The Conventional model predictive control (C-MPC) uses only a single voltage vector in each control period, resulting in poor control performance. In addition, the computational burden in discrete space state generated by three-level inverter cannot be ignored. To improve the control performance, this paper proposes a variable coefficient virtual vector MPC (VC-VV-MPC) method, a five-level virtual space voltage vector is constructed to improve the control performance, and a scaling coefficient is used to further improve the control performance. In addition, a voltage vector pre-selection method is used to reduce the computational burden. The simulation and experiment results shown that, compare with C-MPC and adjacent vector MPC (AV-MPC), the proposed VC-VV-MPC can achieve the improvement of control performance under lower computational burden.