As the proportion of wind power integrated into the grid continues to increase

the frequency and amplitude of grid voltage fluctuations have significantly improved

and even the risk of exceeding limits arises. In response to the high-frequency and amplitude fluctuations and over-limit issues of grid-connected point voltage and the terminal voltage within the wind power cluster which are caused by the integration of wind power clusters into the grid

a real-time hierarchical voltage control method for wind power clusters based on multi-agent deep reinforcement learning is proposed. Firstly

in accordance with the timing coupling relationship of reactive power resource regulation

the online reactive power optimization problem of wind power clusters is transformed into a Markov decision process. Then

the minimum voltage fluctuation range and the minimum network loss are taken as the optimization control objectives

and the multi-agent deep deterministic policy gradient algorithm is adopted to achieve precise real-time control of wind turbines and real-time adjustment control of reactive power compensation devices under "offline centralized learning and online decentralized execution". Finally

through the case analysis of wind farm clusters in a regional high-voltage distribution network

the research results indicate that this strategy can achieve fine real-time control of the grid-connected point voltage and terminal voltage of wind power clusters based on improving the calculation speed

with significant effects on voltage regulation and loss reduction.