You are on page 1of 13

14th IFAC Symposium on System Identification, Newcastle, Australia, 2006

SYSTEM IDENTIFICATION FOR PROCESS CONTROL:


RECENT EXPERIENCE AND OUTLOOK

Yucai Zhu

Faculty of Electrical Engineering


Eindhoven University of Technology
P.O. Box 513, 5600 MB Eindhoven, The Netherlands
y.zhu@tue.nl

Also at: Tai-Ji Control


Grensheuvel 10, 5685 AG Best, The Netherlands
y.zhu@taijicontrol.com

Abstract: This work reports the development of an identification technology and its
application to advanced process control (APC) in the refining/petrochemical industry. We
will introduce model predictive control (MPC) technology which, in the last two decades,
has become the standard tool in industrial APC. Model identification plays a crucial role
in MPC technology and it is also the most time consuming and difficult task in MPC
projects and maintenance. Key issues of identification for MPC will be discussed. The so
called ASYM method of identification is outlined that provides systematic solutions to
problems from plant test to model validation. Based on the method, both off-line and on-
line identification packages have been developed. A large-scale industrial application will
be shown. Considerable benefits are obtained using the new identification technology: 1)
reduction of identification test time and model building time by over 70%; 2) higher
model quality for control; and 3) easier in use. Then, several MPC relevant identification
problems will be introduced. Based on his industrial experience, the author will provide
an optimistic outlook of future APC/MPC and he will point out that identification
technology can play a key role in next generation control systems. Copyright © 2006
IFAC

Keywords: System identification, advanced process control, model predictive control

1. INTRODUCTION 1993, Zhu, 1998). In MPC maintenance, the main


task is model re-identification. Recently, MPC
In the last two decades, model predictive control identification has attracted attentions of both
(MPC) technology has been widely applied in the academic institutes and industry (Zhu, 1998 and Qin
refinery and petrochemical industry and is beginning and Badgwell, 2003).
to spread to other process industries (Cutler and
Hawkins, 1988 and Qin and Badgwell, 2003). Although multivariable control concept is well
Dynamic models play a central role in MPC accepted by the MPC industry, the guideline for
technology. Generally, identified black-box models identification practice, especially identification test,
are used for MPC controllers. The most difficult and until a few years ago, was still single-variable based
time consuming work during an industrial MPC thinking. The identification tests are called step tests,
project is modelling and identification (Richalet, which reflects the fact that each manipulating

20
variable (MV) is stepped separately and some clear petrochemical plants worldwide have implemented
step responses are expected for modelling each MPC controllers for their major processing units.
transfer function. The test time is very long, which Based on the data in Qin and Badgwell (2003), the
occupies much manpower and makes production number of MPC applications in the middle of 2005 is
planning difficult. The tests are done manually, between 7,000 and 10,000, mainly in the
which dictates extremely high commitment of the refining/petrochemical industry. The popularity of
engineers and operators. Worst of all, the models the MPC technology stems from the huge benefit it
derived from this approach are not accurate for brings. An MPC controller of a major refining/
complex and highly disturbed processes. This in turn petrochemical unit, such as a crude unit, a catalytic
causes control performance problems during cracking unit, or an ethylene plant, can create a
controller commissioning. benefit of one to three million US dollars per year.

Recently, under the pressure of refining/ Excellent discussions on algorithms of industrial


petrochemical companies and due to competition, MPC products are given in Qin and Badgwell (2003).
some MPC vendors started working on improving Here the author will introduce a project approach of
their identification techniques. The improvements are MPC and emphasise the need of process
aimed at: 1) to improve identification test method identification.
and 2) to use parametric models, such as subspace
model. For example, it was reported recently that A Project Approach to Industrial MPC
traditional step tests can be automated. Very
recently, automated multivariable test was also In the last two decades, a project approach has been
reported (Mantelli et. al., 2005). It is also reported in developed by MPC vendors and operation
that subspace model is used. companies; see, e.g., Zhu (2001). Although this
procedure is developed within the refining and
In this paper we will introduce the so called petrochemical industry, it can also be used as a
asymptotic method (ASYM) of identification and guideline for other industries.
show its use in MPC applications. The ASYM
method was originally developed for linear robust 1) Benefit Analysis and Functional Design
control; see Zhu et. al. (1991). The method was first
applied in a refinery MPC project in 1993. Since This step involves studying the process unit
then, its popularity in the refining/petrochemical operation practice, reviewing unit operation
industry increased gradually and some major oil economics, identifying process constraints and
companies have used and are starting to use the performing benefit analysis. Much work is spent on-
technology in their MPC projects and maintenances. site with the operation and control personnel for
gathering relevant information. Process drawings are
In Section 2 the MPC technology is introduced and reviewed and process data are studied. Based on the
the requirements for MPC identification are outlined. collected information, a preliminary MPC controller
In Section 3 the identification approach is is designed that reflects the desired operation
introduced. In Section 4, an industrial case study, a improvements. The economical benefits of the MPC
crude unit application, is presented. In Section 5, control are estimated. The results of the study are
several MPC relevant identification problems will be contained in a so called Functional Design Report.
presented. In Section 6, conclusions are draw and a
personal outlook of future MPC technology and 2) Pre-test
control systems are given.
This step is used to obtain process information
needed to carry out process identification and to
2. INDUSTRIAL MPC AND make the unit ready for process identification test.
MPC IDENTIFICATION The work includes:
2.1 A Brief Introduction to Industrial MPC - To identify and fix instrument and actuator
problems such as faulty thermal couples and
Industrial MPC technology is a class of computer sticking valves.
programs that provide control actions of - To inspect and tune MV PID loops.
multivariable processes. The key feature of the MPC - To obtain rough estimate of process time to
technology is the use of an explicit model of the steady state and some process gains.
process for predicting the future response of the - To determine proper step sizes for identification
given process and for calculating the control actions. test.
An MPC controller is used to maintain stable unit
operating, to reduce the effect of disturbances, to To achieve these, step signals are applied to all MVs
reduce unit transition time and to bring the unit to a (one at a time) and CV responses are recorded and
more profitable working range. In the last two studied. The information obtained in the pre-test is
decades, the MPC has become “the killing used to give recommendations for necessary
application” of APC. Most refineries and instrument work and to design an identification test.

21
nonlinearity. The efficiency of MPC commissioning
3) Identification Test and Model Identification depends heavily on process model quality. When
identified model is accurate, the MPC performance
Traditional identification test are carried out will be close to that of the simulation and the
manually using step-like test signals. One MV is commissioning can be done quickly. When model is
tested for a given time in order to avoid correlation. poor and real life MPC performance is also poor, the
For a typical unit such as crude unit or FCC, the test commissioning becomes very difficult.
time is roughly the number of MVs times 24 hours.
So, to test a crude unit MPC with 30 MVs will need 6) Controller Maintenance
30 days around the clock. When a test is finished,
model identification will be performed using the An MPC cannot work forever. The control
collected data. Traditional methods will need several performance degradation is caused by slow process
weeks to perform this task. Somewhat shorter test changes such as heat exchangers fouling and
time is possible if the control engineer has catalysts deactivation, and process modification or
considerable process knowledge and a lot of project revamp. These changes will increase the error of the
experience, and can manually test more than one MV existing model, which in turn leads to poor control
at a given time. performance. Therefore, MPC maintenance is
important to prevent the loss of benefits. The main
4) Controller Tuning and Simulation task of maintenance is to re-identify the process
model and to re-commission the MPC controller. It
The MPC controller is tuned according to the desired may also include MPC design modification when
specifications. The tuning of a MPC controller operating objectives and/or process condition change.
considers first the parameters related to unit Traditionally, in MPC control maintenance, a new
operation requirements, these include MV high/low open loop step test will be performed and model will
limits, MV rate of change limits, MV ideal resting be identified, which is again very time consuming
values (for economic optimization), CV high/low and may disturb the unit operation.
limits (zones) or setpoints. The parameters related to
control performance such as loop speed and The Need of System (Process) Identification
robustness are considered that include MV weighting
factors, CV weighting factors or closed-loop settling There are two approaches of modelling: rigorous
time of the CVs. Some MPC controllers can set CV (physical) modelling and process identification.
priorities in order to solve conflict conditions. Then Rigorous modelling derives process models using
the controller nominal performance is checked in physical laws such as energy and mass balances and
simulation where identified model is used to thermodynamics. Process identification derives
represent the process. In simulation, first the stability process models using testing data. The advantage of
of the controlled system is checked. There should be rigorous models is their ability to describe the
no oscillations and some of the CVs must be process behaviour in a wide operation range. The
decoupled reasonably. The constraint handling is advantage of identified models is their (dynamic)
checked; CV conflict situations are simulated in accuracy for given working points. MPC is the
order to see how they are resolved. application area that process identification is
indispensable and routinely used. Experience has
5) Controller Commissioning shown that rigorous models are in general not
accurate enough for MPC control in the
When simulated system show good control refining/petrochemical industry. In Mantelli et. al.
performance, the developed controller can be loaded (2005), some comparisons of rigorous model and
to the control computer and turned on gradually. identified model of distillation columns are given:
First the controller is set to a simulation or off-line although gains of two models have same signs, the
mode. In this mode, the controller will read the MV sizes of gains, time constants and time delays can
and CV measurements, will simulate the CV differ by a factor of 2 or 3. Major MPC vendors such
movements and show the desired MV actions, but as AspenTech, Honeywell and Invensys all have
the MVs are not yet connected to the MPC control developed and own rigorous models of many refining
outputs. In this way the control engineer can verify if and petrochemical processes. However, these
the model predictions are accurate and control companies use process identification almost
actions are relevant. When proper control exclusively in their MPC modelling.
movements are shown, the MVs will be turned on
meaning they will be connected to controller outputs. With further development and enhancement, process
The controller performance will be tested, which identification will play an increasing role in modern
includes disturbance attenuation, constraint handling MPC and automation systems. See also the
and setpoint following. It may be necessary to adjust discussion in Section 6.2.
the tuning parameters such as MV weights and CV
weights (or CV settling times), because the on-line Inferential Models or Soft Sensors
controller will perform a little differently than the
simulated one due to disturbances, model errors and

22
Product qualities (compositions) are often controlled covering two heaters, an atmospheric tower and a
variables (CVs) of an MPC controller. In the naphtha stabilizer. The primary objectives of the
refining/petrochemical industry, their measurements MPC application are to:
are often too slow. Basically two ways are used to
measure product qualities, online analysers and 1) maximize crude charge (throughput),
laboratory analysis. Many online analyser sampling
2) meet product specifications,
times are between 30 minutes to 60 minutes; typical
sampling times of laboratory analysis are between 6 3) maximize crude preheat train efficiency and
to 24 hours. A typical sampling time of MPC balance heater passes,
controllers is 1 minute. Inferential models or soft
sensors are often used in MPC controllers to provide 4) minimize transient periods during crude switches
real-time prediction of the product compositions that and product grade changes.
are measured at a much lower frequency than the
control frequency. Currently, often static linear or Back in 1998, the MPC controller was commissioned
static nonlinear models are used as inferential using traditional manual step test and FIR model
models. They are obtained either by identification or identification. The following procedure has been
from rigorous models. Inferential models are often taken in the identification of the crude unit:
not accurate and it is desirable to increase their
accuracy. - Manually inspecting all regulatory controllers
for proper controller tuning and valve operation.
Thus requiring the retuning of controllers and/or
2.2 An Application: MPC Control of a Crude Unit fixing/cleaning/replacing valves and or
instruments.
This application was reported in Celaya et. al. - Notify the planning department of upcoming
(2004). The facility is one of the most modern step-tests and request they plan for a more
refineries in the United States and, with crude oil steady diet of crude slate to avoid unnecessary
processing capacity of 495,000 barrels per day, is upsets.
one of the largest refineries in the world. The crude - Notify operations personnel of upcoming step
unit (CDU), Figure 1, is capable of processing tests so they have appropriate and adequate staff
190,000 barrel per day (BPD) of raw crude. Crude is on hand to help with any process problems that
preheated with product heat exchangers before being develop.
heated in two parallel four pass heaters. Heated feed - Perform the step test one variable at a time. This
enters the CDU in the flashzone, where it is distilled is very time consuming. Each variable must be
into seven products plus resid. stepped and held until steady state is achieved,
sometimes up to two hours per variable.
The major disturbance source is the crude switch that - After step testing is complete, data analysis and
takes place every few days. Because the composition model identification begins.
of crude oil differs, crude switch will upset the crude
unit operation to some extend. Other disturbance The test lasted about 30 days around the clock and,
sources include gradual composition change in the after the long step test, data analysis and model
same crude, disturbance from other units due to heat identification took another four weeks of working
integration, changes in temperature and pressure of days. Given the size of the controller, this is typical
heating sources, weather and others. model identification procedure. Note that step test
usually do not disturb normal unit operation.
However, it is a very difficult task to monitor a step
test in order to prevent the test causing too much
disturbance to the unit. It can be seen that a
traditional MPC identification is quit an undertaking.
It is not only a huge task for the MPC project team,
but also involves planning and operation
departments.

Since its commissioning, the MPC controller has


helped the refinery to operate the crude unit more
efficiently and economically and it has brought
considerable benefit to the refinery.

As with any process unit, as run time increases,


Figure 1. The crude unit process problems develop, valves fail, heat
exchangers foul, huge upsets can cause possible tray
damage, process instruments fail, process conditions
The CDU has had MPC control since 1998. The change, crude types change ... All of these caused
CDU MPC application has 33 MVs and 94 CVs the performance degradation of MPC models and the

23
MPC controller. The CDU is running late in its an identification test, the level of disturbances is in
current run cycle and will need to be turned around average above 10% of that of CV variation (in
sometime over the next year(s). The MPC controller power), but it can be as high as 50%. Too large test
was not performing as good as previous years. It was signal amplitudes are not permitted because they will
obvious that maintenance is needed for the MPC cause off-specification of product.
controller which is mainly model re-identification.
Nonlinearity. Although in general linear models are
However, as the crude unit is at the end of its run, the relevant for this class of processes for a given range
MPC re-modelling projects cannot be justified using of operation, some nonlinear behaviour may still
the traditional approach because it is very time show up. Examples are CVs that are very pure
consuming, intrusive to process operations and product qualities and valve positions close to their
costly. This leads the refinery to try an alternative limits. Other nonlinearities are caused by the fact that
identification method that can save testing time, is the same unit is operating at several working points
less intrusive and easy to use. The identification and where process behaviours are different. In a
results of the new approach are shown in Section 4. crude unit MPC, normally a single linear MIMO
model is used for different crude types. Strictly
speaking, however, a crude unit processing different
2.3 Key Issues of MPC Identification crudes will behave differently due to composition
variations. In a polymer plant, different products or
After introducing the crude unit example, we will grades are produced for market requirement. The
discuss the requirements of identification for MPC in process behaves so differently for different grades
the refining/petrochemical industry. However, most that an MPC using a single linear model will
of the discussions also hold for other process normally not work for a polymer plant. A lubricate
industries. This class of processes can be oil unit also produces different grades, which causes
characterised as follows: nonlinearity.

Continuous process. This class of processes operate Based on these observations we will outline the key
around their working points (or stationary points) issues in identification for MPC and give comments
sufficiently long, much longer than their response on the existing methods.
times or time to steady states. Each process can have
multiple working points. This is a "good" character Identification tests. Identification test plays a crucial
of the process class. The fact that these processes are rule for a successful identification. Traditional test
mostly operating in continuous modes makes it method is to use a series of single variable step tests.
feasible to use linear time-invariant dynamic models The tests are carried out manually. The advantage of
for control. this test method is that control engineer can watch
many step responses during the tests and can learn
Large scale and complex. A small MPC controller about the process behaviour in an intuitive manner.
will have 3 to 5 MVs (manipulated variables or The biggest problem of single variable step test is its
inputs) and 5 to 10 CVs (controlled variables or high cost in time and in manpower. The second
outputs); a large size MPC, like the crude unit, will problem is that the data from single variable test may
have more than 40 MVs and 100 CVs. Some CVs, not contain good information about the multivariable
such as product qualities, are very slow (with character of the process and that slow step signals
dominant time constant range from 30 minutes to may not excite enough dynamic information of the
several hours), and other CVs are very fast, such as process. Using automatic multivariable test approach
valve positions (with time constant less than a can solve these problems. If possible, identification
minute). Inverse response (non-minimum phase tests should be done in closed-loop operation. One
behaviour), oscillating behaviour and time delays advantage of closed-loop test is obvious: the
exists. For a large controller such as the crude unit, controller helps to keep CVs within their limits.
over 50% of the process transfer functions are What is less obvious is that a model identified from
practically zero and they need to be located. closed-loop operation gives better control
performance (Hjalmarsson et al., 1996).
Dominant slow dynamics. The time to steady state
of a typical product quality model ranges from one Model structure and estimation method. Several
hour to several hours. This dictates long time for industrial identification software packages are based
identification test. on finite impulse response (FIR) model. FIR models
are called nonparametric models. This model class
High level disturbances. Typical source of has inherent problems for use in identifying
unmeasured disturbances are feed composition processes with slow dynamics. Model variance error
variations and weather changes. These are slow and is high due to its high order; yet bias error is often
irregular variations. Strong and sudden step/pulse not negligible due to truncation. Therefore,
like disturbances also exist such as crude switch in a parametric or compact models are preferred. This is
crude unit, drum switch in a delayed coker and true also for MPC controllers that use non-parametric
product grade change in a lubricate oil unit. During models. Of course, there are challenging problems in

24
parameter estimation and order selection when using 3. ASYMPTOTIC METHOD OF
parametric models. As mentioned before, high level IDENTIFICATION
unmeasured disturbances are present. This means
that when model accuracy is concerned, an The asymptotic method (ASYM) of identification
estimation criterion that includes disturbance model was originally developed for the purpose linear
will be better than a criterion without the disturbance robust control; see Zhu et. al. (1991). The name
model. Note that this is true even when the controller asymptotic method reflects the fact that the approach
does not use disturbance model. Prediction error is based on the asymptotic theory of identification
method belongs to the first class; while output error developed by Ljung and Yuan; see Ljung (1985), and
criterion belongs to the second class. Moreover, Ljung and Yuan, (1985) for single variable version
prediction error method will work properly for and Zhu (1989) for multivariable extension. Given a
closed-loop data (model estimates are consistent multivariable process with m manipulated variables
meaning that the effect of the disturbance will (MVs) and p controlled variables (CVs). Denote the
decrease when test time increases); while output data sequence that is collected from an identification
error method and most of the subspace methods will test as
deliver biased model when using closed-loop data. Z N := {u (1), y (1), u (2), y (2),......,u ( N ), y ( N )} (1)
Several industrial identification software estimate where u(t) is m-dimensional MV (input) vector, y(t)
FIR model using output criterion. This partly is p-dimensional CV (output) vector and N is the
explains why traditional industrial identification tests number of samples.
are mostly carried out in open-loop operation.
We assume that the data is generated by a linear
Model validation. The goal of model validation is discrete-time process as
to test whether the model is good enough for its y (t ) = G o ( z −1 )u (t ) + H o ( z −1 )e(t ) (2)
purpose (MPC control) and to provide advice for -1
where z is the unit time delay operator. Here
possible re-identification if the model is not valid for
its intended use. Commonly used methods of model G o ( z −1 ) is the process model and H o ( z −1 )e(t )
validation are simulation using estimation data or represents the unmeasured disturbances acting at the
fresh data, whiteness test of model residuals and outputs, and e(t) is a p-dimensional white noise
testing the independence between the residuals and vector.
past MVs. These methods only tell how well the
model agrees with the test data. They can neither The model to be identified is in the same structure as
quantify the model quality with respect to the in (2):
purpose of modelling, in our case, closed-loop y(t ) = G ( z −1 )u(t ) + H ( z −1 )e(t ) (3)
control, nor can they give good advice for re- The process model G(z-1) and noise filter H(z-1) will
identification. Trial-and-error approach in this step be parametrized in matrix fraction description (MFD)
makes industrial identification a very expensive with diagonal denominator matrices; see Zhu (2001)
practice. In linear model identification for control, a for details. The model will be calculated by
logical approach for model validation is to quantify minimizing the prediction error cost function; see
model errors in the frequency domain and to find Ljung (1985).
relationship between this kind of error description
and model based control performance. Model error The frequency response of the process and of the
estimation has attracted much interest in the last model are denoted as
decade; this was motivated mainly by linear robust T o (e iω ): = col[G o (e iω ), H o (e iω )]
control theory.
T n (e iω ): = col[G n (e iω ), H n (e iω )]
User-Friendliness. Most MPC users are not where n is the degree of the polynomials of the
identification experts. A desired identification model, col(.) denotes the column operator.
technology should be able to provide simple
guidelines for good identification tests and when test The Asymptotic Properties of High Order Models
data are collected, to carry out various calculations Assume that
automatically without the necessity of user - n → ∞ and n 2 / N → 0 as N → ∞
intervention. Most existing identification software - Test signals have continuous none zero spectra
follows toolbox approach so that the user can try until the Nyquist frequency.
various methods/techniques on each step. This is
certainly very convenient for academic researchers Then (Ljung, 1985 and Zhu, 1989)
and students, but it is rather user-unfriendly for
industrial users. The user-friendliness of an - The estimate is consistent
identification method for industrial users depends T n (e iω ) → T o (e iω ) as N → ∞ (4)
heavily on its ability to solve various fundamental
problems in process identification as outlined before. - The errors of T n (e iω ) follows a Gaussian
distribution, with covariance given as
n
cov[Tˆ n (e iw ) ≈ Φ −T (ω ) ⊗ Φ v (ω ) (5)
N

25
where Φ (ω ) is the spectrum matrix of inputs and The high order model in (8) is unbiased. The
prediction error residual col[u T (t ), ξ T (t )] , Φ v (ω ) is variance of this model is high due to its high order.
One can reduce the variance by performing a model
spectrum matrix of unmeasured disturbances, ⊗ reduction. If we view the frequency response of the
denotes the Kronecker product and -T denotes high order estimates as the noisy observations of the
inverse and then transpose. true transfer function, we can then apply the
maximum likelihood principle. Using the asymptotic
The Procedure of the ASYM Method result of (4) and (5), one can show that the
asymptotic negative log-likelihood function for the
1) Optimal test signal design and test method reduced process model is given by (Wahlberg, 1989)
For an open loop test, the optimal test signal is
designed to minimise the sum of the squares of the p m
ω2
1
simulation error. The spectra of the test signals are ¦¦ ³ | {| Gˆ n
ij (ω ) − Gˆ ij (ω ) |2
[Φ −1 (ω )] jj Φ vi (ω )
}| dω
based on the following formula which can be derived i =1 j =1 ω
1
from (5) (Ljung and Yuan, 1985): (9)
p
 −1
The reduced model G ( z ) is thus calculated by
u (ω ) = μ j Φ u (ω )
Φ opt
j
sim
j
¦Φ
i =1
vi (ω ) (6)
minimizing (2.9) for a fixed order. The same can be
where Φ usim (ω ) is the spectrum of the j-th input done for the disturbance model
H n ( z −1 ) = 1 / A n ( z −1 )
j

movement in a controlled system, μj is a constant


which is adjusted to meet the amplitude constraint of The reduced model is converted to a diagonal Box-
the signal. Jenkins form.

For a closed-loop test, in a SISO case, an 3) Order Selection


approximate optimal spectrum formula of the test The best order of the reduced model is determined
signal at the setpoint of the closed-loop system is using a frequency domain criterion ASYC which is
given as (Zhu, 2001) related naturally to the noise-to-signal ratios and to
the test time. The basic idea of this criterion is to
r (ω ) ≈ μ Φ r (ω )Φ v (ω )
Φ opt (7)
equalise the bias error and variance error of each
where Φr(ω) is the power spectrum of the reference transfer function in the frequency range that is
signal r, Φv(ω) is the power spectrum of the important for control. Let [ω1 , ω 2 ] defines the
disturbance, and µ is a constant adjusted so that the frequency band that is important for the MPC
signal power (or amplitude) is constrained. application, the asymptotic criterion (ASYC) is given
by:
In practice, (6) and (7) are used in combination with
upper error bound (11) to give general guidelines for p m
ω2
n
the test design of various processes. The spectra of ¦¦ ³ | [| Gˆ n
ij (ω ) − Gˆ ij (ω ) |2 −
N
[Φ −1 (ω )] jj Φ v (ω )] | d ω
i
the test signals can be realised by GBN (generalised i =1 j =1 ω
1
binary noise) signals (Tulleken, 1990). The character (10)
of a GBN signal can be determined by its average
switch time and its amplitude. The amplitudes of 4) Model Validation using Error Bound Matrix
GBN signals are determined by a priori knowledge According to the result (4) and (5), a 3ı bound can
of the process from the pre-test. Studies show that be derived for the high order model as follows:
the optimal average switch time of signals can be
related to the longest process response time. One can n
also use other type of signals. Gijo (e iω ) − Gˆ ijn (e iω ) ≤ 3 [Φ −1 (ω )] jj Φ vi (ω ) w.p. 99.9%
N
(11)
Remark. In control applications, it is more realistic
We will also use this bound for the reduced model
to constrain or minimise the output (CV) variance.
because the model reduction will in general improve
model quality.
2) Parameter Estimation
If the controller is linear and time-invariant, the
A) Estimate a high order ARX model
upper bound matrix (11) can readily be used in
A n ( z −1 ) y (t ) = B n ( z −1 )u(t ) + e(t ) (8) robust stability and performance analysis, which is
where A ( z ) is a diagonal polynomial matrix and
n −1
well known in robust control literature. For model
B n ( z −1 ) is full polynomial matrix, both with degree
predictive control where controller is nonlinear and
time-variant, no theory is available yet for the
n polynomials. Denote G n ( z −1 ) as the high order analysis of robust stability and performance (even the
ARX model of the process, and H n ( z −1 ) as the high proof of nominal stability is a challenging problem).
order model of the disturbance. In the following we will give an engineering solution
to the model validation problem using the error
B) Frequency weighted model reduction bounds.

26
Grading the models. This is done by comparing the and carries test design, plant test and model
relative size of the error bound with the model over identification all automatically.
the low and middle frequencies where model quality
is important for MPC control. Then, identified Basically, Tai-Ji Online consists of two modules, the
transfer functions are graded in A (very good), B test device and the identification device; see Figure
(good), C (marginal), and D (poor, or, no model 2. The test device performs test design and plant test;
exists). Based on extensive simulations and project the identification device performs model
experience, usually A grade and B grade models can identification using available testing data. The plant
be used in the controller. C grade and D grade test program is executed, normally, at MPC control
models are treated as follows: sampling interval which is typically 1 minute. The
model identification program is executed at one to
1) Zero them when there are no models expected few hours; it can also been started manually.
between the MV/CV pairs.
2) If a transfer function is expected and needed in
the control, modify the test in order to improve 4. CLOSED-LOOP IDENTIFICATION
the accuracy of these models. OF THE CRUDE UNIT

Using upper bound formula (11) we can easily give About 70 refineries and petrochemical plants have
some guidelines for improving the test design: licensed Tai-Ji ID or Tai-Ji Online for their MPC
identification. The technology has been applied to
• doubling the amplitudes of test signals will half various types of process units: Crude Units, Catalytic
the error over all frequencies; Cracker Units, Aromatics Units, Ethylene Plants and
more. Here the identification of the crude unit
• doubling the average switch time of GBN introduced in Section 2 will be presented.
signals will half the model error at low
frequencies and double the error at high The test was performed with the current MPC
frequencies; halving the average switch time of controller turned on, and, therefore, it was a closed-
GBN signals will do the opposite. loop test. The test program moved all MVs
• Doubling the test time will reduce the model simultaneously with smaller amplitudes and
gradually increased them. After the first 24 hours of
error by a factor of 2 testing, model identification was performed to check
In processes in the refining/petrochemical industry, the model qualities and test signal step sizes were
processes are slow and we can carry out adjusted accordingly. Test signal switch time was
identification in the middle of a test and adjust the also adjusted during the test (modifying signal
ongoing test according to the model validation spectrum). Model identification and test adjustment
results. were done every day. The plant test was stopped after
only 4.5 days when most of the expected models
become “A” and “B” grades.
Process unit
(DCS) The operator had only seen traditional step tests. He
was initially concerned about the multivariable
Test signals MV/DV/CV data
automated test approach. However, after 24 hours of
Tai-Ji Online stable test, the operator was convinced that the new
Testing device test method would not disturb normal production.

step sizes MV/DV/CV data During the test using Tai-Ji Online, the existing MPC
controller has done a good job in stabilizing the unit
Identification
device operation. Operators were also encouraged to move
any MV at any time to keep the unit within the
normal operating range. However, the operator has
Model rarely intervened during the test. Control engineers
file
monitor and make testing adjustments without
intrusion to the operator daily work routine.
Figure 2. Tai-Ji Online
After the test, more detailed data analysis and model
identification and validation took place. It took
Based on the ASYM, two software packages have several days. Remember that data analysis and model
been developed. The first one, Tai-Ji ID, is an off- identification will cost four weeks for a large
line identification package. It performs test signal controller like this in the traditional approach.
design and model identification. A separate plant test
program is needed perform identification test and The identified models have been used in the MPC
data collection. The second one, Tai-Ji Online, is an controller and control performance has increased
online identification package that is connected to considerably. This was achieved while reducing over
process unit via DCS (distributed control systems) 80% time in process identification. After the project,

27
the refinery decided to apply Tai-Ji Online to a
Catalyst Cracker (Celara, et.al., 2005) and to a
delayed coker.

In Figures 3 to 6, part of the MV and CV signals and


model results are shown with some technical
comments. In MPC identification, often an
expectation matrix is used to specify where a model
is expected and where is empty according to process
knowledge. Expectation matrix may be updated
using identification result.

Figure 5. Step responses for selected models. White


background models are expected and have good
qualities; pink background models are expected,
but model qualities are not good enough; grey
background models are not expected.

Figure 3. Selected MV data during the step test. It


can be seen that in a closed-loop test the MV
movements consists of two parts, the test signal
part and control action part. MVs.

Figure 6. Frequency responses (blue solid lines) and


upper bounds (green dashed lines) for selected
models.

5. SOME MPC RELEVANT IDENTIFICATION


PROBLEMS

New development and enhancement in the following


identification problems can make useful contribution
in the MPC technology.

5.1 Multi-Linear Model Identification


Figure 4. Model fit for selected CVs. Green solid
lines are measured CVs and blue dashed lines are Most continuous industrial processes operate at one
model simulations. The large CV dips at the or multiple working points. Examples of multi mode
second half of the test were caused by crude units are: crude units that process different crude
switches and not by the test signals. types, lubricate oil units and polymer plants that
produce different product grades, electrical power
plants that operate on different loads. Therefore,
MPC using multi-linear models is a natural choice
for controlling this type of process units. This can be
viewed as an extension of the linear MPC technology
for controlling severe nonlinear process units and
also as an enhancement of the linear MPC

28
performance for controlling mild nonlinear process In the process control community, maybe also in the
units. identification community, a misconception exists that
says: if the output sample time is longer than the
Some issues arise in various situations. system settling time, then it is impossible to identify
the fast dynamic system using fast input data and
- Known working points without transition slow output data. It is based on a wrong intuition or a
control. In this situation, good control is only misunderstanding of Shannon’s Sampling Theorem.
needed at working points and transition control In fact, it can be shown that dynamic models can be
from one working point to another is not identified using fast input data and slow output data;
important. This is the easiest case and linear see, e.g., Li et al (2001) and Zhu and Telkamp
model identification can be applied for each (2005). This topic is not studied extensively and
working points. A model switching mechanism more research is needed.
can be built in the MPC controller in order to use
the multiple models.
- Known working points with transition control. 5.3 Model Reduction of Rigorous Models for MPC
Now, it is important to have both good control at using System Identification
working points and good transition control.
Polymer plant control and power plant control are Many first principle or, rigorous models exist in the
examples. In this situation, linear models need to process industries. Theses models are usually
be identified for working point control, model described by nonlinear partial differential equations
interpolation is necessary for use in transition (PDE) and algebraic equations of very high order,
control. often higher than 10000. They are useful for process
- Unknown working points. Crude unit control is simulation, process design, and sometimes
an example. While each crude type determines a optimization. However, they are often too complex
working point, the compositions of each type are for real-time MPC control. Model reduction is
varying and cannot be predefined. Also the needed to develop reduced model suitable for
number of types is very large and it is not control.
economical to test and identify models for all
types. A method to solve the problem is to Model reduction of nonlinear systems has received
identify linear models for a number of working much attention in recent yeas; see, e.g., Antoulas
points and then to find and use the nearest linear (2005). A popular approach to nonlinear system
model in MPC for a given working point. model reduction is the method of proper orthogonal
decomposition (POD). The technique uses the
Questions in multi-linear model identification are: orthogonal decomposition of a collection of
measurements of physical quantities in position and
- What test signals are suitable for identification in time for model reduction; see, e.g., Astrid (2004).
test? Should test signals dependent on the The method is based on physical insight of the
working points? original model and on simulated data. The reduced
- How to interpolate linear models for transition model preserves the physical structure of the original
periods? Should transition data be used in model. One disadvantage of the POD method is that
identification and should test signals be used in the reduced model may still be too complex for real-
transition periods? Banerjee and Arkun (1998) time MPC control.
have developed a method to address the problem.
- Can test signals (excitation) be avoided for Some researchers have studied the use of system
process units with frequent transitions? identification for model reduction; see Vargas, A.
- In unknown working point case, what is the and F. Allgöwer (2004) and Wolfrum et. al. (2005).
number of linear models to be identified? In MPC This is a viable approach because:
control, how to pick up the best linear model for
the current (unknown) working point. How to - Identified black-box models are very simple and
find out when is the time that a new working are suitable for real-time MPC control.
point model needs to be identified? - Many numerical techniques have been developed
in system identification for control that can be
borrowed.
5.2 Identification of Dynamic Inferential Models - Test signal design methods can be used in
simulating the rigorous model.
Inferential models or soft sensors play an important - Many black-box model structure can be used for
role in MPC technology. The analyzer or lab model approximation.
sampling time is much longer than the sampling time
of the controller and lab sampling time may also be In general, MIMO nonlinear identification methods
irregular. In process control, most inferential models need to be used or developed, which is not at all an
are static models or static gain plus delay models. easy task. The multi-linear model identification may
One can show that ignoring dynamics in inferential be a candidate method. In identification for model
modelling is one of the major causes of inaccuracy.

29
reduction, only bias error needs to be considered when the data are noisy. Note that the open loop
because disturbance is absent. process is not an oscillating process. Study is needed
to answer following questions:
The development in this direction can also add useful
model structures and numerical methods in nonlinear - What identification methods are more sensitive
system identification. to system oscillation, what are less sensitive?
- Can some data pre-treatment methods help in
improve model quality?
5.4 Model Quality Assessment for MPC Control

MPC maintenance is important for achieving a 6. CONCLUSION AND OUTLOOK


sustained benefit. An automated maintenance scheme
will greatly increase its efficiency. An automatic 6.1 Conclusion
maintenance system will perform three major steps:
MPC technology has become the choice of industrial
1) Controller performance monitor. If, according to APC. System identification is indispensable in MPC
some criterion, the control performance is and it is the most difficult and time consuming part
degraded considerably, perform step 2. of MPC projects and maintenance. Based on the
2) Model quality assessment. If, according to some asymptotic theory, the author has developed an
criterion, model quality is too poor for MPC identification technology that is efficient and user
control, perform step 3. friendly. Many successful applications (over 150)
3) Plant test and model identification. Then, use the have been realised.
new model.
Several reasons have made this possible:
Normally, model assessment is performed using
closed-loop data under MPC control. Important - The drive in developing advanced control
questions are: technology.
- First hand knowledge of industrial control
- How to quantify or benchmark model accuracy environment and its requirements.
for MPC control? - Availability of simple and applicable theory.
- How to schedule the assessment process? Or,
what data length should be used in calculations? There is a need to further increase the efficiency and
- Are test signals (excitation) needed for model reduce the cost of the MPC technology so that it can
assessment? If yes, what kind of signal and how be used for all process industries in this global
long should the excitation be used? economy. Many interesting identification problems
need to be solved in order to develop the next
Methods for fault detection and diagnosis may be generation MPC technology and automation systems.
useful in solving the problem.

6.2 Model Intelligence; an Outlook of Future


5.5 Time Delay Estimation of MIMO Processes Process Control

Process delays exist and delay times range from Challenges in APC and Automation Systems
several samples to half of process response time or
longer. If delays are corrected by properly shifting The automation industry has been in stagnation or
the input/output data before model identification, even in decline (Pinto, 2000). Here, control
lower model orders can be used and variance error algorithms have not produced much technology push.
can be reduced. For large scale MIMO processes, In general, there are two types of control algorithms
efficient algorithms need to be developed for delay used in process control: PID and MPC. PID
estimation. It is also worthwhile to study optimal test algorithms are used in most DCS systems. The PID
design for delay estimation. control algorithms were developed in 1940s and
1950s; PID tuning is mainly done manually in an ad
hoc manner. The MPC technology was developed by
5.6 Closed-Loop Identification of Oscillating control engineers in the industry in the 1980s and it
Systems has created tremendous benefit for the refining and
petrochemical industry. However, other process
The problem arises in MPC maintenance. When the industries do not use MPC due to its high cost.
control of some CVs becomes poor, it is either too
slow or too aggressive. When the controller becomes The impact of academic control community to
too aggressive or too tight, the closed-loop system process control has been very limited, if not
will oscillates. Experience has shown that many negligible. In the last four decades, there are many
identification methods will have poor results when control algorithms developed by the control
using data from a closed-loop oscillating system and community: LQG control, self-tuning or adaptive

30
control, H-infinity or robust control, neuron-network
control and hybrid control. It is unfortunate that none Now, we can draw some pictures of future process
of the algorithms has found a position in process control.
control. One of the important reasons is the high cost
in developing and maintaining process models, Outlook of Future Process Control
because most of these control methods are model
based. A solution is needed to make the technology Next generation MPC
transfer. Although some DCS systems have PID
auto-tuning functionalities developed by the “Get the design right, the rest is automatic.” Using
academia, they are rarely used by control engineers model intelligence, process identification, control
and operators. tuning/commissioning and maintenance can all be
done automatically. The only work left is the design
Model Intelligence and redesign of the controller which cannot be
automated. The main task of an MPC design is the
Definition: Model intelligence is a class of computer selection of MVs, DVs and CVs based on technical
programs and algorithms that can, for a given class and economical requirements of the process unit. An
of processes, automatically develop, maintain and MPC with model intelligence can be called adaptive
use dynamic process models for control, prediction, MPC in a broad sense. The adaptation is not
monitoring and diagnosis. continuous and it is activated when necessary. The
adaptation involves mainly online identification and
A model intelligence system has three components: automated model replacement. An MPC monitoring
program is used to check the MPC performance.
1) Connectivity module. This part takes care of the When model performance degradation is detected,
communication (read, write, error checking) the monitoring program will activate the adaptation
between the system and lower level control mechanism.
systems (DCS, PLC). OPC (OLE for process
control) is becoming the industrial standard for Using model intelligence, the efficiency of MPC
connectivity between Windows® and DCS/PLC implementation and maintenance can be increased by
systems. a factor of 10 when comparing with traditional
2) Modelling module. Typically, this is an online project approach. The MPC is also much more user
identification software that carries out both friendly and easy to use. Due to reduced costs, most
testing and model identification automatically. process industries will be justified to use MPC, not
just the refining/petrochemical industry.
3) Model based application module. This can be a
model based control design and tuning program Nonlinear MPC will be used to control strongly
(e.g., PID tuning), a control design plus nonlinear processes. Multi-linear model
controller (H-infinity control design and control, identification can contribute to nonlinear MPC
MPC tuning and control), a prediction program technology.
(soft sensor), or a model based monitoring and
diagnosis program. Next generation DCS and PLC
The key difference between model intelligence and
other model based methods is automatic model Using model intelligence, all PID loops will have
building and maintenance. System identification self-diagnosis and auto-tuning functions. This will
plays the key role in model intelligence. Model and considerably increase the performance of these
controller adaptation is included in this definition, systems and produce huge benefit.
although it is, in general, not sample-wise adaptation.
Model intelligence is a dream of the many users of Ultimately, MPC controllers and other modern
MPC/APC and automation systems. In process controllers (LQG, H-infinity, ...) can be embedded in
control, at present, hardly any system has model DCS/PLC systems.
intelligence.
MPC and modern control for machines
Rigorous (physical) modelling can certainly be used
in the design of MPC controllers. If a rigorous model Because model intelligence can make the application
is available and it is also simple and accurate enough, of MPC and modern control highly efficient and easy
it can be used as an initial model in the MPC control. to use, they can be used in the control of machines
and equipments. Embedded MPC and modern
Although most modern control methods developed in control (LQG, H-infinity,...) will lead to higher
the last four decades are model based, how to performance and/or reduced costs of machines and
develop models is not solved. A strange situation equipments.
exists for a very long time: modern control methods
can only be used by control experts, often the Bright future for system identification!
developers themselves, not by laymen of control.
Model intelligence can change the situation.

31
Tulleken, H.J.A.F. (1990). Generalized binary noise
REFERENCES test-signal concept for improved identification-
experiment design. Automatica, Vol. 26, No. 1,
Antoulas, A.C. (2005). Approximation of Large- pp. 37-49.
Scale Dynamical Systems. SIAM. Vargas, A. and F. Allgöwer (2004). Model reduction
Astrid, P. (2004). Reduction of Process Simulation for process control using iterative nonlinear
Models: a Proper Orthogonal Decomposition identication. Proc. American Control Conference,
Approach. PhD Thesis, Faculty of Electrical pp. 2915-2920, Boston, , June 2004.
Engineering, Eindhoven University of Wahlberg, B. (1989). Model reduction of high-order
Technology, The Netherlands. estimated models: the asymptotic ML approach.
Banerjee, A. and Y. Arkun (1998). Model predictive Int. J. Control, Vol. 49, No. 1, pp. 169-192.
control of plant transitions using a new Wolfrum, P., A. Vargas, M. Gallivan and F.
identification technique for interpolating Allgöwer (2005). Complexity reduction of a thin
nonlinear models. Journal of Process Control film deposition model using a trajectory based
Vol. 8, Nos 5-6, pp. 441-457. nonlinear model reduction technique. Proc.
Celaya, P., R. Tkatch, Y.C. Zhu and R. Patwardhan American Control Conference, Portland, Oregon,
(2004). Closed-Loop Identification at the 2005.
Hovensa Refinery. NPRA Plant Automation & Zhu, Y.C. (1989). Black-box identification of MIMO
Decision Support Conference, September 19 - 22, transfer functions: asymptotic properties of
2005, San Antonio, Texas. prediction error models. Int. J. Adaptive Control
Celaya, P., V. Marrero, R. Tkatch and R. Patwardhan and Signal Processing, Vol. 3, pp. 357-373.
(2005). Closed-loop identification of a FCC unit. Zhu, Y.C., A.C.P.M. Backx and P. Eykhoff (1991).
NPRA Q&A and Technology Forum, Plant Multivariable process identification based on
Automation & Decision Support Sessions, frequency domain measures. Proceedings of 30th
October 18 - 21, 2005, Grapevine, Texas. IEEE Conference on Decision and Control,
Cutler, C.R. and R.B. Hawkins (1988). Application Brighton, December, 1991, pp. 303-308.
of a large predictive multivariable controller to a Zhu, Y.C. (1998). Multivariable process
hydrocracher second stage reactor. Proceedings identification for MPC: the asymptotic method
of American Control Conference, pp. 284-291. and its applications. Journal of Process Control,
Hjalmarsson, H., M. Gevers, F. de Bruyne (1996). Vol. 8, No. 2, pp. 101-115.
For model-based control design, closed-loop Zhu, Y.C. (2001). Multivariable System
identification gives better performance. Identification for Process Control. Elsevier
Automatica, Vol. 32, No. 12, pp. 1659-1673. Science, Oxford.
Li, D., S.L. Shah and T. Chen (2002). Analysis of Zhu, Y.C. and H. Telkamp (2005). System
dual-rate inferential control systems. Automatica, identification using slowly sampled output data.
Vol. 38, pp 1053-1059. CDC/ECC 2005, 12-15 December 2005, Seville.
Li, D., S.L. Shah and T. Chen (2001). Identification
of Fast-rate Models from Multirate Data.
International J. of Control, Vol. 74, Num. 7, pp
680-689.
Ljung, L. (1985). Asymptotic variance expressions
for identified black-box transfer function models.
IEEE Trans. Autom. Control, Vol. AC-30, pp.
834-844.
Ljung, L. and Z.D. Yuan (1985). Asymptotic
properties of black-box identification of transfer
functions. IEEE Trans. Autom. Control, Vol. AC-
30, pp.514-530.
Mantelli, V, M. Racheli, R. Bordieri, N. Aloi, F.
Trivella and A. Masiello (2005). Integration of
simulation and APC: a case study. ERTC Asset
Maximisation Conference, 24-26 May, 2005,
Budapest.
Pinto, J. (2000). Why is industrial automation
declining? Controls Intelligence & Plant Systems
Report, September 2000.
Qin, S.J. and T.A. Badgwell (2003). A survey of
industrial model predictive control technology,
Control Engineering Practice, Vol. 11, pp. 733-
764.
Richalet, J. (1993). Industrial applications of model
based predictive control. Automatica, Vol. 29,
No. 5, pp. 1251-1274.

32

You might also like