Sei sulla pagina 1di 26

DETERMINATION OF DYNAMIC MODEL PARAMETERS USING CORRELATION TECHNIQUES FOR SMITH PREDICTOR

smail KAYAHAN

Outline

Introduction Smith Predictor Correlation Functions as System Evaluation Tool


Auto Correlation Function Cross Correlation Function

Application of Correlation Functions for Tuning The Smith Predictor Simulation and Results

Open Loop Operation Closed Loop Operation

Conclusion

Introduction

Smith Predictor *

Basically an internal model control (IMC) scheme.

* O.J.M.Smith (1958, 1959)

Smith Predictor

Assuming the process model and delay estimation perfectly match with the actual values as

So our main problem is prediction of the actual values.

Correlation Functions as System Evaluation Tool

The correlation techniques have long been used in many fields of engineering and science, especially in measurement and communication signal processing * The possible correlation or statistical dependence between two different random signals x(t) and y(t)
auto correlation function

cross correlation function

* Turin (1960), Burt et al (1982)

Auto Correlation Function as System Evaluation

Cross Correlation Function

Cross Correlation Function as System Evaluation

The cross correlation function measured between x(t) and z(t) can be proved to be
1 Rxz ( ) h( )Gxx (0) Rxn ( ) 2

where is the cross correlation function between x(t) and n(t), is the power density of the white noise spectrum, is the impulse response of the system under test. when n(t) and x(t) are uncorrelated or independent.

Application of Correlation Functions for Tuning The Smith Predictor

10

Simulation and Results

Open loop operation: The purpose of this operation is to determine system parameters before the loop is closed. This will give an opportunity to tune the model at the outset. Closed loop operation: This operation is executed while the control loop is closed. This operation is assumed to be applied with certain intervals, the period of which is to be determined by the plant dynamics and correlation execution time requirements. We have used both Matlab/Simulink pair to conduct simulations for each case. The open loop simulation employs auto and cross correlation analysis to determine and L respectively system whereas the closed loop simulation uses only the cross correlation analysis to determine both parameters.
11

Open Loop Operation

During the open loop operation plant is isolated from the control loop by opening SW2 and no distortion is assumed to be applied.

The purpose of this operation is to determine the plant parameters so that these parameters would be implemented before the loop is closed.

12

Open Loop Operation for Different Parameters


auto correlation 1.2 1 0.8 0.6 Alpha=2 Alpha=1 Alpha=0.5 Alpha=0.2 Alpha=0.1

cross correlation 0.35 0.3 0.25 0.2 Alpha=2 Alpha=1 Alpha=0.5 Alpha=0.2 Alpha=0.1

Ryy(t)

Rxy(t)
0 200 400 600 800 1000 1200 Time (msec) 1400 1600 1800

0.15 0.1

0.4 0.2

0.05
0

0
-0.2 2000 -0.05 4500

4600

4700

4800

4900 5000 5100 Time (msec)

5200

5300

5400

5500

Normalized Auto-Correlation for various pole positions


cross correlation 0.08 0.07 0.06 0.05 0.04

Normalized Cross-correlation of the system in part (a) with 5sec. of dead time
Lag=2 sec Lag=3.5 sec Lag=5 sec Lag=10 sec Lag=20 sec

Rxy(t)

0.03 0.02 0.01 0 -0.01 -0.02 0 0.5 1 1.5 2 Time (msec) 2.5 3 3.5 x 10 4
4

13

Cross-correlation of the system with different dead time while the plant pole is at -1

Estimation

This equation requires to be studied further to extract 0 corresponding to 1/. We employed exponential fitting function for the estimation of .
0.03 r vs. lags1 fit 0.025 0.02

Rxy

0.015

0.01

0.005

0 -4000

-3000

-2000

-1000

0 Lags

1000

2000

3000

4000

14

Auto Correlation With Disturbance


auto correlation 1 0.9 0.8 0.7 auto correlation 1.2 1 0.8

0.6

Ryy(t)

0.6 0.5

Ryy(t)
0.4 0.2 0 -0.2 0 200 400 600 800 1000 1200 Time (msec) 1400 1600 1800 2000 0

0.4 0.3 0.2 0.2 0.4 0.6 0.8 1 1.2 Time (msec) 1.4 1.6 1.8 x 10 2
4

Distorbance=0.5+0.5*sin(0.2*pi*t); 0.1 Hz , alpha=0.1

Distorbance=N(0,1); 0.1 Hz, alpha=0.1

As it is seen, if distorbance is a sin or periodic function, it can be estimated. However for a random signal because its frequency is close to system cutoff frequency it cannot be estimated exactly.
15

Statistical Evaluation of Open Loop Parameter Estimation

The measurement results for including standard measurement error are given in Table where the auto correlation functions are executed for a period of 180 seconds. Among other error qualities MAPE gives

2 1 0,5 0,333333 0,25 0,2 0,166667 0,142857 0,125 0,111111 0,1 Measured Mean

Pole
0,5 1 2 3 4 5 6 7 8 9 10

(*)
2,023166521 1,021057626 0,516514292 0,339627950 0,245050989 0,203529825 0,166562195 0,143397067 0,123776435 0,111430369 0,099436211

STD 0,274634713 0,084854783 0,039270480 0,016333760 0,011893074 0,007283602 0,007122064 0,009925779 0,003673817 0,003356996 0,002150587

MAD 0,223583 0,076095 0,031721 0,014269 0,010532 0,006305 0,005815 0,008182 0,003101 0,002786 0,001895

MAPE 11,17914 7,609496 6,344241 4,280784 4,212749 3,152326 3,488986 5,727450 2,480563 2,507724 1,894991

RMS 0,261569 0,083209 0,040751 0,016725 0,01232 0,007759 0,006757 0,009432 0,003694 0,003201 0,002117

more significant measure of error. It is clear that when gets larger, CCF for the measurement of exhibits greater error. This is simply because of the decreasing bandwidth of the correlated signal and consequently less data to be correlated. Therefore, when the system is too slow, dither signal source bandwidth may be reduced for more accurate estimation.
2.5 2

Measured Alpha

1.5

(*) Mean of 10 regression each 180 sec STD: Standard Deviation MAD: Mean Absolute Deviation MAPE: Mean Absolute Percentage Error RMS: Root Mean Square

0.5

Mean-Standard Deviation Mean+Standard Deviation Mean 0 0.2 0.4 0.6 0.8 1 Alpha 1.2 1.4 1.6 1.8 2

16

Estimation Dead Time

The cross correlation function requires, a peak finding algorithm to determine the dead time in case of a flat cross correlation functions which may occur in practical system particularly in the presence of disturbance and noise.
cross correlation 0.08 0.07 0.06 0.05 0.04 Lag=2 sec Lag=3.5 sec Lag=5 sec Lag=10 sec Lag=20 sec

Rxy(t)

0.03 0.02 0.01 0 -0.01 -0.02 0 0.5 1 1.5 2 Time (msec) 2.5 3 3.5 x 10 4
4

17

Closed Loop Operation

The closed loop simulation assumes that SW2 is kept closed as SW1 switch is closed with certain periodic intervals

Here the auto correlation analysis cannot be used for the determination of plant pole as it was in the open loop operation since the auto correlation of plant output contains the total closed loop transfer function.
18

Simulink Diagram of Close Loop Operation


0.5 0.4 0.3 0.2 0.1 0
10 x 10
-3

1 kHz

-0.1 -0.2 -0.3 -0.4 -0.5

0.01 Hz
0 10 20 30 40 50 60 70 80 90 100 110 120 130 140 150 160 170 180 190 200 210 220 230 240 250 260 270 280 290 300 time (sec)

-2

-4

20

40

60

80

100

120

PD Controller
19

Estimated Transfer Function Block

Estimated Delay Block

Closed Loop Operation for Different Parameters


cross correlation 0.04 0.03 0.02 0.01
Rxy(t)

cross correlation 0.1


Lag=2 sec Lag=3.5 sec Lag=5 sec Lag=10 sec Lag=20 sec

0.08

Alpha=2 Alpha=1 Alpha=0.5 Alpha=0.2 Alpha=0.1

0.06
Rxy(t)
0 0.5 1 1.5 2 2.5 Time (msec) 3 3.5 x 10 4
4

0 -0.01

0.04

0.02
-0.02 -0.03 -0.04

-0.02 1500

2000

Closed Loop cross correlation functions for Closed Loop cross correlation functions for various values various dead times

2500 3000 Time (msec)

3500

4000

20

Estimation

While actual pole and measured pole determining the plant pole (1/), an exponential fitting algorithm is utilized.
0.03 r vs. lags1 fit 0.025

0.02

Rxy

0.015

0.01

0.005

0 -4000

-3000

-2000

-1000

0 Lags

1000

2000

3000

4000

21

Statistical Evaluation of Closed Loop Parameter Estimation


Pole () 0,5 1 2 3 4 5 6 7 8 9 10 2 1 0,5 0,333333 0,25 0,2 0,166667 0,142857 0,125 0,111111 0,1 Measured Mean (*) 1,96877203 0,9960724 0,49988289 0,32966703 0,25055228 0,19697561 0,16911044 0,14530929 0,12201161 0,10991094 0,10252137 STD 0,703677 0,158167 0,070788 0,037863 0,019768 0,017549 0,013155 0,011583 0,008302 0,007415 0,005119 MAD 0,585444 0,141428 0,053084 0,028759 0,016251 0,011685 0,010008 0,008457 0,006623 0,006378 0,003964 MAPE 29,2722 14,14279 10,61674 8,62765 6,500421 5,842496 6,004929 5,919862 5,298692 5,740305 3,964042 RMS 0,668296 0,150102 0,067155 0,036107 0,018761 0,016921 0,012717 0,011259 0,008424 0,007136 0,005472

Calculated plant pole and associated errors for closed loop This a natural effect of a bandwidth limitation, the smaller the bandwidth the less data to be correlated and consequently flatter cross correlation function and more ambiguity for determining rate of decay.

22

(*) Mean of 10 regression each 180 sec

Estimation Dead Time


cross correlation 0.04 0.03 0.02 0.01
Rxy(t)

Lag=2 sec Lag=3.5 sec Lag=5 sec Lag=10 sec Lag=20 sec

0 -0.01 -0.02 -0.03 -0.04

0.5

1.5

2 2.5 Time (msec)

3.5 x 10

4
4

Closed Loop cross correlation functions for various dead times

Dead Measured Time Dead Time 1 1.002 2 2.002 3 3.002 4 4.002 5 5.002 6 6.002 7 7.001 8 8.002 9 9.001 10 10.002 Measured dead time for closed loop

23

Step response of closed loop


Exact Alpha=2, Measured Alpha=1.878024516 Response with PD Controller 1

0.8

0.6

0.4 Exact Alpha Response Measured Alpha Response 0 2 4 6 8 10 Time (sec) 12 14 16

0.2

the response of the closed loop to step input when actual =2, as the worst mismatch is observed. We read a 5% overshoot while no overshoot is observed in the fully matched system.

24

Discussion and Conclusion

One of the areas for furthering this work is to develop correlation analysis for the second and higher order systems. Another future study is to design an all-in-one unit to be connected to PLC or DCS system as a standard I/O unit housing correlator and plant models so that the Smith predictor compensation can be realized.

25

26

Potrebbero piacerti anche