Ang, Kiam Heong
Evolutionary learning and global search for multi-optimal PID tuning rules.
PhD thesis, University of Glasgow.
Full text available as:
With the advances in microprocessor technology, control systems are widely seen not only in industry but now also in household appliances and consumer electronics. Among all control schemes developed so far, Proportional plus Integral plus Derivative (PID) control is the most widely adopted in practice. Today, more than 90% of industrial controllers have a built-in PID function. Their wide applications have stimulated and sustained the research and development of PID tuning techniques, patents, software packages and hardware modules. Due to parameter interaction and format variation, tuning a PID controller is not as straightforward as one would have anticipated. Therefore, designing speedy tuning rules should greatly reduce the burden on new installation and ‘time-to-market’ and should also enhance the competitive advantages of the PID system under offer.
A multi-objective evolutionary algorithm (MOEA) would be an ideal candidate to conduct the learning and search for multi-objective PID tuning rules. A simple to implement MOEA, termed s-MOEA, is devised and compared with MOEAs developed elsewhere. Extensive study and analysis are performed on metrics for evaluating MOEA performance, so as to help with this comparison and development. As a result, a novel visualisation technique, termed “Distance and Distribution” (DD)” chart, is developed to overcome some of the limitations of existing metrics and visualisation techniques. The DD chart allows a user to view the comparison of multiple sets of high order non-dominated solutions in a two-dimensional space. The capability of DD chart is shown in the comparison process and it is shown to be a useful tool for gathering more in-depth information of an MOEA which is not possible in existing empirical studies.
Truly multi-objective global PID tuning rules are then evolved as a result of interfacing the s-MOEA with closed-loop simulations under practical constraints. It takes into account multiple, and often conflicting, objectives such as steady-state accuracy and transient responsiveness against stability and overshoots, as well as tracking performance against load disturbance rejection. These evolved rules are compared against other tuning rules both offline on a set of well-recognised PID benchmark test systems and online on three laboratory systems of different dynamics and transport delays. The results show that the rules significantly outperform all existing tuning rules, with multi-criterion optimality. This is made possible as the evolved rules can cover a delay to time constant ratio from zero to infinity based on first-order plus delay plant models. For second-order plus delay plant models, they can also cover all possible dynamics found in practice.
Actions (login required)