedded in the Weather Research and Forecasting (WRF) model to simulate the high-resolution urban wind field. Within this machine learning framework, encoders and decoders are trained to compress and reproduce the full-order airflow data. The eXtreme Gradient Boost regression models (XGBoost) are trained to map the boundary conditions to the latent vectors in the encoders and decoders. Thus, high resolution urban airflow can be predicted rapidly by the XGBoost-decoder models with the input of boundary conditions. By comparing four types of encoding-decoding models (i.e., the proper orthogonal decomposition models, autoencoders with linear/nonlinear fully connected neural networks and autoencoders with Convolutional Neural Network (AE-CNN)), we find that the AE-CNN model with advantages of portable model size and convenient updating procedure is capable for high resolution simulations of urban airflow with dynamic boundary conditions at a low computational cost.