Download as doc, pdf, or txt
Download as doc, pdf, or txt
You are on page 1of 5

Process Identification from Overdamped Closed-Loop Response with

Proportional controller using Optimization

Manal Al-Saif

1. Introduction

The dynamic relationships between variables in a process can be determined from


mathematical models of the process. Mathematical equations, based on fundamental physical and
chemical laws, were developed to describe the time-dependant behavior of systems. It is
assumed that the values of all parameters, such as holdups, heat transfer coefficients, reaction
rates, etc., were known. Thus the dynamic behavior was predicted on essentially a theoretical
basis.

For a process that is already in operation, there is an alternative approach that is


essentially empirical. The dynamic behavior can be found from experimental tests. The
experimental approach is sometimes used when the process is though to be complex to model
from the first principles. It is also used to find the values of some parameters that are unknown.
Many of the parameters can be found from steady-state plant data, but some parameters must be
found from dynamic tests.

Experimental identification of process dynamics has been an active area of research in


recent years by many workers in many areas of engineering. The most used models for the
systems are the second order plus dead time (SOPDT). The parameters of the model can be
obtained by performing open-loop or closed-loop test on the process. The open-loop test
requires the process to be perturbed in the manual mode by introducing a change in the process
inputs and then using the resulting process response to obtain the model parameters. This is
undesirable since no control action is taken against disturbances which may occur during the test;
thus closed-loop operation is preferred. This study is based on a similar work done by
Viswanathan and Rangaiah (2000) to identify systems in the form of SOPDT. In that work,
model parameters were found using three optimization techniques, Luus and Jaakola to first find
good estimates of parameters then Nelder and Mead simplex direct search and finally Broyden-
Fletcher-Goldfrad-Shanno method to revover accurate model parametrs. They applied the
identification on systems with PI and PID controller and with overdamped and underdamped
closed-loop responses.

The aim of this work is to present a more simplified method to identify system from the
closed loop response with the case of proportional controller and overdamped response.

2. Process identification by optimization


Consider a simplified feedback loop shown in Figure 1.
R E M C
Gc GP

Figure 1. A simplified feedback control system.


In the step test, a step change in the set point to the closed-loop system under P control, is
introduced. If the controller is already in operation, then its current settings can be used. The
resulting actual closed-loop step response (Ca) along with the step change in step point and
controller setting, are the bases for identification. The dynamics of the process is represented by
a SOPDT model of the form:

this is a good approximation for higher order processes, and it can model underdamped
responses. If the dead time is approximated by Pade’ approximation the process transfer
function will be:

using a proportional controller, the controller transfer function is:

The overall closed loop transfer function is:

From the last equation it is clear that Cc is a function of controller settings as well as the model
parameters. Since the controller setting are assumed to be known, the closed-loop response is a
function of model parameters only.

The goal in process identification by optimization is to determine the best values of


SOPDT model parameters that minimize the sum of squares of errors between the actual
response and the calculated close-loop response:

Here, Ca(ti) and Cc(ti) are respectively the actual and calculated close-loop responses at time ti
(equal to an integer multiple of the sampling period). (In this study, for evaluating the
optimization approach, the actual response is often simulated using the process transfer function,
which may be different from SOPDT model assumed). N is the number of sampling periods
contained in the test duration Tf. the objective function selected is the sum of squares, which is
popular in regression.
In the process identification by optimization, Cc is obtained by simulating the closed-loop
response (eq. 4) with initial values of model parameters for the step change in the set point used
in the test, and then SSQ is calculated. A suitable optimization method is employed to obtain new
valued for the parameters and the closed-loop response is now simulated with these new
parameter values in the model. This is done iteratively until SSQ is minimized.

3. Method for minimization

the objective function (eq.5) for SSQ minimization is non-linear and can be expected to
be non-convex. Extensive simulation tests are necessary to verify the non-convexity and hence
existence of multiple minima. Conventional optimization techniques which satisfy the Kuhan-
Tucher optimality conditions may not find the global minimum and their success depends on the
initial estimates provided.

In this study the initial estimates of the parameters were found by simple approximation to first
order pulse dead time (FOPDT) response if the response is overdamped; the method is outlined
blew.

From eq. 4 the closed loop gain Kov is:

Hence, from the actual closed-loop response and the known value of controller gain, the process
gain can be found. For many cases, the response does not reach to its ultimate value of steady
state, so only the final value is used to get the estimate of Km.

The time constant τm is found from the FOPDT approximation and that is the time corresponding
to 63.2% of the ultimate change. The dead time θm can be easily estimated from the response,
and the damping factor ζm is usually near one so using a value of one is a good estimate. Figure 1
shows the parameters estimate from actual response for a unit step change in set point.

Ca

KcKm
(1+KcKm)

63.2%

θm τm Time
Figure 1: SOPDT estimate parameters from actual response with P controller
Estimates for the model parameters are then used as initial estimates for an optimization
technique to find the accurate parameters. The optimization in this work is done using the Nelder
and Mead simplex direct search method. All computations were carried out using MATLAB and
SIMULINK. The actual response was done with SIMULINK and its output was transferred to an
M-file in MATLAB that calles the Nelder and Mead method from the optimization toolbox. All
computations were performed on Pentium III 800 MHz and some on Pentium I 200 MHz to
compare with the work of Viswanathan and Rangaiah (2000).

Results and Discussion .3

the optimization approach for process identification was tested on four examples which
were earlier used by Viswanathan and Rangaiah (2000) :

Example 1, a true second order overdamped process, was chosen to determine the ability of the
approach to recover the exact model parameters. The second example is a higher order process
with a large dead time, selected to check the robustness of the method in the presence of dead
time. Example 3 is the same as the second example but with a lead term added. The last example
a higher order process with same multiple poles and zero dead time, which will test the
efficiency of the method in model reduction.

4. Process Identification from step response

Closed-loop test was simulated by considering each of the test examples as the process
and response to a step change in set point was obtained. For each example, a P controller was
connected with more than one setting. All responses were overdamped. The estimates were
found as describes above, then the Nelder and Mead method was applied. Table 1 shows the
results.

Estimates Optimization Result


Kc Tf Km,τm, Km τm θm ζm SSQ CPU time(s)
θm,ζm Pen.III(Pen.I)
Example 1
1.000,6.29,
0.4 28 1,1
0.9997 3.4812 0.9972 1.1491 6.2988E-6 4.917(44.21)
0.960,5.45,
0.4 15 0.9,1
0.9954 3.4806 0.9997 1.1399 9.3077E-6 5.488
1.009,5.182,
0.6 15 0.92,1
0.9955 3.4786 1.0017 1.1409 2.1247 E-5 5.007
Example 2
0.999,4.5,
0.01 20 3,1
0.9736 2.4863 3.1054 0.7248 3.2566 E-6 6.92
0.898,3.588,
0.01 10 3,1
0.3302 2.3329 2.5018 0.1446 1.7804 E-6 8.802
Example 3
0.999,3.25,
0.02 20 3,1
0.9815 2.2728 2.8205 0.7650 1.0948 E-5 4.386
0.999,3.875,
0.01 15 3,1
0.9297 2.3998 2.7098 0.6509 2.9231 E-7 5.367
Example 4
1.002,4.7,
0.05 15 0.63,1
0.9766 2.2105 1.3362 0.7738 9.3761 E-6 5.718
0.974,4.75,
0.02 10 0.625,1
0.8428 2.3938 1.1365 0.5300 9.4167 5.208(40.59)

From the results table it is clear that the optimization has converged quickly with small
error. Similar results are found in the studies of Viswanathan and Rangaiah (2000) but with high
CPU time especially for the first optimization method that gets the first estimates of the model
parameters; the computation time was in the order of 10 minutes (on Pentium I) with the first
estimates optimization taking around 80% of the computation time. In our approach, the
simplified method for parameters estimate took a small time, only simple calculations on actual
response.

5. Conclusion

Application of optimization for identification of SOPDT model from closed-loop


responses is proposed. It employs a sum square criteria for estimating model parameters. The
approach is tested on several examples with different test durations and step input. The
advantage of this approach is that it is a simplifies approach to get the estimate of the model
parameters that are used as initial estimates of the optimization model.

6. References

Viswanathan, P.K. and Rangaiah, G. P., “Process Identification from Closed-Loop Response
using Optimization Methods”, IChemE, Vol 78, Part A, May 2000.

You might also like