Active Simulation of Transient Wind Field in a Multiple-Fan Wind Tunnel via Deep Reinforcement Learning

Abstract

The transient wind field during a nonsynoptic wind event (e.g., thunderstorm downburst) presents time-varying mean and nonstationary fluctuating components, and hence is not easy to be reproduced in a conventional boundary-layer wind tunnel with various passive devices (e.g., spires, roughness elements, and barriers). As a promising alternative, the actively controlled multiple-fan wind tunnel has emerged to effectively generate the laboratory-scale, spatiotemporally varying wind flows. The tracking accuracy of target wind speed histories at selected locations in the multiple-fan wind tunnel depends on the control signals input to individual fans. Conventional hand-design linear control schemes cannot ensure good performance due to the complicated fluid dynamics and nonlinear interactions inside the wind tunnel. In addition, the determination of the control parameters involves a time-consuming manual tuning process. In this paper, an accurate and efficient control scheme based on deep reinforcement learning (RL) is developed to realize the prescribed spatiotemporally varying wind field in a multiple-fan wind tunnel. Specifically, the fully connected deep neural network (DNN) is trained using RL methodology to perform active flow control in the multiple-fan wind tunnel. Accordingly, the optimal parameters (network weights) of the DNN-based nonlinear controller are obtained based on an automated trial-and-error process. The controller complexity needed for active simulation of transient winds can be well captured by a DNN due to its powerful function approximation ability, and the “model-free” and “automation” features of RL paradigm eliminate the need of expensive modeling of fluid dynamics and costly hand tuning of control parameters. Numerical results of the transient winds during a moving downburst event (including nose-shape vertical profiles, time-varying mean wind speeds, and nonstationary fluctuations) present good performance of the proposed deep RL-based control strategy in a simulation environment of the multiple-fan wind tunnel at the University at Buffalo.