The stability of DC bus voltage is the key to the normal operation of flexible multi-state switch. The hard switching of multiple working modes of flexible multi-state switch (FMS) will lead to violent fluctuation of DC bus voltage, and the smooth switching is the most effective way to realize bus voltage stability. The principle of smooth switching of steady-state inverse model is analyzed in detail. Then, in view of its shortcomings, the smooth switching technology of steady-state inverse model of radial basis function neural network (RBFNN) is deeply studied, and the principle of using RBFNN to improve the smooth switching of steady-state inverse model is given. Detailed experimental verification is carried out. The theoretical analysis and experimental results show that the modified PI output is controlled by RBFNN to compensate the influence of disturbance on the system, and the modified PI output is superimposed with the steady-state inverse model output to generate the inner-loop reference value, which can effectively smooth the oscillation of bus voltage at the moment of switching, and can realize the advantages of small voltage fluctuation of DC bus, fast response speed, good dynamic characteristics and wide adaptability to working conditions.