Physics-aware deep learning for wind prediction over complex terrain

Abstract

Rapidly computing the wind flow over complex terrain features is a challenging problem with many potential fields of application. Such domains include autonomous UAV path planning, naturally ventilated building design or even wind farm layout optimization. Traditionally, this requires performing Computational Fluid Dynamics (CFD) simulations, which are computationally expensive and time consuming. One recent approach has been to use deep convolutional neural networks as a data-driven simulation alternative which leverages CFD simulation results for training. Albeit fast and relatively accurate, this method does not, however, explicitly take into account flow properties. We believe that doing so could improve prediction accuracy, especially in previously under-predicted areas which usually exhibit strong shearing flows and large velocity magnitudes. In this thesis, we investigate training deep neural networks with physics based loss functions and loss weighting methods, constructed to accentuate desirable flow features and properties. Furthermore, we explore other potential solutions, such as a physics informed adversarial approach, or using the recently proposed SPADE network which reported good spatial expressivity. Comparisons to baseline results are provided for all tested solutions, allowing us to select the combination of methods which provides the most accurate wind flow predictions. We show that training a network with a loss weighted with the velocity gradient magnitude results in a 5% decrease in maximum prediction error while retaining a comparable mean prediction error, when compared to the baseline approach.