A prediction of urban boundary layer using Recurrent Neural Network and reduced order modeling

Abstract

A prediction of urban boundary layer using Recurrent Neural Network (RNN) and reduced order modeling is performed. By employing the simulation of filtered Navier–Stokes equations for the generation of RNN training data, the study provides a reliable foundation for model training, validated against established experimental data. Dimensionality reduction, achieved through singular value decomposition, is conducted with consideration of fluid field reconstruction capability which investigated through energy loss, Root Mean Square Error and Structural Similarity Index Measure of second order statistic of chosen value and Turbulent Kinetic Energy (TKE) budget terms using reconstructed turbulent flow data. For the RNN prediction, various RNN configurations, ultimately selecting an LSTM model with optimized settings, including the LeakyReLU activation function for its superior performance in turbulent flow predictions. This configuration enables precise forecasting of urban boundary layer dynamics, covering both the instantaneous turbulent flow and the profile of TKE budget terms. It also demonstrates a substantial improvement in computational efficiency, achieving speeds up to forty times faster than conventional methods.