Industrial Challenges for the Identification and Control Society

1
Industrial Challenges for the Identification and
Control Society
2
Background
ISMC (Spin Off KU Leuven) IPCOS Technology (TU
Eindhoven / TU Delft)

• Advanced Process Control (APC) products
• All affiliated services consulting, feasibility
  studies, implementation, training, maintenance

3
Academic versus Industrial

• Academic research is mathemically / problem
  driven
• How challenging is the APC problem ?
• How do I make an as good as possible model ?
• Industrial Advanced Process Control (APC)
  applications are economically driven
• How can I make money by solving the APC problem ?
  
• How do I make a good enough model as cheap as
  possible ?

4
Presentation Structure

• Plant Operation Layers
• Typical Advanced Process Control applications
• Low Level Control Tuning
• Soft Sensors
• Multivariable/Predictive Control
• Plant-Wide Dynamic Optimization

• Presentation Goal
• Understand Principles of each layer
• Understand Economics of each layer
• Discuss Academic Challenges

5
Plant Operation Layers
6
Plant Operation Layers
Process Plant Operation is layered
At each layer other technologies timescales
apply and different benefits result

• Low Level Control (PID)
• Supervisory Control (Softsensors MPC)
• Plantwide Optimization (Optimisation)

7
Model Based Control Optimization
Control hierarchy
Technology
Layer
8
1. Low Level Control Layer
Control PID, On-Off Platform DCS,
PLC Timescale seconds Benefits stability
CW
9
2. Supervisory Control Layer
Control MPC, Platform PC,
DCS Timescale seconds - minutes Benefits
operate closer to constraints
Ratio Setpoint for low level control
Steam Flow Setpoint for low level control
10
3. Plantwide Optimization Layer
Platform PC Timescale hours-days Benefits
economical optimization (plant
constraints)
Plantwide Optimization 2 types Static
Optimizer Detects steady state Find optimal
steady state Hands optimal setpoints down to
Supervisory control layer Dynamic
Optimizer Find Optimal Dynamic Trajectories
11
Principles Economics Challenges
12
Model Based Control OptimizationPID Controllers
Technology
Product
13
PID Controller Principles
14
PID Controller EconomicsProcess Performance is
not as good as you think

• PID controllers at lowest level
• PID controllers are the workhorse of Process
  Industry
• 90 of the controllers are PIDs
• More than 30 of PIDs operates in manual
• More than 30 of loops increase short term
  variability
• About 25 of loops use default settings
• About 30 of loops have equipment problems
• APC not useful when PIDs are badly tuned

15
PID Controller Economics Chemical example
Cooling water
Steam

• Where Antwerp
• Company Confidential
• Product Confidential
• Solution optimal PID control for batch
• Benefit 1.000.000 /year/reactor
• Payback 3 weeks
• How was the benefit generated
• Batch time reduction through increased
• throughput in a non saturated market

A B
16
PID Controller Economics Refining Example

• Where Antwerp (Belgium)
• Company BRC
• Product refining
• Solution Optimization of primary loops (ES)
• How was the benefit generated
• More stable operation (operator load)
• Less blending
• Less system load (lifetime)
• First step towards APC

17
PID Controller Challenges (1) Industrial
Requirements

• Industrial Requirements
• Good Load Rejection
• No nervous control signal
• No overshoot
• Fast Tracking (for reference or master/slave
  controllers)
• Robust

18
PID Controller challenges (2)

• Often academically miss-treated because of
  apparent simplicity
• Main Challenge lies in the trade off that must
  be made between performance and robustness with
  a limited PID control structure
• Operators have to tune maintain PID
  Controllers Automate tuning as much as possible
• Use operational (closed loop) data
• Simple operational requirements
• Simple trade off tracking and load rejection
  requirements
• Auto detection of need for re-tuning
• Fast i.e. within ¼ day

19
PID Controller Challenges (3)

• Identification Society
• Fully automatic identification of SISO dynamic
  systems including estimation of delay, poles,
  zeros, integrator
• From badly excited, closed-loop data with low
  frequent disturbances
• Leading to physically acceptable models

20
PID Controller Challenges (4)

• Control Society
• Simple, engineering based statements of the PID
  control problem with operational constraints (MV
  saturation)
• Good and fast optimisation strategies
• Keeping the industrial form of the PID
  controllers in mind
• Pairing of MV to CV
• PID structures paradigms (cascade, split
  range)
• Automatic detection of troublemakers within x00
  PIDs

21
Model Based Control OptimizationSoft Sensors
Technology
Product
22
Soft Sensor Principles
Classical
Concentrations, Density, MI, pH, NOx, CO2
Flows, Pressures, Temperatures
Soft Sensor
On-line
23
Soft Sensor Economics

• Avoid using expensive measurement equipment
• Less use of laboratory
• Closed Loop (High bandwidth)
• Possible to put derived process variables (e.g.
  Efficiency, Emissions) in control loops
• More competitive operation
• Better environmental protection
• Predict possible future problems (pumps, fans,
  valves)
• Less emergency stops
• Maintenance can be planned better
  predictive maintenance

24
Soft Sensor Economics Oil industry example

• Where Asia
• Company Confidential
• Product Oil
• Solution On-line estimation of multi-phase
  flows
• Benefit x.000.000 /year/platform
• Payback weeks
• How was the benefit generated
• Continuous soft-measurement allows for
• on-line monitoring and optimization

25
Soft Sensor Challenges (1)

• Identification Society
• 1) Input Selection How to select a subset of
  inputs (10) for the model from the huge set of
  available inputs (100)
• Need heuristics to avoid computation for years
  by exhaustive search
• PCA, PLS, CCA only determine a linear
  combination
• Avoid over-fitting in the input space
• Huge data sets (1 Mio samples x 300 variables)

26
Soft Sensor Challenges (2)

• 2) Modelling How to make accurate models from
  the data
• Structure Static and dynamic, linear and
  non-linear
• Huge amount of data (1 Mio x 300)
  computationally efficient
• Good initial guesses for optimisation
• Highly correlated historical (closed loop) data
• Automatic trade-off between accuracy and
  generality (overfitting)

27
Soft Sensor Challenges (3)
3) Make accuracy of models depending on local
input densities
4) Allow reasonable extrapolation of Soft
Sensors through a-priori knowledge.
28
Soft Sensor Challenges (4)

• 5) On-line updating based on new information
• Bias correction, Kalman Filter or Receding
  Horizon Estimator
• As long as it is robust and easy to use
• And cheap

6) On-line Track accuracy of model and flag when
model leaves training region and extrapolates
excessively
29
Model Based Control OptimizationModel Based
Predictive Control
Technology
Product
30
MPC Principles

Gas composition change
Gas composition Change (DISTURBANCE)
Y1(t)
Disch. Pres. Air compr
Quality
Y2(t)
CH4 (sec ref.)
Throughput
U2(t)
Feed CH4
U1(t)
T exit Prim Reformer
31
MPC Economics Chemical Example

• Where Burgkirchen-Gendorf (Germany)
• Company Vinnolit
• Product Vinylchlorid
• Solution APC (Products and ES)
• Benefit confidential
• How the benefit was generated generated
• Energy
• Throughput
• Reduced maintenance cost

32
MPC Economics Variance Reduction
33
How?Reduced Energy/cost
Quality
APC off
APC on
APC off
? Profit
time
34
How?Controlled Transitions in Automatic Mode
Process Value
Specification B
Ideal Value
manual transition
controlled transition
? Profit
Specification A
Ideal Value
time
transition start
transition end
transition end
35

MPC Economics on an EDC/VC cracker
36
MPC Economics Variance Reduction

• Visualization of benefit realization by MPC

Standard Control
Model Predictive Control without optimization
Model Predictive Control with performance
optimization
37
MPC Challenges (1)

• Identification
• Need accurate multivariable dynamic models
• With a minimal test time Multivariable Models
  are Expensive
• Up to 40 of an MPC project costs
• Excite multiple input variables at the same time
• Avoid waiting for the process to settle
  (settling times of 24 hours and more)
• Insensitive for low frequent disturbances
• Insensitive for (de-tuned) controller in the
  loop
• Carefully designed experiments (cfr. stiff
  systems)

38
MPC Challenges Testing


39
MPC Challenges Use of Step Response Model
s2
U
Y
s1
s0
Linear Relationship Y F G U

• Holds also for Multiple Input Multiple Output
  system case
• Easy model building
• Low performance (high frequency content)
• Long testing time

40
MPC Challenges Use of State Space Models
y0 Cx0 Du0 y1 CAx0 CB u0 y2
CA2x0 CABu0 CBu1 Du2
xk1 A xk B uk yk C xk D uk
Linear Relationship Y F G U

• Holds also for Multiple Input Multiple Output
  system case
• Easy adaptation for Linear Time Variant model
  (Ak,Bk,Ck,Dk)
• Easy Identification from test data or from
  rigorous process model
• Stiff systems

41
MPC Challenges (2)

• Re-use as much a-prior knowledge as possible by
  using First Principle Models - Multivariable
  Models are Expensive
• Use of model reduction techniques on extremely
  badly conditioned models (2500 states to 10)
• Use of data driven (hybrid) models

42
MPC Challenges (3) Linear MPC

• Origin of Linear MPC lies in plants running in
  one operating point (refineries, large crackers)
• Final challenge is the solution of large scale
  constrained QP problems
• 30 MVs, 30 CVs
• Parameterisation of freedom per MV
• Use of structure in QP problems
• Needs to be solved in limited and predictable
  time

43
MPC Challenges (3) Non-Linear MPC

• New application areas of MPC are
• Transition Control (broad operating regions)
• Batch Control
• Need MPC valid over a non-linear region of the
  plant
• Multiple linear models (more tests)
• Non linear explicit models with fast integration
  time
• Bounded time for non-linear optimisation part of
  the MPC
• Convergence and stability ?
• Simple hybrid models ?

44
Model Based Control OptimizationDynamic
Optimization
Technology
Product
45
Plant-Wide Dynamic Optimisation
RaPID
46
Plant-Wide Economic Dynamic Optimization
Principles
Find dynamic MVs such that objective is
optimized subject to process operation
constraints
47
Dynamic Optimization EconomicsGasphase
Polyethylene reactor
Production
Density--
MI
48
Dynamic Optimization Economics
No optimization
PathFinder
49
Dynamic Optimization Economics Benefits for
polymers

• Extension of the production capacity by
  exploiting the capabilities
• of the process and pushing towards bottle-neck
  constraints
• Range up to 2.5
• Minimizing operating costs by exploiting the
  operation freedom
• Range up to 2.5
• Reduce production losses related to grade
  transitions
• Faster transition policy
• Faster settling in the new grade specifications
• Range up to 20.000 Euro/gradechange
• Minimize off-spec production during normal
  operation
• Range up to 500.000 Euro/year

50
Dynamic Optimization Challenges (1) Economically
Optimal Dynamic Transitions

• Need fast way to do optimise this specific mixed
  problem with a minimum number of iterations
  over the Process Model
• Optimal parameterisation of the MV space

51
Dynamic Optim Challenges (2) State of the Art
52
Dynamic Optimization Challenges (2)

• Optimization Society
• On-line Dynamic Optimization Why not use the
  whole first principle model to optimize and
  control the plant
• Need fast and reliably convergent algorithms to
  solve the on-line optimisation problem
• Need reliable on-line observer algorithms that
  allow tracking of the model when the plant
  drifts
• Need fast computers future

53
Conclusions
54
Conclusions

• Industrial APC projects are economically driven
• Industrially relevant challenges for
  identification and control
• Need algorithms that minimise engineering time
• Need algorithms that allow for simple
  interaction high level of automation, easy to
  use, easy to configure, minimal knowledge
  required to operate, robust, fast
• Models are expensive ! Need algorithms that
  reduce testing time and increasing model
  accuracy
• Need algorithms that allow formulation of the
  identification and control problems as close as
  possible to the operational and economic reality
  
• Need algorithms that can make use of all
  a-priori knowledge available (physical models
  and engineering insight)
• Need engineers with Process Knowledge full
  algorithmic abstraction is a myth !

55
Industrial Challenges for the Identification and
Control Society
Thank you for your attention!