Past Staff and Students
Dr David Dawei Wang
Dr. David Dawei Wang
Institute of Chemical and Engineering Sciences (ICES)
1 Pesek Road, Jurong Island
Tel: +65 67963959
Fax: +65 63166185
- Bachelor (Science), Dept. of Mathematics and Computers, Shenyang Normal University, China
- Master (Engineering), Dept. of Automatic Control, Northeastern University, China
- PhD (Chemical Engineering), University of Sydney, Australia
My current research work, which extends upon my work in PSE@USYD, is a framework for data-based batch process monitoring and recovery
strategy, which will provide on-line information to plant operators for supervision and decision support. This decision includes
control action early in the batch process in order to maintain the final product quality through operation modifications;
and the quick disposing the batch if it will not be recovered - in such a way that time and energy are saved; as well as
diagnosis of assignable causes that may be eliminated from the future batch.
|PHD work in PSE@USYD|
Title: Robust system identification and robust model predictive control with applications to chemical engineering processes.
Industrial processes have unsatisfactory natural behaviour, which may exhibit inadequate performance or even unstable states.
In such a case a controller can be implemented in order to provide the desired stability and performance, which includes set-point
tracking, disturbance attenuation, improved efficiency and accuracy and decreased energy consumption of the process.
A controller can be designed by many methods. In the early days of control theory, control design was carried out by tuning
parameters of a PID controller. Even through this control design method is still very popular in process industry, the drawback
of this method is that it can not be carried out easily in case the process is complex or multivariable, or if high performance
requirements are imposed on the controlled process.
Due to increased complexity, quality, and environmental requirements in process operations, advanced process control strategies
often have to be integrated to fulfill the control needs. The model predictive control (MPC) provides a tremendous opportunity
to improve process efficiency and optimality. It is a control technique where digital computation and modelling play a major
role. It provides the only methodology to handle constraints in a systematic way during the design and implementation of
the controller. Moreover, in its general form MPC is not restricted in terms of model, objective function and/or constraint
Even though the increasing use in industrial processes, MPC requires high maintenance and experts are needed for on-line
tunings to guarantee the performance. One of the reasons is due to its fatal disadvantage, which was recognized later, that
the control design method makes use of the so-called certainty-equivalence principle. This means that a MPC controller is
designed on the basis of a model of the system, as if the model is exactly equal to the system. The control design does not
explicitly take into account the fact that it is practically impossible to obtain a model which is identical to the process
it describes. The theoretically successful features of MPC are based on the assumption that the model employed is a perfect
description of the process. It appears that a controller, designed to have a certain performance for a model, need not show
the same performance for the system. This discrepancy is due to the inevitable modeling error. In general a system is complex,
with high-order, even nonlinear and time-varying, dynamics. A model is generally a more or less crude approximation of the
system, with simplified dynamics, often linear and time-invariant, in order to be tractable for the control design procedure.
Consequently, there is always a model error, or model-plant mismatch, which is defined to be the difference between the system
and the model, even though ones try their best in system identification phase. Due to this model error, a controller designed
for the model need not perform identically for the system. This is the limitation one meet when using a nominal model to
represent a physical system for MPC design. In the face of significant model-plant mismatch, needless to say, such model
predictive control system that provides optimal performance for a nominal model may perform very poorly in practice.
There are two contributions in my PHD work, one of them is to robust system identification, and the other is to robust model
For robust system identification, an alternative solution for designing a robust system identification technique for most
distribution situation is proposed. The theory and implementation results are given about how to design the nonlinear transform
of prediction error which will get a robust estimation not only within given distribution class but also for any error
distribution class, in addition, this robust identification design explicitly include efficiency and robustness degree.
One of the advantages of this approach is that we need not assume prediction error belong to a distribution or a distribution
class in fact it is frequently not so, instead, we estimate the distribution density using a well-performed method
(such as Wavelets in our case) and then design a robust estimator. It is totally data driven. The applications of the proposed
robust approach to on-line and real-time identification are illustrated in chemical engineering examples.
For robust model predictive control, an innovate robust MPC controller design for dealing with model-plant mismatch is
proposed, which is characterized by using a generalized objective function. This technique is ignited by the above results
of robust system identification. It is noted that if the model uncertainty is represent by an error term in the nominal model,
a strong connection can be establish between MPC and parameter estimation. Hence, MPC is the same as parameter identification
problem, they can be regarded as dual problem, and the results of robust identification can be applied to MPC. The stability
of the proposed controller can be proved by using a generalized positive definite objective function instead of the quadratic
one in Rawlings and Muske's approach, which takes the latter as a special case. The robustness of the controller can be
achieved by appropriate choosing the objective function as the counterpart design in robust parameter estimation. The control
move computation can be approximated by a finite horizon problem. It is illuminating in our opinion to compare robust MPC with
robust system identification, and it is also another vehicle to understand robust MPC as well as MPC.
The work on these two areas (RSYSID and RMPC) has been recognised by the process control research community. Two Australia
Research Council (ARC) grants were received during the PHD study(1999,2000) and the work is also published in the open
D. Wang and J. A. Romagnoli, "Generalized T Distribution with its Applications to Process Data Reconciliation and Process
Monitoring", Transactions of the Institute of Measurement and Control, 27, 5, 1-24, (2005)
D. Wang and J. A. Romagnoli, "Robust Multi-Scale Principal Components Analysis with Applications to Process Monitoring",
J. Process Control, 15, 8, 869-882, (2005)
D. Wang and J. A. Romagnoli, "A Framework for Robust Data Reconciliation Based on a Generalized Objective Function",
Ind. Eng. Chem. Res.; 42(13); 3075-3084, (2003)
D. Wang and J. A. Romagnoli, "Robust model predictive control design using a generalized objective function", Computers &
Chemical Engineering; 27(7); 965-982, (2003)
D. Wang and J. A. Romagnoli, "Wavelet-Based Adaptive Robust M-Estimator for Non-linear System Identification", AIChE
Journal, Vol. 46, Issue 8, p1607-1615. (2000)
This site is best viewed with version 5.0 browsers or above (IE & Netscape).www.pseonline.net
Introduction to Process Control, Second Edition