A Neural Network-Based Current Controller for a Three-Phase Grid-Connected Voltage-Source Converter

Abstract

This paper presents a compact neural network (NN) current controller for a three-phase, grid-connected voltage-source converter (VSC). The proposed NN replaces the fixed-gain proportional–integral (PI) compensator within a traditional d-q synchronous frame current-control architecture. An eight-dimensional feature vector of present and past system variables predicts incremental modulation updates through a single-hidden-layer multilayer perceptron (MLP) with 16 hidden neurons and only 178 trainable parameters, requiring 712 bytes of parameter storage. Temporal information is introduced externally through one-step error history and recursive modulation-domain accumulation, avoiding internally recurrent layers such as LSTM cells. Trained on 11 million samples from 10,000 PSIM simulations, the NN controller reduces averaged overshoot by 59%, settling time by 94%, integral absolute error by 27%, and steady-state error by 42% compared to the PI teacher. In simulation-based reference-step tests, the NN controller learns a smoother bounded incremental law from an underdamped PI current-loop teacher, reducing the averaged overshoot, settling time, integral absolute error, and steady-state error magnitude. Hardware-in-the-loop and prototype validation remain outside the scope of this work.