Implementation of a Neural Control System Based on PI Control for a Non-linear Process

Abstract

This paper explores the possibility of using a machine learning algorithm such as artificial neural networks to control a non-linear liquid level system. To achieve this objective, PI controllers were designed for two different scenarios: In the first, a single PI controller was used to control the system at one setpoint. In the second, 4 PI controllers were designed in order to cover a wider operating range of the plant. The input and output signals from the PI controllers were used to train a controller based on artificial neural networks. The neural network that presented greater simplicity and lower computational cost was selected. In this case, a neural network with 3 hidden layers and 20 neurons per layer was the one that best recreated the dynamics of the PI controllers. The root-mean-square error (RMSE) was used to validate the results obtained with the PI controllers and with the controller based on neural networks. In both scenarios the variations of the error were smaller when the neuronal controller was used than when the PI controllers were used. The results show that machine learning algorithms such as artificial neural networks can be used effectively to control processes whose dynamics are complex.