The new power system based on new resources is characterized by low inertia and weak damping. Hence

grid-forming (GFM) converter control represented by virtual synchronous generator (VSG) is proposed to address the above challenges. However

in islanded microgrid dominated by a grid-forming(GFM) converter

the frequency of the common connection point is affected by high-dimensional parameters and there is a high coupling phenomenon between the units

so it is difficult to accurately reveal the operating characteristics of cluster VSGs from the perspective of mathematical modelling

and even more difficult to design a universal controller to enhance the system frequency characteristics. To solve the above problems

this paper proposes a deep reinforcement learning method based on physical model guidance: firstly

a small-signal model of VSG with secondary frequency management link is established

then a variable coefficient control is proposed with the small-signal model as a theoretical support

and it is introduced as a physical model into the deep reinforcement learning environment to improve the training efficiency

and then a reward/punishment mechanism is used to realize the VSG frequency adaptive control. Finally

in this paper

the trained model is extended to multiple VSGs cooperative control using reconstruction and migration neural networks. To verify the effectiveness of the proposed method

this paper is fully validated by MATLAB/Simulink and RT-LAB.