Development of Empirical Models

From Process Data

• In some situations it is not feasible to develop a theoretical
(physically-based model) due to:
1. Lack of information
Chapter 7

2. Model complexity
3. Engineering effort required.
• An attractive alternative: Develop an empirical dynamic
model from input-output data.
• Advantage: less effort is required
• Disadvantage: the model is only valid (at best) for the
range of data used in its development.
i.e., empirical models usually don’t extrapolate very
well. 1
 Simple Linear Regression: Steady-State Model
• As an illustrative example, consider a simple linear model
between an output variable y and input variable u,
y = β1 + β 2u + ε

where β1 and β 2 are the unknown model parameters to be

Chapter 7

estimated and ε is a random error.

• Predictions of y can be made from the regression model,

yˆ = βˆ 1 + βˆ 2u (7-3)
where β̂1 and β̂2 denote the estimated values of β1 and β2,
and ŷ denotes the predicted value of y.
• Let Y denote the measured value of y. Each pair of (ui, Yi)
observations satisfies:
Yi = β1 + β2ui + εi (7-1)
2
 The Least Squares Approach
• The least squares method is widely used to calculate the
values of β1 and β2 that minimize the sum of the squares of
the errors S for an arbitrary number of data points, N:
N N
Chapter 7

2
S =∑ ε12 = ∑ (Yi − β1 − β 2ui ) (7-2)
i =1 i =1

• Replace the unknown values of β1 and β2 in (7-2) by their

estimates. Then using (7-3), S can be written as:
N
S = ∑ ei2
i =1
where the i -th residual, ei , is defined as,
ei Yi − yˆi (7 − 4)
3
 The Least Squares Approach (continued)

• The least squares solution that minimizes the sum of

squared errors, S, is given by:
Suu S y − Suy Su
Chapter 7

β̂1 = 2
(7-5)
NSuu − ( Su )

NSuy − Su S y
β̂ 2 = 2
(7-6)
NSuu − ( Su )
where:
N N N N
Su ∆ ∑ ui Suu ∆ ∑ ui2 S y ∆ ∑ Yi Suy ∆ ∑ uiYi
i =1 i =1 i =1 i =1

4
 Extensions of the Least Squares Approach
• Least squares estimation can be extended to more general
models with:
1. More than one input or output variable.
2. Functionals of the input variables u, such as poly-
Chapter 7

nomials and exponentials, as long as the unknown

parameters appear linearly.

• A general nonlinear steady-state model which is linear in the

parameters has the form,
p
y = ∑β j X j + ε (7-7)
j =1

where each Xj is a nonlinear function of u.

5
 The sum of the squares function analogous to (7-2) is
2
N  p 
S = ∑  Yi − ∑ β j X ij  (7-8)

i =1 

j =1 
which can be written as,
Chapter 7

T
S = (Y - X β ) (Y − X β ) (7-9)

where the superscript T denotes the matrix transpose and:

 Y1   β1 
 
Y =  β= 
 
Yn  β p 
 

6
  X11 X12 X1 p 
 
 X 21 X 22 X2p 
X = 
 
 X n1 X n2 X np 
Chapter 7

The least squares estimates β̂ is given by,

( )
−1
βˆ = X X T
X TY (7-10)

providing that matrix XTX is nonsingular so that its inverse exists.

Note that the matrix X is comprised of functions of uj; for
example, if:
y = β1 + β 2u + β3 u 2 + ε

This model is in the form of (7-7) if X1 = 1, X2 = u, and

X3 = u2. 7
 Fitting First and Second-Order Models
Using Step Tests
• Simple transfer function models can be obtained graphically
from step response data.
• A plot of the output response of a process to a step change in
Chapter 7

input is sometimes referred to as a process reaction curve.

• If the process of interest can be approximated by a first- or
second-order linear model, the model parameters can be
obtained by inspection of the process reaction curve.

• The response of a first-order model, Y(s)/U(s)=K/(τs+1), to

a step change of magnitude M is:
y ( t ) = KM (1 − e −t / τ ) (5-18)

8
 • The initial slope is given by:

d y  1
  = (7-15)
dt  KM t =0 τ
Chapter 7

• The gain can be calculated from the steady-state changes

in u and y:
∆y ∆y
K= =
∆u M

where ∆y is the steady-state change in y.

9
Chapter 7

Figure 7.3 Step response of a first-order system and

graphical constructions used to estimate the time constant, τ.

10
 First-Order Plus Time Delay Model
Ke-θ s
G ( s) =
τs + 1
For this FOPTD model, we note the following charac-
teristics of its step response:
Chapter 7

1. The response attains 63.2% of its final response

at time, t = τ+θ.
2. The line drawn tangent to the response at
maximum slope (t = θ) intersects the y/KM=1
line at (t = τ + θ ).

3. The step response is essentially complete at t=5τ.

In other words, the settling time is ts=5τ.

11
Chapter 7

Figure 7.5 Graphical analysis of the process reaction curve

to obtain parameters of a first-order plus time delay model.
12
 There are two generally accepted graphical techniques for
determining model parameters τ, θ, and K.

Method 1: Slope-intercept method

First, a slope is drawn through the inflection point of the
Chapter 7

process reaction curve in Fig. 7.5. Then τ and θ are

determined by inspection.

Alternatively, τ can be found from the time that the

normalized response is 63.2% complete or from
determination of the settling time, ts. Then set τ=ts/5.
Method 2. Sundaresan and Krishnaswamy’s Method
This method avoids use of the point of inflection
construction entirely to estimate the time delay.
13
 Sundaresan and Krishnaswamy’s Method
• They proposed that two times, t1 and t2, be estimated from a
step response curve, corresponding to the 35.3% and 85.3%
response times, respectively.
Chapter 7

• The time delay and time constant are then estimated from the
following equations:
θ = 1.3t1 − 0.29t2
(7-19)
τ = 0.67 ( t2 − t1 )

• These values of θ and τ approximately minimize the

difference between the measured response and the model,
based on a correlation for many data sets.

14
 Estimating Second-order Model Parameters
Using Graphical Analysis
• In general, a better approximation to an experimental step
response can be obtained by fitting a second-order model to
the data.
Chapter 7

• Figure 7.6 shows the range of shapes that can occur for the
step response model,
K
G (s) = (5-39)
( τ1s + 1)( τ 2 s + 1)
• Figure 7.6 includes two limiting cases: τ 2 / τ1 = 0 , where the
system becomes first order, and τ 2 / τ1 = 1 , the critically
damped case.
• The larger of the two time constants, τ1 , is called the
dominant time constant.
15
Chapter 7

Figure 7.6 Step response for several overdamped second-

order systems.

16
 Smith’s Method
• Assumed model:
Ke −θs
G (s) =
τ 2 s 2 + 2ζτs + 1
Chapter 7

• Procedure:

1. Determine t20 and t60 from the step response.

2. Find ζ and t60/τ from Fig. 7.7.
3. Find t60/τ from Fig. 7.7 and then calculate τ (since
t60 is known).

17
 Chapter 7

18
 Fitting an Integrator Model
to Step Response Data

In Chapter 5 we considered the response of a first-order process

to a step change in input of magnitude M:

( )
Chapter 7

y1 ( t ) = KM 1 − e −t / τ (5-18)
For short times, t < τ, the exponential term can be approximated
by
−t / τ t
e ≈ 1−
τ
so that the approximate response is:

  t   KM
y1 ( t ) ≈ KM 1 −  1 −   = t (7-22)
  τ  τ

19
 is virtually indistinguishable from the step response of the
integrating element
K2
G2 ( s ) = (7-23)
s
In the time domain, the step response of an integrator is
Chapter 7

y2 ( t ) = K 2 Mt (7-24)

Hence an approximate way of modeling a first-order process is

to find the single parameter
K
K2 = (7-25)
τ
that matches the early ramp-like response to a step change in
input.

20
 If the original process transfer function contains a time delay
(cf. Eq. 7-16), the approximate short-term response to a step
input of magnitude M would be

KM
y (t ) = (t − θ ) S (t − θ )
t
Chapter 7

where S(t-θ) denotes a delayed unit step function that starts at

t=θ.

21
Chapter 7

Figure 7.10. Comparison of step responses for a FOPTD

model (solid line) and the approximate integrator plus time
delay model (dashed line).
22
 Development of Discrete-Time
Dynamic Models
• A digital computer by its very nature deals internally with
discrete-time data or numerical values of functions at equally
spaced intervals determined by the sampling period.
Chapter 7

• Thus, discrete-time models such as difference equations are

widely used in computer control applications.
• One way a continuous-time dynamic model can be converted to
discrete-time form is by employing a finite difference
approximation.
• Consider a nonlinear differential equation,

dy ( t )
= f ( y, u ) (7-26)
dt
where y is the output variable and u is the input variable. 23
 • This equation can be numerically integrated (though with some
error) by introducing a finite difference approximation for the
derivative.
• For example, the first-order, backward difference
approximation to the derivative at t = k ∆t is

dy y ( k ) − y ( k − 1)
Chapter 7

≅ (7-27)
dt ∆t
where ∆t is the integration interval specified by the user and
y(k) denotes the value of y(t) at t = k ∆t. Substituting Eq. 7-26
into (7-27) and evaluating f (y, u) at the previous values of y and
u (i.e., y(k – 1) and u(k – 1)) gives:

y ( k ) − y ( k − 1)
≅ f ( y ( k − 1) , u ( k − 1) ) (7-28)
∆t
y ( k ) = y ( k − 1) + ∆tf ( y ( k − 1) , u ( k − 1) ) (7-29)
24
 Second-Order Difference
Equation Models
• Parameters in a discrete-time model can be estimated directly
from input-output data based on linear regression.
• This approach is an example of system identification (Ljung,
Chapter 7

1999).
• As a specific example, consider the second-order difference
equation in (7-36). It can be used to predict y(k) from data
available at time (k – 1) ∆t and (k – 2) ∆t .

y ( k ) = a1 y ( k − 1) + a2 y ( k − 2 ) + b1u ( k − 1) + b2u ( k − 2 ) (7-36)

• In developing a discrete-time model, model parameters a1, a2,

b1, and b2 are considered to be unknown.
25
 • This model can be expressed in the standard form of Eq. 7-7,
p
y = ∑β j X j + ε (7-7)
j =1

by defining:

β1 a1 , β2 a2 , β3 b1 , β4 b2
Chapter 7

X1 y ( k − 1) , X2 y ( k − 2) ,
X3 u ( k − 1) , X4 u ( k − 2)

• The parameters are estimated by minimizing a least squares

error criterion:
2

N p 
S = ∑  Yi − ∑ β j X ij  (7-8)

i =1 

j =1 
26
 Equivalently, S can be expressed as,
T
S = (Y - X β ) (Y − X β ) (7-9)

where the superscript T denotes the matrix transpose and:

Chapter 7

 Y1   β1 
 
Y =  β= 
 
Yn  β p 
 

The least squares solution of (7-9) is:

( )
−1
βˆ = X T X X TY (7-10)

27