Otimização de controladores PID e Fuzzy-PID em sistemas de regulação térmica com aprendizagem de máquina

Abstract

This research is situated within the scope of Industry 4.0, where accuracy and efficiency in thermal control systems have become increasingly essential. The main objective of this work is to apply and evaluate different Machine Learning models to optimize the control action in systems based on Proportional-Integral-Derivative (PID) and Fuzzy-PID controllers. The methodology followed a supervised learning workflow focused on regression, using a public dataset extracted from an intelligent manufacturing system. After data preparation - including cleaning, standardization, and feature selection - nine regression models were trained and evaluated, such as Random Forest, Multi-Layer Perceptron, and different variants of Linear Regression (Linear, Ridge, and Lasso), SVR, Gradient Boosting, XGBoost and LightGBM. The prediction target corresponded to the control action generated by the Fuzzy-PID controller. The results show that Machine Learning can effectively capture the behavior of the adaptive controller, even in nonlinear thermal systems subject to hysteresis and environmental variations. Among the models ested, Lasso Regression achieved the best overall performance, reaching '2 = 0.9407 and the lowest error metrics, indicating that the relationship between process variables and control action follows a sufficiently regular structure to be learned by a regularized linear model. These findings reinforce the potential of Machine Learning as a support tool for controller tuning, offering greater predictability, computational efficiency, and reduced calibration effort. The study highlights that modern regression techniquescan complement - and in some scenarios, outperform - traditional approaches, directly contributing to the advancement of intelligent solutions in industrial environments aligned with Industry 4.0 principles.