The large-scale access of new energy with strong randomness to the power grid brings more and more weak control performance to the power grid. The reinforcement learning with the characteristics of Markov stochastic process has advantages in solving random problems. However

when facing the large-scale access of new energy

it still faces the problem of failing to obtain the optimal solution and the control performance is not ideal. Therefore

a kind of soft actor-critic multi-agent cooperative deep reinforcement learning algorithm with value estimation correction is proposed to obtain the multi-region cooperative optimal solution. The problem of value overestimation is alleviated by the state-action distributed value function

thus the global optimal solution is obtained. The improved IEEE standard two-region model and southwest three-region power grid model are simulated to verify the effectiveness of the proposed algorithm

and it has better control and frequency stability than many other control methods.