Neural Network-Based Voltage Control for Safer, Smarter Droop-Controlled Inverters

Abstract

This This paper presents the design, simulation, and comparative evaluation of an Artificial Neural Network (ANN)- based internal control strategy for three-phase grid-connected droop-controlled distributed generators (GCDCDGs). The proposed architecture integrates a novel droop-based power and current limiting strategy with an ANN-enhanced voltage control loop that replaces conventional Proportional-Integral (PI) controllers. The ANN is trained offline using 400,000 data samples captured from a supervisory PI controller operating across four comprehensive dynamic disturbance scenarios— upstream grid frequency steps, voltage magnitude sags, balanced three-phase load injection, and induction motor switching—with Levenberg-Marquardt Backpropagation (LMBP) employed as the optimization algorithm. Simulation studies in MATLAB/Simulink evaluate performance under six case studies spanning frequency drops (50→49 Hz), severe voltage sags (50→22.5 V), combined disturbances with resistive-inductive loads, and dynamic motor starting. For rigorous benchmarking, a Model Predictive Controller (MPC) is also implemented as a performance upper bound. Results demonstrate that the ANNbased controller achieves approximately 50% reduction in Total Harmonic Distortion (THD)—from 2.8–4.0% with PI to 1.6– 2.0%—and 12–24% improvement in transient rise/fall times, attributable to the ANN's ability to approximate the optimal nonlinear voltage-to-current mapping while generating smoother control signals. The proposed droop-based limiting strategy successfully constrains output currents (iod ≤ 12 A, ioq ≥ −7.2 A) and active/reactive power within safe inverter ratings under all disturbance scenarios, without requiring external limiters or anti-windup schemes. While MPC achieves the lowest steadystate THD (<1%), its computational burden is approximately one order of magnitude higher than PI or ANN inference, limiting suitability for resource-constrained embedded implementations. The ANN controller thus offers the optimal practical trade-off: near-MPC power quality, superior dynamic response relative to PI, and deployment-ready computational efficiency. This work validates ANN-based control as a viable, high-performance solution for next-generation grid-connected inverters requiring compliance with harmonic standards (IEEE 1547, IEC 61727) and robust operation under grid disturbances. Future work will explore MPC-supervised ANN training, Lyapunov-based stability analysis, and hardware-in-the-loop experimental validation.