Aerodynamic optimization of wind turbine blades via surrogate-assisted deep reinforcement learning

This study employs the Soft Actor-Critic (SAC) deep reinforcement learning algorithm to optimize the aerodynamic performance of wind turbine blades. A surrogate model based on a multilayer perceptron is constructed to predict aerodynamic responses, enabling rapid prediction of aerodynamic characteristics and significantly improving optimization efficiency. The aerodynamic design optimization problem is formulated as a Markov Decision Process, where the SAC agent learns to optimize airfoil geometries through interaction with the environment. The results demonstrate that the surrogate model exhibits high accuracy in predicting lift and drag coefficients, with mean absolute errors of 0.42% and 0.74%, respectively. During the SAC training process, the optimization gradually converges, and the final optimized airfoils achieve significant improvements in lift-to-drag ratios under thickness constraints. For airfoils with low, medium, and high initial lift-to-drag ratios, the optimized lift-to-drag ratios improve by 69.05%, 28.77%, and 12.65%, respectively, while satisfying aerodynamic constraints. Further analysis reveals that the optimal strategy evolves from broad exploration to fine-tuned local adjustments, achieving efficient convergence. Additionally, the optimization process demonstrates strong physical consistency and engineering applicability by effectively regulating geometric and flow characteristics. This work provides an efficient, intelligent, and physically sound optimization approach for complex aerodynamic design problems.