to our dear friend : David CLARKE ! .Defence : How to control a - - PDF document

INDUSTRIAL APPLICATIONS OF PREDICTIVE FUNCTIONAL CONTROL

to our dear friend : David CLARKE !

Oxford 09/01/2009 Jacques Richalet

• jacques. richalet @ wanadoo.fr

End of the Sixties:

• ORIGIN OF MODEL BASED PREDICTIVE CONTROL
• PETROLEUM INDUSTRY / DEFENCE INDUSTRY

.Refineries : How to handle constraints on MV’s and CV’s ? How to control a multivariable process ? .Defence : How to control a process with a non stationary set point with no lag error (follow-up servo) ? PROFIT OPTIMIZATION CONSTRAINTS* PREDICTION MODEL M.B.P.C / PREDICTIVE CONTROL

WHY MPC ? WHY MPC ?

Gasoil Residue Crude oil

CONTROL ON CONSTRAINT CONTROL ON CONSTRAINT

Viscosity

• f residue

LIMIT ! t Ko Ok

1975:Total Normandy Refinery. D11column 30.000 T/day!

O D H

Tracking servo system Tracking servo system

P Heading Target trajectory t X

NO LAG ERROR NO LAG ERROR

εΤ = 0 !

….the past…

1968 : Basic principles of Predictive Control 1973 : 1st version of PFC/ IDCOM 1974 : 1st industrial applications of PFC/IDCOM Steam Generator/ Reactor/ Distillation 1978 : 1st contribution MPC “Automatica “ ( problems..!) 1980-85 : HIECON with Set Point ( Houston) 1989 : Extension from PFC to PPC: Parametric Predictive Control (control by reactor flow..) 1998 :1st implementation of an MPC in the native library of a PLC : Modicon. Schneider-Electric.Momentum.Concept

FEATURES

…the oldest MBPC… 1968 A “few” thousand applications in many fields

DEFENCE

Mine hunter Ariane 5 attitude control Missile autopilot Laser guided missile (t=62 microsecond…) Gun turrets Radar antennas Infra Red camera High speed Infra Red Missile launch Camera mount Radar antennas Laser mirror Tank turret (T. 55) Mine sweeper auto-pilot Aircraft carrier auto-pilot

METALLURGICAL INDUSTRIES

Coating lines aluminium Thickness control - Roll eccentricity Mono-multi-stands Rolling mills Continuous casting (slab) Push-ovens Coke furnace Hot/cold/Thin rolling mills Steam generators Level, temperature, pressure……;

MISCELLANEOUS (Chem reactors and distillation excluded)

Bioreactor Melissa ESA Plastic extrusion robot (high speed) Temperature control of gas furnace Temperature control of TGV train carriage River dam level control (T=1 hour ) Powder milk dryer Electric furnace brazing etc….

AUTOMOTIVE

Gear box test bench Dynamic test bench engine Fuel injection Idle fuel injection Clutch antistroke Gear box (tank) Hybrid car (electric-fuel) Air conditioning

FIRST INDUSTRIAL APPLICATIONS OF P.F.C. FIRST INDUSTRIAL APPLICATIONS OF P.F.C.

EXOTIC APPLICATIONS…

North : Molde (Norway) South : Plaza Huincul ( Patagonia-Argentina) East : Hokaido ( Japan) West : Mobile ( USA.DEGUSSA) High : Eiffel Tower elevator ! Space : European Space Agency : Mars project bioreactors ( 2029 !) Depth : Mine hunter boat Speed : Temperature cabin ( TGV : 574.8 km/h W.record ) Fast : Laser guided bomb ( Tsampling : 65

Slow : River dam level ( Tsampling : 1 hour) Ecology: Diester from colza Animal : Dog food pellet dryer Plants : Greenhouses etc…

AT WHAT LEVEL DO WE OPERATE ?

Back to Basics:

• Level 0 : Ancillary processes e.g: FRC/ Pid

(the valve is the nightmare of control !)

• Level 1 : Dynamic control with constraints
• Level 2 : Optimization of working conditions
• Level 3 : Production planning

.Level N is operative if level N-1 works well…! . »There are no technical problems, but only technical aspects of economic problems … »

50 100 150 200 250 300 350 400 450 0.5 1 1.5 2 QUALITY N Histogram Q1 Sigma 1 OFF SPEC

50 100 150 200 250 300 350 400 450 0.5 1 1.5 2 OFF SPEC QUALITY N Histogram Q1 Sigma 1 Sigma 2 LEVEL 0 / LEVEL 1 Sigma 1 -- Sigma 2 50 100 150 200 250 300 350 400 450 0.5 1 1.5 2 OFF SPEC QUALITY N Histogram Q1 Sigma 1 Q2 Sigma 2 LEVEL 0 / LEVEL 1 Sigma 1 -- Sigma 2 LEVEL 2 Q1 -- Q2 50 100 150 200 250 300 350 400 450 0.5 1 1.5 2 OFF SPEC QUALITY N Histogram Q1 Sigma 1 Q2 Sigma 2 DELTA COST (W2-W1) LEVEL 0 / LEVEL 1 Sigma 1 -- Sigma 2 LEVEL 2 Q1 -- Q2 W1 W2 COST FUNCTION

« SQUEEZE AND SHIFT » 4 BASIC PRINCIPLES OF PFC : J. PIAGET 4 BASIC PRINCIPLES OF PFC : J. PIAGET – Operating Image – Target – Sub Target – Action – Comparison : – Predicted / Actual

• Internal independant Model

– Reference Trajectory – Solver. Functional basis Structured future MV – Error compensator

• Natural Control : “You would not drive your car using a PID scheme”

STRATEGY…

. Control of industrial processes : PI(D) ? but PID cannot solve all control problems, but any candidate controller should « look like a PId »: EASY TO UNDERSTAND, TO IMPLEMENT, TO TUNE .. consistent with floor instrumentists’ habits so: . Tuning parameters should have a clear, physical interpretation. . Elementary mathematics. . No explicit integrator in the loop. . No matrix calculation. . No quadratic minimization on-line. . No iterative computation on-line. . Open technology…

TARGETED PERFORMANCES

. Control « all » types of processes: time delay, unstable, non-minimum phase, some non-linear . Handles all constraints on the MV and on internal variables of the process CV. . No lag error on dynamic set points with no integrator Industrial settings: . Transparent and Cascade Control with transfer of constraints from inner loop to outer controller (« Back Calculation ») . True feed-forward . Split-range control with different dynamics (reactors!) . Control with 2 cooperative MV’s, e.g., big valve / small valve . Extendable to 2 MV/ 2 CV….. . Dual control: total elimination of harmonic disturbances . Full Robustness analysis: easy trade-off ,

. Immediate tuning / Open technology

REFERENCE TRAJECTORY REFERENCE TRAJECTORY

Past n Future TRBF

Set point Process output

ε (n)

Predicted output

C(n+H)

Model output

∆ ∆

Manipulated variable

H 1 H 2 MV i MV

Prediction horizon

y P

Reference trajectory

yref y M

Coïncidence horizon

y P ε (n+H) H M

Independent models are selected: . Input /Output models valid for all math. structures

. No permanent errors in steady-state modes!

( never mix process variables with model parameters..)

Re-aligned Model

sp sm

e Process

Model

perturbation

+ +

e

sm sp

Process Model

perturbation

+ +

Independent Model

Solver

• The future MV is projected on a Functional Basis:

– e.g. Taylor expansion / polynomials

• Thus, the MV is restricted a priori and not damped

afterwards….

• MV(n+i) =

• U0(i) = 1, U1(i) = i U2(i)= i2 etc …
• Find the

– At a finite number of points, the predicted model output coincides with the desired reference trajectory

• (i.e., coincidence points)

Why structuring the future MV?

• Determing all future MVs is of little benefit:

– many projects lead to the same future behaviour

• (low-pass process)
• Computation time increases ( PLC !)
• Introduction of a damping term on the speed
• f the MV a priori is difficult to tune !
• Projection on « Eigen » functions of linear

dynamic processes insures no lag error on dynamic set points

100 200 300 400 500 600 700 800 900

• 20

20 40 60 80 100 120 140 OPEN LOOP dCV CV MV

exp( 20 ).(1 15 ) ( ) (1 25 )(1 55 )(1 80 ) s s H s s s s − − − − − − − − = = = = + + + + + + + + + + + +

100 200 300 400 500 600 700 800 900 20 40 60 80 100 120 140 time ClOSED LOOP WITH CONSTRAINT MV CONSTRAINT SETPOINT CV CLTR DESIRED =410 h=125 CLTR ACTUAL =424

100 200 300 400 500 600 700 800 900

• 20

20 40 60 80 100 120 140 REFERENCE TRAJECTORY at ti=20 Setpoint=100 REF TRAJ.(green) CV PREDICTED(black) CV PASSED(red) MV tinit=20 100 200 300 400 500 600 700 800 900

• 20

20 40 60 80 100 120 140 REFERENCE TRAJECTORY at ti=60 Setpoint=100 REF TRAJ.(green) CV PREDICTED(black) CV PASSED(red) MV tinit=60 100 200 300 400 500 600 700 800 900

• 20

20 40 60 80 100 120 140 REFERENCE TRAJECTORY at ti=100 Setpoint=100 REF TRAJ.(green) CV PREDICTED(black) CV PASSED(red) MV tinit=100 100 200 300 400 500 600 700 800 900

• 20

20 40 60 80 100 120 140 REFERENCE TRAJECTORY at ti=150 Setpoint=100 REF TRAJ.(green) CV PREDICTED(black) CV PASSED(red) MV tinit=150 100 200 300 400 500 600 700 800 900

• 20

20 40 60 80 100 120 140 REFERENCE TRAJECTORY at ti=200 Setpoint=100 REF TRAJ.(green) CV PREDICTED(black) CV PASSED(red) MV tinit=200 100 200 300 400 500 600 700 800 900

• 20

20 40 60 80 100 120 140 REFERENCE TRAJECTORY at ti=300 Setpoint=100 REF TRAJ.(green) CV PREDICTED(black) CV PASSED(red) MV tinit=300

ACCURACY

• Basic Eigen Function property:
• If the set point can be projected on a polynomial basis
• f order N, and if the non integrative model is of order

N, then there is no lag error

• Accuracy depends on the choice of the functional basis

and not on the gain of the controller….

Pr H (C - (n)) (1 - ) (n) y y M u(n) + H GM G M (1 - ) λ α =

P = G - s e 1 + Ts M y u θ = = y(n) = α y(n+1) + (1- α) . u(n-1-r) . G α = e -Tsamp T θ = Tsamp . r

ELEMENTARY TUTORIAL EXAMPLE ELEMENTARY TUTORIAL EXAMPLE

Trajectory expo. : λ λ λ λ 1 coincidence point: H 1 Base function: step

Target : ∆P(n+H) = (C - yPr) (1 - λH) Processus with delay : yPr (n) = yP (n) + yM (n) - yM(n-r) Model : Free mode Forced mode : M y (n) H α H u(n) . GM 1 - α

     

Model increment :

M M

P H H H (C - (n )) (1 - ) (n ) + u (n ) G M (1 - ) - (n ) y y y r λ α α =

Control equation :

P(n+H) M(n+H) ∆ ∆ =

Increment = Free(n+h) + Forced(n+h)- ymodel(n)

The only mathematical problem for trainees …!

DUAL CONTROL

• Problem:

– Classical control:

• Constant set-point + harmonic disturbance rejection

. Complex algebra controller : much simpler !

• Examples:

– Helicopter : 20mm artillery / vibration damping – Continuous casting steel industry :

• scillation of the mould and roll swell effect

– Wave Energy Converter etc..

• 6.5
• 6
• 5.5
• 5
• 4.5
• 4
• 3.5
• 3
• 2.5
• 2
• 1.6
• 1.4
• 1.2
• 1
• 0.8
• 0.6
• 0.4
• 0.2

0.2 0.4 SENSITIVITY / CLTR LOG(omega) LOG(CV/DV) CLTR=30 CLTR=300 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 0.05 0.1 0.15 0.2 0.25 Mod|CV/DV| b withoutDual / r with Dualcontrol tauprocess=50

• meg=1 // tau =150)

without dual with dual wave pulsation 1=2pi/150 damping tauprocess=50 CLTR=200 tauprocess=50

• meg=1 // tau =150)

without dual with dual wave pulsation 1=2pi/150 damping tauprocess=50 CLTR=200

Temperature control with internal variable constraints Tsurf

Constraints on internal variables: the multiple controller approach !

Extension to different dynamic constraints …

Cons2=CV2max Cons1

P2 R2

MV2 CV2

P1

MV1

Superviseur

M2 M1 R1

contrainte CV2

(a) (b)

Cons1 CV1 SM2 SM1 t t CV1 Cons2=CV2max Cons1

P2 R2

MV2 CV2

P1

MV1

Supervisor

M2 M1 R1

constraint CV2

(a) (b)

Cons1 CV1 SM2 SM1 t t CV1

100 200 300 400 500 600 700 800

• 20

20 40 60 80 100 120 140 160 SETPOINT CV1=100 / CONSTRAINT CV2 = 129 CV1 CV2 MV2 active MV1 active MV1 MV2 CONSTRAINT CV2 PERT

20 40 60 80 100 120 140 160 4 4.5 5 5.5 6 6.5 7 OLTR CLTR GAIN MARGIN GAIN MARGIN / CLTR

TUNING

Dynamics Robustness Accuracy 2 Base Function CLTR Horizon 2 1- 1+ 2

Goodwin ?

Types of model:

1) Black Box

• Physical analysis . Choice of model ?
• Application of a Test signal protocol ?.
• Optimisation of the protocol?
• TOUGH CONSTRAINTS from THE PRODUCER.!
• Pseudo-random binary noise to be avoided in practice …
• Action on-site and off-site : Cost / Man Hours
• No trade ! : to be repeated on all processes
• Weak capitalisation.
• BUT: no investment, easy to perform….

Identification of Industrial Processes The target is not to minimize the quadratic error between the process output yp(n) and the model

• utput ym(n):

Cstate = ∑ ( yp(n)-ym(n) )2 but to minimize the Structural Distance between the process parameters Pp(j) and the model parameters Pm(j) ( Lyapunof function !..) Cstructure= ∑ ( Pp(j)-Pm(j) )2 because Cstate depends on the input of the process and can be small with wrong parameters so:

• Thus the problem is :
• How to define the most pertinent input signal

leading to the minimum uncertainty of the parameters taking into account the industrial constraints ? :

• Limited duration of test protocole: Hmax
• Limited amplitude of test signals : Amax
• Limited frequency spectrum :

<

• The design of the « best protocol» becomes, mainly,

deterministic : iterative learning procedure !

10 20 30 40 50 60 70 80 90 100 0.4 0.6 0.8 1 1.2 1.4 1.6 H=50 Taum GAIN H=150 H=400 H=2000 H=50 ISO-DISTANCE GAIN / TAU H=150 H=400 H=2000

2) First principles models : knowledge based model

• Physical analysis of the process:
• « Brain juice »:

– Tough - expensive – risky ... – Expertise needed in Physics, Chemical engineering ,Mechanics, etc.

• but ….

– Generic. Fewer calendar days and man-hours

– PHYSICAL ADAPTATION OF THE CONTROLLER

– Strong capitalisation

the future………

Transfer / Training

• Vendors :

– SET POINT (USA Houston) – P.Latour / P.Grosdidier / Tom Badgwell / Greg Martin – . YOKOGAWA (Japon. Mitaka) – . SOTEICA (Argentina / USA) – . SAMSUNG (Korea) – . REPSOL (Spain) – . JGC (Japon / Yokohama) – . SCHNEIDER ELECTRIC (France ) – . INSTITUT TECH NAGOYA (Japan) – . UNIV . of Hanghzou (China)

Users….

• AMOCO ( CAN)
• MOBIL OIL ( USA / F)
• TOTAL ( F)
• ATO CHEM ( F)
• LUBRIZOL ( USA. F)
• SOFIPROTEOL( B. F)
• BASF ( B. D)
• DEGUSSA. EVONIK.( D )
• PECHINEY/ALCAN (F )
• ARCELOR-MITTAL (L )
• SANOFI- AVENTIS ( F.D )
• IRA (F) etc….
• VEOLIA (F)
• ..and others..

PID + LOGIC

PFC without Logic

Péchiney Aluminium 1988 Cold Rolling Mills : 16 / 18

Thickness control Ts =5ms Flatness control Ts=20ms F.P. models : Self.tuning = speed/ thickness/ metal + Low level on line identification of model (lubrification !)

Speed Stdev PID 250 m/mn 4.5 m PFC 1450 m/mn 0.7 m

Pay Out Time ?

Fichier nivocc1.mat

7000 8000 9000 10000 11000 12000 13000 60 65 70 75 80 85 nivocc1 MV(k) CV(r) Cons(b) sec

• ident. 1
• ident. 2

0.05 0.1 0.15 0.2 0.25 0.3 0.35 4.5 4.6 4.7 4.8 4.9 5 5.1 5.2 5.3 5.4 5.5

R=11

nivocc1 ISO-D tau K

R=11

ISO-DISTANCES IN THE PARAMETRIC SPACE

• Ident. 2
• Ident. 1

Comparaison PID / PFC

Mesure du niveau d’acier (70%) Mesure position du vérin ~88 mm Mesure de la vitesse ~1.3 m/mn

39 La centrale : une chaudière

1 2 3 4

• TS

70 t/h TS 70 t/h

HF HF

VENT 270 000 M3/H VENT 270 000 M3/H

Utilités usine

Détente 1 Détente 2 prélèvement prélèvement Vapeur moyenne pression 13 bars 250 ° C

• 40

4 Steam Generators

500 1000 1500 2000 2500 3000 3500 4000 4500 5000 362 364 366 368 370 372 374 376 378 t30095a-Constdes1(k) Tdes1(r) tsur1(m) Qdes1+360(b) sec

Qdes t/h Tdes ° C

Set Point Protocol for Tdes (1GV)

Superheater Temperature - Cascade Control

PFC2

kd

PFC1 .p T 1

1

+ 1 p T 1 G.e

3.p R

• .

3 + tdes tvapint Tsur (° C) Ctdes

gdes

• Transf. de

contraintes

Constvap 500 ° C

Procédé Tdes Procédé Tvap

Qvap (t/h)

+ +

Tp + 1 1

• nov 05

300° C 410° C d/dt=50/2 ° C/s 0-40 t/h d/dt=20 t/h/s T1=21.5s Le FRC est supposé parfait : Kp=1.35 Ti=100s. gdes ~1

PFC1

TRBF1=150s h=2 s Tech=2s kd=a-b*qvap+c*qvap² Contraintes

x Cqdes

Régulateurs FRC

Cqdes*

rcpy rcpy

+

Charge calorifique

Charg (%)

MI :K=1 T=5s p Rch

e Gch

.

. −

PFC2

h =(Tm1+Tm2+Tm3)*0.3 (=30.3 s) Tech=2s Zone : largeur = ±5 ° C TRBF2 L = 35 s TRBF2 H =(Tm1+Tm2+Tm3)*3+retard (=350 s) MI du 3 ème ordre Tdes en boucle fermée : cte de temps = T2

1000 2000 3000 4000 5000 6000 7000 8000 200 300 400 500 600 sec tvap (° C) et Qvap (t/h) GV-TEMPERATURE VAPEUR en PFC- Consigne:b - Tvap:r - Qvap*2:k t22043a.mat

• 30
• 20
• 10

10 20 30 2 4 6 8 10 Histogramme de l'écart TVAP - t22043a Frequence en % ECART TVAP en ° C Tvapmoyen=499.7504 ecart-type=2.352 Min/maxTvap=492.328 507.983 tmin/tmax (s)=3000 7000

PFC Histogram

1000 2000 3000 4000 5000 6000 7000 8000 300 350 400 450 500 550 sec tvap (° C) et Qvap (t/h) GV-TEMPERATURE VAPEUR en PID- Consigne:b - Tvap:r - Qvap*2:k t22043a.mat

• 30
• 20
• 10

10 20 30 2 4 6 8 10 Histogramme de l'écart TVAP - t22043a Frequence en % ECART TVAP en ° C Tvapmoyen=488.2418 ecart-type=3.7752 Min/maxTvap=478.784 500.757 tmin/tmax (s)=3000 7000

PID Histogram

ECONOMIC EVALUATION

Efficiency of turbines increases with steam temperature Above 519 °C the blades of the front compressor starts Creeping ……. and hit the case Classical problem : « 2 S’s » « Squeeze the variance and Shift the target » 3°C -------------- 1M€ Original set point : 493°C -- Today: 506°C Pay Out Time : meaningless…!

• Heating / Ventilation / Air-conditioning
• Europe : 10% increase 2006 - 2007
• Market : 14 MM€
• Energy saving / Global warming
• Ease of implementation

4.1 Moyens mis en oeuvre

OVERHEATING HP

CLASSICAL COOLING PROCESS: Mollier diagram

LOAD CHANGE

PID

• 10
• 5

5 10 15 9:36 10:04 10:33 11:02 11:31 12:00 12:28 12:57 Temps T° C 20 40 60 80 100 120 140 kWe Platelec Surchauffe Consigne Puissance Electrique P Platelec

!

Danger !!!!

LOAD CHANGE

PFC

• 10,00
• 5,00

0,00 5,00 10,00 15,00 10:48 11:16 11:45 12:14 12:43 13:12 13:40 14:09 Temps T° C 0,00 20,00 40,00 60,00 80,00 100,00 120,00 140,00 kWe Platelec Surchauffe Consigne Puissance Electrique P Platelec

Danger!

Model Predictive Control and Real- time Optimization Implementation Policy and Best Practice ( from ARC )

• Increasing

Staying Constant Decreasing 69.8 % 28.8 % 1.4 % !!"

• !!"

120 110 100 90 80 70 60 50 40 30 20 10 80.0 % 20.0 % ## $" %&' ( % !&! !) * !&$++ # , # &' ,

• .

/&0. , !

• #(

# (

• (
• (
• Refining

Chemicals Oil & Gas Other Pulp & Paper Power Electronics Plastics & Rubber Metals Cement & Glass Mining Automotive Aerospace Food & Beverage Machinery Pharmaceuticals 24.3 % 20.6 % 15.4 % 11.0 % 5.1 % 5.1 % 5.1 % 2.9 % 2.2 % 2.2 % 1.5 % 1.5 % 0.7 % 0.7 % 0.7 % 0.7 %

PFC/PPC in the Chemical Industry

• r : « heat exchange based industry »

a) The higher the added value: the lower the interest in production improvement.! b) Batch reactor specific problem: Cooling fluid unstable but Mass Temp. stable ! c) Gap between : Chemical Engineering / Automatic Control « School reproduces Industry / Industry reproduces School »

Why Chemistry lags behind Petroleum ?

.. the trend…!

Constraints :

– Oil barrel price is uncertain. – Lethal Asian competition. – Quality norms (FDA) – Ecology

But fortunately :

– Controlers and software are more accessible – Incoming of a new generation of personnel – The « crisis » will force Industry to improve!

75

R TM Fi, Ti Te

R Reagent

∞

ρ

M C P M V M

dT

M

dt = UA T

e - T M + ∆Hx

ρe CP

e Ve dTe

dt = ρe Fi Ti - Te + UA TM - Te

θ(Fi) TM+ TM= Ti

.

1) Fi =ct /MV= Ti CV=TM / level 0 =Ti ? :PFC 2) Ti =ct (!) MV=Fi Parametric Control non linear :PPC 3) MV : Ti and Fi : Enthaplic control (power) :PPC+ 4) MV : Pressure of reactor :PFC 4 CONTROL STRATEGIES OF BATCH REACTORS

F1, T1 F2, T2 F, T

F.T =F1.T1+F2.T2 T=λ λ λ λ.T1 +(1- λ λ λ λ).T2 λ λ λ λ = F1/(F1+F2) 0 ≤ ≤ ≤ ≤ λ λ λ λ ≤ ≤ ≤ ≤ 1

Convexity Therorem Equivalent Temperature Procedure

• . dT/dt +T = T1+( 1 -

) T2

• . dT/dt +T = TEQ
• PFC computes TEQsol ( sol)
• sol is connected to the physical MV by

a non-linear relation

• Non-linear solver outside the dynamics!

tsf qf , tef qp , tep tsp

)] 1 1 ( . exp[ 1 )] 1 1 ( . exp[ 1 ) ( Ff Fp A U Ff Fp Ff Fp A U Qf − − − − − − = Γ

Fp, Ff : Thermal flows Fp = (ρ ρ ρ ρ.Cp)p.Qp Ff = (ρ ρ ρ ρ.Cp)f.Qf.

Physical Model: Heat Exchanger/ Reactor

• Mass balance/ Enthalpy balance
• Parameters : Geometry / Fluid

Characteristics / Flows and Temp

• Heat exchange coeff. U : Wm2/K as a

function of Reynolds/ Prandtl/ Nusselt

• Function of temperature and flows

NATURAL PHYSICAL SELF ADAPTATION OF THE INTERNAL MODELS OF PFC

100 200 300 400 500 600 700 10 15 20 25 30 35 40 45 50 55 60 TSP d° C QP L/H/5 QL L/H/5 TEP d° C

20 40 60 80 100 120 20 40 60 80 100 120 TEMPERATURE DE MASSE d° min 92° Tmax=108.4° Tinit=42°

Figure 4

EXOTHERMICITY ESTIMATOR by INVERSE PFC

Pert

R1 R2

Cons1 MV1

P

SP SP*

M≡ ≡ ≡ ≡P

MV2 Cons2=SP + + B + +

Pert*

• +

STRUCTURE and STATE DISTURBANCES

• On line estimator: an example!

DV (n) = actual disturbance DV*(n) = estimated disturbance Gp = 1/ gain of process Gm = 1/ gain of model (same dynamics…) DV* (n)= DV(n) +Setpoint.( Gm-Gp ) State and structure mismatches in the same equation !

1 0 2 0 3 0 4 0 5 0 6 0 7 0 8 0 2 0 4 0 6 0 8 0 1 0 0 1 2 0 E S T IM A T IO N D E L ’E X O T H E R M IC IT E T m a s s e / T m a s s e e s tim é e T S - T E T S E X O T H E R M IC IT E m in

Figure 3

31

Reactant Flow T1 Reactor Cooler Steam T2 T4 T3 Heat Exchanger

BASF BASF BASF

B A S F B A S F

B A S F

• DEGUSSA. EVONIK

World n°1 Specialty Chemistry n° 3 German Chemical industry 49.000 employees > 100 production sites WW Technical Center Hanau ( Hesse D.)

Adopts Predictive Control PFC in 2003

Training of Control Staff Standardization of software and procedures

PFC / PPC at DEGUSSA

• Batch reactors:Heat exchangers
• Esterification : Batch time reduction : approx. 10%
• Polymerization: Temperature control error / 10
• Continuous reactors

Polymerization . Improved reproductibility. Quality Other Units: Increase of production . Less energy consumption

• 1/01/2009 : > 15 sites WW / 22 workshops / >180 PFC/PPC

8 engineers : work load..?!

Control Systems

INVENSYS PLS 80E (ex Eckhart) FOXBORO IAS ( HLBL..!) SIEMENS PLCS7 ABB Freelance 2000 YOKOGAWA DCS 3000 EMERSON DeltaV In all DCS PFC/PPC is implemented with a generic procedure : « Structured Text » but mainly by « Block programming » (IEC1131.3) compatible with Simulink !

• Increase of production
• No more off-specification products
• Economy : Energy / Solvant
• Maintenance cost lower
• Operators : happy : less work load !
• Conclusion : large dissemination…
• But where are the problems ?.......;
• Safety . Repeatability

ECONOMIC IMPACT… Pay out time : « very short »

• 04. April 2006

Dr.-Ing. Wolfgang Deis (S-TE-VT- A-I) 97

Anwendungsbeispiel 1: Exotherme Batchreaktion

FC TI TI Standard-Regler für Kühlwasser- menge MPC-Regler für Temperatur

Neue Reglerstruktur:

TC

• 04. April 2006

Dr.-Ing. Wolfgang Deis (S-TE-VT- A-I) 99

Anwendungsbeispiel 1: Exotherme Batchreaktion

Entwurf, Simulation und Vorparametrierung in Matlab (Auszug):

Control structure

PFC PPC plant Q Q Q Qf,Stp

f,Stp f,Stp f,Stp = f(

= f( = f( = f(τ τ τ τ) ) ) ) PID TM,Stp Qf V Teq τ τ τ τ TM Qf,Stp

f f

Q Q b a ⋅ + = τ

trbfPFC = trbfPPC / 3 : Back calculation of constraint τ τ τ τ* = f-1(Qf) Teq*

• 04. April 2006

Dr.-Ing. Wolfgang Deis (S-TE-VT- A-I) 101

Anwendungsbeispiel 1: Exotherme Batchreaktion

20 40 60 80 100 120 65 70 75 80 85 Zeit [min] Temperatur [° C]

+/- 0,2 °C neue Fahrweise (MPC-Regler)

hier liegen 35 Batches übereinander !

Fortschritt [%]

Anwendungsbeispiel 1: Exotherme Batchreaktion

20 40 60 80 100 120 65 70 75 80 85 Zeit [min] Temperatur [° C] Fortschritt [%]

+/- 2 ° C alte Fahrweise (PI-Regler)

Anwendungsbeispiel 2: Endotherme Batchreaktion

0.2 0.4 0.6 0.8 1 1.2 1.4 20 40 60 80 100 120 PID: Reaktortemperatur [% des Sollwertes] MPC: Reaktortemperatur [% des Sollwertes] 0.2 0.4 0.6 0.8 1 1.2 1.4 20 40 60 80 100 120 PID: Dampfventil [%] MPC: Dampfventil [% ]

PID

50 100 150 200 250 300 350 400 87 88 89

[% ]

T REACTOR SETPOINT T REACTOR [%] 50 100 150 200 250 300 350 400 60 80 100 120

[% ]

MV [%] 50 100 150 200 250 300 350 400 72 74 76 78 80 82

time [%] [% ]

LOAD [%]

• Identification

50 100 150 200 250 300 350 79 80 81 82

m a s s [% ]

LOAD 50 100 150 200 250 300 350 60 80 100

T [° C ] P U M P S IG N A L [% ]

T INPUT REACTOR [%] MV [%] 50 100 150 200 250 300 350 87 88 89 90

Zeit [min.] T [% ]

T REACTOR [%] T MODEL [%]

PFC

145 150 155 160 165 170 175 180 185 190 87 88 89

[% ]

T REACTOR SETPOINT T REACTOR [%] 145 150 155 160 165 170 175 180 185 190 60 80 100 120

[% ]

MV [%] 145 150 155 160 165 170 175 180 185 190 72 74 76 78 80 82

time [%] [% ]

LOAD [%]

Enthalpic Control Reaching the temp. plateau

SANOFI-AVENTIS

• Pharmaceutical industry
• World n°3 / Continental Europe n°1
• World n°1 :Vaccines
• Personnel > 140.000
• >100 production sites
• Turnover : 29MM euros
• Drugs, vaccines…

SANOFI AVENTIS PREDICTIVE CONTROL

• Target:

– Improvement of quality of product and repeatability

• f recipes. Energy consumption.

– Start : November 2005 .Training of personnel:PFC

• Flexible reactor of the IRA institute in ARLES

Controllers:

• Premium Unity Schneider-Electric
• PFC implemented in native Control Library

Identification Mass Temperature

20 40 60 80 100 120 140 30 40 50 60 70 80 90 100 min Tmasse /mod° / k Tmasse /processus° /r Temp.sortie calo estimée !?/g Temp entrée calo° /b PRE MIE RE ID ENTIFCA TION TMA S S E 20 25 30 35 40 45 50

Endothermicity Estimator and Feed-forward

Internal temperature Jacket input temp Endothermicity estimator 1 2 3 4 5 6 7 8 9 x 10

4

10 20 30 40 50 60 70 80 90 100 Commande en pression à TEC constante TEC TMASSE PRESSION/10 TSC

YES BUT: …..! « Ok .. But my problem is more complex and very specific… » Industrial resistance ?

• lack of basic training in modelling
• no clear understanding of hierarchical control
• a priori year budget and no « Pay Out Time » approach

No conflict with PID : both techniques are used: If PID works well (e.g, FIC) - use it If not, use another tool…

ASSESSMENT ?

• MATURE TECHNIQUE
• CONTROLERS and TECHNOLOGY : OK !
• Validated by numerous world-wide applications in most

industrial domains

• Short PAY OUT TIME…..
• but…

85% of the work and 95% of the adrenalin

DYNAMIC FP MODELLING

Who is going to Implement Predictive Control ?

Since the main problem lies in Modelling : The expert in Modelling ? ( Education ?)

• The local industrial Control Instrumentist ?
• The DCS/ PLC vendor ?
• An Engineering Company ?
• -------------------------------------------------------------

..since Control is made « simple », the Control expert has eliminated himself !

Industrial Prospective

Integrated design: To take into account the posibilities of Advanced Control in the early design of the process (e.g: CCV in Aircraft design) : less steel, less concrete…. : more « brain juice » Example in Chemical Engineering !

BATCH REACTOR: Enthalpic Control

TIT 3 TIT 2 TI 1 Heating rod : P Temperature source : ts SIC 2

1 Pump 1 Variator 2 Temp. sensors

2030€

Teaching control in Technical schools

• 39 Technical schools in France
• ≠ 660 Young technicians per year
• PFC was taught to 23 profs of 4 academia
• Lectures and laboratory works on PLCs

and laboratories processes.

• The first 125 young technicians went out

to industry in 2008, completely at ease with PFC… To be continued….

Thank you for your attention….

• jacques. richalet @ wanadoo.fr

There is a post-script…!

• ACADEMIC CHALLENGE.!

A little more difficult ? Time delay Non minimum phase Integrative Unstable pole Pure oscillator Time constant Set point : ramp with no lag error Constraints on MV and a state variable CV

H(s)= K.exp(-T1.s).(1-T2.s) / s.(1-T3.s).(s2+

2).(1+T4.s)

bon appétit…