Upload
others
View
3
Download
0
Embed Size (px)
Citation preview
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
From Process Unit to Plantwide Control & Optimization Page 2
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
From Process Unit to Plantwide Control & Optimization Page 3
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-$2M
Honeywell
From Process Unit to Plantwide Control & Optimization Page 4
Typical Payback Time by Industry
Typical Payback Time:
Oil and Gas 1 to 2 months
Refining 3 to 6 months
Chemicals
Petrochem 4 to 6 months
MMM
Pulping/Paper 6 to 8 months
Industrial
Power 10 to 12 months
Honeywell
From Process Unit to Plantwide Control & Optimization Page 5
Unit-Wide MPC Control Objectives
Feed ($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
From Process Unit to Plantwide Control & Optimization Page 6
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
From Process Unit to Plantwide Control & Optimization Page 7
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
From Process Unit to Plantwide Control & Optimization Page 8
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
From Process Unit to Plantwide Control & Optimization Page 9
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
From Process Unit to Plantwide Control & Optimization Page 10
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
From Process Unit to Plantwide Control & Optimization Page 11
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
From Process Unit to Plantwide Control & Optimization Page 12
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
From Process Unit to Plantwide Control & Optimization Page 13
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
From Process Unit to Plantwide Control & Optimization Page 14
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
From Process Unit to Plantwide Control & Optimization Page 15
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
From Process Unit to Plantwide Control & Optimization Page 16
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
From Process Unit to Plantwide Control & Optimization Page 17
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
From Process Unit to Plantwide Control & Optimization Page 18
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
From Process Unit to Plantwide Control & Optimization Page 19
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
TRDI TRDI TRDI TRDI TRDI TRDI ADI 03 DEMAND
DIE PREP
TRDI TRDI TRDI TRDI TRDI TRDI ADI 04 DEMAND
DIE PREP
TRDI TRDI TRDI TRDI TRDI TRDI ADI 05 DEMAND
A model predictive control application for inventory and production management.
Honeywell
From Process Unit to Plantwide Control & Optimization Page 20
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
Die xxx_27 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
Die xxx_47 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
From Process Unit to Plantwide Control & Optimization Page 21
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
Economic Optimization
Multivariable Control
Economic Optimization
Multivariable Control
Slave MPC 1 Slave MPC 2 Slave MPC n
Economic Optimization
Multivariable Control
Master MPC
Human Coordination
APC
A set of production schedules
(e.g., feed blending schedule) Blending/shipment schedules
Honeywell
From Process Unit to Plantwide Control & Optimization Page 22
(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
Slave MPC 1 Slave MPC 2 Slave MPC n
Master MPC
two-way connection two-way connection
Honeywell
From Process Unit to Plantwide Control & Optimization Page 23
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
From Process Unit to Plantwide Control & Optimization Page 24
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
From Process Unit to Plantwide Control & Optimization Page 25
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
From Process Unit to Plantwide Control & Optimization Page 26
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
From Process Unit to Plantwide Control & Optimization Page 27
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
From Process Unit to Plantwide Control & Optimization Page 28
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
Slave MPC 1 Slave MPC 2 Slave MPC n
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
From Process Unit to Plantwide Control & Optimization Page 29
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
From Process Unit to Plantwide Control & Optimization Page 30
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
From Process Unit to Plantwide Control & Optimization Page 31
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
From Process Unit to Plantwide Control & Optimization Page 32
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
From Process Unit to Plantwide Control & Optimization Page 33
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
From Process Unit to Plantwide Control & Optimization Page 34
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
From Process Unit to Plantwide Control & Optimization Page 35
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
Crude Unit 2 deltaFeed
Crude Unit 3 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
Crude Unit 1 deltaFeed
Crude Unit 2 deltaFeed
Crude Unit 3 deltaFeed
Combined Crude deltaFeed
Average crude
saving is ~ 0.5% for
meeting the same
product demand.
Huge cost saving!
Honeywell
From Process Unit to Plantwide Control & Optimization Page 36
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
Slave MPC 1 Slave MPC 2 Slave MPC n
Master MPC
two-way connection two-way connection
Multivariable
MPC Control
Economic
Optimization
Multivariable
MPC Control
Economic
Optimization
Multivariable
MPC Control
Economic
Optimization
Multivariable
MPC Control
Economic
Optimization
Slave MPC 1 Slave MPC 2 Slave MPC n
Master MPC
two-way connection two-way connection
Multivariable
MPC Control
Economic
Optimization
Level-3 Master
two-way connection two-way connection
Multivariable
MPC Control
Economic
Optimization
Multivariable
MPC Control
Economic
Optimization
Multivariable
MPC Control
Economic
Optimization
Multivariable
MPC Control
Economic
Optimization
Slave MPC 1 Slave MPC 2 Slave MPC n
Master MPC
two-way connection two-way connection
Honeywell
From Process Unit to Plantwide Control & Optimization Page 37
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