Active flow control of square cylinder adaptive to wind direction using deep reinforcement learning

This paper concentrates on implementing active flow control (AFC) through computational fluid dynamics simulations involving square cylinders at various wind attack angles using deep reinforcement learning (DRL). Specifically, the soft actor-critic algorithm is utilized to govern the control configurations of multiple jets positioned at the four corners of a square cylinder submerged in a two-dimensional flow domain. Surface pressure sensors are strategically placed to monitor the flow state, providing a practical engineering solution. The learning environment accommodates four distinct flow configurations with wind attack angles (α) of 0°, 15°, 30°, and 45°, respectively. The results demonstrate the significant effectiveness of DRL control in substantially mitigating fluctuations in both lift and drag coefficients, resulting in notable reductions in mean drag coefficients of 59.2%, 51.8%, 53.7%, and 36.9% at α = 0°, 15°, 30°, and 45°, respectively. Notably, this control mechanism effectively minimizes drag for wind attack angles not encountered previously, highlighting the generalization capabilities of deep neural networks. This achievement marks a significant milestone in advancing the practical applications of DRL in the field of AFC.