Industrial Practice of Model Predictive Control

Honeywell

From Process Unit to Plantwide Control & Optimization Page 1

Honeywell

Workshop at CPC-FOCAPO, Tucson

January 8th, 2017

Joseph Lu

Honeywell

Industrial Practice of Model Predictive Control*

* Primarily through the lens of Honeywell practices in the process industries

Honeywell


How are MPCs typically justified in process industries?

Model predictive control applications are typically financially

evaluated and justified by its benefits on a per-application basis.

More often, MPC Control provides only the necessary feasibility

condition for process optimization, whereas the latter provides

the sufficient financial justification for its inception.

Typical Sources of Benefits:

• Increased Throughput

• Improved Yields

• Reduced Energy Consumptions

• Decreased Operating Costs (e.g. catalyst)

• Improved Product Quality Consistency

• Increased Operating Flexibility

• Improved Process Stability

• Improved Process Safety

• Improved Operator Effectiveness

• Better Process Information

Honeywell


MPC Applications and Benefits

• 5000+ of running Honeywell applications across many industries

• Industry-wide, the number could exceed 10,000+

• Generally accepted ranges for MPC and optimization benefits (per process unit)

Oil Refining

Crude Distillation (150 KBPD)

Coking (40 KBPD)

Hydrocracking (70 KBPD)

Catalytic Cracking (50 KBPD)

Reforming (50 KBPD)

Alkylation (30 KBPD)

Isomerization (30 KBPD)

Benefits ($0.01/bbl)

5-13

15-33

13-30

13-30

10-26

10-26

3-17

Benefits (US$/yr)

2.7M - 7.0M

2.2M - 4.8M

3.3M - 7.6M

2.4M - 5.4M

1.8M - 4.7M

1.1M - 2.8M

.3M - 1.8M

Industrial Utilities

Cogeneration/Power Systems

Benefits (/yr)

2-5% Decrease in Operating Costs

Petrochemicals

Ethylene

VCM

Aromatics (50KBPD)

Benefits (/yr)

2-4% Increase in Production

3-5% Increased Capacity / 1-4% Yield Improvement

3.4M - 5.3M US$

Chemicals

Ammonia

Polyolefins

Benefits (/yr)

2-4% Increased Capacity / 2-5% Less Energy/Ton

2-5% Increase in Production/Up to 30% faster grade transition

Pulping

Bleaching

TMP (Thermo Mechanical Pulping)

Benefits (/yr)

10-20% Reduction in Chemical Usage

$1M-$2M

Honeywell


Typical Payback Time by Industry

Typical Payback Time:

Oil and Gas 1 to 2 months

Refining 3 to 6 months

Chemicals

Petrochem 4 to 6 months

MMM

Pulping/Paper 6 to 8 months

Industrial

Power 10 to 12 months

Honeywell


Unit-Wide MPC Control Objectives

Feed ($y)

Gas ($x1)

Raw Gasoline ($x2) Naphtha ($x3) Kerosene ($x4) Light Gas Oil($x5) Heavy Gas Oil($x6)

Residue ($x7) Fuel ($z)

Sim

pli

fied

Pro

cess

Un

it

Products:

Operating Constraints

Three Common Objectives: • Produce more product(s) of same (or better) quality

• Satisfy all safety/quality/operating constraints

• Maximize the product value and/or minimize operating cost

Honeywell


HILO

RQsu

ysy

usyuAu

~minarg22

,*

Model Predictive Range Control

Versatile control needs in MIMO applications Regulatory control

Reference tracking

Disturbance rejection / minimum variation

Traditional SISO control

Constraint handling

Prevents another set of variables from exceeding their limits

Transition control

Moves the operating point of a process unit from x to y

Role of individual output can change

Degrees of freedom

Model Predictive Range Control

Output prediction:

The set-point is changed into a ―set-range‖

A slack variable is introduced in the MPC formulation

Additional degrees of freedom are used for

Local disturbance rejection/absorption

Optimization: e.g.,

(independent tuning vs. control response)

-2 -1 0 1 20

20

40

60

80

100

120

pe

na

lty

control variable

-2 -1 0 1 20

20

40

60

80

100

120

control variable

pe

na

lty

-2 -1.5 -1 -0.5 0 0.5 1 1.5 20

20

40

60

80

100

120

control variable

penalty

22~minarg*RQu

uryuAu

Regulatory

Control

Range

Control

yuAy ~

ry HILO yyy

HILO yyyy ;min

s

Honeywell


Range MPC Design and Solution

Range control benefits

Robustness

Minimum effort control (when the control solution is non-unique)

Smoother control, minimum high frequency excitation

Range control algorithm

Two-stage analytical approach

a) solve the range control problem

b) If soln to (a) is not unique, solve for the minimum effort control

Online algorithm similar to the active-set QP: (here assume R = 0)

low limit

high limit

correction horizon

time

prediction horizon

manipulated variable

current time

controlled variable

T s

correction horizon

time

prediction horizon

manipulated variable

current time

controlled variable

T s

set-point

HILORuyyuAyuu

~ ;minarg2

*

HILORQsuysyusyuAu

;~minarg

22

,*

yywyywuAu HILOQwu

~~ ;minarg2

,*

2

,minarg*

Qfree

act

free

act

wu w

wu

A

Au

2

2minarg* act

Tkkk

uwuVRUu

T

kkkact VRUAQ k )(21

Let

Performance Specification (via a funnel):

When the solution is not unique, a

secondary optimization can be performed

to find the minimum effort solution.

Honeywell


Non-Square Control Example—Solvent Dewaxing Unit

. .

Feed

Steam

Double

Pip

e

Chill

er

Refrig

Refrig

Refrig

Dewaxed Oil

WFO Mix Htr

Reflux

Reflux

Reflux

Reflux

Steam WF

O

Dehyd

rato

r

Vacuum

Double

Pip

e

Chill

er

Refrig

Reflux

Reflux

Reflux

Reflux

Steam

Wax

Steam

Keto

ne

Str

ipper

Steam

Refrig

WF

O A

tmos

Fla

sh T

ow

er

WF

O P

ress

Fla

sh T

wr

WF

O F

inal

Fla

sh T

wr

WF

O S

trip

per

PW

Fin

al

Fla

sh

Tw

r

PW

Pre

ss

Fla

sh

Tw

r

PW

Atm

os

Fla

sh T

wr

PW

Mix

Surg

e

Dry Wet

Slo

p

Solv

ent Repulp

Filtrate

Repulp

Filt

Feed

Primary

Filtrate

Prim

ary

Filt

Feed

Double

Pip

e

Chill

er

Double

Pip

e

Chill

er

Do

ub

le

Pip

e

Exch

an

ge

r

Dou

ble

Pip

e

Exch

an

ge

r

Dou

ble

Pip

e

Exch

an

ge

r

Dou

ble

Pip

e

Exch

an

ge

r

PW

Str

ipper

Honeywell


Solvent Dewaxing Unit Illustration

1 2 3

6 5 4

Tank 1: Mixture with solvent

Tank 2: Cooled mixture (wax slurry)

Tank 3: Chilled solvent (wax slurry)

Tank 4: Filtrate with wax removed

Tank 5: Wet solvent (oil removed)

Tank 6: Dried solvent for reuse

Wax + Oil

Mixture

Solvent

Wax

Lubricant Oil

Filtering

Solvent Inventory Control

Honeywell


Model Dynamics of Solvent Control

Primary

Filtrate Flow

Repulp

Kickback Flow

Slop Solvent

Flow

Slack Wax

Flow

5th

Dilution

Flow Ratio

Primary

Filtrate Level se

142.1 3

se

142.1 3 se

s

10113.1

Repulp Filtrate

Level se

139.1 3

s

111.1

Dry Solvent

Drum Levelse

se 53 1

6.1

Moist Solvent

Drum Levelse

se 53 1

62.1

se

116.2 3 se

se 53 1

08.1

s

1173.0

Slop Solvent

Tank Level se

123.1 3

Slack Wax

Tank Level se

11 4

From the application reported by Hall, et. al., 1995.

Honeywell


Simulation Results

0 100 200 300 400 500 600 700 800 900 1000-15

-10

-5

0

5

10

15Controlled Variables

Time[Min]

1

2

3

4

5

6

0 100 200 300 400 500 600 700 800 900 1000-100%

-50

0

50

100%Manipulated Variables

1

2

3

4

5

Min

5

Filter Wash Grade Change

Honeywell


Pulp and Paper Example: Thermomechanical Pulping (TMP) Refiner

TMP refiner – a naturally non-square 4x5 interacting system

• Objective: consistent quality from coordinated control (decoupling)

Customer estimated benefits: ~$13.75M/Year

Honeywell


Ethylene Plantwide Control and Optimization

Fig 1. Aerial photo of an ethylene plant

Fig 2. Process overview – flow diagram

Feed System

FURNACES

OIL QUENCH

WATER QUENCH

DIS

TIL

LA

TE

S

TR

IPP

ER

TCG

CA

US

TIC

CO

ND

EN

SA

TE

S

TR

IPP

ER

DE

ME

TH

AN

IZE

R

DE

ET

HA

NIZ

ER

DE

PR

OP

AN

IZE

R

DE

BU

TA

NIZ

ER

ET

HY

LE

NE

S

PL

ITT

ER

ACETYLENE RECOVERY

UNIT

MAPD CONVERTER

HYDROGEN

FUEL GAS

COLD BOX

ETHYLENE REFRIG

PROPYLENE REFRIG

ETHYLENE

ETHANE COCRACK

PR

OP

YL

EN

E

SP

LIT

TE

R

PROPYLENE

PROPANE COCRACK

MIXED C4

TCG

Profit

CGC

DE

BU

TA

NIZ

ER

Feed System

FURNACES

OIL QUENCH

WATER QUENCH

DIS

TIL

LA

TE

S

TR

IPP

ER

TCG

CA

US

TIC

CO

ND

EN

SA

TE

S

TR

IPP

ER

DE

ME

TH

AN

IZE

R

DE

ET

HA

NIZ

ER

DE

PR

OP

AN

IZE

R

DE

BU

TA

NIZ

ER

ET

HY

LE

NE

S

PL

ITT

ER

ACETYLENE RECOVERY

UNIT

MAPD CONVERTER

HYDROGEN

FUEL GAS

COLD BOX

ETHYLENE REFRIG

PROPYLENE REFRIG

ETHYLENE

ETHANE COCRACK

PR

OP

YL

EN

E

SP

LIT

TE

R

PROPYLENE

PROPANE COCRACK

MIXED C4

TCG Controllers

CGC

DE

BU

TA

NIZ

ER

Process and Operating Characteristics

Universal:

► No product blending

► Stringent quality requirements

► Slow dynamics from gate to gate

► Gradual furnace and converter coking

► Furnace and converter decoking

► Frequent furnace switching

Site specific:

► Feed quality variations

► Product demand changes

► Sensitivity to ambient conditions

► Periodic switching (e.g., dryers)

Advanced Control and Optimization Goals

► Stabilize operation

► Minimize product quality give away

► Maximize selectivity & yield

► Minimize converter over-hydrogenation

► Minimize ethylene loss to methane/ethane recycle.

Ethylene Plant Overview

• Feed and products

Ethylene plants are very flexible and can process a wide variety of feed stocks, ranging from gases (Ethane,

Propane, LPG etc), naphthas (light and full range) to distillates and gas oils. Main products are polymer grade

ethylene and propylene.

• Typical market driving forces

Ethylene being a bulk commodity is sensitive to market conditions. Depending on the local economics, operations

objectives are a combination of yield improvement, production maximization and energy intensity reduction.

• Operating flexibilities (variables for optimization)

Main operating degrees of freedom include feed selection, furnace feed rates, cracking severity, dilution steam,

CGC suction pressure, typical column variables (reflux, reboiler and pressure), converter temperature and H2 ratio

and refrigeration compressor suction pressure.

• Complex chemical reactions and yield prediction

Ethylene yield from the furnaces cannot be directly measured. Therefore we use a yield model to estimate the

yields (& derivatives) for control and optimization. Technip’s on-line SPYRO, the industry standard, has been

integrated into Honeywell’s UniSim Design as an extension Unit Op.

Honeywell


Fra

ctio

na

tor

DeC

1

DeC

2

DeC

3

C3

Sp

litt

er

C2H2

Converter

MAPD

Converter

Cold

Box

Qu

ench

C3

R

RMPCT 1

RMPCT 2

RMPCT 8

RMPCT 9

RMPCT 11

C2

Sp

litt

er

C2

R

RMPCT 10

RMPCT 12

RMPCT 13

RMPCT 14

RMPCT 15

Profit Controller Envelopes

Honeywell


Plantwide Dynamic Optimization

Fract

ion

ato

r

DeC

1

P

DeC

2

DeC

3

C3 S

pli

tter

C2H2

Converter

MAPD

Converter

Cold

Box

Qu

ench

C2

R

C3

R

Profit Optimizer

RMPCT 1

RMPCT 2

RMPCT 8

RMPCT 9

RMPCT 11

RMPCT 14 RMPCT 15

RMPCT 10

RMPCT 12

RMPCT 13

Honeywell


Typical Results of Ethylene Plant APC

• Execution Frequencies:

- MPC Controllers: ½ to 1 minute

- Global Optimizer: 1 minute

- Gain updating: 5 minutes

• Typical Benefits

- Production Increase: $1.5-3Millions

- Energy Savings: Additional

Implementation Details

Design Typical design comprises of individual Profit Controllers for each of the units, a

Profit Optimizer on top of all the controllers, a UniSim model for inferentials and

gain mapper, which is a Profit Bridge app that calls the UniSim model in real

time to update the gains in both Profit Controllers and Profit Optimizer once

every 5 mins.

Modeling Dynamic controller models are identified, in either open loop or closed loop,

using Profit Stepper. Profit Optimizer model is automatically aggregated from all

Profit Controllers. Bridge Models are identified from historical data (no step

testing is required) for modeling the dynamic interactions among all MPC

controllers. The UniSim modeling is fairly simple as only material balance

calculations are required for gain updating.

Commissioning Profit Controllers are commissioned first to provide stable unit operations. Profit

Bridge is then commissioned to provide gain updating for the furnace controllers.

Next, Profit Optimizer is commissioned, and Bridge Model predictions are

verified against data. Finally, a plantwide objective function is populated and

optimization moves are validated. Continuous operation of the controllers and

optimizer starts after operator training.

Maintenance Typically little maintenance is required to keep the ethylene APC/Optimization

running. Clients either dedicate a parttime control engineer to monitor and

perform minor services as needed, or depend on Honeywell’s BGmax quarterly-

visit services.

90

92

94

96

98

100

102

104

106

108

110

Before APC

After APC

Decoking campaigns; one furnace out of service

Normalized Olefin Production Rate Trend Production is increased

and product variability

reduced...

Ethane Impurity Content (in ppm) in Product Ethylene

Base

+100

+200

+300

+400

+500

+600

Before APC After APC

Tighter control and

less quality give-away

Honeywell


Semiconductor

Multi-factory

Supply Chain

vendor1

vendor2

Fab2

P1,P2

Fab1

P1

Fab3

P2

Sort3

P2

Sort2

P1,P2

Sort1

P1

Asm3

Asm2

Asm1

vendor3 vendor4

vendor5 vendor6

1.1 si

1.2 si

2.1

2.2

2.3

3.1

3.2

3.4 pp 3.3 pp

3.7

3.8 pp 3.9 ram

3.10 3.11

3.12

3.5

3.6

3.5

A

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

26

27

24

25

28

29

30

31

32

Box1

7.2

7.1

Box2 7.6

7.5

F

43

7.4

pp

45

44 vend8

46

7.3

pp

Test3

Test1

Test2 Fuse2

Fuse1

4.3

5.2

5.1 33

34

ven

d

7

4.1

4.2

6.2

6.1

6.3

C

B

D

35

36

37

38

39

40

41

42

Blue = Intel

Red = Mat. Sub.

Green = Cap. Sub.

= Inv Hold

= Prod Mfg

= Transport

= Mats Mfg

Nontraditional Space: Discrete Manufacturing

Semiconductor

Courtesy of Karl Kempf – Intel

Honeywell


TPT Distribution

0

0.2

0.4

0.6

0.8

1

1.2

0

0.05

0.1

0.15

0.2

0.25

Data

Model

DataPDF

ModelPDF

Semiconductor Discrete manufacturing – Sorting Assembly

T3

T2

T1

Fab/Sort

Die/Package Inventory

Assembly/Test

Semi-Finished Goods Inv.

Finish/Pack

Components Warehouse

Shipping

Demand

(Orders)

C1

C2

C3

M1

M2

M3

I10

I20

I30

C4

M4

D1 D2

D3

t

T4

Simple Supply

Chain Sequence

sort test

wafer

Honeywell


MPC Controller Design

One controller per product

124 controlled variables

5 fab/sort End-of-Line Inventories

7 Assembly Die Inventories (ADI)

7 die prep capacities (Work-in-Progress)

56 Tape and Reel Die Inventories (TRDI)

48 ADI + Die-Prep + TRDI inventories

90 manipulated variables (Green)

35 ships from sorts

48 ships from a hub TRDI to assembly test

manufacturing (ATM) sites

7 starts from ADI

293 feedforward variables (Blue)

5 sort supply w/forecasts

48 ATM TRDI demands w/forecasts

240 TRDI cross-ships w/forecasts

Current system design consists of 34 such

controllers (products) per computer node

One dynamic optimizer covers all 34 MPC

controllers.

DIE PREP

TRDI TRDI TRDI TRDI TRDI TRDI ADI 01 TRDI TRDI TRDI TRDI TRDI TRDI

01 ADI

FAB

DIE PREP DEMAND

SORT 02

FAB

SORT 03

FAB

SORT 04

FAB

SORT 01

DIE PREP

TRDI TRDI TRDI TRDI TRDI TRDI ADI 02 DEMAND

DIE PREP


DIE PREP


DIE PREP


A model predictive control application for inventory and production management.

Honeywell


Chip Inventories at an Unnamed Site

APC used

APC used

APC used

Die xxx_06 Inventory

0

500

1000

1500

2000

2500

3000

5/21/06 5/28/06 6/4/06 6/11/06 6/18/06 6/25/06 7/2/06 7/9/06

Date

K D

ie

High Limit

Low Limit

Inventory


0

500

1000

1500

2000

2500

3000

5/21/06 5/28/06 6/4/06 6/11/06 6/18/06 6/25/06 7/2/06 7/9/06

Date

K D

ie

High Limit

Low Limit

Inventory


0

250

500

750

1000

1250

1500

5/21/06 5/28/06 6/4/06 6/11/06 6/18/06 6/25/06 7/2/06 7/9/06

Date

K D

ie High Limit

Low Limit

Inventory

Honeywell


Refinery-wide (Production) Control/Optimization

Customer Demand

(downstream

supply chain)

Supply Side

(upstream

supply chain)

Raw Materials Execution Products

month ahead days ahead hours ahead now hours ahead days ahead months ahead

Production Plan

(monthly) What Feed to buy? What products to sell?

Average daily production Average daily delivery Generate a month-plan

based on an average model

Product to

be Deliveried Feed Received What Feed to use? What products to deliver?

Translate the plan into production

activities to satisfy the plant logistics

Production Schedule

(weekly)

Economic Optimization

Multivariable Control





Slave MPC 1 Slave MPC 2 Slave MPC n



Master MPC

Human Coordination

APC

A set of production schedules

(e.g., feed blending schedule) Blending/shipment schedules

Honeywell


(Multiscale) MPC Cascade Structure

1-to-n MPC cascade – One Master MPC and multiple Slave MPCs

The Master MPC employs a dynamic model which can be, at user’s discretion, a coarser-scale model

e.g., similar to the scale-level of a planning model

The Slave MPCs employs a dynamic model which can be a finer-scale one

If Slave MPCs are deployed at every process unit, then the Master MPC can solve the plantwide

production control problem dynamically and in closed loop

Currently, a steady-state version of this problem is controlled by planning solutions in open-loop

Multivariable

MPC Control

Economic

Optimization

Multivariable

MPC Control

Economic

Optimization

Multivariable

MPC Control

Economic

Optimization

Multivariable

MPC Control

Economic

Optimization


Master MPC

two-way connection two-way connection

Honeywell


Refinery-wide Production Control : JIT manufacturing

Straig

ht-R

un an

d/o

r Blen

din

g

Blending Components

Straight-run products

• The intermediate tanks here are for illustration and may vary from plant to plant.

Honeywell


JIT manufacturing viewed as a control problem

Component

Tank 2

Component

Tank 3

Component

Tank 1

Light Distillate

Hydrotreater

Heavy Distillate

Hydrotreater

Hydrocracker

Crude Unit

1

Crude Unit

3

Crude Unit

2

Vac Unit

1

Vac Unit

2

Blending

5 KBarrels

8 KBarrels

10 KBarrels

7 KBarrels

10 KBarrels

Material Balance &

Inventory Control

Crude Feed 1

Crude Feed 2

Crude Feed 3

Imports Moving Avg

daily delivery:

Scope: the distillate production pool in an oil refinery It is simpler but still captures most of the important issues

Honeywell


Just-in-Time Manufacturing – Quality Control (1)

Component

Tank 2

Component

Tank 3

Component

Tank 1

Light Distillate

Hydrotreater

Heavy Distillate

Hydrotreater

Hydrocracker

Crude Unit

1

Crude Unit

3

Crude Unit

2

Vac Unit

1

Vac Unit

2

Blending

Sweet

0.1% Sulfur

Sour

1.2% Sulfur

Medium

0.3% Sulfur

12 ppm S

10 ppm S

10 ppm S

10 ppm S

10 ppm S

1~2 ppm S

Quality Control - Sulfur

4~20 ppm S

Imports

4~20 ppm S

100~600

ppm S

50~300

ppm S

100~400

ppm S

Control

handles Legend:

Honeywell


Jet Cut

Stove Cut

Just-in-Time Manufacturing – Quality Control (2)

Component

Tank 2

Component

Tank 3

Component

Tank 1

Light Distillate

Hydrotreater

Heavy Distillate

Hydrotreater

Hydrocracker

Crude Unit

1

Crude Unit

3

Crude Unit

2

Vac Unit

1

Vac Unit

2

Blending

Sweet

Sour

Medium

Cloud Spec

(winter)

-36 ⁰C

-44 ⁰C

-40 ⁰C

-30 ⁰C

-42 ⁰C

-40 to -55 ⁰C

= Feed Cloud

Jet Cloud

-30 to -40 ⁰C

Stove Cloud

-5 to -10 ⁰C

Quality Control - Cloud Point

-5 to -45 ⁰C

Jet Cut

Stove Cut

Jet Cut

Stove Cut Gasoils

Control

handles Legend:

Imports

Few if any (closed-loop) control solutions are available for this class of problems!

Honeywell


Each (i, j) entry contains a dynamic model :

Often simple dynamics such as or would suffice

Notice that there is no setpoint in this control problem – just constraint control

Control Model Matrix for the Master MPC

MVs

CVs

Crude1

Feed

Crude2

Feed

Crude3

Feed Hydrocracker

Conv

Cr Unit

Cuts Hydrotreater

Sulfur

Product-i

delivery

CompTank2.Level

CompTank2.Sulfur

CompTank2.Cloud

CompTank3.Level

CompTank3.Sulfur

CompTank3.Cloud

More CVs…

1) Unit feed rates

2) Control factors that affect the product quality

1)

Pro

du

ctio

n a

mo

un

t

2)

Pro

du

ct q

ual

ity

DVs

)(, sg ji

se

ss

k

)1(

se

bsas

k

12

Manipulated Variables (MVs)

Co

ntr

oll

ed V

aria

ble

s (C

Vs)

1) Demands

2) Others

Honeywell


Master MPC for JIT Manufacturing in closed loop?

Multivariable

MPC Control

Economic

Optimization

Multivariable

MPC Control

Economic

Optimization

Multivariable

MPC Control

Economic

Optimization

Multivariable

MPC Control

Economic

Optimization


Master MPC

Yes, if the solution for JIT manufacturing from the Master

MPC is implementable by the Slave MPCs.

No, if the solution from the Master MPC can cause one or

more Slave MPCs to wind up (i.e. sustained CV constraint

violations).

The Real Challenge: How to get the Master to understand the Slave’s constraints

well enough so that the Master’s solution will respect them

MVs

CVs

Crude1

Feed

Crude2

Feed

Crude3

Feed

Hydrocr

acker

Conv

Cr Unit

Cuts

Hydrot

reater

Sulfur

Produc

t-i

deliver

y

Tank2.Level

Tank2.Sulfur

Tank2.Cloud

Tank3.Level

Tank3.Sulfur

Tank3.Cloud -

More CVs…

MV1 MV2 MV3 MV4 MV5 MV6 MV7 MV8 Feed

110 5 16 0 105 10.4 0.68 27.4 33.5

CV1

89.3

CV2

7.23

CV3

89.3

CV4

•

•

•

Honeywell


Master-Slave Models: a side-by-side view

1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 2.a 2.b

Below is an illustration of the model structure for 2 combined units:

1 2 3 4

Slave MV index

Master MV/DV index

Mas

ter

CV

s S

lave1

CV

s S

lave2

CV

s

Plantwide Production Control

• Inventory control

• Quality control

• Cost and profit optimization

Unit Process Control

• Safe/smooth operation

• Unit efficiency

• Intermediate Quality

• Local Optimization

Different Tasks

The relationship is a bit convoluted, so let’s first look at the simplest case: 1 conjoint variable per Slave MPC

Slave MPC 1

Slave MPC 2

Master MPC

Conjoint

MVs

Free MVs

Honeywell


Straig

ht-R

un

and

/or B

lend

ing

Blending Components

Straight-run products

nyd

yd

yd

yd

sg

sg

sg

sg

)(

)(

)(

)(

3

2

1

Products

Hydrocracking

Feed

The Concept of Proxy Limit – An Illustration for 1 Conjoint Variable Case

Opt Value: 38 4.5 2.3 10 40 10.7 1.25 28.3 114

Opt

Value 100

6.20

75.0

Refinery Steps in the hydrocracker example:

1) The HC feed is configured as MV3 in

the Master.

2) Its current value is 33

3) The Master makes an inquiry call to the

Slave to maximize the feed, subject to

the Slave constraints

4) The call returns the following:

a) the maximum feed = 38

b) 4 active HC constraints are shown

below, which are CV1, CV3, MV4 &

MV5 (others not shown)

5) The master uses the maximum feed

limit of 38 as a proxy limit for all Slave

constraints.

MV3 = 33

Feed MV2 MV3 MV4 MV5 MV6 MV7 MV8 MV9

33 5 16 0 105 10.4 0.68 27.4 110

CV1

89.3

CV2

7.23

CV3

89.3

CV4

•

•

•

Honeywell


k k+1

MV 2 Low

MV 2 High

MV1 High

Master CMV(k)

Master CMV(k+1)

MV1 Low Lmt

Approximate the feasible region defined by the Slave constraints

for 2 (or more) conjoint variables

Call # c t d t

1 -1 0 0 1

2 1 0 0 1

3 0 -1 1 0

4 0 1 1 0

5 -1 1 1 1

6 1 -1 1 1

7 1 1 1 -1

8 -1 -1 1 -1

Free MVs Conjoint MVs

Current Operating Point (0,0)

Master MVs

Slave MPC (i) 0

min

conj

t

HIfreefconjcLO

conj

t

x

xd

bxAxAb

xc8 inquiry calls:

I

GA Where,

Honeywell


Proxy Limits – Customer Example

70.000

75.000

80.000

85.000

90.000

95.000

100.000

1 8

15

22

29

36

43

50

57

64

71

78

85

92

99

106

113

120

127

134

141

148

155

162

169

176

183

190

197

204

211

218

225

232

239

246

253

260

267

274

281

288

295

302

309

316

323

330

337

344

351

358

365

Hydrocracker Conversion - Actual vs Maximum

To what extent is the Master MPC allowed to change the Slave unit’s operations?

Honeywell


Proxy Limits – Customer Example (2) 1 8

15

22

29

36

43

50

57

64

71

78

85

92

99

106

113

120

127

134

141

148

155

162

169

176

183

190

197

204

211

218

225

232

239

246

253

260

267

274

281

288

295

302

309

316

323

330

337

344

351

358

365

Heavy Distillate HTR Feed Rate - Actual vs Maximum

Honeywell


50 100 150 200 250 300 350

-48

-47

-46

-45

-44

-43

-42

-41

-40

Tank3.Cloud Point in Deg C

Tank3.Qualities (After - Master MPC)

Historical Operation Data (Before)

50 100 150 200 250 300 350

48

50

52

54

56

58

Tank3.Flash Point in Deg C

50 100 150 200 250 300 350

245

250

255

260

Tank3.90pct Point in Deg C

0 100 200 300 4003

4

5

6

7

8

9

10

11Tank3.Sulfur in PPM

Master’s Control and Optimization Results:

Component Quality Changes in Tank #3

Higher cloud point

means that more

low-cost (heavier)

components are

used to produce the

same products

Higher 90pct point

means that more

low-cost (heavier)

components are

used to produce the

same products

Lower flash point

means that more

low-cost (lighter)

components are

used to produce

the same products

Sulfur is controlled at

its high limit which

means much lower

catalyst/utility costs

Component production costs are significantly reduced!

Honeywell


0 50 100 150 200 250 300 350 400-5

-4

-3

-2

-1

0

1

2

3

4Incremental Changes in Crude Unit Feed Rate

Crude Unit 1 deltaFeed



Combined Crude deltaFeed

Feed Cost Saving from Optimizing the Volumetric Gains

Net Result: Crude Reduction

0 50 100 150 200 250 300 350 400-5

-4

-3

-2

-1

0

1

2

3

4Incremental Changes in Crude Unit Feed Rate




Combined Crude deltaFeed

Average crude

saving is ~ 0.5% for

meeting the same

product demand.

Huge cost saving!

Honeywell


Future Work

A 3-level MPC Cascade for a larger part of the supply chain?

For example: a combination of multiple refineries and fuel/crude depots

Multivariable

MPC Control

Economic

Optimization

Multivariable

MPC Control

Economic

Optimization

Multivariable

MPC Control

Economic

Optimization

Multivariable

MPC Control

Economic

Optimization


Master MPC


Multivariable

MPC Control

Economic

Optimization

Multivariable

MPC Control

Economic

Optimization

Multivariable

MPC Control

Economic

Optimization

Multivariable

MPC Control

Economic

Optimization


Master MPC


Multivariable

MPC Control

Economic

Optimization

Level-3 Master


Multivariable

MPC Control

Economic

Optimization

Multivariable

MPC Control

Economic

Optimization

Multivariable

MPC Control

Economic

Optimization

Multivariable

MPC Control

Economic

Optimization


Master MPC


Honeywell


Closing Remarks

MPC is the control solution of choice in the process industries

Economic optimization is often applied in tandem with MPC

Close to 10 thousand MPC controllers have been implemented in different industries.

Oil refining, petrochemical, pulp & paper, coal-to-fuel, alumina, LNG, shale gas – just to name a few

>100 plantwide optimization solutions have been implemented

A majority of them are for ethylene plants

Estimated benefits delivered: $10 billions in total ($5 billions by Honeywell since 1996)

Solution Operability and User Preference

Monolithic solutions are not desirable

Decentralized control – flexible to turn parts of control ON or OFF when addressing process upsets,

maintenance issues, etc.

Gate-to-gate, plantwide optimization – greater benefits and no need to use any transfer prices

Multiscale MPC cascade solution is being developed, which

Performs gate-to-gate, plantwide control and optimization

It can capture a significantly greater amount of benefits (~$1/barrel)

Leverages the existing planning model (structure)

Streamlines the current workflow from planning, scheduling to APC

Honeywell


Questions?

Documents

Industrial Practice of Model Predictive Control