Rule-based PI tuning with neural networks for first order process plus dead time

Abstract

This Bachelor’s thesis explored the function approximation capacity of multi-layer perceptrons (MLPs) for tuning PI controller parameters when the process can be modelled to a first-order delayed process and shows varying dynamics with time. Traditional PI tuning rules provide a solid approach to insure a good initial guess with desired input, but their performance is often limited based on an important ratio referred to as normalized dead time. A neural network training methodology will be presented that learns from a comprehensive dataset to identify when each tuning rule is optimal and applies the most suitable rule accordingly.

The solution proposed focuses on achieving optimal performance-robustness trade-offs across a range of normalized dead time values, and based on a Matlab analysis it demonstrated good generalization capacity, though performance varies across different operating regions.