See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/259124422

Predictive Functional Control

Article · January 2009

CITATIONS READS

32 673

2 authors:

Jacques Richalet Donal O'Donovan

93 PUBLICATIONS 4,653 CITATIONS

Cork Institute of Technology
19 PUBLICATIONS 157 CITATIONS
SEE PROFILE
SEE PROFILE

All content following this page was uploaded by Donal O'Donovan on 15 October 2021.

The user has requested enhancement of the downloaded file.

 Introduction to
Model-based Predictive Control

Presented by:
Donal O’Donovan
Munster Technological University
([email protected])
 What Can I Expect From Today’s Lecture?
• Gain an intuitive feel for what Model Predictive Control (MPC) is
• Understand the benefits of MPC
• Where and when MPC may be applied
• Appreciate the implementation requirements
• Why isn’t MPC everywhere?
• Is there an alternative MPC that is accessible for ‘smaller’ plants?
• Predictive Functional Control

Donal O'Donovan - Munster Technological University

 Agenda
• Model Predictive Control (MPC) • Why don’t I see MPC Everywhere?
• Why and where does it fit into a typical plant • Predictive Functional Control (PFC)
• Brief history and commercial availability • Benefits of using PFC
• Underlying Principles
• What is Model Predictive Control?
• PFC Examples
• Optimal Model-based Control
• Summary
• Model-based Predictive Control
• What is it and how controllers formulated/solved? • Questions
• Application: Idle-speed control
• Decentralised MPC

Donal O'Donovan - Munster Technological University

 Why Model Predictive Control (MPC)?
Traditional PID Control Model Predictive Control

• No knowledge of Constraints • Constraints systematically included in the design phase

• Consequently, setpoint further from constraints • Consequently, setpoint closer to constraints
• Not optimal process operation • Improved process operation
• SISO (possibly, low-order MIMO) systems • SISO/MIMO systems
• A mathematical model is not needed • A mathematical model is needed

Donal O'Donovan - Munster Technological University

Operational Hierarchy Before and After MPC

Optimization Level Unit 1: Conventional Structure Unit 2: MPC Structure

Global Steady-state optimization
(Every day)
Plant-wide Optimisation

Local Steady-state optimization

(Every hour)
Unit 1: Local Optimisation Unit 2: Local Optimisation

Dynamic Constraint Control

High/Low Select Logic
(Every minute)
Model Predictive Control (MPC)
PID Lead/Lag PID

Supervisory Dynamic Control

(Every minute) Sum Sum

Unit 1: Distributed Control System (PID) Unit 2: Distributed Control System (PID)
Basic Dynamic Control
(Every second)
FC TC FC TC

PC LC PC LC

Donal O'Donovan - Munster Technological University

 History
• Model Predictive Control (MPC) algorithms were born in industrial
environments (mostly refining companies) during the 70’s:
• IDCOM (Adersa-Gerbios, France) [Richalet et al., 1978]
• DMC (Shell, USA) [Cutler and Ramaker, 1979]
• Born out of necessity to satisfy the more stringent production requests:
• Economic optimization
• Maximum exploitation of production capacities
• Minimum variability in product qualities

Donal O'Donovan - Munster Technological University

 Industry and Academia

• Nowadays, most complex plants, especially in refining and (petro)chemical

industries, use MPC systems
• After an initial reluctance, the academia “embraced” MPC contributing to:
• Establish theoretical foundations
• Develop new algorithms

Donal O'Donovan - Munster Technological University

Commercial Product Evolution
• Initial Commercial Products
• [DMC] Dynamic Matrix Control) DMC Corporation (USA)
• [SMCA] (ex IDCOM) Multivariable Control Architecture – Set-Point Inc. (USA)
• [PCT] (RMPCT) Predictive Control Technology – Honeywell Profimatics (USA)
• [OPC] Optimum Predictive Control – Treiber Controls, Inc. (Canada)
• [MVPC] Multivariable Predictive Control – ABB Ind. System Corp. (USA)
• [IDCOM-Y]) Johnson Yokogawa Corp. (USA)
• [MVC] Multivariable Control – Continental Control, Inc. (USA)
• [C-MCC] Contas-Multivariable Constrained Control – CONTAS s.r.l. (Italy)

Donal O'Donovan - Munster Technological University

 Commercial Product Evolution
• Mergers
• In middle of the 90’s,many acquisitions and merges occurred
• The situation became quite steady with two main competitors (DMC+ and RMPCT)
• Other less diffused technologies (Connoisseur, SMOC, PFC, etc.)

Donal O'Donovan - Munster Technological University

From [Qin and Badgwell, 2003]
Current Linear MPC Vendors and Packages
• Aspentech
• DMCplus
• DMCplus-Model
• Honeywell
• Robust MPC Technology (RMPCT)
• Adersa
• Predictive Functional Control (PFC)
• Hierarchical Constraint Control (HIECON)
• GLIDE (Identification package)
• MDC Technology (Emerson)
• SMOC (licensed from Shell)
• Delta V Predict
• Predictive Control Limited (Invensys)
• Connoisseur
• ABB
• 3d MPC Donal O'Donovan - Munster Technological University
 Model Predictive Control (MPC)
• Use a receding horizon control strategy i.e. Model Predictive Control
• At each time step, compute control by solving an open-loop optimisation problem for the prediction horizon
• Typically, using linear or quadratic programming-based techniques
• Apply the first value of the computed control sequence
• At the next time step, get the system state and re-compute

Source: Wikipedia
Source: [8] Donal O'Donovan - Munster Technological University
 Chess Game
(𝒌 + 𝑯)
Analogy
Etc. Horizon (H) (𝒌)
(𝒌 + 𝑯 + 𝟏) Horizon (H) (𝒌 + 𝟏)
Opponent’s Move My Move
(Your Move) (Plant) (Controller)
(My Move) (Your Move) Plant
Controller Plant 𝑦(𝑘 + 2)
⋮
⋮

⋮
⋮
𝑘+𝑁 𝑘+𝑖 (My Move)
⋮ Controller
⋮
⋮

𝑢(𝑘 + 1) (Your Move)

⋮
Plant
⋮ ⋮ 𝑦(𝑘 + 1)
⋮ ⋮
⋮ ⋮ ⋮ (My Move)
⋮ Controller New State
Etc. ⋮
⋮
⋮
⋮
⋮

⋮ u(𝑘)
⋮
⋮ ⋮ Opponent’s
My Move
⋮ Move
⋮ ⋮ (Plant) (Controller)

⋮ ⋮ ⋮
⋮
⋮

⋮ ⋮ ⋮ (My Move)
⋮ ⋮ Controller
⋮ ⋮
⋮
⋮
⋮
⋮

Donal O'Donovan - Munster Technological University

MPC for Idle Speed Control
• Ford Pickup truck, V8 4.6L petrol engine
• Process:
• 1 output (engine speed) to regulate
• 2 inputs (airflow, spark advance)
• Input delays

• Objectives and specs:

• Regulate engine speed at constant rpm
• Saturation limits on airflow and spark
• Lower bound on engine speed ≥ 450 rpm
• Problem suitable for MPC design (Hrovat,1996)

Donal O'Donovan - Munster Technological University

Example: Load Torque in Power Steering
Convergence Speed relative to a typical Linear Controller

(Di Cairano, Yanakiev, Bemporad, Kolmanovsky,Hrovat,2011)

Convergence 10 s faster
Peak reduced by 50%
Donal O'Donovan - Munster Technological University
Example: Temperature setpoint tracking in Air Conditioner

Donal O'Donovan - Munster Technological University

 Centralised vs Decentralised/Distributed Control
• Centralised Control
• Need a global model of the overall system (and its maintenance)
• Increase in complexity not scalable with plant size
• Computational complexity may become prohibitive
• Control design is hard to commission, start-up and maintain
• Many tuning ‘knobs’

• High risk: a single controller is running the whole plant

• Good theoretical properties (e.g. closed-loop stability)

Donal O'Donovan - Munster Technological University

 Centralised vs Decentralised/Distributed Control
• Decentralized Control
• Local sub-models of system components are enough
• Computational tasks are parallelised, each task is simple
• Data gathering ins simpler (local measurements are used locally)
• Commissioning, start-up and maintenance are more practical
• e.g. Controller updates do not require a whole plant shutdown
• Global properties (stability, performance) harder to assess, especially in the
presence of input/state constraints
• Careful coordination of controllers is needed to ensure global requirements are met
• e.g. Stability and constraint fulfilment

Donal O'Donovan - Munster Technological University

 Typical Decentralised Approach
• Measure/estimate local states
(Centralised) Supervisor
• Compute control actions locally
• Exchange decisions with neighbours,
possibly reiterate local computations Controller Controller Controller
#1 #2 #3
• Apply the current command input to
local actuator(s)
Large-scale Process
• Possible interact with upper level of
decision making Main issues:
• Hierarchical control • Global closed-loop stability?
• Feasibly of global constraints?
• Loss of performance w.r.t. centralised control?
Donal O'Donovan - Munster Technological University
Example: Distributed MPC of Barcelona Water Network

• Benefits evaluated on 3 days historic data set

• ~20% direct cost savings
• (Pumping and water sources)
• Indirect savings by smooth MV’s operation
• Leakage prevention by small pressure
surges
• Reduced equipment tear & wear
(Trnka, Pekar, Havlena, IFAC 2011)

Donal O'Donovan - Munster Technological University

 Why Isn’t MPC Everywhere?...
• As we’ve seen, MPC uses a complex, numerical optimization of the actual and future MV
sequence, taking into account different constraints
• The main drawback of MPC is that the algorithm considering constraints is complex, so:
1. The implementation requires expert knowledge
2. Commercial programs have to be installed with a correspondingly high-cost license fee
3. A lot of effort is required in developing models for plants and may require expert knowledge to
construct a successful model
• However, when the goal is to improve the behaviour of low-level, basic controllers only,
Predictive Functional Control (PFC) is a good choice
• Possesses both calculation and tuning simplicity
• Easy to implement and capable of handling constraints
• The PFC algorithm is not intended to replace a multivariable, constrained MPC
• Both algorithms have their different field of application
Donal O'Donovan - Munster Technological University
 Predictive Functional Control - History
• Predictive Functional Control (PFC) can be used as an alternative when
dealing with complex dynamics, non-minimal phase (inverse dynamics),
time delay, etc.
• The first PFC controller was implemented in 1968 by Jacques Richalet
• The first industrial commissioning taking place in 1973
• PFC has been successfully applied to a large variety of processes
• Thousands of successful applications world-wide

Donal O'Donovan - Munster Technological University

Predictive Functional Control – The Basic Concepts
• PFC combines four simple concepts and ideas to create a model-based
controller
• PFC is mainly concerned with single input, single output (SISO) processes
• But, an extension to 2 MVs / 2 CVs is also possible
• Sampling periods ranging from:
• 1 hour (level control of river dam)
• 63 micro seconds (guided missile)
• Numerous applications in the production industry for both continuous and
batch processes

Donal O'Donovan - Munster Technological University

Predictive Functional Control – The Basic Concepts
• Consider the everyday activity of driving a car…
• We learn, through experience, how the car will generally responds to a particular depressions
and releases of the accelerator pedal
• We then check if the response of the car is as expected
• We adjust the pressure on the accelerator, according to previously-learned, associative memory,
to track a notional acceleration profile we have in our heads

• Compare this to the behaviour of a PID controller

1. Constantly looking at the current error (proportional error)
2. A historical windowed estimate of previous error values (integrator error)
3. The rate of change of the error to form decisions about the future? (derivative error)
• All three are combined, with different emphasis, to decide what to do next

• Humans employ a fundamentally different control strategy to that of PID…

Donal O'Donovan - Munster Technological University
 Predictive Functional Control – The Basic Concepts
• When driving a car, we use our previous experience of the car’s dynamic behaviour to determine the
extent to depress, or release, the accelerator pedal
1. In PFC, a model replaces the experience of the plant’s dynamic behaviour (associative memory)
• The model provides predicted, future plant responses to specified inputs
• An independent model, i.e. solely based on inputs and outputs, rather than a realigned model (on the
process output after each time step), approach is used
• Results in significant improvements in sensitivity to measurement noise over the state realignment
equivalent
• As we’ll see, PFC restricts the MV to weighted combinations of pre-defined input (basis)
functions
• At each time instant, the model is used to compute the best combination of weighted basis
function values that constitute the manipulated Variable (MV) output of the PFC controller
Donal O'Donovan - Munster Technological University
Predictive Functional Control – The Basic Concepts
• When driving the car, we periodically compare the car’s acceleration
progress, against the acceleration profile we have in our heads, and
make adjustments, if we are not happy with the comparison
2. In PFC, we refer to this as the reference trajectory and it is updated
at each new sampling point
• The reference trajectory is the connection between the actual speed (plant
output) and the to the desired speed (setpoint)
• This reference trajectory fixes the closed-loop dynamics of the controller
• The choice of trajectory is open
• It may be calculated on-line, stored in a look-up table, adapt to the environment, etc.
• Generally, an exponential relationship connecting plant and the setpoint is chosen

Donal O'Donovan - Munster Technological University

 Predictive Functional Control – The Basic Concepts
• When driving the car, the decision-making process involves deciding what
changes in the depression, or release, of the accelerator pedal are necessary
to produce the acceleration profile (reference trajectory) we desire
3. In PFC, the objective of the control equation is to determine the required
pedal actions necessary to produce the desired behaviour
• In effect, it is the inverse of the associative memory, discussed previously

Donal O'Donovan - Munster Technological University

 Predictive Functional Control – The Basic Concepts
• When driving the car, we evaluate the behaviour of the car over a period
of time to determine whether the previous set of adjustments are
producing the desired results
4. Comparison of real / predicted outputs = Disturbance Estimator
• Two types of error may account of the mismatch
• The image (model) of the world is not exact
• Other unknown disturbances are acting on the process
• Over short time windows it is difficult/impossible to distinguish between the two

• This is acknowledging that our model of the world is not perfect

Donal O'Donovan - Munster Technological University

 Basic PFC Control Equation
At time n, the error signal ε(n) is given by:
𝜀(𝑛) = 𝑦𝑠𝑝 (𝑛) − 𝑦𝑝 (𝑛)
At time n+h it is desired that:
𝜀(𝑛 + ℎ) = 𝜀 𝑛 𝜆ℎ (i.e. the desired reference trajectory)
3𝑇𝑠
−
where 𝜆 is a function of the desired CLTR typically taken as 𝜆 = 𝑒 𝐶𝐿𝑇𝑅

ℎ is the desired coincidence point

At time n+h, the desired increment Δp of 𝑦𝑝 is such that:
𝜀 (𝑛) = Δ𝑝 + 𝜀 𝑛 𝜆ℎ
where Δ𝑝 = 𝜀 (𝑛) − 𝜀 𝑛 𝜆ℎ = 𝜀 (𝑛)(1 − 𝜆) The future output of 𝑦𝑚 (𝑛 + ℎ) is the sum of the free and
forced responses:
Find a manipulated variable 𝑢(𝑛) that delivers an increment Δ𝑚, at time 𝑛 + ℎ,
𝑦𝑚 𝑛 + ℎ = 𝑦𝑚 𝑓𝑟𝑒𝑒 𝑛 + ℎ + 𝑦𝑚 𝑓𝑜𝑟𝑐𝑒𝑑 𝑛 + ℎ
(i.e. 𝑦𝑚 𝑛 + ℎ = 𝑦𝑚 + Δ𝑚) where 𝚫𝒎 = 𝚫𝒑
= 𝑦𝑚 𝑛 𝑎ℎ + 𝑢 𝑛 𝐾𝑚 1 + 𝑎ℎ
Recall, the difference equation describing a first-order process, with a gain 𝐾𝑚 ,
The increment Δm at time n+h is Δ𝑚 = 𝑦𝑚 𝑛 + ℎ − 𝑦𝑚 𝑛
time constant 𝜏𝑚 and sampling period 𝑇𝑠 , is given by:
𝑇
− 𝑠 Recall, 𝑢(𝑛) is chosen such that Δ𝑚 = Δ𝑝
𝑦𝑚 𝑛 = 𝑦𝑚 𝑛 − 1 𝑎 + 𝑢 𝑛 − 1 𝐾𝑚 (1 − 𝑎) where 𝑎 = 𝑒 𝜏𝑚
Substituting, and reorganising, gives the basic control
equation:
𝒚𝒔𝒑 𝒏 − 𝒚𝒑 𝒏 𝟏 − 𝝀𝒉 + 𝒚𝒎 𝒏 𝟏 − 𝒂𝒉
𝒖 𝒏 =
Donal O'Donovan - Munster Technological University 𝒌𝒎 (𝟏 − 𝒂𝒉 )
 Future MV Hypotheses - Basis Functions
• The basic control equation makes the implicit assumption that the
future MV values over the coincidence horizon will remain the same
i.e. constant
𝑢 𝑛 + 𝑖 = 𝑢 𝑛 ∀ 𝑖 > 0 where 0 ≥ 𝑖 ≥ 𝐻

• In other words, the computed MV(0) represents the best choice, of

all available MV values, if the same value were to be applied over the
complete horizon
• However, recall that only 𝑀𝑉(0) is applied to the plant and the whole
process is repeated with the new state of the system taken into
account
• This works fine if we are tracking a constant setpoint
Donal O'Donovan - Munster Technological University
 Future MV Hypotheses - Basis Functions
• But, what about a setpoint with a time-varying profile?
• It turns out that, in order to track setpoints with a known profile, with zero tracking error, an
MV of a similar profile needs to be adopted
• This implies that the MV should consist of a set of basis functions, which are eigenfunctions
• i.e. input signals that elicit a similar-shaped response in the plant output, in the longer term

• The goal of the control equation is to estimate the projection weighting's of the functional basis
set on to the future 𝑢(𝑛 + 𝑖)
• i.e. 𝑀𝑉(0) is the weighted sum of the functional basis set
• For example:
• If the setpoint is a ramp, 𝑢(𝑛) should contain step and ramp basis function components
• The step is needed to deal with constant offsets
Donal O'Donovan - Munster Technological University
 Future MV Hypotheses - Basis Functions
• The weighting coefficients of the functional basis are computed as the unknown variables of
the control equation
• The 𝑢(𝑛 + 𝑖) is then, by extension, projected on to a polynomial basis: 𝑃𝑗 𝑛 , where 𝜇𝑗 are
the unknown weighting coefficients of the polynomial 𝑃, which is a function of the future
time 𝑖, i.e.

𝑴𝑽 𝒏 + 𝒊 = ෍ 𝝁𝒋 𝑷𝒋 𝒊 with 𝒋 = 𝟏, 𝟐, … , 𝒋𝒎𝒂𝒙
𝒋
• For instance,
𝑴𝑽(𝒏 + 𝒊) = 𝝁𝟏 𝟏 + 𝝁𝟐 𝒊
where 𝑷𝟏 = 𝟏 i.e. step
𝑷𝟐 = 𝒊 i.e. a ramp
𝝁𝒋 are the unknown coefficients to be determined
Donal O'Donovan - Munster Technological University
 Future MV Hypotheses - Basis Functions
• If two basis functions are selected, then there should be two equations and, at least,
two coincidence points
• This implies that it is possible to structure the future 𝑢(𝑛 + 𝑖) for 𝑛 = 1, … , ℎ using a
limited number of weighted functions of the functional basis set
• Also, it is not necessary to compute all future 𝑢(𝑛 + 𝑖) over the coincidence horizon
• For instance, in the step and ramp case, it is sufficient to choose only two coincidence
points ℎ1 and ℎ2 over the horizon
• If the functional basis is limited to ramp, step, and parabola, a simple algebraic ‘one-
step’ solver may be used – no iterations required
• In summary, the eigenfunctions, chosen as the functional basis set for the MV, dictate
the ability of track of the controller to track the setpoint
• i.e. It dictates the ‘accuracy’ of the closed-loop control
Donal O'Donovan - Munster Technological University
 Implicit Integrator and Time Delay Compensation
• A steady-state analysis of the basic PFC equation shows that there is no
steady-state error present
• It may be easily shown that there is an implicit integrator present in the controller

• The absence of an explicit integrator has large implications

• i.e. No integrator initialization, no integral wind-up issues, etc.

• If the process possesses a pure time delay 𝜃 the reference trajectory is simply
initialized, using the predicted value of the process output 𝑦ො𝑝 at the future
time 𝑛 + 𝜃 to compensate

Donal O'Donovan - Munster Technological University

PFC Controller Configurations
• Can be used in similar configurations to PID i.e.
1. Feed-forward
2. Split Range
3. Cascade

Donal O'Donovan - Munster Technological University

 Constraints on the Manipulated Variable
• There is always a practical limit to the value of 𝑢(𝑛)
• Disturbances and/or setpoint specifications may result in 𝑢(𝑛) reaching one of its
limits
• So, how is are constraints handled?
1. Apply a constrained MV
2. Apply a constrained CV
• i.e. if an MV value is to violate the DV, then a second PFC controller computes an MV using a constrained
setpoint as a reference. Hopefully, this wil
3. The normal practice for constraint handling is to tune up two separate control strategies:
I. Well tuned controller
II. When required, detuned controller such that it has low input activity and output overshoot is not expected
4. Place structure (prestabilisation) into the predicted future control trajectory to bring the unstable
dynamics under control and substitute for ii above
I. Has an implicit capacity to ignore/modify set point changes where this is judicious
5. For Cascade loops, there is the ‘back-calculation’ method to map the inner-loop MV constraint
onto the outer-loop MV
I. Ensure that the constrained Donal
MV is supplied
O'Donovan to the
- Munster internalUniversity
Technological models
PFC Tuning
• Tuning an industrial system is always a compromise between the
conflicting objectives of dynamics (CLTR) and robustness (stability
margins)

• Easier, and more intuitive, to tune

• Tuning knobs are not tightly-coupled like PID
Basis Reference Coincidence
Function trajectory Point (h)
ACCURACY 2 0 0
DYNAMICS 0 2 1
ROBUSTNESS 0 1 2

Donal O'Donovan - Munster Technological University

 Elementary PFC Examples

G. Valencia-Palomo and J.A. Rossiter [7]

Auto-tuned PI(D) controller described in (Clarke,2006; Gyöngy and Clarke, 2006) MPC solved using active set method proposed by Fletcher (1987) PFC: one coincidence point
 Elementary PFC Examples

G. Valencia-Palomo and J.A. Rossiter [7]

Auto-tuned PI(D) controller described in (Clarke,2006; Gyöngy and Clarke, 2006) MPC solved using active set method proposed by Fletcher (1987) PFC: one coincidence point
 Elementary PFC Examples

G. Valencia-Palomo and J.A. Rossiter [7]

Auto-tuned PI(D) controller described in (Clarke,2006; Gyöngy and Clarke, 2006) MPC solved using active set method proposed by Fletcher (1987) PFC: one coincidence point
PFC Temperature Control of a Chemical Batch Reactor
• Temperature control of an exothermic acid–base neutralization
chemical reaction between the hydrochloric acid (HCl) and the sodium
hydroxide (NaOH)
• The control system must manipulate the inlet jacket reactor
temperature (Tj,i) using the electrical resistance and the two plate heat-
exchangers (PHE), to control the reaction mixture temperature (Tr)
• A cascade control structure using the PFC technique is implemented

[9] H. Bouhenchir, M. Cabassud, M.V. Le Lannc, Computers and Chemical Engineering (2006)
 Setpoint Profile of Chemical Batch Reactor
• Test the robustness of the control system when there are
dynamic changeovers, due to the heat release, during
constant setpoint stage
• The reactor was fed with a solution of hydroxide sodium
(NaOH/water = 57.14 g/485.71 g) at temperature of 22 ℃
• The set-point temperature profile is composed of:
• Stage 1: Heating from 22 to 45 ℃ for 1500 s (0.92 ℃ / min)
• Stage 2: Constant temperature at 45 ℃ during 3500 s in
which solution of hydrochloric acid (HCl/water = 44.03
g/74.97 g) is fed during 2000 s
• Stage 3: Cooling from 45 to 30 ◦C during 1500 s (−0.6 ℃/
min) and fourth stage: maintain at 30 ◦C during 500 s
• In order to eliminate discontinuities resulting from change of
stage, the set-point profile was filtered by a procedure called
“docking procedure” which transforms the temperature
profile as shown
 Results for PFC-controlled Chemical Batch Reactor
• Set point (Tr,set)
• Adjusted set point (Td,set)
• Reaction mixture (Tr) temperatures (left y-axis)
• The manipulated variable (β) computed by the slave controllers (right y-axis)
• The manipulated variable (β) denotes the fraction (compared to its
maximum) of the electrical power value or the mono-fluid flow rate fraction
dispatched to one of the two PHE
 Results with Model Mismatch

Temperature and manipulated variable (model mismatch).

The results show the performance of cascade control when there are differences in the dynamics of the different heating/
cooling system elements.
Results with No Model Mismatch

Temperature and manipulated variable (no-model mismatch).

 What Did I Learn From Today’s Lecture?
• Gain an intuitive feel for what Model Predictive Control (MPC) is
• MPC is a model-based, receding horizon control scheme that may be used to regulate both SISO and MIMO
systems
• Understand the benefits of MPC
• MPC can systematically take hard and soft plant constraints into account
• Regulate the plant more effectively, permitting the plant to operate at a higher level of efficiency
• “Squeeze and shift” principle
• Appreciate where and when MPC may be applied
• MPC is best used at the MIMO level, as an overarching controller, or in situations where PID is sub-optimal,
or unable to meet the plant’s requirements
• Why isn’t MPC everywhere?
• A model of the plant is required and this is where most of the implementation ‘effort’ will be invested
• At the plant level, MPC implementation requires expert knowledge and commercial programs have to be
installed with a correspondingly high-cost license fee
• Is there an alternative MPC that is accessible?
• Predictive Functional Control (PFC) allows you to take advantage of the benefits of MPC at a SISO / low-scale
MIMO level where PID does not work well, or not at all
Donal O'Donovan - Munster Technological University
Questions

Donal O'Donovan - Munster Technological University

 References
• [1] “Predictive Functional Control: Principles and Industrial Applications”, J. Richalet, D. O'Donovan, Springer
London, 2009, DOI: 10.1007/978-1-84882-493-5
• [2] “An Alternative for PID control: Predictive Functional Control - A Tutorial” R. Haber, J.A. Rossiter, K. Zabet
• [3] “PID + Kalman Filter + MPC LabVIEW Applications”, https://www.halvorsen.blog
• [4] “Handling constraints with Predictive Functional Control of unstable processes”, J.A. Rossiter, J. Richalet,
American Control Conference, 2002
• [5] “Generalized Predictive Control (GPC) - Ready for use in drive applications?” February PESC Record - IEEE
Annual Power Electronics Specialists Conference 4:1839 - 1844 vol. 4, 2001 DOI: 10.1109/PESC.2001.954389
• [6] “Model Predictive Control”, 5th HYCON2 PhD School, A. Bemporad
• [7] “Comparison between an auto-tuned PI controller, a predictive controller and a predictive functional controller in
elementary dynamic systems”, G. Valencia-Palomo¤ and J.A. Rossiter
• [8] Comparison between Model Predictive Control and PID Control for Water-level Maintenance in a Two-tank System
• [9] “Predictive functional control for the temperature control of a chemical batch reactor”, H. Bouhenchir, M. Cabassud, M.V.
Le Lannc, Computers and Chemical Engineering 30 (2006) 1141–1154

Donal O'Donovan - Munster Technological University

View publication stats