Comparative analysis of neural network architectures for adaptive control of nonlinear thermal systems

Abstract

This paper presents a comparative study of control strategies for nonlinear thermal systems with significant time delay and thermal inertia. Three controllers are evaluated: a classical PID controller, a multilayer perceptron (MLP)-based adaptive controller, and a recurrent neural network controller based on a gated recurrent unit (GRU). A nonlinear thermal process model is used as a benchmark, incorporating heat transfer dynamics, radiative losses, actuator saturation, and external disturbances. The controllers are assessed based on temperature tracking performance, tracking error, control signal behavior, and mean squared error (MSE). The results demonstrate that neural network-based controllers outperform the classical PID approach. The MLP-based controller improves transient response and reduces tracking error. The GRU-based controller further enhances performance by capturing temporal dependencies, leading to improved tracking accuracy and disturbance rejection.
However, the improved performance of the GRU controller is accompanied by increased variability in the control signal, highlighting a trade-off between accuracy and smoothness. The findings confirm the potential of recurrent neural networks for adaptive control of nonlinear and delay-dominant thermal processes.