Industrial Process Modelling and Control Ton Backx Emeritaatsviering Joos Vandewalle
Dec 26, 2015
Industrial Process Modelling and Control
Ton Backx
Emeritaatsviering Joos Vandewalle
Outline
• History• Process performance and process control• Model predictive control essentials• Process modeling• Current developments• Future perspective
Emeritaatsviering Joos Vandewalle Page 24 september 2013
Model Predictive Control History
Early developments of Model Predictive Control (MPC) technology were initiated by two pioneers:
• Dr. Jacques Richalet (Adersa, 1976) ‘Model Predictive Heuristic Control’ (MPHC) using IDCOM
as the MPC software for process identification (IDentification) and for control (COMmand)
Use of Finite Impulse Response (FIR) models Control inputs computed by minimization of a finite horizon
quadratic objective function without consideration of constraints
Plant output behavior specified by reference trajectories
Emeritaatsviering Joos Vandewalle Page 34 september 2013
Model Predictive Control History (cont’d)
• Dr. Charles Cutler (Shell Oil, 1979) ‘Dynamic Matrix Control’ (DMC) Use of Finite Step Response (FSR) model Linear objective function subject to linear inequality
constraints using a finite prediction horizon (LP) Plant output behavior specified by setpoints Optimum inputs calculated by solving a Linear Programming
problem
Emeritaatsviering Joos Vandewalle Page 44 september 2013
Process performance and process control
Process performance is governed by:• Critical process and product variables –”Controlled
Variables”- need to meet specifications• During startup, shut-down and product changeovers off-
spec products are produced Need for minimization of transition losses
• During production disturbances cause variations in critical variables Need for disturbance rejection
Emeritaatsviering Joos Vandewalle Page 54 september 2013
manipulatedvariables
disturbances
controlledvariablesProcess
Process performance and process control
Emeritaatsviering Joos Vandewalle Page 64 september 2013
Model predictive control is the supervisory control layer that enables process optimization by minimization of production costs ensuring product specifications and production quantities
– the model predictive control system realizes targets set by the optimizer MPC
Targets (setpoints,setranges, …)
Operatinginformation
Process
setpointsProcessvalues
PID
– optimum operating conditions are determined by an optimizer (setpoints, set ranges, priorities and weights, operating constraints)
Optimizer
Costs andSpecifications
Targets (setpoints,setranges, …)
Operatinginformation
Operatinginformation
0
0.51
1.52
2.53
prob
abili
ty d
ensi
ty fu
nctio
n
probability density
Cpk
= 0
.96
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
x 104
19
19.2
19.4
19.6
19.8
20
20.2
20.4
20.6
20.8
21Measured process signal
time
valu
e
Process performance and process control
Visualization of benefit realization by MPC
0246810
12
Cpk
= 0
.96
Cpk
= 4
.3
0246810
12
Cpk
= 0
.96
Cpk
= 1
.6
Economicbenefit
Standard ControlModel Predictive Control without optimization
Model Predictive Control with performance optimization
4 september 2013Emeritaatsviering Joos Vandewalle Page 7
• to predict future process output behavior
• to determine the best future input manipulations to drive the process to optimum conditions
Model Predictive Controller
SetpointsSet ranges
Controller
DisturbanceModel
measureddisturbances
OperatingConstraints
Optimizationand
constraint handling
Model predictive control essentials
MPC strength is based on the explicit use of (a) (set of) model(s):
• to feedforward compensate disturbances
• to respect operating constraints and to determine optimum conditions
• To handle non-linearities
UnitProcess
manipulatedvariables
disturbances
controlledvariables
ProcessModel
ProcessModel
f g
+-ProcessModel
84 september 2013
Model predictive control essentials
Emeritaatsviering Joos Vandewalle Page 94 september 2013
Time (t)
Time (t)
Past Future
Deadtime
Prediction horizon
Predicted future process responses
Setpoint valuePast process responses
Past control manipulations Future control manipulations
Control horizon
Output horizon applied for optimization
Present moment
Model predictive control essentials
Linear models are used to calculate the responses to past and future process input manipulations and similarly to predict future responses to known disturbances
Emeritaatsviering Joos Vandewalle Page 104 september 2013
),(),(),(),(
),,(),,(),(
cfcfpppf
cfffpffpff
NtUNNTNtUNNH
NNtYNNtYNtY
In this expression:• Yfp denotes the part of the future outputs stemming from past input
manipulations• Yff denotes the part of the future outputs resulting from future input
manipulations
Past
Cannot be influenced any more
Past
Future
Still to be determined by future inputs
Future
Emeritaatsviering Joos Vandewalle Page 114 september 2013
Process modeling
Process application example
Process modeling
Emeritaatsviering Joos Vandewalle Page 124 september 2013
CONCEPT OPERATIONDESIGN
Detailed design and optimisation of process equipment
Model-based automation applications for decision support
Troubleshooting with detailed
predictivemodels
Process flowsheeting
Detailed design of complex units
Design of optimal operating procedures
Simultaneous equipmentand control design and
optimisation
E V O L V I N G M A S T E R M O D E L
Laboratory experiment design and optimisation
Operator training
DESIGN
Model Predictive control
Equipment performance monitoring
Process Health monitoring
New process design
Pn + M Pn+1….
Process modeling
System identification is the modeling technique applied in industry for sufficiently accurate modeling of the relevant process dynamics for MPC
• Data driven modeling Model set: Non-parametric, semi-parametric, parametric Model structure Parameter estimation criterion: Output error, equation error,
input error
Emeritaatsviering Joos Vandewalle Page 134 september 2013
Emeritaatsviering Joos Vandewalle Page 144 september 2013
Process modeling
Required capabilities of models
1. Accuracyon-line assessment of model validity
2. Adaptabilityflexible on-line updating of models (dynamics and interconnection structure)
3. Active data-driven learningdemands on accuracy, autonomy, robustness
active probing for information
Emeritaatsviering Joos Vandewalle Page 154 september 2013
Process modeling
Example of current limitations:
• MPC projects in industry highly depend on accurate plant models and well-tuned controllers
• Controllers and models are verified (identified) during commissioning
• When during operation process behavior changes: MPC’s are switched to “manual”
• Loss of performance• Expensive experimental campaign to re-identify the
models is the only way out
Process modeling
G +
v
u y
Emeritaatsviering Joos Vandewalle Page 164 september 2013
Back to the core of the problem of data-driven modeling / identification of Linear Time Invariant (LTI) models
+ G +
v
C
- u yr
Emeritaatsviering Joos Vandewalle Page 174 september 2013
Process modeling
The classical identification problems:
G +
v
u y
open loop closed loop
Identify a plant model on the basis of measured signals u, y (and possibly r)
• Several classical methods available (Prediction Error, subspace, Output Error, non-parametric,..)
• Well known results for identification in known structure (open loop, closed-loop, possibly known controller)
+ G +
v
C
- u yr
Emeritaatsviering Joos Vandewalle Page 184 september 2013
Current developments
Next step in the development:
Autonomous economic model-based operation of industrial process systems
• Bring plant operation / automation to higher level of autonomy
• Monitor plant performance and detect changes on-line
• Generate probing signals when necessary and based on economic considerations (least costly experiments)
• Re-identify models and retune controllers on-line• Keep high performance control• Use economic performance criteria
Thank you for your attention
Emeritaatsviering Joos Vandewalle Page 194 september 2013