Professional Documents
Culture Documents
1, December 2013
Abstract
In this paper an extensive review of the theoretical
aspects of the self-tuning model predictive controllers
[MPC] is presented. Despite the existence of many
computer-based modern control methods, it is shown that
such formulation is still quite suitable for the
implementation in many types of industrial processes.
Key Words:
1- Introduction
Self-tuning control has increased in popularity in recent
years, as a research topic, and as a new adaptive method
for application to certain control problems. Much work
has been done on the theoretical aspects of self-tuning, in
defining and validating algorithms, but less has appeared
on the practicalities of the technique. This article covers
the link between the theory of the model predictive
control (MPC) and the application of self-tuning control,
and the practicalities between these two. The point has
been made that self-tuning affords the user powerful
means of on-line fixed control laws designed for tuning
the system to any of a range of more or less acceptable
mode. Although the theoretical background of this
method of control is quite old but recent implementations
have appeared in the literature (Mohammed, et. al., 2010,
Fissore, 2009, Nikolaou, 2009, Daau, et. al., 2008,
Muske, 1995).
1- Background and Literature Review
The idea of adaptive control grew largely after the 2nd
World War with the development of supersonic aircraft,
whose control surface dynamics displayed considerable
changes with flight altitude and speed. The first mention
of the term adaptive control appeared in Draper and
Lis classic paper (Draper and Li 1951) on the control of
the internal combustion engine.
controllers, in which control coefficients are calculated as
if the estimation result was correct and cautious
controllers in which some weighting is given to the
NUJES Al Neelain University Journal of Engineering Sciences, VOL 1, No. 1, December 2013
NUJES Al Neelain University Journal of Engineering Sciences, VOL 1, No. 1, December 2013
Kallstrom, et. a1., 1977, Kallstrom 1974) (ship steering),
Clarke et.al (Clarke, et. a1., 1973) (multivariable boiling
ring), Borrison and Wittenmark (Borrison and
Wittenmark
1973) (paper machine), Dumont and
Belanger (Dumont 1977, Dumont and Belanger 1978)
(titanium dioxide kiln), Morris et.al. (Morris, Fenton and
Nazer 1977, Morris Nazer and Chisholm 1979)
(Distillation columns etc.) and Horn (Horn 1978) (batch
chemical reactors).
Work on stability and convergence in self-tuning systems
has been covered by Ljung and Wittenmark in
1974(Ljung 1974a, Ljung 1974b, Ljung and Wittenmark
1974). They formulated a complicated differential
equation to be solved and their method has been much
simplified by Thiruarooran (Thiruarooran 1978).
Convergence properties of the pole shifting self tuner
have been covered by Wellstead, et. a1., 1979, Wellstead,
et. a1., 1978). Ljungs work on the stability of self -tuning
systems has continued (Ljung 1977a, Ljung 1977b).
Practical applications of self-tuners have been reported by
Jensen and Hansel (Jensen and Hansel 1974) (enthalpy
exchanger), Cegrill and Hedqvist (Cegrell and Hedqvist
1975) (paper machine), Wouters, 1974, 1977
(Continuously Stirred Tank Chemical Reactor), Borrison
and Syding (Borrison and Syding1976) (ore crusher),
Sastry and Seborg (Sastry, et. a1., 1977) (distillation
column), Keviczky (Keviczky, et. a1., 1978, Csaki, et.
a1., 1978), (cement making, multivariable), Kallstrom
et.al. (Kallstrom, et. a1., 1978, Kallstrom 1979,
Kallstrom, et. a1., 1977, Kallstrom 1974) (ship steering),
Clarke et.al (Clarke, et. a1., 1973) (multivariable boiling
ring), Borrison and Wittenmark (Borrison and
Wittenmark
1973) (paper machine), Dumont and
Belanger (Dumont 1977, Dumont and Belanger 1978)
(titanium dioxide kiln), Morris et.al. (Morris, Fenton and
Nazer 1977, Morris Nazer and Chisholm 1979)
(Distillation columns etc.) and Horn (Horn 1978) (batch
chemical reactors).
The authors have knowledge of a number of other
projects active at the time of writing. D. Clarkes group at
Oxford University has developed an MPU based portable
self-tuner (Cope 1976) and has done some work on steel
soaking pits. A.J. Morris groups at Newcastle
University have worked on auto electrolysis, and T.
Fortiscue at Imperial College, London has worked on
distillation columns. A. Johnson at Delft University,
Holland is working on batch fomenters.
The authors have (unofficial) knowledge of two
independent studies into the self-tuning control of gas
turbine engines, and knowledge of a study into the selftuning control of petrol engines.
Wittenmark in 1971(Wittenmark 1971) published a
optimal way, help the estimator to identify the system
(Wittenmark 1975).
Within the class of self-tuners, individual algorithms
differ from each other (i) in the identification method
(distillation
survey of adaptive control methods, including MRAC,
hill climbers and stochastic methods and in
1975(Wittenmark 1975) published a very thorough
survey including 86 references. Bar-Shalom and
Gershwin in 1978(Bar-Shalom and Gershwin 1978)
published a brief trends and opinions article on
adaptive control and its applications. Kurz et.al in 1978
published a survey covering self-tuning and alternative
certainty equivalence controllers (Clarke and
Gawthrop1975) Vaneeck in (Vaneeck1978) published a
survey article briefly covering several topics in self
adaptive control and Clarke and Gawthrop (1979)
published a paper on self-tuning control ( Mendes,
1971).
Self-tuning applications include studies on unclear
power plant (De Kayser and Van Cauwenberghe1978)
and DHulster and Van Cauwenberghe (DHulster and
Van Cauwnberghe 1978) have applied a complex new
algorithm to an ethane/propane/naphtha cracking
furnace.
1- Self-Tuning Control
Self- tuning control is the name given to a class of
autonomic control algorithms in which three levels of
activity are combined:
(i) A digital, sampled data, model of an unknown
system, using measurements of the systems output(s)
and input(s).
(ii) Some controller synthesis algorithm is employed to
generate the parameters of a digital control law, based
entirely upon the parameters of the system as estimated
above.
(iii) The control law specified above is used to generate
the plant controller which is then self-tune the output.
The numerical complexity of recursive estimation
schemes means that self-tuners can practically only be
implemented as part of a digital computer control
scheme.
Self-tuning control differs from other
automatic control methods in two notable ways:
(i) It is primarily a digital sampled data control scheme
for application to analog continuous systems.
(ii) Its parameter estimation methods are based upon
maximum extraction of information from the
measurements, rather than upon hill climbing around
functions of the system responses.
Even so, it can intuitively be compared with hillclimbing techniques in that it involves updating a
control law with an optimization routine which
attempts to minimize some error signal.
Historically, self-tuning is regarded as a stochastic
adaptive control method, developed around the optimal
regulation and optimal identification problems. It falls
into the class of Certainty equivalence controllers,
since no allowance is made in the control law for the
uncertainty in the parameter estimates, and no
NUJES Al Neelain University Journal of Engineering Sciences, VOL 1, No. 1, December 2013
NUJES Al Neelain University Journal of Engineering Sciences, VOL 1, No. 1, December 2013
uniquely defined. Should the true system time delay ever
exceed k then the estimator will become under
parameterized and the control solution will become suboptional, and self-tuning, in the strict mathematical sense,
will be lost. A change in the systems pure time delay can,
therefore, be regarded in the same light as any other
change in the systems numerator Bs polynomial, when the
pole shifter is applied.
The equations given can be seen to represent a set of
simultaneous equations to be solved on line for each
iteration of the self-tuning algorithm. These equations
impose a severe computation time penalty over the
minimum variance algorithms. The equation may be
solved by any standard technique (Gaussian elimination
etc.), or by using the RLS algorithm, thus saving some
computer space. The equations can be formulated as a
matrix, in which a structure becomes apparent. Such
structure may be exploited in the future to construct a
fast algorithm for solving the equation set (Kailath
1974).
3.2 The Extended Pole Shifter
The extended pole shifter represents the pole
shifting version of the Astrm-Wittenmark algorithm, and
has been mentioned by Wellstead et.al. (Wellstead, et.
a1., 1978). The algorithm was introduced in parallel with
the pole shifter , with the object of studying their relative
merits. The two methods have proved to display very
similar properties both mathematically and practically.
The pole shifter, however, consumes less computation
time than does the extended pole shifter, thus is generally
more useful.
3.3 Modifications to Wittenmarks Method
Wittenmark drew his method (Wittenmark 1973)
from the intuitions of classical control (Horowitz 1963),
and similarly other schemes suggest themselves.
Consider, for example, the problem of controlling a type
1 system which displays a small, slowly changing and
unpredictable drift input to its constituent integrator
(position control of a system having analog gyro rate
feedback, for instance). Classical control tells us that
simply cascading an integrator would make the system
difficult to control (double integrator), and that P + I are
the obvious answer.
Returning to two basic system model equations, can
replace ut by any xt defining
xt
1+ X
ut
Y
Such that
(1 + A) (1+ X) yt = z-k B.Y xt + (1+ C) (1+ X)et (1 + A)
(1 + X)rt
k
xt = z [(1 + A)(1 + C ) (1 + C )](1 + X ) y
B(1 + C )Y
yt
or
ut = k xt + Vt
where Vt = (1 k)xt + Vt-1
with Vt kept explicitly as the state of the integrator,
with the usual considerations of initialization, limiting
and prevention of winding up.
A second appeal to engineering
intuition suggests that if k, the proportional gain,
approaches unity the effect of the integral term on the
estimator should be negligible, i.e. the estimated A
polynomial need not be extended to allow for the
integrator and the sub optimality due to estimating too
few parameters should in this case be hardly noticeable,
Furthermore, it could save multiplication by re-writing
ut as
Vt = Xt + Vt-1
ut = xt + Vt
1
the function Y/(1+X) then becomes (1 + ) z
1 z 1
G
A'' T ' '
A' ' T ''
yt = e '' yt = e
yt
f0 + F
Be
BS (1 + C )
(1 + AS ) 1 + z k AS' '
=
zBS
zBS''
-1
and SP = z
G
in Equation - Properties of the servo Self Tuners
f0 + F
The servo self-tuners formulation doubles the number
NUJES Al Neelain University Journal of Engineering Sciences, VOL 1, No. 1, December 2013
E3.60
[(1 + AS )(1 + C ) + Z 1 (T ' ' Ae' ' )] yop t = Z ( k +1) [(1 + As )(1 + C ) + Z k (T ' ' Ae' ' )] sp t + (1 + C )(1 + C ) et
[( 1 + A s )( 1 + C ) + Z
(T
''
A e' ' )]
( 1 + C )( 1 + Z
''
thus
yop
( k +1)
. sp
1 + C
1 + Z kT
''
et
G
u~t =
yt
1+ F
and
1 + AS
B (1)
SP =
Bs
B(1)
= z-k
BS
(1 + F )
.SPt +
et
B (1)
(1 + T )
''
NUJES Al Neelain University Journal of Engineering Sciences, VOL 1, No. 1, December 2013
fluctuations. Tuning times are usually longer and
transients worse than in the feedback self-tuners.
(iv) Saturation limits
The importance of reflecting back saturation limits in
feedback self-tuners was discussed and the same
arguments apply to servo self-tuning. A complication
arises when using laws. The control, ut is made up of two
parts; ut and xt. Limits must be reflected back to one or
both of these before and/or after they are added to form
ut. The best procedure is by no means certain, but the
method adopted in the examples in the next section was
to limit ut after it was formed and reflect the limit back to
u t.
Choice of Control Law
The minimum variance control laws can only be
applied to systems, which are minimum phase (z-plane)
over their entire operating ranges. NMP systems occur
very frequently in practice, most systems of third order
and over are NMP and most systems with significant time
delays are NMP. Most systems with non integer time
delays over 50% of the sampling interval are NMP and
such time delays can frequently be attributed to the
computation time of the algorithm. The pole shifting
controller can cope with NMP systems and with varying
or unknown pure time delays hence is a most practical
general purpose algorithm. The extended pole shifter has
no advantages over the pole shifter and need not be used
in practice.
NMP system can be tackled with the Lambda
controller, which requires much less computation effort
than does the pole shifter. Closed loop pole positions can,
however, vary with changing system dynamics and might
even drift into instability. The minimum variance
algorithms are most useful for controlling fast, low order
systems with negligible pure time delays. The shortened
minimum variance algorithm runs slightly faster than the
Astrm Wittenmark law and tunes the same control law
for ke= 0 or 1. The Astrm Wittenmark algorithm is
optimal for any ke value, most systems with large time
delays, however tend to be NMP.
Feed forward versions of the regulation laws may be
used where servo responses are important but require
considerably more computational effort than do the
feedback laws.
Tuning Behavior and Block Traces
The degree of tuning of a self-tuning controller, in
practice must be a tradeoff between precision of
regulation and control effort, optimality at an operating
point and tracking ability, and disturbance rejection and
step responses, with saturation limits, slow rate limits and
sampling rate limits to be taken into consideration.
NUJES Al Neelain University Journal of Engineering Sciences, VOL 1, No. 1, December 2013
NUJES Al Neelain University Journal of Engineering Sciences, VOL 1, No. 1, December 2013
Linking the Model with Self-tuning:
The model equations provided in the above mentioned
section were used to design the different Graphic User
Interfaces (GUI) used in this simulation. The MATLAB
help file provided the procedure steps on the formulation
and design of the GUI (MATLAB User Guide 2004). The
designed GUI was then provided with the model
equations and the industrial Parameters.
The application was made on two refinery units mainly
the Atmospheric Distillation Unit (ADU) and the
Residual Fluid Cracking Unit (RFCC) by making
different manipulations to the inputs, namely temperature
and feed concentration. The results of this
implementations was carried out and the results are
explained in details elsewhere (Nikolaou, Muske, 1995,
Daau, et. al., 2008, Fissore, 2009, Mohammed, et. al.,
2010).
Conclusion
Extensive studies were made into the problems of
initialization of self-tuners and results and guidelines
have been presented. Two methods of handover of control
have been presented, namely handover by prior
simulation, and handover by open loop tuning to the
system under fixed-term digital control.
The point has been made that self-tuned control laws can
be fixed as conventional digital control laws. It has been
asserted that self-tuning can be used as an online
controller design tool, to good effect. The control
problems of some non-linear systems were looked at, and
some distinctions drawn to classify those non-linearity
with which self-tuning can and cannot cope.
The
application of these schemes when coupled with the MPC
and the model and design equations of the industrial
equipment under investigation showed clear and
progressive success on using these type of controlling
systems for industrial purposes.
References
Astrm K.J. and Wittenmark B.,2008, Adaptive Control
second edition.pp51-52; Dover.
Alster J. and Belanger P.R., 1974, A Technique for Dual
Adaptive Control Automatica 10 pp627-634.
Astrm K.J., 1965 Optimal Control of Markov Processes
within complete State Information.
Part 1, J.
Math. Anal. Applic. 10 pp 174-205; Part 2, J.Math. Anal.
Applic.36 pp 403-406 (1969)
Astrm K.J. 1970, Introduction to Stochastic Control
Theory Academic Press. New York.
Astrm K.J., Borrison U. Ljung L. and Wittenmark
B.,1977, Theory and Applications of Self-Tuning
Regulators, Automatica 13 pp. 457-476.
Astrm K.J., and Eykhoff P., 1971 Systems
Cope S.N. 1976, The Microprocessor Implementation
of a Self-tuning Controller, IEE Colloquium on Selftuning and Adaptive Systems.
10
NUJES Al Neelain University Journal of Engineering Sciences, VOL 1, No. 1, December 2013
NUJES Al Neelain University Journal of Engineering Sciences, VOL 1, No. 1, December 2013
Control., 94 pp 119-132.
Landau I.D., 1974, A Survey of Model Reference
Adaptive Techniques Theory and Application,
Automatica, 10, pp 353-379.
Lindorf D.P. and Carroll R.L., 1973, Survey of Adaptive
Control using Liapunov Design, Int. J. Control, 18, pp
897-914.
Ljung L., 1974a, Convergence of Recursive Stochastic
Algorithms Lund Report 7403, Div. Aut. Control, Lund.
Inst. Tech., Sweden.
Ljung L., 1974b, Convergence of Recursive Stochastic
Algorithms, Proc. IFAC Symposium on Stochastic
Control, Budapest, pp 215-223.
Ljung L., 1977a,Analysis of Recursive Stochastic
Algorithms, IEEE Trans. Aut. Control ,AC-22,pp 551575.
Ljung L., 1977b, On Positive Real Transfer Functions
and the Convergence of some Recursive Schemes, IEEE
Trans. Aut. Control, AC-22,pp 539-550.
Ljung L. and Wittenmark B., 1974, Asymptotic
Properties of Self-Tuning Regulators, Report 7404,Div.
Aut. Control, Lund Inst. Tech., Sweden.
McGreavy C. and Gill P.J., 1975, Self-Tuning State
Variable Methods for DDC, 2nd IEE Conference on
Trends in on Line Computer Control Systems, Sheffield,
UK.( IEE Conference publ.127)
Mendes M., 1971, An On-Line Adaptive Control
Method, Automatica, 7, pp 323-332.
Morris A.J., Fenton T.F. and Nazer Y.,1977,
Application of Self-Tuning Regulators to the Control of
Chemical Processes, IFAC Congress on Digital
Computer Applications to Process Control, The Hague,
Holland.
Morris A.J., Nazer Y. and Chisholm K.,1979, Single
and Multivariable Self-tuning Microprocessor-Based
Regulators, 3rd IEE Conference on Trends in On Line
Computer Control Systems, Sheffield, UK, ( IEE
Conference publ. 172).
Morris E.L. and Abaza B.A., 1973, Adaptive Digital
Control of a Steam Turbine, Proc. IEE, 123,pp 549-553.
Murphy W.J., 1968, Optimal Stochastic Control of
Discrete Linear Systems with Unknown Gain IEEE
Trans. Aut. Control, AC-13 pp 338-344.
Muske, K.R., Linear Model Predictive Control of
Chemical Processes, PhD Dissertation, The Univ. of
Texas at Austin, May 1995.
Nikolaou, M. Model Predictive Controllers : A Critical
Synthesis of theory and Industrial Needs Chemical Eng.
Dept., University of Houston , Houston TX, U,S,A.
2009.
Parks P.C. 1966, Lyapunov Redesign of a Model
Simulation Council Conference on Computer Simulation,
Bowness- on-Windermer.
Wellstead P.E. and Edmunds J.M., 1975, On Line
Process Identification and Regulation, 2nd IEE
Conference on Trends in On Line Computer Control
11
12
NUJES Al Neelain University Journal of Engineering Sciences, VOL 1, No. 1, December 2013