In order to solve the problem that the current-based midpoint-clamped three-level inverter open-circuit fault diagnosis is easily affected by the load changes, this paper mainly improves the accuracy from the fault feature differentiation. Firstly, VMD improves the modal aliasing phenomenon, and its optimal modal number K and penalty coefficient α are optimized by SMA, which improves the independence of fault features. Second, due to the relatively smooth distribution of wavelet packet energy, which can effectively overcome the characteristics of load influence, the maximum value of the two-layer wavelet packet energy of each IMF is taken as the fault feature quantity, so that the time-frequency feature information is more centralized which further improves the fault feature differentiation without the influence of varying loads. Finally, the fault features are applied to the neural network for training, and the weights and thresholds of the model are optimized by SSA, which solves the problem of local optimum of the model and improves the accuracy of fault identification. Through the simulation experimental results of 17 open-circuit faults in the NPC three-level inverter circuit model, the diagnostic accuracy of this method reaches 98.99%, which is applicable to the online fault diagnosis of NPC three-level inverters under variable load conditions.