Design and comparison of a PI controller and an ANN controller for a three-phase three-wire grid-connected NPC inverter

In this study, two different control systems are proposed for a three-phase, three-level, three-leg, three-wire (3P3L3L-3 W), grid-connected (GC) neutral point clamped (NPC) voltage source inverter. The first controller is a proportional-integral (PI)-based technique developed using a detailed mathematical model, which includes a virtual closed loop created between the inverter and grid neutral points via Kirchhoff’s voltage law. This design offers decoupled dq axes current control (CC) as well as capacitor voltage balancing (CVB) combined with 0-axis CC. The second controller is an artificial neural network (ANN)-based technique trained with the simulation data of the PI-based controller. Three independent ANN controllers are constructed for d-axis, q-axis, and CVB control, respectively. Each ANN is designed with minimal structure to achieve low computational cost without compromising performance. The originality of the study lies in the unified modeling-based design of the PI-based controller including a novel CVB method, and the proposed low-complexity multi-ANN structure that replicates and enhances this control behavior. Both controllers are evaluated through nine test cases, including noise injection, sensor errors, grid disturbances, non-linear power sources, and component aging. The simulation results show that both approaches successfully regulate the inverter. The ANN-based controller outperforms the PI-based one in terms of steady-state accuracy and robustness under uncertain conditions, while also reducing computational burden compared to similar ANN controllers reported in the literature.