To improve power overshoot, frequency deviation and transient oscillation in the grid-connected control of modular multilevel converters (MMC) with virtual synchronous control (VSG), a grid-connected power control method combining radial basis function supervisory tuning and model predictive power compensation is [ ]proposed. First, an active-power closed-loop model of the modular multilevel converter with virtual synchronous control was established, and the effects of virtual inertia and damping parameters on dynamic response were analyzed to determine the feasible parameter ranges. Then, a radial basis function (RBF) neural network was used to reconstruct the virtual inertia and equivalent damping parameters according to the frequency deviation and its rate of change. Finally, model predictive control (MPC) was introduced to perform receding-horizon optimization compensation for the active-power error. Simulation results show that, under active-power step, frequency drop and power-flow reversal conditions, the proposed method can reduce power deviation and frequency deviation, shorten transient recovery time, and maintain the stability of submodule capacitor voltage and three-phase AC current. The method improves the dynamic regulation capability and operational stability of the grid-connected modular multilevel converter system with virtual synchronous control.