A learning algorithm for a data-driven controller based on fictitious reference iterative tuning

Abstract

In this study, a learning algorithm for a data-driven proportional-integral-derivative (DD-PID) controller that uses a database for tuning control parameters is considered. PID controllers are still used in many process systems. However, if systems exhibit nonlinearity, PID controllers with fixed PID parameters cannot achieve a desired control performance when a system's equilibrium points are changed by set point changes. To solve this problem, the DD-PID controller was proposed. The DD-PID controller can maintain good control performance for nonlinear systems because it learns the PID parameters in its database so that the control performance around each equilibrium point has a desired characteristic. In this research, a fictitious reference iterative tuning (FRIT) method is applied as the learning method of the DD-PID controller. This method can perform offline learning and obtain a desired tracking property of a closed-loop system by storing one-shot operating data given by a PID controller with fixed PID parameters into a database using the concept of the FRIT. This paper also shows that the DD-PID controller can achieve good control performance for unlearned system changes and disturbances by performing online learning with the same criterion used in offline learning. The effectiveness of the proposed method is evaluated using simulation results.