Simulation Model Emulation in Control System Designmams.rmit.edu.au/rs0z8nwgz8ou.pdf · Simulation Model Emulation in Control System Design ... control systems but the control techniques

Simulation Model Emulation in ControlSystem Design

A Power Plant ApplicationC. Lu , N.W. Rees and P.C. Young

[email protected], [email protected]

School of EET, UNSW, Australia. CRESS, Lancaster University, UK

Workshop-Young RMIT Melbourne Dec 2009 – p.

Quotation - Profos 1959

The only remedy is increased teamwork: The controltheorist should make it his obligation to discuss theassumptions on which he is going to base hiscalculation with a experienced control or power plantengineer before erecting a great theoretical buildingupon an unfit foundation, whereas the practicalengineer should take time and pains for suchdiscussions. This would enhance the practical value ofcorresponding studies and promote the mutualunderstanding between theoreticians and practician.And this is certainly one of the most importantconditions that control theory will be of practical help.


Power plant control is very important

• Modern power stations have excellent computercontrol systems but the control techniques usedare still mainly SISO/PID

• Attempts have been made at MIMO control(especially LQR) but they have not caught on

• This study proposes a new advanced controlscheme that seems to have potential for large andfast rate load changes in wide range of power plant


Power Plant Layout


Control problems

• Power plants need to produce MW as efficiently aspossible without violating temperature, pressureand water level constraints.

• Plant is highly nonlinear with strong interactions,between MW, steam pressure P, and water levelDL.

• Schematic shows the three major inputs andoutputs.


Major inputs/outputs

mill trans

u1 - MWsp

u3 - Psp

u2 - DLsp

pump trans

freq sync

y1 - MW

y3 - TSVP

y2 - DL G(s)

(highly coupled)


Step input interactions

0

0.5

1

1.5From: In(1)

To: O

ut(

1)

-0.5

0

0.5

1

1.5

2

To: O

ut(

2)

0 500 1000 1500 2000 2500-1

-0.5

0

0.5

1

1.5

2

To: O

ut(

3)

From: In(2)

0 500 1000 1500 2000 2500

From: In(3)

0 500 1000 1500 2000 2500

Step Response

Time (sec)

Am

plit

ud

e


Control problems

• Stiff system - 10 time’s difference in time delaysbetween the major variables

• The existing control scheme cannot meet themarket demand on fast load change to powerindustry, causing instability even trips to the plant

• One of the key control issue is the drum boiler,where the steam is generated. Water level shrinkand swell caused by load change.


Drum boiler

Natrual/Forced Circulation

DOWNCOMER

RISER

Q - heat input

DRUMVALVE

Wf - feedwater flow

L - drum water level Steam

Saturated Mix

Pd - drum pressure

Ws - steam flow

TVSP - steam pressure


Unit Controller

UNITCONTROLLER

Load/Pressure

Control

Combustion

Control

Drum Level

Control

Temperature

Control

SC SCSCSC SC SCSCSC SCSCSCSC

UC

PRC

Workshop-Young RMIT Melbourne Dec 2009 – p. 10

Control loops of power plant

fueldemand

Throttle Valve & Superheater

Turbine /Generator

tf

30 s

15 s +1

TSVP FB

Sensor

TSVP

Sum 9

Sum 5

Sum 10

Yt

Pd

Steam Flow

TSVP

Steam Flow FB

Sensor

Sliding Pressure

Sliding Pressure

PressureController

PID

Mannual TSVP setpoint

MW output

MW manual setpoint

MW manual

MW demand with rate limiter

MW demand

MW controller

PID

MW FB

Sensor

Level Controller

PID

Kyt

1

Kup

1

Kpff 11

Kp FF

−K−

Kmw FF

−K−

Kfqs

1

qs MW Out

Fuel System

FeedwaterPump

FeedwaterController

PID

FW Flow FB

Sensor

Drum −Boiler

qs

qf

tf

Q

Drum L

Drum P

Drum Level setpoint

L_init

DL FB

Sensor

DL

2

Trans

1

Q_init

qs _init


UNSW Simulator

DRUMBOILER

MILLSOR

COMPRESSOR

DEAERATOR GENERATOR

FURNACE

FORCED DRAUGHT

FAN

CONDENSERCONDENSATE

PUMP

FEEDWATER FEED PUMP

AIR HEATER

THROTTLEVALVE

CONTROLLER

UNITCOORDINATOR

TURBINE & REHEATER

SUPERHEATERTEMPERATURECONTROLLER

Superheater

SUPERHEATER& DESUPERHEATER

RESERVETANK

DEAERATORLEVEL

CONTROLLER

MANUALSETUP

CONDENSATE PUMP

CONTROLLER

FEEDPUMPCONTROLLER

MW OUTPUT

DRUM LEVELCONTROLLER

Governor Valve

TV

FAN SPEEDCONTROLLER

FUELCONTROLLER

FEEDER

(LP) CLOSED FEEDWATER

HEATER

ECONOMISER

(HP) CLOSED FEEDWATER

HEATER

22

22

2

6

62

22

2

2

2

2

2

2

2

3

2

2

2

22

2

8

8

2

2

2

2

2

2

2

2

2

2

2

2

2

2

2

2

22

2

2

2

2

2

3

32

2


UNSW Simulator

• A nonlinear UNSW Simulator is established tosimulate the complex nonlinear power plant.

• A mixture of Knowledge based and Interpretationmodels are created in Matlab/Simulink.

• The extension of key section model - Åström andBell nonlinear drum boiler model

• together with other models for all key power plantsections makes the simulator truly a complex,physically based simulation system.


UNSW simulator

• The simulator also includes the existing plantcontrol systems (PID loops) for MW, pressure andlevel

• plus the feedforward of unit demand, slidingpressure and overfiring/underfiring signals.

• All major actuator nonlinearity and constrains areimplemented

• With all above features UNSW simulator can beconfigured and tuned against a real plant data.

• The simulator needs to be benchmark tunedagainst real plant data, before it can be used forcontrol study.


Simulator benchmark tuning

• Set-up real plant operational parameters - MW,Pressure, etc.

• Tune individual PID control to match the plant keysections’ dynamic benchmark test data.

• Internal saturations (such as PID control outputs)need to be removed under small (10 per cent)perturbation and normal operation ramp inputs.

• Validate the simulator against plant inputs outputsdata.

• One this is done the simulator represent thecomplex nonlinear power plant.


Benchmark data - Drum level


Benchmark Model response

2800 2900 3000 3100 3200 3300 3400 3500 3600 3700 3800

0.52

0.54

0.56

0.58

0.6

0.62

0.64water level


Advanced control implementation

• Add-on structure (keep conventional control inside)• Trim actions to unit control setpoints

Y

U_adv

G(s)

U_con

YUnit Demand

SwitchSum 2Sum1

3

POWER PLANT PROCESS

2

ADVANCED CONTROL

1 − Fall back0 − Adv control

1

1

CONVENTIONAL CONTROL


Advanced control implementation

uu

setpoints

outputs Y

mux2

Unit Coordinater

Boiler Demand

MW Demand

Sliding P SP

To Workspace2

uu

To Workspace1

sfp

To Workspace

ycl

Switch

S & Y

OutputScaling

In1

In2

In3

Out1

Non−linear Power Plant Subsystemwith PID control loops

mw demand

Sliding Pressure

boiler demand

In1

In2

In3

MW Output

Drum level

Pressure at TV

Manual Operation

Out1 Input Scaling

In1

Out1

Out2

Out3

Fallback

FBK

4.587e−005

0.0007938

0.09794

Constant

0Bypass

MW BP

TSVP BP

MWd

TSVPd

Demand

Advanced Controller

Set

FB

Out3

33

3

3

3

3

3

33

3

3

3

3

3

3

33

6

3


Control design - Why Captain

• It has paid special attention to the TransferFunction (TF) models using robust unbiasedRefined Instrumental Variable (RIV) and SimplifiedRefined Instrumental Variable (SRIV) algorithms

• It can establish both DT and CT well fit TF models.• It can identify the TF models in a multiple-input,

single output (MISO) manner.• It handles multivariable NMSS formation and

control design in a streamlined way.• It has multivariable PIP control implementation.


Control design - Why Captain

• It has a specialised function rivid which allows theautomatic search for a group of the best fit modelsover a user defined range of different modelstructures and time delays which makes thedetermination of the TF models very efficient.

• Its model fitness statistical criteria includes theCoefficient of Determination (R2

T ), AkaikeInformation Criterion (AIC) and especially theYoung Information Criterion (YIC) which give veryclear indication of how well the model describesthe data.


Dominant Mode Analysis(DMA)

• The response of high order linear dynamic modelsis always dominated by a small number of modes.

• If these modes can be detected, then they formthe basis of an accurate reduced order model.

• Through Dominant Mode Analysis (DMA) [Young,1999] such a dominant mode model can beobtained from the real data, as a reduced order’emulation’ of the high order simulation model.

• Often such dominant mode model can mimic thelarger perturbation responses.



.

.

.

.

.

.

.

.

.

.

.

High Dynamic

Order Model

Estimated

Parameters

Reduced

Dynamic

Order Model

Parameters

High Order

Model:(e.g. 17th order and

circa 180 parameters)

Reduced Order

Model:(e.g. 3rd order and

27 parameters)

Parameterized

State-Dependent

Parameter

Regression

(SDR)

Relationships



• The resulting nominal emulation model canproduce continuous and discrete SISO and MISOtransfer function models.

• Further CAPTAIN functions are used to formNMSS models which can be used to designfeedback controllers.

• In particular the multivariable LQ-PIP controls canbe formed.


Control design

• Define 3 x 3 input/output discrete-time (DT)transfer function (TF) model, a linear model of theabove nonlinear process, as the suitable model forthis application

• Refined Instrumental Variable (RIV) algorithmsfrom Captain is used for the identification ofnominal TF models.


Identification of nominal TF models

• The process excitation and data logging, in orderto be able to apply to a real plant, is in SIMO form,as it is impossible to perturb 3 major power plantsetpoints simultaneously.

• The model is then established in MISO way, afterall data collected, provided that the plant’soperating points are steady and not time varyingduring the logging.



• Two different sets of models are necessary in ourcontrol design procedure. They are design modelDM and process model PM.

• The idea is that the DM used for control designcannot be used to verify the controller (idealresult).

• Yet at design stage, it is difficult to tune control onan nonlinear simulator, or the real process. So PMis used to verify control result first before furthertests on a nonlinear model.



• It is desirable that DM is as simple as possible, yetit is necessary to capture the major processdynamics with correct phase.

• With quantitative analysis procedure DM and PMdesign method gives the designer much betterchances to identify and judge the ’right’ DM modelefficiently.

• DM’s model structure requires commondenominator, while PM does not have suchrestrictions.



• DM - as low order as possible with commondenominator TFs in a row of TF matrix. [2 3 1] and[3 3 1] models have been established as the lowerorder models - for control design.

• PM - no order or common denominator restrictionson TF. [4 4 1] model is the highest order structurefits the process very well - for control tuning.

• Through DMA two reduced order linear dynamicmodels are estimated. (Both have lower order thanthat of nonlinear simulator)


NMSS and PIP control design

• NMSS presentation of MIMO TF model for controldesign is extremely relevant to industrialapplications, for its explicit use of only measurablevariables and their past values that are availablefrom DCS system.

• The non-minimal states as past measurable valuesin the state space matrix makes the resultingcontrol with inherent model predictive controlaction, that is significant feature for problems withlong time delay, such as power plant control.


NMSS and PIP control design

• In DT LQR design, system output vector Y in thecost function has horizontal history values from theDM model (dynamics of the process) it isinteresting to note the non-adaptive modelpredictive action of the resulting state feedbackcontrol law.

• PIP control implementation, with introducedintegral of error in NMSS, is a multivariablecontroller with extra integral action to eliminate thestatic error in the system. It is important feature forreal control applications.


PIP control

The block diagram for such a PIP control system isshown. The negative sign associated with isintroduced to allow the integral states to take on thesame structural form as multivariable PI/PID. In thismanner, the PIP controller can be interpreted as alogical extension of conventional PI and PIDcontrollers, with additional dynamic feedback andinput compensator introduced automatically when theprocess has second order or higher dynamics or morethan a single sample pure time delay.


PIP forward path control

I u(k) y(k)

M(z)+I

L(z)

+-

y (k)d

Process+-

S(z )^ -1

z(k)K(I)


Design flowchart

System

Identification

Data Collected

from Power Plant

2 sets of

models

DM

Design Model

PM

Process Model

END

PIP-LQR Design

Control Simulation

Using DM

Captain Tool Box

To establish models

PM – The most fit model

with/without common

denominators

DM – Lower order model

with common denominators

Desired control

results?NO

Choose

Weights

Non-linear System

Simulation and

Real Plant Test

YES

Implement control structure

with LQR gains.

Either FB or FP structure

Desired control

results?

DCS

Implementation

YES

NO

Program DCS with

Controller Structure

Benchmark Tuned

Non-linear Simulator

PIP-LQR Design

– Young, et al

Better Control -

Advanced controller manipulates all 3 set

points to achieve much reduced DL and P

errors while the MW is least disturbed

Control Simulation

Using PM

NO

Real Issues –

A. Add-on structure

B. Scenario tests

C. Disturbance tests

D. Starting up ...


Advanced controller tuning

• Follow the design procedure DM, PM, andnonlinear simulator are utilized to carry design,initial control test and nonlinear control test

• LQ-PIP forward path (FP) control structure is foundto be able to stabilise the process and to improvethe performance.

• While standard feedback structure (FB) givesidentical results when DM is used as process, butfailed the test when PM or nonlinear models isused.


Advanced controller tuning

• Extensive tuning is carried out at each stage of thedesign and test, through diagonal weight vectorswy, wu and wz. NMSS representation gives a senseof physical meaning on weights but the link to thecontrol performance is still not direct due to LQ’snature.

• A numerical procedure with quantitative errorindicators is created to tune the weights one at atime, which makes the tuning process moreefficient.


Control results

• This multivariable linear control solution, whenapplied to a nonlinear process, cannot totallydecouple the interactions, rather it is a optimalmanipulation of the interaction to reduce theinteraction so the errors can be significantlyreduced.

• Control results show very good error deduction,especially drum water level is well within the alarmlines. Shrink and swell has be reduced more than5 times. Pressure delay has been much reduced.


Tuning results - 231DM

• The weights are tuned directly in favour of MW andDL, so we have good tracking performance,

• P is tuned to its derivation because pressurederivation has the immediate effect on water level,as shown in both nonlinear and linear models.

• The resulting steam pressure is moving aroundsetpoint during the transient. In fact in practice thepressure is never tuned too tight.


Tuning results - 231DM

• Tuning is progressive - from on PM and on NS(nonlinear simulator)

• Results shown are from the tuning on NS.

wy =

2900

1

500

; wu =

125

1

1

; wz =

16

1

1

;


Tuning results MW step

Blue - open loop, Red-dotted - unity weights, Black - weights tuned

200 400 600 800 1000 1200 1400 1600 1800 2000-0.05

0

0.05

0.1

0.15

200 400 600 800 1000 1200 1400 1600 1800 2000-0.05

0

0.05

0.1

0.15

200 400 600 800 1000 1200 1400 1600 1800 2000-0.05

0

0.05

0.1

0.15

MW-unityMW -tunedMW-PID

DL-unityDL -tunedDL-PID

TSVP-unityTSVP-tunedTSVP-PID


Tuning results MW ramp


200 400 600 800 1000 1200 1400 1600 1800 2000-0.05

0

0.05

0.1

0.15

200 400 600 800 1000 1200 1400 1600 1800 2000-0.05

0

0.05

0.1

0.15

200 400 600 800 1000 1200 1400 1600 1800 2000-0.05

0

0.05

0.1

0.15





Tuning results TSVP step


200 400 600 800 1000 1200 1400 1600 1800 2000-0.05

0

0.05

0.1

0.15

200 400 600 800 1000 1200 1400 1600 1800 2000-0.05

0

0.05

0.1

0.15

200 400 600 800 1000 1200 1400 1600 1800 2000-0.05

0

0.05

0.1

0.15





Tuning results TSVP ramp


200 400 600 800 1000 1200 1400 1600 1800 2000-0.05

0

0.05

0.1

0.15

200 400 600 800 1000 1200 1400 1600 1800 2000-0.05

0

0.05

0.1

0.15

200 400 600 800 1000 1200 1400 1600 1800 2000-0.05

0

0.05

0.1

0.15





Original control performance

y1 MW - Red, y2 Drum level - Green, y3 TSVP - Blue

2000 3000 4000 5000 6000 7000 8000 9000 10000 11000 12000-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8


Closed loop wide range load - 231DM


2000 3000 4000 5000 6000 7000 8000 9000 10000 11000 12000-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8


Control signals - 231DM

u1 for MW - light blue, u2 for Drum level - magenta, u3 for TSVP - yellow

2000 3000 4000 5000 6000 7000 8000 9000 10000 11000 12000-0.25

-0.2

-0.15

-0.1

-0.05

0

0.05

0.1

0.15

0.2

0.25


Closed loop wide range load - 331DM


2000 3000 4000 5000 6000 7000 8000 9000 10000 11000 12000-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8


Control signals - 331DM

u1 for MW - light blue, u2 for Drum level - magenta, u3 for TSVP - yellow

2000 3000 4000 5000 6000 7000 8000 9000 10000 11000 12000-0.25

-0.2

-0.15

-0.1

-0.05

0

0.05

0.1

0.15

0.2

0.25


Discussion and Future Works

• Multiple range control designs• Nominal TF models from the data collected at

different load range , design and tune controlleragainst each model

• Test each controller on wide range operation• Compare the performances of each controller

over wide range load condition• Use the best performed control or if none of

them can cover the full load range of powerplant - gain scheduling?



• Multiple range control designs• Nominal TF models from the data collected at

different load range , design and tune controlleragainst each model

• Test each controller on wide range operation• Compare the performances of each controller

over wide range load condition• Use the best performed control or if none of

them can cover the full load range of powerplant - gain scheduling?

• Control sensitivity test - Varying plant parameters(Monte Carlo Simulation)



• Develop a complete nonlinear State-DependantRegression (SDR) TF model for this nonlinearsystem.




• Control design issues




• Control design issues• Mapping the relationship between PM and DM.

(PM and DM are all nominal models obtained fromthe nonlinear data.) To judge how well DMrepresents the dominant dynamics of the originalsystem.


Conclusions


Conclusions


Documents

Simulation Model Emulation in Control System Designmams.rmit.edu.au/rs0z8nwgz8ou.pdf · Simulation Model Emulation in Control System Design ... control systems but the control techniques