Honeywell

Industrial Practice of Model Predictive Control*

Workshop at CPC-FOCAPO, Tucson

January 8th, 2017

Joseph Lu
Honeywell

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

From Process Unit to Plantwide Control & Optimization Page 1

 How are MPCs typically justified in process industries?
Honeywell
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

From Process Unit to Plantwide Control & Optimization Page 2

 MPC Applications and Benefits Honeywell
• 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)
Petrochemicals Benefits (/yr)
Ethylene 2-4% Increase in Production
VCM 3-5% Increased Capacity / 1-4% Yield Improvement
Aromatics (50KBPD) 3.4M - 5.3M US$

Chemicals Benefits (/yr)

Ammonia 2-4% Increased Capacity / 2-5% Less Energy/Ton
Polyolefins 2-5% Increase in Production/Up to 30% faster grade transition

Industrial Utilities Benefits (/yr)

Cogeneration/Power Systems 2-5% Decrease in Operating Costs
Pulping Benefits (/yr)
Bleaching 10-20% Reduction in Chemical Usage
TMP (Thermo Mechanical Pulping) $1M-$2M
Oil Refining Benefits ($0.01/bbl) Benefits (US$/yr)
Crude Distillation (150 KBPD) 5-13 2.7M - 7.0M
Coking (40 KBPD) 15-33 2.2M - 4.8M
Hydrocracking (70 KBPD) 13-30 3.3M - 7.6M
Catalytic Cracking (50 KBPD) 13-30 2.4M - 5.4M
Reforming (50 KBPD) 10-26 1.8M - 4.7M
Alkylation (30 KBPD) 10-26 1.1M - 2.8M
Isomerization (30 KBPD) 3-17 .3M - 1.8M

From Process Unit to Plantwide Control & Optimization Page 3

 Typical Payback Time by Industry Honeywell

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
10 to 12 months
Power

From Process Unit to Plantwide Control & Optimization Page 4

 Unit-Wide MPC Control Objectives Honeywell

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
Products:

Gas ($x1)

Raw Gasoline ($x2)

Simplified Process Unit

Operating
Constraints Naphtha ($x3)

Kerosene ($x4)

Feed ($y) Light Gas Oil($x5)

Heavy Gas Oil($x6)

Fuel ($z)
Residue ($x7)

From Process Unit to Plantwide Control & Optimization Page 5

 Model Predictive Range Control
Honeywell
Versatile control needs in MIMO applications Regulatory
 Regulatory control Control
 Reference tracking
 Disturbance rejection / minimum variation
u*  arg min Au  ~
y r  u
2 2
 Traditional SISO control u Q R
 Constraint handling
 Prevents another set of variables from exceeding their limits
 Transition control 120 120

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

100 100

 Role of individual output can change

80 80

penalty
penalty
60 60

 Degrees of freedom 40 40

20 20

Model Predictive Range Control 0

-2 -1.5 -1 -0.5 0 0.5
control variable
1 1.5 2
0
-2 -1 0
control variable
1 2

~
 Output prediction: y  Au  y
 The set-point y  r is changed into a ―set-range‖ y LO  y  y HI u*  arg min Au  ~
y s  u
2 2

u , s Q R
 A slack variable s is introduced in the MPC formulation
y LO  s  y HI
 Additional degrees of freedom are used for 120

100 Range
 Local disturbance rejection/absorption 80
Control
 Optimization: e.g., y LO  y  y HI
penalty

min y; 60

40

(independent tuning vs. control response) 20

0
-2 -1 0 1 2
control variable

From Process Unit to Plantwide Control & Optimization Page 6

 Range MPC Design and Solution Honeywell
 Performance Specification (via a funnel): Range control benefits
 Robustness
prediction
horizon  Minimum effort control (when the control solution is non-unique)
correction
horizon controlled
variable
 Smoother control, minimum high frequency excitation
high limit

low limit Range control algorithm

manipulated
 Two-stage analytical approach
variable a) solve the range control problem

u*  arg min Au  ~

y s  u R ; y LO  s  y HI
2 2
T s
u , s Q
current time
time b) If soln to (a) is not unique, solve for the minimum effort control

u*  arg min u yLO  Au  ~

y  yHI
2
;
prediction u R
horizon
correction
controlled
set-point
horizon
variable  Online algorithm similar to the active-set QP: (here assume R = 0)

u*  arg min Au  w Q ; y LO  ~

y  w  y HI  ~
2
y
manipulated u , w
variable
2
 Aact   wact 
u*  arg min  u  
T
 
s

u , w
current time  A free   w free  Q
time

 U k RkVkT
1
 When the solution is not unique, a
(k )
Let Q 2 Aact
secondary optimization can be performed
arg min U k RkVkT u  wact
2
u* 
to find the minimum effort solution. u 2

From Process Unit to Plantwide Control & Optimization Page 7

 Feed
Steam

Dewaxed Oil
WFO Atmos WFO Press
Flash Tower Flash Twr

Htr
Mix
WFO
Reflux
Reflux
Double Double
Pipe Pipe
Exchanger Exchanger

WFO Stripper WFO Final

Flash Twr
Double Double
Refrig

Pipe Pipe
Exchanger Exchanger
Wet

Reflux

Reflux

Steam

From Process Unit to Plantwide Control & Optimization

Dry

Vacuum
WFO
Dehydrator
Double Double
Pipe Pipe
Chiller Chiller
Ketone
Stripper
Slop
Refrig

Solvent
Refrig

Wax
Steam
Refrig

Reflux

Reflux

Steam
Refrig

PW Final Double
PW Stripper Pipe
Flash Twr Double
Chiller
Pipe
Chiller
Reflux

Primary
Reflux

Repulp
Steam

Filt Feed Filt Feed

PW Atmos PW Press
Flash Twr Flash Twr
Non-Square Control Example—Solvent Dewaxing Unit

Filtrate
Repulp
Filtrate
Primary

Page 8

PW Mix
Surge
Honeywell
 Solvent Dewaxing Unit Illustration Honeywell

Tank 1: Mixture with solvent

Tank 2: Cooled mixture (wax slurry)
Tank 3: Chilled solvent (wax slurry)

Filtering
Wax + Oil Wax
Mixture
1 2 3

Solvent

Lubricant Oil
6 5 4

Tank 4: Filtrate with wax removed

 Solvent Inventory Control Tank 5: Wet solvent (oil removed)
Tank 6: Dried solvent for reuse

From Process Unit to Plantwide Control & Optimization Page 9

 Model Dynamics of Solvent Control Honeywell

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

Primary Repulp Slop Solvent Slack Wax 5th Dilution
Filtrate Flow Kickback Flow Flow Flow Flow Ratio
Primary 1 1 1
Filtrate Level  1.42e 3 1.42e 3 1.13 e 10 s
s s s
Repulp Filtrate 1 1
Level  1.39e 3  1.11
s s
Dry Solvent 1
Drum Level 1.6e 3 e 5 s
s
Moist Solvent 1 1 1 1
Drum Level 1.62e 3 e 5 s 2.16e3 1.08e 3 e 5 s  0.173
s s s s
Slop Solvent 1
Tank Level  1.23e 3
s
Slack Wax 1
Tank Level 1e4
s

From Process Unit to Plantwide Control & Optimization Page 10

 Simulation Results Honeywell
Controlled Variables
15

10
4
5
3 5
0
6 2
-5

-10
1
-15
0 100 200 300 400 500 600 700 800 900 1000

Manipulated Variables
100%
Time[Min]
2 5
50
3
1
0

4
-50

5
-100%
0 100 200 300 400 500 600 700 800 900 1000
Min

Filter Wash Grade Change

From Process Unit to Plantwide Control & Optimization Page 11

 Pulp and Paper Example: Thermomechanical Pulping (TMP) Refiner
Honeywell
TMP refiner – a naturally non-square 4x5 interacting system
• Objective: consistent quality from coordinated control (decoupling)

Customer estimated benefits: ~$13.75M/Year

From Process Unit to Plantwide Control & Optimization Page 12
 Ethylene Plantwide Control and Optimization
Fig 1. Aerial photo of an ethylene plant
Honeywell
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.

Process and Operating Characteristics

Universal: ETHYLENE PROPYLENE
► No product blending REFRIG REFRIG

► Stringent quality requirements ETHYLENE

► Slow dynamics from gate to gate COLD HYDROGEN
► Gradual furnace and converter coking

CAUSTIC

ETHYLENE
BOX

SPLITTER
OIL FUEL GAS
WATER
► Furnace and converter decoking FURNACES QUENCH QUENCH ACETYLENE
RECOVERY
► Frequent furnace switching UNIT

DEMETHANIZER

DEETHANIZER
Site specific: ETHANE
► Feed quality variations Feed
System CGC
COCRACK

► Product demand changes MAPD

CONVERTER PROPYLENE
► Sensitivity to ambient conditions

CONDENSATE

DEPROPANIZER
DISTILLATE
STRIPPER

PROPYLENE
STRIPPER

SPLITTER
► Periodic switching (e.g., dryers)
Advanced Control and Optimization Goals
► Stabilize operation MIXED

DEBUTANIZER
DEBUTANIZER
C4
► Minimize product quality give away
PROPANE
► Maximize selectivity & yield TCG COCRACK
► Minimize converter over-hydrogenation
► Minimize ethylene loss to methane/ethane recycle. Profit
Controllers Fig 2. Process overview – flow diagram
TCG

From Process Unit to Plantwide Control & Optimization Page 13

 Profit Controller Envelopes
Honeywell
RMPCT 11 RMPCT 13

C3R
C2R
RMPCT 9
Cold

Quench
Box

C2H2

C2 Splitter
Converter

RMPCT 1 RMPCT 12
Fractionator

RMPCT 15

DeC1

DeC2
RMPCT 2

C3 Splitter
DeC3
RMPCT 8 RMPCT 10
MAPD
Converter
RMPCT 14

From Process Unit to Plantwide Control & Optimization Page 14

 Plantwide Dynamic Optimization
Honeywell
Profit Optimizer
RMPCT 13
RMPCT 11

C2R

C3R
RMPCT 9
Cold

Quench
Box

C2H2
Converter

P
RMPCT 1 RMPCT 12
Fractionator

DeC1

DeC2
RMPCT 2

C3 Splitter
DeC3
RMPCT 10
RMPCT 8
MAPD
Converter
RMPCT 14
RMPCT 15

From Process Unit to Plantwide Control & Optimization Page 15

 Typical Results of Ethylene Plant APC Honeywell
• Execution Frequencies: Production is increased
- MPC Controllers: ½ to 1 minute Normalized Olefin Production Rate Trend and product variability
- Global Optimizer: 1 minute reduced...
110
- Gain updating: 5 minutes
108
• Typical Benefits 106 Before APC

- Production Increase: $1.5-3Millions 104

102
- Energy Savings: Additional After APC
100
98
Implementation Details
96
Decoking campaigns; one furnace out of service
 Design 94
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 92
gain mapper, which is a Profit Bridge app that calls the UniSim model in real
90
time to update the gains in both Profit Controllers and Profit Optimizer once
every 5 mins.

Ethane Impurity Content (in ppm) in Product Ethylene

Tighter control and
 Modeling
Dynamic controller models are identified, in either open loop or closed loop,
less quality give-away
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 +600
calculations are required for gain updating.

+500
Commissioning
Profit Controllers are commissioned first to provide stable unit operations. Profit
Bridge is then commissioned to provide gain updating for the furnace controllers.
+400
Next, Profit Optimizer is commissioned, and Bridge Model predictions are
verified against data. Finally, a plantwide objective function is populated and +300
optimization moves are validated. Continuous operation of the controllers and
optimizer starts after operator training. +200

Maintenance +100
Typically little maintenance is required to keep the ethylene APC/Optimization
running. Clients either dedicate a parttime control engineer to monitor and Base Before APC After APC
perform minor services as needed, or depend on Honeywell’s BGmax quarterly-
visit services.

From Process Unit to Plantwide Control & Optimization Page 16

 Nontraditional Space: Discrete Manufacturing
Honeywell

Semiconductor 3.1
Asm1 28
4.1
Test1 33
5.1
Fuse1
Box1
Multi-factory 34 37
7.1

18 3.2
Supply Chain B 39 43 7.2
20 24 6.1 40
7.4 7.3 44

vend
11 vend8
3.3 pp 3.4 pp

7
Fab1 7 Sort1 41 pp pp 46
2.1 6.2
P1 P1 vendor3 vendor4 C 42 45
vendor1 21 7.5
12 25
1
1.1 si 3.5 35 38
2 A 19
3 13 7.6 Box2
Fab2 8 Sort2 Asm2 29 Test2 Fuse2
14 4.2 5.2
2.2 3.6 36
4
P1,P2 P1,P2
15
Semiconductor30 F
5
1.2 si 3.5 3.7

vendor2
6
16
Blue = Intel
Fab3 9 Sort3
2.3
P2
vendor5 22 26 vendor6
Red = Mat. Sub.
P2 3.8 pp 3.9 ram
10
17
27
Green = Cap. Sub.
23
= Mats Mfg
3.10 3.11 31
= Inv Hold D

= Prod Mfg 3.12 Asm3 4.3 Test3 6.3

32
= Transport
Courtesy of Karl Kempf – Intel
From Process Unit to Plantwide Control & Optimization Page 17
 Semiconductor Discrete manufacturing – Sorting Assembly
Honeywell

wafer

test
sort
C1 Fab/Sort
M1 T1
I10 Die/Package Inventory
Simple Supply
C2 Assembly/Test
Chain Sequence M2 T2
Semi-Finished Goods Inv.
I20
Finish/Pack
1.2 0.25 C3
TPT Distribution
1
0.2
M3 T3 Components Warehouse
0.8 I30 Shipping
Data
0.15
Model
0.6 C4
DataPDF
0.1 Demand
0.4 ModelPDF M4 T4
0.05
(Orders)
0.2

0 0
D1 t
D3 D2

From Process Unit to Plantwide Control & Optimization Page 18

 MPC Controller Design
Honeywell

 One controller per product FAB

 124 controlled variables TRDI

ADI DIE PREP
TRDI
TRDI
TRDI
TRDI
TRDI
TRDI DEMAND
SORT TRDI
TRDI 01
 TRDI
ADI DIE PREP TRDI
TRDI
5 fab/sort End-of-Line Inventories 01 01

 7 Assembly Die Inventories (ADI)

 7 die prep capacities (Work-in-Progress)
 56 Tape and Reel Die Inventories (TRDI) ADI
TRDI
TRDI
TRDI
TRDI
TRDI
02 DIE PREP TRDI DEMAND
 48 ADI + Die-Prep + TRDI inventories
SORT
 90 manipulated variables (Green) 02

 35 ships from sorts

FAB
 48 ships from a hub TRDI to assembly test ADI
TRDI
TRDI
TRDI
TRDI
TRDI
03 DIE PREP TRDI DEMAND
manufacturing (ATM) sites
 7 starts from ADI SORT
03
 293 feedforward variables (Blue)
 5 sort supply w/forecasts FAB
ADI
TRDI
TRDI
TRDI
TRDI
DIE PREP TRDI
TRDI DEMAND
 48 ATM TRDI demands w/forecasts 04

 240 TRDI cross-ships w/forecasts

SORT
 Current system design consists of 34 such 04

controllers (products) per computer node FAB

ADI
05
DIE PREP
TRDI
TRDI
TRDI
TRDI
TRDI
TRDI DEMAND

 One dynamic optimizer covers all 34 MPC

controllers. A model predictive control application for inventory and production
management.

From Process Unit to Plantwide Control & Optimization Page 19

 Chip Inventories at an Unnamed Site Honeywell
Die xxx_06 Inventory
3000
High Limit
2500
APC used
2000
K Die

Inventory
1500
1000
500
Low Limit
0
5/21/06 5/28/06 6/4/06 6/11/06
Die xxx_27 6/18/06
Inventory 6/25/06 7/2/06 7/9/06
3000 Date
High Limit
2500
2000
K Die

Inventory
1500 APC used
1000 Low Limit
500
0
5/21/06 5/28/06 6/4/06 6/11/06
Die xxx_47 6/18/06
Inventory 6/25/06 7/2/06 7/9/06
1500 Date
1250
High Limit
1000
K Die

750
APC used
Inventory
500
250
Low Limit
0
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
From Process Unit to Plantwide Control & Optimization Page 20
 Refinery-wide (Production) Control/Optimization
Honeywell
Production Plan
Supply Side (monthly) Customer Demand
(upstream What Feed to buy? What products to sell? (downstream
supply chain) supply chain)

Average daily production Generate a month-plan Average daily delivery

based on an average model

Production Schedule
(weekly)
What Feed to use? What products to deliver? Product to
Feed Received be Deliveried

Translate the plan into production

activities to satisfy the plant logistics

A set of production schedules Blending/shipment schedules

(e.g., feed blending schedule) Master MPC
Economic
Optimization

Multivariable
APC
Control

Economic Economic Economic

Optimization
Human Coordination
Optimization Optimization

Multivariable Multivariable Multivariable

Control Control Control

Slave MPC 1 Slave MPC 2 Slave MPC n

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

Raw Materials  Execution Products 

From Process Unit to Plantwide Control & Optimization Page 21

 (Multiscale) MPC Cascade Structure
Honeywell
Master MPC
Economic
Optimization
Multivariable
MPC Control

two-way connection two-way connection

Economic Economic Economic

Optimization Optimization Optimization
Multivariable Multivariable Multivariable
MPC Control MPC Control MPC Control
Slave MPC 1 Slave MPC 2 Slave MPC n

 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

From Process Unit to Plantwide Control & Optimization Page 22

 Refinery-wide Production Control : JIT manufacturing
Honeywell
Blending Components
Straight-run products

Straight-Run and/or Blending

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

From Process Unit to Plantwide Control & Optimization Page 23

 JIT manufacturing viewed as a control problem
Honeywell
Scope: the distillate production pool in an oil refinery
 It is simpler but still captures most of the important issues

Crude Imports
Unit Light Moving Avg
1 Distillate daily delivery:
Crude Feed 1 Hydrotreater
Component
Tank 1
5 KBarrels
Vac
Unit
8 KBarrels
1
Crude
Component
Unit Heavy Tank 2
2 Distillate 10 KBarrels
Hydrotreater
Crude Feed 2 7 KBarrels
Component
Tank 3
10 KBarrels
Crude Vac
Unit Unit Hydrocracker Blending
3 2
Crude Feed 3

Material Balance &

Inventory Control
From Process Unit to Plantwide Control & Optimization Page 24
 Just-in-Time Manufacturing – Quality Control (1)
Honeywell
Quality Control - Sulfur Legend:
Control
handles

50~300
Crude Imports
Sweet ppm S Light
Unit
1 Distillate
Hydrotreater
4~20 ppm S
0.1% Sulfur
Component
Tank 1
12 ppm S
Vac
Unit
10 ppm S
1
Crude 100~400 Component
Medium Unit ppm S Heavy 4~20 ppm S Tank 2
2 Distillate 10 ppm S
0.3% Sulfur Hydrotreater
10 ppm S
Component
Tank 3
10 ppm S

Crude Vac
Sour Unit Unit Hydrocracker 1~2 ppm S Blending
3 2 100~600
1.2% Sulfur ppm S

From Process Unit to Plantwide Control & Optimization Page 25

 Just-in-Time Manufacturing – Quality Control (2)
Honeywell
Quality Control - Cloud Point Control
Legend: handles

Crude Jet Cloud Imports

Sweet Unit Jet Cut -30 to -40 ⁰C Light Cloud Spec
1 Distillate = Feed Cloud (winter)
Stove Cut
Hydrotreater
Component -36 ⁰C
Tank 1
Vac
Unit Stove Cloud -44 ⁰C
1 -5 to -10 ⁰C
Crude
Component
Medium Unit Jet Cut Heavy Tank 2
-40 ⁰C
2 Distillate
Hydrotreater -5 to -45 ⁰C
Stove Cut
-30 ⁰C
Component
Tank 3
-42 ⁰C
Crude Jet Cut Vac
Sour Unit Unit Hydrocracker -40 to -55 ⁰C Blending
3 Stove Cut 2 Gasoils

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

From Process Unit to Plantwide Control & Optimization Page 26

 Control Model Matrix for the Master MPC
Honeywell
1) Unit feed rates 1) Demands
2) Control factors that affect the product quality 2) Others

Manipulated Variables (MVs) DVs

MVs Crude1 Crude2 Crude3 Hydrocracker Cr Unit Hydrotreater Product-i

CVs Feed Feed Feed Conv Cuts Sulfur delivery

CompTank2.Level
Production amount

Controlled Variables (CVs)

CompTank2.Sulfur
Product quality

CompTank2.Cloud

CompTank3.Level

CompTank3.Sulfur

CompTank3.Cloud
1)
2)

More CVs…

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

k  s k
 Often simple dynamics such as
 s

as  bs  1
2
e or s (s  1)
e would suffice
 Notice that there is no setpoint in this control problem – just constraint control

From Process Unit to Plantwide Control & Optimization Page 27

 Master MPC for JIT Manufacturing in closed loop?
Honeywell
MVs Crude1 Crude2 Crude3 Hydrocr Cr Unit Hydrot Produc
Feed Feed Feed acker Cuts reater t-i

CVs
Conv Sulfur deliver
y
Master MPC
Tank2.Level
Economic
Tank2.Sulfur Optimization
Tank2.Cloud Multivariable
Tank3.Level MPC Control
Tank3.Sulfur

Tank3.Cloud -

More CVs…

Economic Economic Economic

Optimization Optimization Optimization
Multivariable Multivariable Multivariable
MPC Control MPC Control MPC Control

Slave MPC 1 Slave MPC 2 Slave MPC n

 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 MV1
110
MV2
5
MV3
16
MV4
0
MV5
105
MV6
10.4
MV7
0.68
MV8
27.4
Feed
33.5

more Slave MPCs to wind up (i.e. sustained CV constraint CV1

89.3

violations). CV2

 The Real Challenge:

7.23

CV3

 How to get the Master to understand the Slave’s constraints 89.3

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

CV4
•
•
•

From Process Unit to Plantwide Control & Optimization Page 28

 Master-Slave Models: a side-by-side view Honeywell
 Below is an illustration of the model structure for 2 combined units:
Different Tasks
1 2 3 4 Master MV/DV index

Plantwide Production Control

Master MPC • Inventory control
Master CVs

• Quality control
• Cost and profit optimization
Conjoint
MVs
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
Slave MV index
Slave1 CVs

Slave MPC 1 Free MVs

Unit Process Control

• Safe/smooth operation
• Unit efficiency
• Intermediate Quality
Slave2 CVs

• Local Optimization
Slave MPC 2

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

From Process Unit to Plantwide Control & Optimization Page 29

 The Concept of Proxy Limit – An Illustration for 1 Conjoint Variable Case
Honeywell
Steps in the hydrocracker example: Refinery Blending Components
Straight-run products
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

Straight-Run and/or Blending

Slave to maximize the feed, subject to
the Slave constraints
4) The call returns the following:
MV3 = 33
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.
Hydrocracking
Opt Value: 38 4.5 2.3 10 40 10.7 1.25 28.3 114
 g yd ( s )1 
 g (s) 
Feed MV2 MV3 MV4 MV5 MV6 MV7 MV8 MV9
Opt 33 5 16 0 105 10.4 0.68 27.4 110
Value
100 CV1  yd 2 
89.3
 g yd ( s ) 3 
6.20 CV2
7.23 Feed  
75.0 CV3    Products
89.3
 g yd ( s ) n 
CV4  
•
•
•

From Process Unit to Plantwide Control & Optimization Page 30

 Approximate the feasible region defined by the Slave constraints
for 2 (or more) conjoint variables Honeywell
Master
CMV(k+1)
MV1 Low Lmt
MV1 High

MV 2 High

Master MVs
k k+1

Current Operating Point (0,0)

MV 2 Low

Conjoint MVs Free MVs

Master CMV(k)

8 inquiry calls:
min c t xconj
x
Call # ct dt
1 -1 0 0 1 bLO  Ac xconj  A f x free  bHI
2 1 0 0 1
3 0 -1 1 0 d t xconj  0 Slave MPC (i)
4 0 1 1 0
5 -1 1 1 1
6 1 -1 1 1
7 1 1 1 -1 Where, A 
G 
8 -1 -1 1 -1  I 
From Process Unit to Plantwide Control & Optimization Page 31
 Proxy Limits – Customer Example Honeywell
To what extent is the Master MPC allowed to change the Slave unit’s operations?

Hydrocracker Conversion - Actual vs Maximum

100.000

95.000

90.000

85.000

80.000

75.000

70.000
57
15
22
29
36
43
50

64
71
78
85
92
99
1
8

106

260

309
113
120
127
134
141
148
155
162
169
176
183
190
197
204
211
218
225
232
239
246
253

267
274
281
288
295
302

316
323
330
337
344
351
358
365
From Process Unit to Plantwide Control & Optimization Page 32
 Proxy Limits – Customer Example (2) Honeywell

Heavy Distillate HTR Feed Rate - Actual vs Maximum

15
22
29
36
43
50
57
64
71
78
85
92
99
1
8

134

169

204

239
106
113
120
127

141
148
155
162

176
183
190
197

211
218
225
232

246
253
260
267
274
281
288
295
302
309
316
323
330
337
344
351
358
365
From Process Unit to Plantwide Control & Optimization Page 33
 Master’s Control and Optimization Results:
Component Quality Changes in Tank #3 Honeywell
Tank3.Cloud Point in Deg C Tank3.Flash Point in Deg C
-40

-41 58 Lower flash point

Higher cloud point -42
means that more
means that more 56 low-cost (lighter)
-43
low-cost (heavier) components are
54
components are -44 used to produce
used to produce the -45 52
the same products
same products -46
50
-47

-48 48

50 100 150 200 250 300 350 50 100 150 200 250 300 350
Sulfur is controlled at
Tank3.Qualities (After - Master MPC)
its high limit which
Historical Operation Data (Before)
means much lower
Tank3.90pct Point in Deg C Tank3.Sulfur in PPM
11 catalyst/utility costs
10
260
9

Higher 90pct point 8

means that more 255
7
low-cost (heavier)
components are 250 6
used to produce the 5
same products
245 4

3
50 100 150 200 250 300 350 0 100 200 300 400

Component production costs are significantly reduced!

From Process Unit to Plantwide Control & Optimization Page 34
 Feed Cost Saving from Optimizing the Volumetric Gains
Honeywell
Net Result: Crude Reduction
Incremental Changes in Crude Unit Feed Rate
4
Crude Unit 1 deltaFeed
Crude Unit 2 deltaFeed
3 Crude Unit 3 deltaFeed
Combined Crude deltaFeed

Average crude
saving is ~ 0.5% for -1
meeting the same
product demand.
-2
Huge cost saving!

-3

-4

-5
0 50 100 150 200 250 300 350 400

From Process Unit to Plantwide Control & Optimization Page 35

 Future Work
Honeywell
 A 3-level MPC Cascade for a larger part of the supply chain?
 For example: a combination of multiple refineries and fuel/crude depots

Level-3 Master

Economic
Optimization

Multivariable
MPC Control

two-way connection
two-way connection

Master MPC Master MPC Master MPC

Economic
Economic Economic
Optimization
Optimization Optimization

Multivariable Multivariable Multivariable

MPC Control MPC Control MPC Control

two-way connection two-way connection two-way connection two-way connection two-way connection two-way connection

Economic Economic Economic Economic Economic Economic Economic Economic Economic

Optimization Optimization Optimization Optimization Optimization Optimization Optimization Optimization Optimization

Multivariable Multivariable Multivariable Multivariable Multivariable Multivariable Multivariable

Multivariable Multivariable
MPC Control MPC Control MPC Control MPC Control MPC Control MPC Control MPC Control
MPC Control MPC Control
Slave MPC 1 Slave MPC 2 Slave MPC n Slave MPC 1 Slave MPC 2 Slave MPC n Slave MPC 1 Slave MPC 2 Slave MPC n

From Process Unit to Plantwide Control & Optimization Page 36

 Closing Remarks Honeywell
 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

From Process Unit to Plantwide Control & Optimization Page 37

 Honeywell

Questions?

From Process Unit to Plantwide Control & Optimization Page 38