CHE 509: Process Dynamics: Empirical Modeling of Dynamical Systems

CHE 509: Process Dynamics
MODULE 6:
Empirical Modeling of Dynamical Systems
Ayorinde Bamimore, Ph.D.

Department of Chemical Engineering
Obafemi Awolowo University
Module Objective
 To develop empirical models for dynamical
systems from input/output data.
The area of research in Process Systems

Engineering (PSE) that deals with this is
called “System Identification”.
Obafemi Awolowo University 2

Chemical Engineering
What is Process Identification?
Process Identification involves constructing a process model
strictly from experimentally obtained input/output data, with
no response to any laws concerning the fundamental nature
and properties of the system.

The Process Identification Problem
The process identification problem can be posed

thus:
Given u(t), the input to a dynamic process and y(t), the
process response to this input function, what is g(.), the
process transfer function model?
u (t) y (t)
?
Black box
Fig. 6.1: The Problem statement of process identification

Empirical vs Theoretical Modeling
Table 6.1: comparison between empirical and theoretical modeling

Theoretical modeling Process identification
1 Usually involves fewer measurements; Requires extensive measurements, because it
requires experimentation only for the relies entirely on experimentation for the
estimation of unknown model parameters model development
2 Provides information about the internal Provides information only about that portion
state of the process of the process which can be influenced by
control action (i.e., the observable part)
3 Promotes fundamental understanding of Treats the process like a “black box”
the internal workings of the process
4. Requires fairly accurate and complete Requires no such detailed knowledge; only
process knowledge that output data be obtainable in response to
input changes
5 Not particularly useful for poorly Quite often proves to be the only alternative
understood and/or complex processes for modeling the behavior of poorly
understood and/or complex processes
6. Naturally produces both linear and Requires special methods to produce
nonlinear process models nonlinear models

Types of Empirical Models
Step response
Impulse response
Linear Linear autoregressive
Empirical
Models Neural networks
Non-Linear Fuzzy models
Non-linear autoregressive
Hammerstein
Wiener
Fig. 6.2a: Types of empirical models

System Identification Procedure
System Identification entails the following four (4) basic

steps (See Page 3 of Yucai Zhu):
(1) Identification tests or experiments
(2) Model order/ structure selection
(3) Parameter estimation
(4) Model validation

Page 9 of the book: Ljung, L. (1987,

2nd Edition, 1999). System
Identification: Theory for the User.
Prentice Hall, Englewood Cliffs, N.J
Fig. 6.2b: The system identification loop

(1) Identification tests or experiments
 The input-output data are usually collected from an

identification test or experiment that is designed to make the
measured data maximally informative about the system
properties that are of interest to the user.
 Central to conducting an identification test is the issue of choice
of input/excitation signal. The input function must be designed
in such a way that the process to be identified will be forced to
exhibit its inherent characteristics.

The typical input functions usually employed in system identification includes:

 Step
 Impulse
 Pulse (rectangular or arbitrary)
 Sine waves
 White noise
 Pseudo random binary sequences
After collecting the input-output data, usually this is divided into two: the
identification and the validation data. The identification data will be used
for process identification while the validation data will be used for validation.

(2) Model order/ structure selection
A set of candidate models is obtained by specifying their

common properties; a suitable model is searched for within
this set. This is the most theoretical part of the system
identification procedure. It is here that a priori knowledge and
engineering intuition and insight have to be combined with the
formal (mathematical) properties of models.

(3) Parameter estimation

 When the data are available and the model set is
determined, the next step is to find the best model in this
set. For model parameter estimation, an error criterion
(loss function) is specified.
 Often the sum of the squares of some error signals
(residuals) is used as the criterion. The values of the
parameters are determined by minimizing the loss function.

(4) Model validation
When a model is identified, the next step is model validation. This

step is to test whether the estimated model is sufficiently good for
the intended use of the model. First of all, a check to see if the
model is in agreement with the a priori knowledge of the system.
Then a check if the model can fit the test or experimental data well,
preferably using a data sequence that has not been used in model
estimation. The final validation of the model is the application of
the model.
Tools for System Identification
The following are the common tools usually employed for process
identification:
 Graphical tool
 Linear regression
 Non-linear regression
 MATLAB System Identification Toolbox

Linear Regression
a) The general linear model is given by:
[6.1]
Experimentally observed variable is given by:
= [6.2]
Writing out this equation for each observation, , i = 1, 2,…, n
y1 = [6.3]
y2 = [6.4]

Linear Regression
y3 = [6.5]
:
:
yn = [6.6]
In Matrix form we have:
= + [6.7]
Y=Xβ+ϵ [6.8]
Linear Regression
Using the method of least squares of error, we can determine the model
parameters.
[6.9]
i = ith data
j = jth function
S = (Y – Xβ)T(Y – Xβ) [6.10]
To find the value of which minimizes S, differentiating the function wrt
and setting it zero gives
[6.11]
[6.12]

Linear Regression
For a polynomial function
[5.13]
[5.14]
N.B:
The tricks in solving Linear regression problem is in populating matrices X
and Y
Linear regression = if the empirical model is linear with respect to the
model parameters
Nonlinear regression = if the empirical model is nonlinear with respect

to the model parameters

Nonlinear Regression
A general nonlinear model can be written as
y = f (u1, u2, u3, …, β1, β2, β3, …) [6.15]
where y = model output; uj = inputs; βj = model
parameters
For example in , where y = , u = CA
This model is linear with respect to parameter and nonlinear with
respect to n.

Nonlinear Regression
The model parameters can be determined by
[6.16]
where Yi = output measurement at point i
= model prediction at point i

Methods of Identification
Time-domain fitting of step test data
In this method the process to be identified is forced by a unit step input. The
process response (i.e., the output) to this input is recorded. An appropriate
model is then fitted to the step response data.
The candidate models to chose from are:
(a) First Order Plus Time Delay
If the candidate model considered for the process is FOPTD, given below:
[6.17]

Methods of Identification
Transient response to a step input function of size M is given

by:
[6.18]
Model parameters to identify: , and .

Methods I: From step response graph
u (t)
= steady state gain

Δu
= Time for the output to

t=0 t reach 62.3% of its final
y (t) value
= Time delay, this can

Δy
easily be read from the
0.623 Δy
graph
ϴ τ
t=0 t=θ t=τ+θ t
Fig. 6.3a: Step response of a FOPTD system

Method II
The equation can be rewritten as:
[6.19]
where = ultimate value of the step response data

Model parameters to identify: , τ, θ
= MK
=
Eq(6.19) can be rearranged in the linear form as follows:
[6.20a]
y = mx + c [6.20b]

Method II
[6.20a]
c=
y=mx + c [6.20b]
m=
Fig. 6.3b: Plot of vs t

Second-Order-Plus-Time-Delay (SOPTD)
(b) Second-order-plus-time-delay (SOPTD)
Given SOPTD model:
[6.21]
The transient response of the model to step input function of size M is given be:
[6.22]
Parameters: , , ,

Second-Order-Plus-Time-Delay with Numerator Dynamics
(c) Single zero, two poles, plus time delay

Given the model:
[6.23]
The transient response of the model to step input function of size M
is given be:
[6.24]
Parameters: , , , ,

Worked Examples
Example 1: The Figure below (Fig. 6.3c) shows the response
of the temperature T in a continuous stirred-tank reactor to a
step change in feed flow rate w from 120 to 125 kg/min.
Find an approximate first-order model for the process at
these operating conditions.

Example 1
Fig. 6.3c: Temperature response of a stirred-tank for a step change in feed flow rate.

Solution to Example 1
First note that
Because
the steady-state gain is

The time constant obtained from the tangent construction shown in

the (Fig.6.4) is Note that this result is consistent with the
“63.2% method,” because
Consequently, the resulting process model is

where the steady state gain is 4°C/kg/min.

Example 2
Example 2: Identification of an Industrial Distillation Column: The
temperature on Tray #16 of a certain industrial distillation column
is regulated by the flowrate of the underflow reflux to the tray just
below it. The typical steady-state operating values for these
variables are 380oF and 25,000 lb/hr respectively. The result of a
step test carried out on this column is represented below in Table
6.2.

Table 6.2: Step Response Data

Table 6.2: Step Response Data

A candidate model being entertained for this process is:
[6.25a]
Use the information provided to estimate the unknown parameters
It is to be noted here that the data given are in terms of actual process
variables; they must be expressed in terms of deviations from initial
steady-state values. Thus, from Table 6.2, we obtain the output y(s) and
input u(s) as:

y(t)=T 380; u(t)=R 25000 [6.25b]
A careful inspection of the step-response data indicates that the ultimate
change in the Tray #16 temperature as a result of the implemented
reflux flow change is, on the average, 15.0oF, i.e.,
Observe from Table 6.2 that the magnitude of the step change in reflux
rate = (27500 25000) = 2500(lb/hr)

The steady state gain is then estimated as:
We next obtain the variable and set it up in Table
6.3 below:

Solution to Example 2 (Table 6.3)


Applying linear regression on these data against the

expression
[6.26]
we obtain ,
Assignment
(a) Use graphical method to estimate the parameter
(b) Use nonlinear regression to estimate the parameters

Example 3
Example 3: Consider the step response data y(k) in Table

6.4 below. At t=0, a unit step change in u occurs, but the first
output change is not observed until the next sampling instant.
(a) Estimate the model parameters in the second order
difference equation from the input-output data.
[6.27]
(b) Fit a second order transfer function model to this data
using nonlinear regression
(c) Compare the models obtained in (a) and (b) with the
original data.

Example 3
Table 6.4
Table 5: Step Response Data

Example 3
First, generate the matrices thus

6.27
[6.28]
N.B: y(k) = 0 for k < 0; u(k) = 0 for k < 0

Example 3
Y=X [6.29]
Y(K – 1 ) Y(K – 2 ) U(K – 1 ) U(K – 2 )
[6.30]

Example 3
Table 6.4: Data regression for Example Using

3
we have

Example 3
The steady state gain for a second-order difference equation

is obtained thus:
[6.31a]
let and denote the new steady-state values after a step

change in u. Substituting these values in the above equation
gives:
Gain [6.31b]

Example 3
(b) Given a second order transfer function model
g(s) = [6.32]
The transient response of this model to step input of size M is given

by:
[6.33]

Example 3
In discrete time k, the model can be rewritten as:
[6.34]
The problem can now be posed as
[6.35]

MATLAB Solution to Example 3
STEP 1: Create a MATLAB function file and name it “fun_che509”

Table 6.5: MATLAB function “fun_che509”
function f = fun_509(x)
t1=1, ymeas_1=0.058
t2=2, ymeas_2=0.217
t3=3, ymeas_3=0.360
t4=4, ymeas_4=0.488
t5=5, ymeas_5=0.600
t6=6, ymeas_6=0.692
t7=7, ymeas_7=0.772
t8=8, ymeas_8=0.833
t9=9, ymeas_9=0.888
t10=10, ymeas_10=0.925

MATLAB function to obtain model parameters
%%tau1=x(1), tau2=x(2) and K=x(3)

y1=x(3)*(1-((x(1)*exp(-t1/x(1))-x(2)*exp(-t1/x(2)))/(x(1)-x(2)))) ;

f = (ymeas_1-y1)^2+(ymeas_2-y2)^2+(ymeas_3-y3)^2+(ymeas_4-
y4)^2+ (ymeas_5-y5)^2+(ymeas_6-y6)^2+(ymeas_7-y7)^2+(ymeas_8-
y8)^2+ (ymeas_9-y9)^2+(ymeas_10-y10)^2; % Cost function
STEP 2: Invoke the function “fun_che509” from the command window

using
[x,fval] =patternsearch(@fun_509,[3,1,1])
We obtain the results:

The analytical solution
[6.36]
Corresponds to the difference equation:
6.37
where

, , ,

Example 3
Comparison of Results
Linear Regression Nonlinear Regression

Example 4
Example 4: An experiment has been performed to
determine the steady-state power delivered by a gas
turbine-driven generator as a function of fuel flow rate.
The following normalized data were obtained:

Example 4
Fuel Flow Rate Power Generated
Table 6.5: Data

1.0 2.0 regression for Example 4
2.3 4.4
2.9 5.4
4.0 7.5
4.9 9.1
5.8 10.8
6.5 12.3
7.7 14.3
8.4 15.8
9.0 16.8

Example 4
The linear model , should be satisfactory
because the data trend reveal a monotonic trend. Compare
the best linear and quadratic model
Solution
Recall, Y = X [6.8]
[6.38]

Example 4
[6.38]
y 1 u
Using , we
have ,

Example 4
Y=X [6.8]
[6.39]

Example 4

Example 4
Table 6.5: A Comparison of Model Predictions from Example 7.1 (SSE = Sum of squared errors)


CHE 509: Process Dynamics: Empirical Modeling of Dynamical Systems

Uploaded by

Copyright:

Available Formats

CHE 509: Process Dynamics: Empirical Modeling of Dynamical Systems

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

CHE 509: Process Dynamics: Empirical Modeling of Dynamical Systems

Uploaded by

Copyright:

Available Formats

CHE 509: Process Dynamics

Ayorinde Bamimore, Ph.D.

The area of research in Process Systems

Obafemi Awolowo University 2

Process Identification involves constructing a process model

strictly from experimentally obtained input/output data, with

no response to any laws concerning the fundamental nature

and properties of the system.

Obafemi Awolowo University 3

The process identification problem can be posed

Fig. 6.1: The Problem statement of process identification

Table 6.1: comparison between empirical and theoretical modeling

Obafemi Awolowo University 5

Linear Linear autoregressive

Fig. 6.2a: Types of empirical models

Obafemi Awolowo University 6

System Identification entails the following four (4) basic

(1) Identification tests or experiments

(2) Model order/ structure selection

(3) Parameter estimation

(4) Model validation

Obafemi Awolowo University 7

Page 9 of the book: Ljung, L. (1987,

Fig. 6.2b: The system identification loop

Obafemi Awolowo University 8

(1) Identification tests or experiments

 The input-output data are usually collected from an

Obafemi Awolowo University 9

The typical input functions usually employed in system identification includes:

Obafemi Awolowo University 10

(2) Model order/ structure selection

A set of candidate models is obtained by specifying their

Obafemi Awolowo University 11

(3) Parameter estimation

Obafemi Awolowo University 12

(4) Model validation

When a model is identified, the next step is model validation. This

 MATLAB System Identification Toolbox

Obafemi Awolowo University 14

a) The general linear model is given by:

Experimentally observed variable is given by:

Writing out this equation for each observation, , i = 1, 2,…, n

Obafemi Awolowo University 15

In Matrix form we have:

Obafemi Awolowo University 17

For a polynomial function

Nonlinear regression = if the empirical model is nonlinear with respect

Obafemi Awolowo University 18

A general nonlinear model can be written as

y = f (u1, u2, u3, …, β1, β2, β3, …) [6.15]

where y = model output; uj = inputs; βj = model

For example in , where y = , u = CA

This model is linear with respect to parameter and nonlinear with

Obafemi Awolowo University 19

The model parameters can be determined by

where Yi = output measurement at point i

= model prediction at point i

Obafemi Awolowo University 20

Time-domain fitting of step test data