Hybrid Modeling

1
Hybrid Modeling Model Predictive Control  

• The Potential to Leverage Success in Other
  Industries

2
Discussion Overview

• Model Predictive Control 15 years of Reducing
  Variability Improving Quality
• Process Understanding Leveraging Fundamental
  AND Empirical Knowledge
• Design Space ensuring model accuracy with
  multi-dimensional boundaries
• Lessons Learned and Success Achieved in other
  Industries
• A Practical Risk Based Approach to Implementation
  

3
Pavilion Technologies

• Our Mission
• Deliver the worlds leading model-based software
  to improve our customers profitability
• Founded in 1991
• Combined intellectual property of DuPont and
  Eastman Chemical Company
• Global Presence
• Offices in North America, Europe, China, and
  Pacific Rim
• Financials
• Acquired on November 1, 2007 by
• Commitment to Innovation
• Team of researchers, computer scientists and
  industry experts leveraging more than 155
  patents in the field of modeling, control and
  optimization

4
What is a Model?
A model explains or emulates the behavior of a
process ...
using a set of computations

• A model provides predictive capability through
  computational experimentation
• Manufacturing assets can be understood
• Manufacturing assets can be managed
• Note that the "process" need not be physical

5
Linear vs Non-Linear Models

• Sophisticated mathematics
• Accurate process representation
• Universal applicability

• output (y)

• y a1 x a0

• Non-Linear

• output (y)

• y a3 x3 a2 x2 a1 x a0

• input (x)

• Linear

• Mathematically simple
• Limited representation

• input (x)

6
Model Development

• Model
• Develop Statistical Models
• Characterize Process
• Trace of Outputs vs. Inputs
• Trace of Output Sensitivities
• Interactive Set-Point Analysis
• Multivariate Data Analysis

7
What is a Good Model?

• Desired Model Features
• Offer prediction accuracy and maintain
  computational efficiency
• Provide global validity over the entire operation
  region
• Enable optimal combination of empirical data,
  first-principles models, and process knowledge
• Remain physically meaningful
• Offer robustness to modeling inaccuracies and
  disturbances by enabling optimization-based
  modification of the models
• Enable solution standardization by simplifying
  template building

8
Established Modeling Paradigms aren't Enough

• First Principles Modeling
• Leverage explicit knowledge based on scientific
  principles
• Strengths
• Global validity
• Parameters have physical meaning
• Not data dependant
• Weaknesses
• Typically incomplete
• Slow evaluation (solver)
• Could contain implicit outputs

• Empirical Modeling
• Leverage implicit knowledge based on
  historical/test data
• Strengths
• Typically explicit
• Fast evaluation (no solver)
• Wide applicability and quick to develop
• Weaknesses
• Valid for observed data
• Often lacks physical meaning
• Requires rich data

9
Building Parametric Hybrid Models

• Specify Hybrid Model Structure
• Often a composite model that includes both
  empirical and FP components

EmpiricalModel
First-PrinciplesModel

• Train the Model using Constrained Optimization
• Often constraints are imposed to ensure model
  parameters reflect first-principles knowledge

EmpiricalModel
First-PrinciplesModel
ConstrainedTrainer
increasingconstrainteffects
decreasingconstrainteffects
10
Flexible Composite Modeling Example
Fundamental Models
11
Pavilion8 in Drying Processes The Primary
Objectives

• Plant Obedience is an advanced form of steady
  State
• APC manages the dynamics rapid disturbances
  that occur

• Consistency of operation
• Enables optimal targets to be achieved with
  confidence
• Preserves safety - plant, people product

• Plant Obedience APC on when the model is in the
  valid zone approx _at_ 30 solids

12
Fermentation Control Application

• Capabilities
• 50 reduction in batch EtOH and Dextrose/residual
  sugars variability
• Continuously manage enzymes to maximize
  throughput and ethanol yields
• Optimal target on temperature and pH for
  fermentation
• Manage fermentations to match production targets

• Benefits
• Increase in batch drop ethanol yield (MMGPY) by
  0.5-1.0
• Increase in fermentation capacity by 5-12
  (MMGPY)
• Increase in batch yields (gal/bu) by 2-5
• Reduce enzymes/gal ( enzymes/gal) by 5-10

13
Bio-Fermentation Batch A-1022
14
Bio-Fermentation Batch A-1145
15
Reactor Control Application

• Benefits
• Faster transition times
• Reduce off-spec
• Increase production
• Reduce variability
• Improve catalyst and monomer efficiency

• Capabilities
• Adjust reactor conditions to maintain at-grade
  quality control of the resin properties
• Control concentrations within the reactor for
  improved reactor stability
• Control and/or maximizes production rates within
  the reactor
• Optimize transitions
• Standardize reactor best practices for complex
  procedures

16
4 Steps of PAT Implementation by John R. Davis,
PE and John Wasynczuk, PhD

• Discover
• Analyze Existing Processes
• Develop
• Develop Statistical Models
• Characterize Process
• Demonstrate
• Substantiate Statistical Correlations
• Deploy
• Perform Continuous Monitoring Analysis
• Better Process Knowledge
• Through Analytics

17

• 4 Steps of PAT Implementation by John R. Davis,
  PE and John Wasynczuk, PhD

• Discover
• Analyze existing processes
• Data Consolidation Visualization
• DCS/SCADA Historical Data
• PAT Sensor Array Data
• Quality
• LIMS
• MES/ERP
• Data Preprocessing
• Remove Outliers
• Signal Conditioning
• Begin analysis in hours not days weeks

18

• 4 Steps of PAT Implementation by John R. Davis,
  PE and John Wasynczuk, PhD

• Develop
• Develop Statistical Models
• Characterize Process
• Trace of Outputs vs. Inputs
• Trace of Output Sensitivities
• Interactive Set-Point Analysis
• Multivariate Data Analysis
• Pavilions modeling technology can incorporate
  fundamental models into empirical models,
  leveraging the benefits both provide for Linear,
  Non-Linear, Monotonic and Non-monotonic processes
  

19

• 4 Steps of PAT Implementation by John R. Davis,
  PE and John Wasynczuk, PhD

• Demonstrate
• Substantiate Statistical Correlations
• Predicted vs. Actual Scatter-Plots
• Interactive What-Ifs Tools
• Provide different input values to tool predicts
  output values (Prediction)
• Provide desired output values to tool determines
  best input values (Optimization)

20
Moisture Inferential Accuracy
Powder Moisture Inferential Accuracy
21

• 4 Steps of PAT Implementation by John R. Davis,
  PE and John Wasynczuk, PhD

• Deploy
• Perform continuous monitoring analysis

• Optimization

• Advanced
• Control

• Standard
• Control

22
Lessons Learned

• A systematic modeling approach enables
  cross-divisional collaboration (research,
  operations, quality, engineering)
• There are a number of unit processes
  (fermentation, drying, mixing, etc.) that have
  proven results improvements leveraging this
  approach and technology in other industries