Economic Optimization

The Next Generation in Control

Standards David March

Certification
Education & Training Process Capital Productivity
Publishing
Conferences & Exhibits Specialist 1
RE-THINK

Traditional Control Philosophy is a

constraint to your process control
efficiency and your profitability.

You need to RE-THINK the basics of

Process Control.

#
 What Are We Trying to Achieve?
Increased Profitability
Minimize operational variance

DV CV

Transition effectively SPECIFICATION OR LIMIT

KEY TARGET

CV

WITH
Follow operational plan BEFORE APC WITH APC OPTIMIZATION

TIME

CV
Non-Symmetrical Economics

#
There Are 2 Parts:

• First. We must control and minimize

process variance.

• Second. Drive (control) the process

by the economics of the process and
business needs, not by functional
parameters.

#
The Dirty Little Secret

Control Systems do not control a process.

They create a box.

The
Process
Runs
Inside
the Box.
#
 How Our Time is Spent:
PID Rodeo

Most of our work and effort is focused on controlling the control.

#
 Process Responses #1 (Very High Feed Rate)
Recipe requirements: External
Blended Mix Temperature 110 - 140 Disturbance:
Setpoint degF
Ambient Temperature 85
Dry Mix Feed Rate Setpoint 50 lb/minute degF
Water Flow Rate Setpoint 2.56
lb/minute
degF degF
Temperature Setpoint degF

Temperature (PID) degF

100.000 %
%

Temperature (Supervised PID)

Steam Valve Manipulation (PID)

Steam Valve Manipulation (Supervised PID)

#
 Process Responses #2 (High Feed Rate)
Recipe requirements: External
Blended Mix Temperature 110 - 140 Disturbance:
Setpoint degF Ambient Temperature 85
Dry Mix Feed Rate Setpoint 35 lb/minute degF
Water Flow Rate Setpoint 2.56
lb/minute
degF degF
Temperature Setpoint degF

Temperature (PID) degF

%
%

Temperature (Supervised PID)

Steam Valve Manipulation (PID)

Steam Valve Manipulation (Supervised PID)

#
 Process Responses #3 (Low Feed Rate)

Recipe requirements: External

Blended Mix Temperature 110 - 140 Disturbance:
Ambient Temperature 85
Setpoint degF degF
Dry Mix Feed Rate Setpoint 20 lb/minute
Water Flow Rate Setpoint 2.56
lb/minute
degF degF
Temperature Setpoint degF

Temperature (PID) degF

%
%

Temperature (Supervised PID)

Steam Valve Manipulation (PID)

Steam Valve Manipulation (Supervised PID)

#
 Process Responses #4 (Very Low Feed Rate)

Recipe requirements: External

Blended Mix Temperature 110 - 140 Disturbance:
Ambient Temperature 85
Setpoint degF degF
Dry Mix Feed Rate Setpoint 10 lb/minute
Water Flow Rate Setpoint 2.56
lb/minute
degF degF
Temperature Setpoint degF

Temperature (PID) degF

%
%

Temperature (Supervised PID)

Steam Valve Manipulation (PID)

Steam Valve Manipulation (Supervised PID)

#
MPC: A Different Philosophy

Traditional Philosophy: Control where it is now and it will get to where

we need it to go. No response until something causes a shift away
from target (not predictive). Each loop acts as an island (box)
controlling a simple, isolated objective (not systematic).

MPC Philosophy: Modify the trajectory of where it is heading so it will

converge to where we need it to go. Effective and highly efficient.
Predict where it is going, respond to disturbances fed forward as
their predicted response, control and coordinate a system of
interacting measurements, leverage dynamic models directly.

MPC Is Based on GOAL Seeking Optimization (typically

economic)

#
 Model Predictive Control

Driving By Looking Forward

Rather Than In The Rear
View Mirror

#
Deep Blue

Think of the process like the way a computer plays chess.

Computer Chess takes the current state of the board and
looks at all combinations of future moves and chooses
the best.

MPC applies the current state to the process and the model
of the process (how it works/responds), to see where it is
going and then plans a series of small changes to guide
the process’s calculated trajectory to the goal.

#
What?

Yes, We Are Upside Down and

Backwards to traditional control.

We control where it is going rather than

where it is!

Guiding the process to the economic

goal!

#
“Controlling the Trajectory Reduces Variability

Weight fraction

Manual
PID/Advanced Control
MPC

Product Property
• Reduction in variability
• Reduction in off-spec More Profit
• More Consistent Products
Think about the Rat. You don’t completely control the Rat by confining it
#
First Comes the Model
Porsche vs Geo Metro
static model captures "steady-state" operating conditions ...
steady-state response
Traditional Control
Static Model
(Steady-State)

no change in response

Accelerator Position Predicted Speed

Porsche dynamic response

action no longer changing

Geo Metro dynamic response

Dynamic Model
Predictive Control

dynamic response still changing

You don’t drive a Porsche and a Geo the same way, unless you want to die
#
 First You Develop A Model of The Process

• Multivariable
– Control strategy based on a cross
correlated matrix of key process
variables (The gofer game. Manipulate
one and others respond)
SPECIFICATION OR LIMIT

KEY TARGET
• Predictive
– Math based models developed
through process testing
– Control moves mapped to “drive” the
process to the optimum through an BEFORE MODEL WITH MODEL PREDICTIVE CONTROL
PREDICTIVE CONTROL AND OPTIMIZATION
anticipated process control horizon
TIME

• Adaptive
– Control is dynamic, utilizing process
feedback for continuous performance
enhancement

An Effective Model Must Have Contain All 3

Attributes #
Example: Combustion Optimization Solution

Complex Relationships of Multiple Variables Are

Continuously Changing … Dramatically Impacts Combustion
Ambient Conditions
MW Demand
Uncontrolled Influences Coal quality

Operators

Emissions (CO, NOx)

Excess Air
Boiler Efficiency (Heat Rate)
Coal flow Distribution Boiler
Control System Unburned carbon-in-ash
Air flow Distribution
Process Temperatures
Burner tilts, PA etc.
Mill Performance

Process Setpoints Desired Objectives

#
MPC Process

#
Building The Model

• A Model Represents “Knowledge”

– A model emulates the behavior of the process
– A model provides predictive capability through “computational
experimentation”

• Model Types
– Static vs. Dynamic
– Linear vs. Non-linear
– First principle vs. Empirical, “Hybrid”

#
Modeling Challenges for Optimization

Predictive
PredictiveAccuracy
Accuracy Computational
ComputationalEfficiency
Efficiency
•• Nonlinear
NonlinearInteractions
Interactions •• Fast
FastSolution
Solution
•• Process
ProcessDynamics
Dynamics •• Deterministic
DeterministicResults
Results

Conflicting
ConflictingRequirements
Requirements
•• More
Moreaccurate
accuratemodels
models
Less

Lesscomputationally
computationallyefficient
efficient
•• More
Morecomputationally
computationallyefficient
Less
efficient
Lessaccurate
accuratemodels
models

Desired
DesiredModel
ModelFeatures
Features
•• Must
Mustsystematically
systematicallyincorporate
incorporateempirical
empiricaldata,
data,first-
first-
principles
principlesmodels,
models,and
andprocess
processknowledge
knowledge
•• Must
Mustremain
remainphysically
physicallymeaningful
meaningful
•• Must
Mustallow
allowrobust
robustonline
onlineadaptation
adaptation(facilitating
(facilitatingmodel
model
deployment,
deployment,maintenance)
maintenance)
#
Linear vs. Non-Linear Models

Linear Models Non-Linear Models

capture "straight-line" relationships capture more complex relationships

output (y)
output (y)

input (x) input (x)

• Mathematically simple • Mathematically sophisticated

• Over simplification • Accurate representation
• Localized representation • Global representation

#
Modeling Paradigms

First
FirstPrinciples
PrinciplesModeling
Modeling Empirical
EmpiricalModeling
Modeling
•• Leverage
Leverageexplicit
explicitknowledge
knowledge •• Leverage
Leverageimplicit
implicitknowledge
knowledge
based
basedononscientific
scientificprinciples
principles based
basedon
onhistorical
historicaldata
data
•• Strengths
Strengths •• Weaknesses
Weaknesses
•• Global
Globalvalidity
validity ••Valid
Validfor
forobserved
observedoperation
operation
(good
(goodextrapolation)
extrapolation) (poor
(poorextrapolation)
extrapolation)
•• Parameters
Parametershave
havephysical
physical ••Parameters
Parametersdo donot
nothave
havephysical
physical
meaning
meaning meaning
meaning
•• Not
Notdata
datadependant
dependant ••Subject
Subjecttotopoor
poordata
data
••Weak
Weak(closed-loop)
(closed-loop)data
data
••Correlated
CorrelatedData
Data
•• Weaknesses
Weaknesses •• Strengths
Strengths
•• Typically
Typicallyimplicit
implicit ••Typically
Typicallyexplicit
explicit
•• Slow
Slowevaluation
evaluation(solver)
(solver) ••Fast
Fastevaluation
evaluation(no(nosolver)
solver)
•• Specific
Specificmodeling
modelingframework
framework ••General
Generalmodeling
modelingframework
framework
•• Not
Notalways
alwaysavailable
available ••Wide
Wideapplicability
applicability
•• Complex
Complexdevelopment
development ••Quick
Quick todevelop
to develop
#
Fundamental vs. Empirical Models

Fundamental Models Empirical Models

Explicit knowledge Implicit knowledge

Based on scientific principles Based on historical data

output (y)
∂u 2
∂ 2u
=c
∂t ∂ t2
+1
Uo ( x - υ )2
⌠ exp dυ y = f ( x , p1 , p2 , ... )
2c√π t ⌡ 4c2 t
-1 input (x)

• Slow to design and simulate • Quick to develop & execute

• Must understand process • Leverage available information
• May not capture unique attributes • Based on an actual process
• Can extrapolate for process design • Robust for closed-loop operation

#
Empirical Models Fit Data

output (y)

training data
ei model prediction

input (u)
representation structure

u y = a1 u + a0 y • obtain representative data

• choose model structure
{ a1, a0 }
• adjust model parameters
inputs outputs • minimize deviation from data
parameters

#
Some items are too complex for either FP or
Standard “Empirical Regression”. Then we turn nto
Non-Linear Neural Models
A mathematical representation
based on neuro-biology, …

output

input B
input A

… neural networks can fit

non-linear relationships

#
To accurately represent a process: You
need to combine all model types.

First Principal Model

Empirical Model

Combining Model Types is Called Hybrid Modeling #

 Fermentation Reactor: Multiple Models
MPC SP
recipe profile SP [rpm]
DO
DO controller constant SP
motor
Air flow Drive current Pressure
controller controller
DO … Dissolved Oxygen [%]
air flow exhaust
pressure

MPC SP CO2 gas Foaming reduction = pressure elevation, aeration reduction, mixing reduction
recipe profile SP Feeding: Neural Net CO2 Concentration
Glucose feeding

steam
CPR … CO2 Production Rate

water
MPC SP
recipe profile SP
SP
Glucose
foam Heating: Empirical
~ Biomass Growth Rate

Heating/cooling jacket
pH
pH controller auto feed
controller constant SP
feed Temperature
CO2 dissolved
Yeast Growth First
O2 dissolved
Broth mash controller temperature

pH
pHH
Acid
& NaOH
Principal
pHL dosing Acid
reagents agitation (mixing]
acid

base
NaOH aeration Agitator: Vendor
DO … Dissolved Oxygen Concentration

Supplied Empirical
pH Nutrient Substrate Derived:
Loading
Broth mass Glucose Consumption Rate
(Batch in)
OUR … Oxygen Uptake Rate
Broth
pHHH upstream
Bacteria
pHH Preparation
Optimal living
pHL
conditions Broth Glucose conc.
pHLL
sampling Titre
Lab analysis
Harvesting
Optional partial harvesting during the batch Off-loading
(Batch out)

Downstream processing: Purifying compounds from a fermentation broth

#
Main Components in Hybrid Modeling

Hybrid
HybridModeling
Modeling
•• Obtain
Obtaincomposite
compositemodel
modelthat
thatincludes
includesboth
bothempirical
empiricaland
andFP
FPmodels
models
•• Framework
Frameworkallows
allowsfor
forcreating
creatingvarious
variousmodel
modelstructures
structures

Series Network
Parallel
Model A Model B
Model A Model B
Model A

Model C Model D
Nested
Model B
Model A
Model E Model F
Model B

Model C
Structures
Structurescan
canbe
becombined
combinedallowing
allowing
great
greatmodeling
modelingflexibility
flexibility
#
Knowledge Refined Modeling

Derivative Response

Model trained Model trained

without FP using FP
knowledge knowledge

#
30
Once You Have The Model; Where Does It Fit?

MPC
Subordinates
Local Control.
Local Loops get
there set points
from MPC

#
Now The Control Process

#
Control Process Horizon

#
Control Tactics during the horizon

#
REMEMBER: There Are 2 Parts

• First. We must control and minimize

process variance.

• Second. Incorporate the specific

quality and economics into the
objective function and optimize that
function by managing the process
trajectories.
#
 What is Multi-Variable Economic
Optimization?
thermostat

ceiling fans

windows

• Control multiple factors using multiple

interacting “control handles”

• How to leverage degrees of freedom …

optimization !
outside temperature

room temperature

energy cost

#
Optimization Must Look Ahead to Find
the Best Trajectory

DEN
$250 $600

$150 $700
SF LV NY

$200 $500
DAL

“short-sighted” decision (over time)

#
Supervisory Economic Control

The objective function is expressed as an equation of a

number of economic and operational parameters. Real
Time control of this function is optimized to meet or
exceed these parameters or goals.

Subject to:
• Changes in all process and economic input variables
(know variance)
• Real Time Perturbations (unknown and event driven
variance)
Via
• Real Time Dynamic control of set points and loop
parameters.
• Process Control Loops Become Subordinate to ESCADA
#
 Economic Optimization Architecture

Process Control MPC Optimizer Business Sys.

MMI MPC Server
PLC
RM LP RIO RIO EI
328K1MEG 10-17 00-07
1 2 1

L
E
N

E
N
E
T

Process Historian

#
Using Economics To Constrain System

#
Typical Improvement for Biotech Producers

Revenue
Increase production 4 – 10%

Operating Cost
Reduce energy costs 3 – 6%
Improved yields 2 – 4%

Typical Project Payback 4 – 10 mo.

Typical Project ROI > 120%
Project Duration 4 - 12 mo.

#
41
 What Are We Trying to Achieve?
Increased Profitability
Minimize operational variance

DV CV

Transition effectively SPECIFICATION OR LIMIT

KEY TARGET

CV

WITH
Follow operational plan BEFORE APC WITH APC OPTIMIZATION

TIME

CV
Non-Symmetrical Economics

#
Karl Fisher Distribution

+ 0.08 wt%
#
43
What’s it worth? Improved Rx Stability

#
Distillation Column Results

Reduced Product Variability by 85%

EB - Before EB - After

#
Boiler Optimization NOx & Heat Rate
0.55

0.52
NOx (lbs/MMBtu)

Reduce NOx
0.49 10 - 35%
0.46

0.43

0.4
BEFORE AFTER
10050
HEATRATE (Btu/kWhr)

10025
Improve
10000 Heat Rate
9975 .5 - 2%
9950

9925
TIME

#
Using PHNL Model for Batch Optimization/Control

24 hour
reduction
per batch
220 hrs 244 hrs

#
Distillation / Molecular Sieve Application

Deliverables Benefits
– Integrate non-linear models, – Increase production capacity 4.6%
multivariable control and
advanced optimization in one – Reduce energy use/gallon by 6.8%
algorithm
– Increase yield by .13%
– Control distillation to optimum
moisture
– Control end-product moisture to
the desired specifications

#
When to use MPC

• When control problem is larger than 2-3 variables

• When variables are tightly coupled
• When dynamics must be more complex than first order
plus lag
• When process nonlinearities are coupled with a
multivariate problem
• When process dynamics vary measurably with rate,
levels or other capacitance
• When process model is unknown and challenging to
identify
• When economic optimization is nonlinear and integral to
benefits

#
 Model Predictive Control

Driving By Looking Forward

Rather Than In The Rear
View Mirror

#
 Thank You
And
Good Luck
For more information please contact:
David March
704-778-1909
[email protected]
#