Data-driven PID controller tuning using Bayesian neural networks: A study on automatic voltage regulator control

Abstract

In this paper, an effective data-driven tuning method for the PID controllers of automatic voltage regulators (AVRs) is proposed by Bayesian neural networks (BNNs). In this method, the AVR is simulated as a dynamic control system consisting of an amplifier, exciter, generator and sensor. A full dataset is obtained by varying the PID parameters in a systematic way and recording the corresponding closed-loop performance indices, such as settling time, and maximum overshoot. Then, a nonlinear mapping between controller parameters and performance of the system is built on this data set. Then, a Bayesian neural network (BNN)  was trained to learn the inverse mapping from the desired performance specifications to the optimal PID parameters. Differently from traditional neural networks, the BNNs take advantage of probabilistic inference, which allows uncertainty measurement in predictions and better generalization ability. This is especially useful in controlling applications where system parameters may be uncertain or corrupted by noise. Arduino Uno-based simple real-time simulation studies verify the proposed method. The real-time simulation results demonstrate that the BNN-based tuning approach yields better transient performance, with lower overshoot and quicker settling time, than the traditional PID control. The proposed method is a robust, efficient and data-driven approach for the optimal PID controller design in AVR system, especially in the presence of model uncertainty and parameter variations.