Deep Learning for Realistic Wind Field Prediction in a typical Urban Morphology for Application to Small Unmanned Aerial Systems

Abstract

This works aims to generate realistic wind data in urban spaces, which is essential in developing, testing and ensuring the safe operations of Small Unmanned Aerial Systems (sUAS) using Deep Learning (DL). This provides an alternative to existing turbulence models, specifically aimed at urban air spaces. We devise and utilize a Non-Intrusive Reduced Order Model (NIROM) approach to replicate and realistically predict wind fields in urban spaces. The method uses Large Eddy Simulation data from well-established computational fluid dynamics solvers like OpenFOAM to devise the NIROM. High-fidelity data generated from OpenFOAM is decomposed using Proper Orthogonal Decomposition (POD) into its orthogonal modes and basis. These orthogonal modes obtained over time are trained on specialized Recurrent Neural Networks like Long-Term Short Memory (LSTM) to complete the NIROM formulation. This method combined the traditional reduced order modeling approach with deep learning techniques to devise a framework for easy building and application of Machine Learning (ML) based Reduced Order Models (ROMs). A typical urban morphology subject to the wind is chosen and considered a test case for demonstrating the method.