Low-altitude unmanned aerial vehicles (UAVs) often encounter strong turbulence during flight, such as gusts, leading to operational instability. To ensure stable flight of low-altitude UAVs, deep reinforcement learning (DRL) and co-flow jet (CFJ) control are utilized. This paper proposes DRL-based and attention-based CFJ control to obtain appropriate control strategies for a CFJ airfoil in strong turbulence, achieving lift stabilization, lift enhancement, drag reduction, and maintaining jet energy, while improving the visual explanation of intelligent flow control. The cylinder wake flow is used to simulate the actual unsteady working environment of low-altitude UAVs at the Reynolds number of 2.6 × 106. Based on the deep deterministic policy gradient algorithm, first, utilizing the standard fully connected neural network with lift stabilization and enhancement as control objectives, generalization is applied to different initial unsteady flow fields, reducing the standard deviation and increasing the average of the lift coefficient by 74% and 87%, respectively. Then, based on the attention branch network (ABN), while ensuring control effectiveness and generalization performance, the transparency of the agent's decision-making process is improved, showing that the agent primarily focuses on regions with significant pressure changes in the flow field. Finally, based on the ABN, the reward function additionally considers drag and jet energy consumption. The agent is still able to find a reasonable control strategy, achieving the control objective of lift stabilization and a 13% drag reduction, while keeping the jet energy almost unchanged.
Yin et al. (Thu,) studied this question.