Industrial Control Chapter 4

------------------------- ---------------

Chapter 4

EMPIRICAL PROCESS MODELING

The mathematical (or mechanistic) modeling approaches introduced in the previous

chapters leads to one of the most explicit and useful set of models to describe plant or
process behaviour. Nevertheless, such approaches are difficult and very time consuming.

Therefore various empirical modelling approaches are oftenly used in practice. These
methods could be very easy, fast, and accurate, though they give a black box
representation of the input-output behaviour.

Selecting a structure for the model and then finding the best parameters of that structure to
fit the input-output data we have already collected from real experiments or simulations is
called empirical modeling.

In the following, a few examples of such metods are introduced. More useful tools for
developing such methods through collected input-output data from the plant based on
available software are also introduced in other parts of this course.

4.1 FIRST ORDER PLUS TIME DELAY MODELS

This method is based on the step response of the process. The step response of most
industrial processes has the general S-shaped curve shown in Figure 5.7 which is also
called the process reaction curve and can be generated experimentally or from a
dynamic simulation of the plant.

Figure 4.1. Process reaction curve (i.e a simple step response) .

The shape of this curve is characteristic of high order systems, and the plant input-
output behavior may be approximated by

Y ( s) K e − t d s
G ( s) = = ,
U ( s ) τs + 1

1
 Industrial Control Chapter 4
------------------------- ---------------

which is simply a first order system plus a transportation delay (i.e. or lag. It should be
noted that time delay in s-domain is equal to time shift in time-domain).

The constants parameters of the above equation can be determined from the unit step
response of the process shown above.

If a tangent is drawn from the inflection point of the reaction curve, then the time
constantτ, and the time delay td can be approximated as shown on the Figure.

Though the above equation often provides an adequate model for the plant, if the
actual plant output does not fit this simple model, other structures may be selected as
is described below.

4.2 HIGHER ORDER MODELS

Using the step response data, the following approach can be invoked to develop models
without any pre-assumption on the order of the best model to fit data in a hierarchical
way.

We may assume that the transient unit step response of the plant is a combination of some
elementary transient responses as

y (t ) = y (∞) + Ae −αt + Be − βt + Ce −γt + L (4.1)

Figure 4.2 depicts a typical step response of this system.

Figure 4.2. A typical step response.

Subtracting off the final value y(∞) and assuming -α as smallest (the slowest) root, we
may write
y (t ) − y (∞) = Ae −αt + L
log10 [ y (t ) − y (∞)] ≅ log10 A − αt log10 e (4.2)
≅ log10 A − 0.4343αt
If [y(t)-y(∞)] is negative then we multiply both sides by -1 and we will have

log10 [ y (∞) − y (t )] ≅ log10 (− A) − 0.4343αt

This is the equation of a line.

2
 Industrial Control Chapter 4
------------------------- ---------------

If we fit a line to the plot of log10 [y(∞)-y(t)], or log10 [y(t)-y(∞)], then we can estimate A
and α . Once these two were estimated we follow the same procedure except that we have

log10 [ y (∞) − y (t ) + Ae −αt ] ≅ log10 (− Be − βt ) , (4.3)

to estimate B and β. This procedure will be continued until a model with appropriate order
is obtained.

As an example this method is appled to the data which has been collected in Table 4.1.

Table 4.1. A set of step response data.

Following the procedure described above and using the data given in Table 4.1 we will
come up with the graph of Figure 4.3.

Figure 4.3. Plot of step response data based on the procedure explained.

3
 Industrial Control Chapter 4
------------------------- ---------------

Now from the line fitted by eye on this figure the values are

log10 (− A) = 0.125 ⇒ A = −1.33

1.602 − 1.167 0.435
0.4343α = = ⇒ α =1
∆t 1
Now, if we subtract this line from the previous log plot (following Equation (4.3)), we
will obtain Figure 4.4.

Figure 4.4. Plot of log10 [ y (∞) − y (t ) + Ae −αt ] for the step response data.

From Figure 4.4 we estimate B and β,

log10 (− B) = −.45 ⇒ B = −0.35
− 0.45 − (−1.7)
0.4343 β = = 0.25 ⇒ β = 5 .8
0 .5
Combining these results, we estimate output
^
y (t ) = 1 − 1.33e −t + 0.35e −5.8t

Where its corresponding Laplace form is

^ 1 1.33 0.35 0.02s 2 − 0.56s + 5.8 1
Y ( s) = − + = ⋅
s s + 1 s + 5.8 ( s + 1)( s + 5.8) s
− 0.56s + 5.8
⇒ G ( s) ≈
( s + 1)( s + 5.8)

If this model is not accurate enough, we may continue to obtain higher orders of this
model. But it does not seem necessary here.

Experimetal modelling approaches presented in this section and the previous section are
only two simple examples of experimental modelling. However, since you will be
familiar with MATLAB and other modelling and control software, you will have access to
various modelling tools and there is no need to learn hand-driven methods anymore.