Funded Projects

Wind tunnels usually serve as a valuable tool to ensure the safety of critical civil structures under extreme winds. The non-synoptic wind events such as downbursts usually exhibit spatiotemporally wind profile and nonstationary fluctuations. Hence, it is not feasible to use passive devices (e.g. roughness elements, spires) in conventional boundary layer wind tunnels to reproduce these wind events. Multiple-fan wind tunnels, where customized control of individual fan is realized, are promising candidates for physical simulation of non-synoptic winds. The design of controller usually requires the knowledge of the dynamics of controlled system, which is challenging considering the complicated dynamics of the wind tunnel, for example, the fluid interactions among the flows injected from individual fan. As a result, controller design practice for the multiple-fan wind tunnels still needs time-consuming hand tuning. The accuracy and efficiency of tuning controllers largely rely on the experience of control experts, which may not be always available in wind engineering community. Hence, it is necessary to develop novel schemes to facilitate the controller design and take full advantage of the capacity of multiple-fan wind tunnels to simulate complicated wind fields. Recent years have witnessed the fast development of Artificial Intelligence (AI), which has shown great potential in solving complicated tasks. However, the great potential of utilizing AI to control multiple-fan wind tunnels has not been investigated yet. This proposed project will utilize deep reinforcement learning, which belongs to an important category of AI, for controller design of multiple-fan wind tunnels. In particular, the optimal controller for accurately tracking the target wind speeds at sensor locations will be obtained by directly interacting with the wind environment, which eliminates the need of complicated model of system dynamics and the experience from control experts.