<p>With the increasing penetration of renewable energy and power electronic equipment into the power system, the characteristics of the grid have undergone significant changes. Therefore, the equivalent inertia and damping of the grid have been significantly weakened, escalating the risk of oscillations. However, the traditional adaptive strategy for virtual synchronous generators (VSG) with fixed parameters struggles to meet the requirements of inertia-damping coordinated regulation under complex disturbances. To address this issue, this paper proposes an improved parameter adaptive method for VSG based on the Deep Deterministic Policy Gradient (DDPG) algorithm in deep reinforcement learning. Specifically, the method reconstructs the fixed parameters in traditional adaptive control as learnable dynamic variables, and designs a reward function targeting frequency and voltage stability. Then, it employs the DDPG algorithm to realize real-time optimization of these parameters. Thus, the proposed method achieves nonlinear dynamic matching of inertia and damping coefficients. Simulation results demonstrate that, compared with the traditional adaptive strategy, the proposed method has significant advantages. Specifically, under power disturbances, the frequency response overshoot is reduced by 0.06% points, and the regulation time is shortened by 25%. Meanwhile, the active power output overshoot decreases by 1.25% points, and the regulation time is reduced by 28%. Through MATLAB/Simulink simulations, the control performances of fixed inertia-damping, traditional adaptive strategy, and DDPG-based adaptive strategy are compared. It has been verified that the proposed control strategy is feasible and effective.</p>