Multivariable and Closed-Loop Identification

for Model Predictive Control

Yucai Zhu*
Faculty of Electrical Engineering, Eindhoven University of Technology
P.O. Box 513, 5600 MB Eindhoven, The Netherlands
Phone +31.40.2473246, y.zhu@tue.nl

Firmin Butoyi
Dow Benelux N.V.
P.O. Box 48, 4530 AA Terneuzen, The Netherlands
*
Also with: Tai-Ji Control
Hageheldlaan 62, NL-5641 GP Eindhoven, The Netherlands
Phone +31.40.2817192, fax +31.40.2813197, y.zhu@tip.nl

Abstract: In this work we will study multivariable and closed-loop identification of large
scale industrial processes for use in model predictive control (MPC). The advantages of
closed-loop identification will be discussed and related problems of identification are
outlined. Then, asymptotic method (ASYM) of identification is introduced. The four
problems, test signal design for control, model order/structure selection, parameter estimation
and model validation, are solved in a systematic manner. The method provides accurate
input/output model and unmeasured disturbance model, model errors are quantified by an
upper error bound matrix that can be used for model validation and test redesign. To
demonstrate the use of the method, the identification of a deethanizer for use in MPC will be
presented.

1 Introduction
Model predictive control (MPC) technology has been widely applied in refinery and
petrochemical industries and is beginning to attract interest from other process industries.
Dynamic models play a central role in MPC technology. Generally identified black-box
models are used for MPC controllers. Industrial project experience has shown that the most
difficult and time-consuming work in an MPC project is modelling and identification. Typical
identification test will take several weeks. These long tests make production planning
difficult. The tests are done manually, which dictates extremely high commitment of the
engineers and operators and the quality of collected data depends heavily on the technical
competence and experience of the control engineer and the operator. The data collected from
these long tests may not be good enough for model identification due to not enough excitation
and process nonlinearity. After the test, due to the lack of systematic identification approach,
it can take another few weeks to analyse the data and to identify the models. Some people
even believe that the long time needed for tests and data analysis is the price one must pay for
a good model.
The high cost of identification of current industrial approach is caused by several factors. First
single variable tests make the test time unnecessarily long. Secondly, open-loop tests are used

1
in industrial MPC projects. When the process is sensitive and/or non-linear, such as a high
purity distillation column, it is very difficult to carry out open-loop tests on the process.
Upsets often occur and it is difficult to fit a linear model when controlled variable variations
are too large. Finally, most industrial identification packages are based on FIR model that is a
nonparametric model with a large number of parameters.
In control community, system (process) identification has been one of the most active
branches in the last three decades. An astonishing fact is that most of the identification results
developed in the last 30 years are not used by industrial control engineers, although there is an
urgent need for efficient and effective identification methods in process control industry.
In the last decade, there is a renewed interest in closed-loop identification and control-relevant
identification; see van den Hof (1997) for a recent summary of the work. Many identification
schemes have been proposed. Unfortunately, most researchers involved in closed-loop
identification assume that the existing controller is linear and the process is single variable.
This makes their result not relevant for use in the framework of MPC, because a MPC
controller is nonlinear and the process under control is multivariable. Although called
“conventional” by many researchers, the prediction error method of Ljung (1987) is still a
more powerful methodology than many newly proposed schemes for use in closed-loop
identification for MPC.
Recently, Zhu (1998) has developed a so-called ASYM method that uses automated
multivariable tests and parametric models. Better models can be obtained with much shorter
test and over 60% time saving is reported.
In this work we will continue the development of Zhu (1998) and study closed-loop
identification. In Section 2 the characteristics of refinery/petrochemical processes are
discussed and closed-loop identification is motivated. In Section 3 the ASYM method is
introduced where closed-loop test design and model validation will be emphasised. In Section
4, the ASYM is used to identify a deethanizer for MPC control. Section 5 contains the
conclusions and perspectives.

2 Closed-Loop Identification, Why and How?

Closed-loop identification means that process model is identified using data collected from a
closed-loop test where the underlying process is fully or partly under feedback control. In this
section we will motivate the use of closed-loop identification from both process operation
point of view and control theoretical point of view and discuss various issues around closed-
loop identification. We will focus our discussion on hydrocarbon process industry (HPI)
processes where MPC application is most challenging and also most beneficial. This class of
processes can be characterised as follows:
1) Large scale and complex. A large size MPC will have 10 to 20 MVs and 20 to 40 CVs.
Some CVs, such as product qualities, are very slow (with dominant time constant range
from 30 minutes to several hours), and other CVs are very fast, such as valve positions
(with time constant in few minutes). There exist inverse responses, oscillations and time
delays.
2) Dominant slow dynamics. The time to steady state of a typical product quality model
ranges from one hour to several hours. This dictates long identification test.
3) High level and slow disturbances. Typical sources of unmeasured disturbances are feed
composition variations, weather changes and disturbances from other part of the unit.

2
 These are slow and irregular variations. During an identification test, the level of
disturbances is in average above 10% of that of CV variation (in power), but it can be
much higher. Too large test signal amplitudes are not permitted because they will cause
off-specification of product and/or will excite nonlinearity.
4) Local nonlinearity. Although in general linear models are relevant for MPC for this class
of processes for a given range of operation, some nonlinear behaviour may still show up.
Examples are CVs that are very pure product qualities and valve positions close to their
limits.
Based on these observations we will outline the special needs in HPI process identification for
MPC and give comments on the existing methods. The discussion will be around the four
problems of identification: test design, model structure and parameter estimation, order
selection and model validation.
1) Identification Test
A good identification test plays a key role in a successful identification. Current practice of
MPC industry is to use a series of open loop and single-variable step tests. The tests are
carried out manually. The advantage of this test method is that control engineer can watch
many step responses during the tests and can learn about the process behaviour in an intuitive
manner. The biggest problem of single variable step test is its high cost in time and in
manpower. The second problem is that the data from single variable test may not contain good
information about the multivariable character of the process (ratios between different models)
and that step signals do not excite enough dynamic information of the process.
Using automatic multivariable test can solve these problems; see Zhu (1998). In an open-loop
multivariable test, many, or, all MVs are perturbed using some test signals such as PRBS
signals. If the process behaves linearly in the operating range, levels of disturbance are low
and the operation constraints are not very tight, one can use open-loop test without problems.
However, as mentioned before, HPI processes often suffer from high level disturbances and
will be nonlinear in a wide range. In such cases, identification tests can be done in closed-loop
operation with part or all the CVs under feedback control. There are many advantages of
closed-loop test:
• Reduce the disturbance to process operation and eliminate product off-specification.
When a multivariable open-loop test is used, some of the CVs may drift away and operator
needs to intervene in order to prevent product qualities from off-specification. In a closed-
loop test, however, one can specify the amplitude of the setpoint movement and the
controller will help to keep the CVs within their operation limits.
• Easy to carry out. During an identification test, to obtain sufficiently high signal-to-noise
ratio and to maintain all the CVs in their operation constraints are often in conflict. It is
most difficult to perform manual single-variable step tests. It demands very high
commitment of the control engineer and of the operator if the process is sensitive and has
tight constraints. An automatic multivariable open-loop test will be much easier, but
operator intervention may still be necessary to keep the CVs within their limits. The work
will become easiest if the process is under feedback control.
• Better model for control. This can be explained in several aspects. Under the same CV
variance constraints, the control performance degradation caused by model errors will be
less if closed-loop test is carried out; see Gevers and Ljung (1986) and Hjalmarsson et. al.
(1996). In fact, CV variances in a closed-loop test can be actually larger than those in an

3
 open loop test, because the process operation constraints are in CV amplitude limits, not in
variance limits and the controller can remove the slow drifts of the CVs. This means that
the signal-to-noise ratio can be higher in a data set from a closed-loop test. The effect of
feedback will have additional advantage if the process is ill conditioned meaning that
several CVs are strongly correlated. High purity distillation columns are often ill
conditioned where top and bottom compositions have strong correlation. For the control of
ill-conditioned processes, it is important to identify the model that has good estimate of
low-gain direction. In an open loop test where MVs are moved independently, the
information of low gain direction will have very low power and cannot be estimated
accurately from noisy data. In order to amplify the power of low-gain direction, certain
correlation between MV movement is needed. This correlation can be created naturally by
(partial) feedback control; see Koung and MacGregor (1993) and Jacobsen (1994).
Remark on the feedback controller
The controller is used to keep important CVs within their operation constraints during the
identification test. It can be one or several PID loops or it is an existing MPC controller. For
example, when testing a distillation column, it is often sufficient to control the top and bottom
compositions or (pressure compensated) temperatures using two PID controllers. If only one
quality is sensitive and has tight constraints, one PID controller will be sufficient. These
controllers often exist for HPI processes due to the evolution of advanced process control in
HP industry.
Remark on process identifiability in closed-loop test
Some researchers and engineers have mistakenly believed that the process is only identifiable
when open-loop test is performed and when MVs are moved independently. This may explain
why so little closed-loop identification applications have been reported. It has been shown
long time ago that, if persistent excitation signals are added on the MVs and/or on the CV
setpoints, the process will be identifiable in a closed-loop test, provided a right model
structure is used; see Gustavsson et. al. (1977). It is true that some model structures and some
estimation methods will be biased and not consistent if used for closed-loop identification; see
the next paragraph.
Test signal design is another important issue that will be discussed in the next section.
2) Model Structure and Parameter Estimation
Before further discussion, let us introduce some notations. Given a multivariable process with
m manipulated variables (MVs or inputs) and p controlled variables (CVs or outputs). Denote
the data sequence collected from an identification test as
Z N := {u (1), y (1), u (2), y (2),......, u ( N ), y ( N )} (2.1)
where u(t) is m-dimensional input vector (MVs), y(t) is p-dimensional output vector (CVs)
and N is the number of samples.
We assume that a linear discrete-time process generates the data
y ( t ) = G o ( z −1 ) u ( t ) + H o ( z −1 ) e ( t ) (2.2)
where z-1 is the unit time delay operator, Go(z-1) is the process transfer function matrix, Ho(z-1)
is the noise filter and e(t) is a p-dimensional white noise vector. Here the term Ho(z-1)e(t)
represent the unmeasured disturbances acting at the process outputs.
The model to be identified is in the same structure as in (2.1):

4
 y ( t ) = G ( z −1 ) u ( t ) + H ( z −1 ) e ( t ) (2.3)
Depending on how to parameterise the model in (2.3), different parameter estimation methods
studied in literature can be derived.
FIR (finite impulse response) model
nb
y (t ) = B( z −1 )u (t ) + e(t ) = (∑ Bk z − k )u (t ) + e(t ) (2.4)
k =0

where Bk is a constant matrix.

Several industrial identification software packages are based on finite impulse
response (FIR) model. This model class has inherent problems for use in identifying
processes with slow dynamics. Model variance error is high due to its high order; yet
bias error is often not negligible due to truncation. Therefore, model errors of FIR
models are considerably larger than that of parametric or compact models; see Zhu et.
al. (1998) for comparative studies.
ARX (AutoRgressive with eXternal input) model, or, least-squares model
y (t ) = A −1 ( z −1 ) B( z −1 )u (t ) + A −1 ( z −1 )e(t ) (2.5a)
or
A( z −1 ) y (t ) = B( z −1 )u (t ) + e(t ) (2.5b)
n n
where A( z −1 ) = I + ∑ Ak z − k and B( z −1 ) = ∑ Bk z − k are called polynomial matrices.
k =1 k =0

Output error (OE) model

y ( t ) = A −1 ( z −1 ) B ( z −1 ) u ( t ) + e ( t ) (2.6)

where A( z −1 ) and B( z −1 ) are polynomial matrices.

ARMAX (AutoRgressive Moving Average with eXternal input) model
y (t ) = A −1 ( z −1 ) B( z −1 )u (t ) + A −1 ( z −1 )C ( z −1 )e(t ) (2.7a)
or
A( z −1 ) y (t ) = B( z −1 )u (t ) + C ( z −1 )e(t ) (2.7b)

where A( z −1 ) , B( z −1 ) and C ( z −1 ) are polynomial matrices.

Box-Jenkins model
y (t ) = A −1 ( z −1 ) B( z −1 )u (t ) + D −1 ( z −1 )C ( z −1 )e(t ) (2.8)

where A( z −1 ) , B( z −1 ) , C ( z −1 ) and D( z −1 ) are polynomial matrices.

The model parameters are determined by minimising the sum of squares of the error
e(t). In literature, ARX, OE, ARMAX and Box-Jenkins models are called parametric
models and FIR model is called nonparametric model. The difference between the two
types of models is that parametric models are much more compact than FIR models
and need much less parameters to describe the same dynamic behaviour. Let us use the

5
 degree of polynomial matrix as a measure of model compactness. Then a model is said
to be more compact if the polynomial degree is lower.
Subspace method
In recent years, the so called subspace method of parameter estimation has been
proposed and studied in the literature; see van Overschee and de Moor (1994),
Verhaegen (1994) and Larimore (1990). Subspace methods estimate state space model
of a multivariable process directly from input/output data. A state space model of a
linear process with disturbance can be given as
x(t + 1) = Ax(t ) + Bu (t ) + Ke(t )
(2.9)
y (t ) = Cx(t ) + Du (t ) + e(t )
where x(t) is the state vector of dimension nA, the constant matrices A, B, C and D
form the state space description of the process and constant matrix K is the Kalman
gain that characterise the state noise, e(t) is a p dimensional white noise vector.
For closed-loop identification, the choice of model structure depends on three and often
conflict issues:
1) the compactness of the model
2) the numerical complexity in parameter estimation
3) the consistency of the model in closed-loop identification.

When noisy data is used in identification, a more compact model will be more accurate
provided that the parameter estimation algorithm converges to global minimum and the model
order is selected properly; see Section 3.
Among the parametric models, one would like to have the model that is most accurate or, in
identification term, minimum variance. In general a model structure or an estimation method
that includes a disturbance model will be better than a method without the disturbance model;
see Ljung (1987) and Söderström and Stoica (1989). Prediction error criterion and maximum
likelihood criterion belongs to the first class; while output error criterion belongs to the
second class. At present it is not clear what criterion subspace methods use. Moreover,
prediction error method and maximum likelihood method will give consistent estimate for
closed-loop data meaning that the effect of the disturbance will decrease when test time
increases; while output error criterion will deliver biased model when using closed-loop data.
The consistency of subspace methods can only be proven in for open-loop test case; see
Jansson and Wahlberg (1996).
However, a more compact model needs more complex parameter estimation algorithms. To
estimate OE models, Box-Jenkins models and ARMAX models, nonlinear optimisation
routines are needed which often suffer from local minima and convergence problems when
identifying a multivariable process. In FIR and ARX models, the error term e(t) is linear in the
parameters; see (2.4) and (2.5b). Due to this property, linear least-squares method can be used
in parameter estimation that is numerically simple and reliable, and no problem of local
minimum and convergence will occur. This explains partly why FIR model is often used in
industrial identification. The subspace methods are exceptional: they estimate a parametric
model and they are numerically efficient. The main part of a subspace method consists of
matrix singular value decomposition (SVD) and linear least-squares estimation which are
numerically simple and reliable.

6
Recently, orthonormal basis functions are studied for use in identification; see, e.g., Wahlberg
(1991) and van den Hof et. al. (1995). One of the motivations of using this class of models is
to estimate an output error model using the numerically simple least-squares method.
To summarise the discussion, the following table compares the advantages and disadvantages
of various model structures or parameter estimation methods. It is clear that Box-Jenkins
model is the best candidate for closed-loop identification, provided that parameter estimation
and order selection problems can be solved.
Table 2.1 Comparison of various model structures and estimation methods
Model structure or Numerical Compactness Consistency in
estimation method difficulty closed-loop test
FIR Low Low No
ARX (high order) Low Medium Yes
OE High Highest No
ARMAX High High Yes
Box-Jenkins High Highest Yes
Subspace method Low High No
Orthonormal basis function Medium Medium No

3) Order Selection
In identification literature, when prediction error is used in parameter estimation, it is also
used in order selection. We will argue that, although prediction error is a good choice for
parameter estimation, it is not the best criterion for order selection for control. For the purpose
of control, it is most important to select the model order so that the process model G ( z −1 ) is
most accurate. In the time domain, this requires that the simulation error or, output error of the
model be minimal; see Zhu (1988b) for more details and an example. In the frequency
domain, this requires that the bias part and the variance part of the model are nearly equal so
that the total error is minimal; see Section 3.
4) Model Validation
The goal of model validation is to test whether the model is good enough for its purpose and
to provide advise for possible re-identification if the identified model is not valid for its
intended use. Commonly used methods of model validation are simulation using estimation
data or fresh data, whiteness test of model residuals and testing the independence between the
residuals and past MVs. These methods only tell how well the model agrees with the test data.
They can neither quantify the model quality with respect to the purpose of modelling, in our
case, closed-loop control, nor can they give good advice for re-identification. In this way, the
fundamental problem of model validation is basically unsolved in both literature and practice,
especially for multivariable processes. Trial-and-error approach in this step makes industrial
identification a very expensive practice. In linear model identification for control, a logical
approach for model validation is to quantify model errors in the frequency domain and to find
relationship between this kind of error description and model based control performance.
Model error estimation has attracted much interest in the last decade, which is motivated by
linear robust control theory.
The above mentioned four problems are fundamental in industrial process identification.
Solving these problems in a systematic manner is a challenging task. In the next section the so

7
called ASYM method of identification is introduced that provide industrial solutions to these
problems.

3 Asymptotic Method of Identification

The asymptotic method (ASYM) of identification was developed for the purpose model based
control; see Zhu (1998).
Denote frequency responses of the process (2.2) and of the model (2.3) are denoted as
T o (e iω ) := col[G o (e iω ), H o (e iω )]

T n (e iω ): = col[G n (e iω ), H n (e iω )]
where n is the degree of the polynomials of the model, col(.) denotes the column operator.

The Asymptotic Properties of Prediction Error Method

Assume that
 n → ∞ and n 2 / N → 0 as N → ∞
 Test signals have continuous none zero spectra until the Nyquist frequency.
Then (Ljung, 1986 and Zhu 1989)
 T n (e iω ) → T o (e iω ) as N → ∞ (3.1)

 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 ) (3.2)
N

where Φ(ω ) is the spectrum matrix of inputs and prediction error residual col[u T (t ), ξ T (t )] ,
Φ v (ω ) is spectrum matrix of unmeasured disturbances, ⊗ denotes the Kronecker product and
-T denotes inverse and then transpose.

Note that the theory holds for the general case of closed-loop test. For open-loop test, the
following simplified results can be obtained for process model
n
cov[ col[ G n (e iw )] ≈ Φ −u T (ω ) ⊗ Φ v (ω ) (3.3)
N

where Φ u (ω ) is the spectrum matrix of inputs.

In the following, we will outline the ASYM method that makes extensive use of the
asymptotic theory. In order to minimise the presentation complexity, we will outline the
method for open-loop experiment and also assume that the test inputs are mutually
independent.

The Procedure of the Asymptotic Method

1) Identification Test

8
One should realise that a theory for practical identification test design does not exist and,
therefore, project experience and control engineering intuition are combined with the
asymptotic theory in developing the ASYM test scheme. The following are the important
features of the ASYM test:
a) Duration of identification test. The duration of ASYM is determined by several
factors: the validation of the asymptotic expression, the process time to steady state
(settling time), and the number of MVs. The experience has shown that when
compared to traditional step test, 70% of test time can be saved using multivariable
test. Note that this test is designed for parametric model identification. The test time
may not be long enough for nonparametric FIR models, because the identification of
nonparametric models needs much longer test time; see Zhu et. al. (1999).
b) (Partial) Closed-loop test. When the level of disturbance is not high and all the CVs
have large operation ranges, it will be easy to test the process in open loop operation.
But often some CVs are sensitive and have tight constraints; open loop test may be
difficult to carry out. In this situation, closed-loop will be used with these sensitive
CVs being controlled by some single loop PID controllers or by an existing MPC
controller.
When single loop controllers are used, the test signals are applied to the MVs which
are not under feedback control and the setpoints of the CVs which are under control. If
an MV under closed-loop test does not move sufficiently due to too slow control
action, additional test signal can be added on that MV.
c) Spectra of test signals. The optimal test signal is designed to minimise the sum of the
squares of the simulation error (often called prediction error in MPC term). In open-
loop test, the spectra of the test signals are based on the following formula which can
be derived from (3.6) assuming open-loop test (Ljung and Yuan, 1985):
p

Φ opt
uj (ω ) = µ j Φ sim
uj (ω ) ∑Φ vi (ω ) (3.4)
i =1

where Φ usimj (ω ) is the spectrum of the j-th input movement in a controlled system, µj is
a constant which is adjusted to meet the amplitude constraint of the signal.

In closed-loop test, the following formula can derived for the single variable case (Zhu
and van den Bosch, 1999)

r (ω ) ≈ µ Φ r (ω )Φ v (ω )
Φ opt (3.5)

where Φ r (ω ) is the reference signal at setpoint of the closed-loop system.

In practice, (3.4) or (3.5) is used in combination with upper error bound (3.9) to give
general guidelines for the test design. The spectra of the test signals can be realised by
PRBS (pseudo-random-binary-sequence) signals or filtered white noises.

9
Before model identification, some data pre-treatment is carried out. This includes removing
the spikes, data slicing, hi-pass filtering (de-trending), known delay correction and signal
scaling. This part of the work is well known; see Ljung (1987) and Zhu and Backx (1993).
2) Estimate a high order ARX (equation error) model
A n ( z −1 ) y (t ) = B n ( z −1 )u(t ) + e(t ) (3.6)

where A n ( z −1 ) is a diagonal polynomial matrix and B n ( z −1 ) is full polynomial matrix, both

with degree n polynomials. Denote G n ( z −1 ) as the high order ARX model of the process, and
H n ( z −1 ) as the high order model of the disturbance.
3) Perform frequency weighted model reduction (ML estimate)
The high order model in (3.6) is practically unbiased, provided that the process behaves linear
around the working point. The variance of this model is high due to its high order. Here we
intend to reduce the variance by perform a model reduction on the high order model. If we
view the frequency response of the high order estimates as the noisy observations of the true
transfer function, we can then apply the maximum likelihood principle. Using the asymptotic
result of (3.1) and (3.3), we can show that the asymptotic negative log-likelihood function for
the reduced process model is given by (Wahlberg, 1989, Zhu and Backx, 1993)
p ω2
m
V= ∑ ∑ ∫ |{| G n
ij (ω ) − G ij (ω )|2 [ Φ −1 (ω )] −jj1 Φ v−i1 (ω )}|d ω (3.7)
i =1 j =1 ω
1

The reduced model G ( z −1 ) is thus calculated by minimizing (3.7) for a fixed order. The same
can be done for the disturbance model H n ( z −1 ) = 1 / A n ( z −1 ) .
There are some theoretical evidence that the high order plus model reduction approach will
produce more accurate models than that obtained using direct estimation of low order models;
see Tjärnström and Ljung (1999).
4) Use asymptotic criterion (ASYC) for order selection
The best order of the reduced model is determined using a frequency domain criterion ASYC
which is related naturally to the noise-to-signal ratios and to the test time; see Zhu (1994) for
the motivation and evaluation. The basic idea of this criterion is to equalise the bias error and
variance error of each transfer function in the frequency range that is important for control.
Let [ω 1 , ω 2 ] defines the frequency band that is important for the MPC application, the
asymptotic criterion (ASYC) is given by:
p ω2
m
n
ASYC := ∑ ∑ ∫ |[| G n
ij (ω ) − G ij (ω )|2 − [ Φ −1 ] jj (ω )Φ v (ω )]|d ω
N i
(3.8)
i =1 j =1 ω
1

5) Model validation using error bound matrix

According to the result (3.1) and (3.3), a 3-δ bound can be derived for the high order model as
follows:
n
Gijo (e iω ) − G ijn (e iω ) ≤ 3 [Φ −1 (ω )] jj Φ v (ω ) w.p. 99.9% (3.9)
N i

10
We will also use this bound for the reduced model because the model reduction will in general
improve model quality.
For use in MPC, model validation is done by comparing the relative size of the bound with the
model over the low and middle frequencies. More specifically, identified transfer functions
are graded in:
A, very good, when bound ≤ 30%model
B, good, when 30%model < bound ≤ 60%model
C, marginal, when 60%model < bound ≤ 90%model
D, poor, or, no model exists, when bound >90%model).
1) Zero them when there are no transfer between the MV/CV pairs. This can be determined
by using the process knowledge and cross checking.
2) If a transfer function is expected and needed in the control, redesign a test in order to
improve the accuracy of these models.

Using upper bound formula (3.9) we can easily give some guidelines for improving the test
design:
• doubling the amplitudes of test signals or quadruple the test time will half the error over
all frequencies;
• doubling the average switch time of PRBS signals will half the model error at low
frequencies and double the error at high frequencies.
Because ASYM provides systematic solutions to all the four identification problems, it can be
made very user friendly for non-expert users. Recently, ASYM has been implemented in
automated identification software Tai-Ji ID; see Zhu and Ge (1997). Using the software, the
time needed for model identification of HPI processes ranges from few hours to few days.

4 Closed-Loop Identification of a Deethanizer for MPC

The process is a Deethanizer which separates C2 and lighter from C3 and heavier. The light
product leaves the column overhead as vapor distillate and the heavy product exits the column
as bottom liquid flow. The column operates in a high purity range.
A DMC controller was designed and was going to be commissioned as part of an APC and
optimization project. The purpose of the deethanizer DMC is to reduce the variations of
product qualities while respecting process operation constraints.

The MVs of the controller:

Reflux: Reflux flow setpoint
Steam: Reboiler steam flow setpoint
Preheater: Feed preheater flow setpoint

DV:
Feed: Column feed flow

Main CVs of the controller:

OverheadC3: Overhead C3 composition
DeltaPress: Column pressure difference

11
 BotTemp: Bottom temperature
TopTemp: Top temperature
TrayTemp: A tray temperature

The experience of column operation has shown that the tray temperature TrayTemp is an
important variable. It should be controlled in a range of 6 degrees for normal operation. When
the try temperature is outside this range, it will be difficult to control the column.

Initially, an open loop multivariable test has been carried out; see Figures 4.1 and 4.2. PRBS
signals are used as test signals. The following have happened during the open loop test:

- The tray temperature varied in a range of about 20 degrees, which is far beyond the
normal operation range. The accepted moves to apply on MV’s and determined during
the pretest session turned out to be too big.
- The tray temperature became too high at about sample 600. The operator closed the PI
control loop in order to bring it back. The control action reduced the steam flow to a
very low level, which may cause nonlinearity in the data.
- The preheater flow could not be moved during the most part of the test period due to
high level of disturbance introduced by reflux and steam.
- The column pressure difference became too high after sample 2200, which indicates
the column was in flooding. The overhead C3 composition increased during this
period.
- The second half of the test has disturbed normal unit operation.

The data have been used for identification using several software packages. Data slicing is
used to remove the portion during column flooding. When the identified models are used in
the DMC controller, the closed-loop performance was not satisfactory. During the control
commissioning it was quickly decided that the best solution would to retest the plant to get
better data set.

Remark: The difficulty of the test is caused more by the high purity characteristics of the
column than by the multivariable test approach. A single variable open test will face the same
problem for high purity distillation columns. In general a well designed multivariable test will
not cause more disturbance than a conventional single variable step test; see Zhu (1998).

Finally, it was decided to carry out a partial closed-loop test with the tray temperature
controlled by the steam flow using an existing PI controller; see Figures 4.1 and 4.2 for the
test data. The test can be summarised as follows:

- CV variations are much smaller than the CV variations of the open loop test. The tray
temperature is within a range of 7 oC during the test.
- All the three MVs could be tested according to the plan. The test did not disturb
normal operation.
- Much less operator intervention has taken place, which implies easy test.
- During the second half of the test, the feed flow had to be reduced by about 20% due
to production planning. The operator could handle this easily, thanks to the tray
temperature controller.

12
The ASYM method is used to identify the models using the closed-loop data. The DMC
commissioning using the closed-loop model went smoothly and the controller has been online
for over half year without major problems.

For comparison, the identification results of the open loop data and the closed-loop data are
shown in Figures 4.3, 4.4 and 4.5. Note that the tray temperature TrayTemp is identified as an
integral process and the models of its derivative are shown. One can see that the quality of
closed-loop model is higher than that of open loop model according to upper error bounds.
The difference can also be seen using the process knowledge. For example, the gains between
the reflux and both top and bottom temperatures should be negative. The closed-loop model
correctly determines these. However, the gain of the open loop model is positive for the
bottom temperature and is nearly zero for the top temperature. The model fit of the closed-
loop model is slightly better that that of the open-loop data. The difference is partly due to the
different operation range and partly due to uncertainty caused by disturbance and/or
nonlinearity.
Due to the change of feed flow during the closed-loop test, the usable data for identification is
less than desired. Fortunately, the model from this very short test period is good enough for
control. This shows the accuracy of parametric models.

MV signals, open loop test MV signals, closed−loop test

5 5
0 0
Reflux

Reflux

−5 −5
−10 −10
0 1000 2000 3000 0 1000 2000 3000

5 5
Steam

Steam

0 0

−5 −5

0 1000 2000 3000 0 1000 2000 3000

10 10
Preheater

Preheater

0 0

−10 −10

0 1000 2000 3000 0 1000 2000 3000

4 4
2 2
0 0
Feed

Feed

−2 −2
−4 −4
0 1000 2000 3000 0 1000 2000 3000
Samples Samples

Figure 4.1 MV plots of the open loop test and closed-loop test. The data are normalised by
subtracting their mean values and by dividing by some factors. For each MV, the same scaling
factor is used for both open loop test and closed-loop test.

13
 CV signals, open loop test CV signals, closed−loop test
4 4
OverheadC3

OverheadC3
2 2
0 0

0 1000 2000 3000 0 1000 2000 3000

DeltaPress

DeltaPress
2 2
1 1
0 0
−1 −1
0 1000 2000 3000 0 1000 2000 3000

2 2
BotTemp

BotTemp
0 0
−2 −2
−4 −4
0 1000 2000 3000 0 1000 2000 3000
4 4
TopTemp

2 TopTemp 2
0 0
−2 −2
0 1000 2000 3000 0 1000 2000 3000
4 4
TrayTemp

TrayTemp

2 2
0 0
−2 −2
0 1000 2000 3000 0 1000 2000 3000
Samples Samples

Figure 4.2 CV plots of the open loop test and closed-loop test. The data are normalised by
subtracting their mean values and by dividing by some factors. For each CV, the same scaling
factor is used for both open loop test and closed-loop test.

14
 openid: Step Responses

Reflux Steam Feed

D B D
OverheadC3

−592.8525 2634.1249 −8.6078

C B C
DeltaPress

11.6178 383.759 −29.5231

C B C
BotTemp

0.12585 1.0515 0.16275

A A C
TopTemp

−0.0069949 1.7092 0.059531

D B C
d−TrayTemp

0.014099 0.55079 −0.058329

Time (MIN): 1:200

closeid: Step Responses

Reflux Steam Preheater Feed

B A B D
OverheadC3

−590.2981 3516.4541 1297.0848 −127.7527

D A D C
DeltaPress

0.46384 242.5045 25.5465 −18.4637

D A C C
BotTemp

−0.19631 1.0918 0.47129 −0.10944

C A C D
TopTemp

−0.17415 0.95786 0.4052 −0.018708

D A D D
d−TrayTemp

−0.019959 0.37361 0.072788 −0.041174

Time (MIN): 1:200

Figure 4.3 Step responses of open loop model (up) and closed-loop model

15
 openid: Frequency Responses (solid) & Error Bounds

Reflux Steam Feed

D B D
OverheadC3

−562.4309 2855.0308 −41.0974

C B C
DeltaPress

20.2271 442.8802 −29.8915

C B C
BotTemp

0.18735 1.2357 0.16493

A A C
TopTemp

0.023416 1.8434 0.085283

D B C
d−TrayTemp

0.02465 0.53511 −0.054227

Frequency (rad/MIN): (LOG) 0.000314:3.14

closeid: Frequency Responses (solid) & Error Bounds

Reflux Steam Preheater Feed

B A B D
OverheadC3

−589.4091 3458.7164 1246.8705 −143.0344

D A D C
DeltaPress

−3.1146 238.1602 30.5064 −15.4426

D A C C
BotTemp

−0.21724 1.0357 0.51006 −0.13884

C A C D
TopTemp

−0.15857 0.79818 0.40014 0.0069508

D A D D
d−TrayTemp

−0.018277 0.35552 0.058949 −0.045548

Frequency (rad/MIN): (LOG) 0.000314:3.14

Figure 4.4 Frequency responses and error bounds of open loop and closed-loop models

16
 openid: Model Fit, CVs (solid) & Simulated CVs
6000
5000
4000
OverheadC3

3000 Error%
2000
1000 18.1242
0

3050
3000
DeltaPress

2950
2900
Error%

2850
2800 44.365
2750
84

83
BotTemp

82 Error%

81
19.6131
80

−8

−9
TopTemp

Error%
−10
22.681
−11

0.5
d−TrayTemp

0 Error%

−0.5
64.9077

0 200 400 600 800 1000 1200 1400 1600

Time (MIN)

closeid: Model Fit, CVs (solid) & Simulated CVs

10000

8000
OverheadC3

6000 Error%

4000
36.7326
2000

0
2900

2800
DeltaPress

Error%
2700

13.5747
2600

84

83
BotTemp

Error%
82
19.1112
81

80

−8

−9
TopTemp

Error%
−10

23.0571
−11

0.6
0.4
d−TrayTemp

0.2
0 Error%
−0.2
−0.4 48.8301
−0.6

0 200 400 600 800 1000 1200 1400 1600 1800 2000
Time (MIN)

Figure 4.5 Model fit of open loop model (up) and closed-loop model

17
5 Conclusions and Perspectives
In this work we have studied multivariable and closed-loop identification for use in MPC. The
ASYM method is introduced to solve the problem. An industrial application is reported. We
have shown that closed-loop tests have many advantages over open-loop tests. When
comparing to the conventional open-loop step test approach, the following improvements
have been made by the ASYM method:
• The model quality is higher due to well-designed test, the use of parametric models and
ability to keep the process in a linear range.
• The disturbance to uint operation is much smaller, due to closed-loop control and
encouraged operator intervention when necessary.
• The cost of identification is significantly reduced. Reduction of both test time and data
analysis time by 70% can be realised. The method requires less user knowledge and
experience on identification.
A powerful and efficient identification method can make MPC economically feasible for more
process units.
From application point of view, the next logical step is to use full closed-loop test with an
MPC controller online. The identified models can be used for MPC controller diagnosis and
maintenance. When sufficient industrial experience of closed-loop identification is obtained, it
would be natural to try adaptive MPC. Closed-loop identification of nonlinear models can be
very useful for difficult processes such as high purity distillation columns and more research
is needed for this challenging topic.

References
Cutler, C.R. and R.B. Hawkins (1988). Application of a large predictive multivariable
controller to a hydrocracher second stage reactor. Proceedings of American Control
Conference, pp. 284-291.
Gustavsson I., L. Ljung and T. Söderström (1977). Identification of processes in closed loop
— identifiability and accuracy aspects. Automatica, Vol. 13, pp. 59-75.
Hjalmarsson, H., M. Gevers, F. de Bruyne (1996). For model-based control design, closed-
loop identification gives better performance. Automatica, Vol. 32, No. 12, pp. 1659-1673.
Jacobsen, E.W. (1994). Identification for Control of Strongly Interactive Plants. Paper 226ah,
AIChE Annual Meeting, San Francisco.
Jansson, M. and B. Wahlberg (1996). On consistency of subspace system identification
methods.
Koung, C.W. and J.F. MacGregor (1993). Design of identification experiments for robust
control. A geometric approach for bivariate processes. Ind. Eng. Chem. Res., Vol. 32, pp.
1658-1666.
Larimore, W.E. (1990). Canonical variable analysis in identification, filtering and adaptive
control. Proceedings of 29th IEEE Conference on Decision and Control, Honolulu,
Hawaii, pp. 596-604.
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. (1987). System Identification: Theory for the User. Prentice-Hall, Englewood Cliffs,
N.J.

18
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.
Qin, S.J. and T.A. Badgwell (1996). An overview of industrial model predictive control
technology, Internet: www.che.utexas.edu/~qin/cpcv/cpcv14.html.
Richalet, J. (1993). Industrial applications of model based predictive control. Automatica,
Vol. 29, No. 5, pp. 1251-1274.
Tjärnström, F. and L. Ljung (1999). L2 model reduction and variance reduction. Report LiTH-
ISY-R-2158, Department of Electrical Engineering, Linköping University, June 1999.
van den Hof, P.M.J. (1997). Closed-loop issues in system identification. In Proceedings of the
11th IFAC Symposium on System Identification, Fukuoka, Japan, Vol. 4, pp. 1651-1664.
Van Overschee, P. and B. De Moor (1994). N4SID: subspace algorithms for the identification
of combined deterministic-stochastic systems. Automatica, Vol. 30, No. 1, pp. 75-93.
Verhaegen, M. (1994). Identification of the deterministic part of MIMO state space models
given in innovations form from input-output data. Automatica, Vol. 30, No. 1, pp. 61-74.
Wahlberg, B. (1989). Model reduction of high-order estimated models: the asymptotic ML
approach. Int. J. Control, Vol. 49, No. 1, pp. 169-192.
Zhu, Y.C. (1989). Black-box identification of MIMO transfer functions: asymptotic properties
of prediction error models. Int. J. Adaptive Control and Signal Processing, Vol. 3, pp.
357-373.
Zhu, Y.C. (1998). Multivariable process identification for MPC: the asymptotic method and
its applications. Journal of Process Control, Vol. 8, No. 2, pp. 101-115.
Zhu, Y.C. and T. Backx (1993). Identification of Multivariable Industrial Processes: for
Simulation, Diagnosis and Control. Springer-Verlag, London.
Zhu, Y.C. and X.H. Ge (1997). Tai-Ji ID: automatic system identification package for model
based process control. Journal A, Vol. 38, No. 3, pp 42-45.
Zhu, Y.C. and P. van den Bosch (1999). A simple closed-loop test design formula for internal
model control. Submitted to Automatica.
Zhu, Y.C., E. Eduardo, F. Butoyi and F. Cortes (1999). Parametric versus nonparametric
models in industrial process identification for MPC. To appear in Hydrocarbon
Processing.

19