Abstract: China has proposed the carbon peaking and carbon neutrality goals which based on a new power system domicated by renewable energy sources. Wind power is the key to achieve the goal, and there is still need to be improved in terms of cost reduction and energy efficiency increase. We study the optimisation control strategy based on the read wind farms data, to improve the output power of the wind farms, so as to further increase the utilisation rate of wind power. This study mainly focuses on the problem of wake effects among wind turbines and proposes a multivariate optimal control strategy for wind farms based on model-free Deep Reinforcement Learning(DRL). The strategy adopts the Proximal Policy Optimization (PPO) algorithm to optimize multiple variables such as yaw angle, tilt angle, blade pitch angle, and tip speed ratio (TSR) in a dynamic wind farm. Using data generated by the agent during operation for intelligent learning, the optimal control strategy is obtained, overcoming the limitations of traditional mathematical optimization methods. Simulation results show that compared to existing wind turbine control algorithms, the model-free DRL-based multivariable optimization control strategy significantly improves computational efficiency, reduces the difficulty of parameter optimization, optimizes wake direction and intensity, and it has a 37.08% increase in average output power after optimisation.

Key words: Wind farm, Deep reinforcement learning, Wake control, Multivariable control, Wind generation power