[IEEE 2012 UKACC International Conference on Control (CONTROL) - Cardiff, United Kingdom...
-
Upload
john-anthony -
Category
Documents
-
view
220 -
download
0
Transcript of [IEEE 2012 UKACC International Conference on Control (CONTROL) - Cardiff, United Kingdom...
A systematic selection of an alternativeparameterisation for predictive control
Bilal KhanAutomatic Control and Systems Engineering
The University of Sheffield,UKEmail: [email protected]
John Anthony RossiterAutomatic Control and Systems Engineering
The University of Sheffield,UKEmail: [email protected]
AbstractβAlternative parameterisations have been shown toimprove the feasible region for a predictive control law whenthe number of degrees of freedom is limited. One questionyet to be resolved is: which alternative parameterisation isbest for a particular problem and what choice of parameter(s)within each parameterisation will lead to an improved feasibleregion with good performance? This paper tackles this questionand demonstrates two systematic approaches to select the bestalternative parameterisations. These approaches are based onmultiobjective optimisation and a pragmatic selection. Numericalexamples demonstrate the efficacy of both methods.Keywords: Alternative parameterisation, Feasibility, multiobjec-tive optimisation.
I. INTRODUCTION
Model Based Predictive Control (MPC) or receding horizoncontrol (RHC) or embedded optimisation or moving horizon orpredictive control [2], [3], [1], is the general name for differentcomputer control algorithms that use past information of theinputs and outputs and a mathematical model of the plant tooptimise predicted future behaviour. MPC has been developedwidely both in the process industry and control researchcommunity and has reached a high degree of maturity in itslinear variant. Currently MPC research focuses on stochasticand nonlinear scenarios, robustness and fast optimisation orrelated computational aspects.
All algorithms to some extent form a trade-off betweenfeasibility, optimality and inexpensive optimisation, but maynot have systematic tools for doing this trade-off. The successof earlier industrial heuristic MPC algorithms motivated theresearch community to develop several algorithms with im-proved performance and feasibility. There are several success-ful theoretical approaches but few of them are exploited com-mercially for real-time implementation. One important issuefor real-time implementation is the ability to do a systematictrade-off between feasibility, performance and computationalburden when choosing from currently available algorithms.In recent years many authors have focussed on parametricsolutions [5] or efficient implementations of online optimisers(e.g. for quadratic programming). However, this paper willfollow a different route and consider the underlying structureof the MPC algorithm. Specifically, this paper proposes a sys-tematic selection of an alternative parameterisation to enlargethe feasible region without too much detriment to performance(or optimality) and the computational burden.
In an MPC problem formulation, stability can be ensuredusing dual mode prediction by including a terminal constraintat the expense of the computational load [16]. In dual modeprediction, different formulations of the degrees of freedom(d.o.f.) or decision variables have been utilized using interpola-tion techniques [8] to enlarge feasible region without detrimentto performance. There are different variants of interpolation,however, these methods are limited to small dimensionalsystems. Another approach in the literature is a concept oftriple mode control [6], [7]; the triple mode strategy introducesan extra mode which may reduce t he online computationalburden with good performance and feasibility. However, inthis strategy, the challenge is to find a suitable linear timevarying (LTV) control law which enlarge feasible region withimproved performance. One pragmatic solution to finding thislaw is tackled in [12] which used Laguerre parameterisations.However, this paper does not pursue Triple mode approachesas the offline complexity and decision making is substantiallyincreased as compared to dual mode approaches.
More recently, alternative parameterisation based on La-guerre [17] and Kautz functions have been proposed to sim-plify the trade-off without increasing computational burdenwithin dual mode prediction paradigm [9], [10], [11]. Themain idea is to form the degrees of freedom in the predictionsas a combination of either Laguerre or Kautz functions. Thesefunctions have proven to be a very effective for improvingthe volume of the feasible region with a limited numberof d.o.f. These alternative parameterisations are generalisedusing orthonormal basis functions with Laguerre and Kautzfunctions as special cases [13].
The key question left to be resolved is, what is the bestprediction dynamics to assume for predicted inputs, that is thed.o.f., to allow for a large feasible region without detrimentto the closed loop performance and computational burden? In[15], different ways of parameterising the d.o.f. were inves-tigated and a mechanism based on a Monte Carlo approachto define the best parameterisation using optimal sequencesfor numerous search directions was considered. However, asystematic choice for the underlying dynamics remains anopen question. This paper focuses on systematic approaches toidentify the best parameterisation dynamics using: (i) a multi-objective optimisation and (ii) a pragmatic choice.
This paper assumes that Laguerre, Kautz and Generalised
246
UKACC International Conference on Control 2012 Cardiff, UK, 3-5 September 2012
978-1-4673-1560-9/12/$31.00 Β©2012 IEEE
parameterisation are able to achieve large feasible regionswhile maintaining local optimality and a relatively low com-putational complexity [9], [10], [11], [12], [13], [14] andextends the earlier studies in [11], [13] in order to identifya systematic selection of the best parameterisation dynamics.Section II will give the necessary background about an optimalMPC, Laguerre MPC, generalised parameterisation for optimalMPC. Section III discusses two proposed schemes to identifythe best parameterisation dynamics based on a muti-objectiveoptimisation and a pragmatic approach. Numerical examplesare presented in section IV and paper finishes with conclusionsand future work in section V.
II. BACKGROUND
This section will summarise the background informationrelated to nominal dual-mode MPC and the use of alternativeparameterisations within MPC.
A. Problem formulation for MPC
Assume a discrete time linear time invariant (LTI) statespace model of the form
π₯π+1 = π΄π₯π +π΅π’π (1)
where π₯π β π ππ₯ and π’π β π ππ’ which are the state vectors andthe plant input respectively. Assume that the states and inputsat all time instants should fulfill the following constraints:
π’ β€ π’π β€ π’; Ξπ’ β€ Ξπ’π β€ Ξπ’; π₯ β€ π₯π β€ π₯ (2)
B. Nominal MPC algorithm
In dual-mode MPC it is assumed that one has total freedomin the choice of the input signal π’π up to horizon ππ subjectto constraints. Beyond horizon ππ, a terminal control law withan asymptotic stabilising optimal feedback gain πΎ is assumed.The βpredictedβ control law [3], [16] takes the form
π’π = βπΎπ₯π + ππ, π = 0, ..., ππ β 1,
π’π = βπΎπ₯π, π β₯ ππ, (3)
where only first control move is ever implemented, ππ aredegrees of freedom (d.o.f.) available for constrained handling.The terminal control law defines an invariant set π0 for statevector π₯π [8]. This invariant set is also known as maximumadmissible set (MAS) which satisfies all polytopic constraintswith recursive use of the terminal control law π’π = βπΎπ₯π βπ0. The MAS is defined as:
π0 = {π₯0 β π ππ₯ : π₯ β€ π₯π β€ π₯, π’ β€ βπΎπ₯π β€ π’,
π₯(π + 1) = π΄π₯(π) +π΅π’(π), βπ β₯ 0} (4)
In compact form defined as π0 = {π₯π : ππ₯π β€ π} forsuitable π and π. Performance, either predicted or actual, willbe assessed by the cost
π½ =ββπ=0
π₯ππππ₯π + π’π
ππ π’π (5)
Substituting the nominal model and predicted control values(3) into (5) and ignoring terms that do not depend upon the
d.o.f., one finds from [3] that the optimisation problem in (5)can be reformulated as:
minπΆ
π½π = πΆπππΆ π .π‘. ππ₯π +ππΆ β€ π; (6)
where πΆ = [πππ , . . . , πππ+ππβ1]. Details are in the literature [2],
[3], [4]. The maximal control admissible set (MCAS) ππ, thefeasible set for optimal control problem in (6) that satisfies allpolytopic constraints, is defined as:
ππ = {π₯π : βπΆ,ππ₯π +ππΆ β€ π} (7)
The optimal MPC (OMPC) algorithm with guarantees ofrecursive feasibility and convergence, for the nominal caseis given by solving QP optimisation (6) at every samplinginstance then implementing the first component of πΆ, that isππ in the control law of (3). The algorithm is formulated as:
Algorithm 2.1: OMPC [16], [3]
πβπ = πππ minππ
π½π π .π‘. ππ₯π +πππ β€ π;
Implement π’π = βπΎπ₯π + ππ1 πβπ where ππ1 = [πΌ, 0, . . . , 0].
When the initial states π₯π β π0 then the optimising πβπ is zeroso the terminal control law π’π = βπΎπ₯π is implemented.
Remark 2.1: There is a well understood set of potentiallyconflicting objectives using OMPC e.g. between the desire forgood performance and large feasible regions with the equallyimportant desire to keep the number of d.o.f. small. In order forOMPC to obtain a large feasible region and good performance,a large number of d.o.f. or ππ is required.
C. Generalised parameterisation for Optimal MPC
Alternative parameterisation techniques for the d.o.f. inthe future control values have been developed to improvethe feasible region in nominal case. Laguerre, Kautz andgeneralised parameterisations have been proposed in [9], [11],[13] as an effective alternative to the standard basis set. Thissection will summarise Generalised optimal MPC.
1) Generalised functions: The generalised parameterisation[13] is defined using a higher order network such as:
π’π(π§) = π’πβ1(π§)(π§β1 β π1) . . . (π§
β1 β ππ)
(1β π1π§β1) . . . (1β πππ§β1); (8)
0 β€ ππ < 1, π = 1 . . . π
With π’1(π§) =
β(1βπ2
1)...(1βπ2π)
(1βπ1π§β1)...(1βπππ§β1) . The generalised functionwith ππ, βπ = 1, . . . , π gives [13]
πΏπππ’ππππ : π’π = βπ, ππ ππ = [π]
πΎππ’π‘π§ : π’π = π¦π, ππ ππ = [π, π]. (9)
The Laguerre function is a special case of a generalisedfunction and may be computed using the following state-space
247
dynamics model:
π’(π + 1) =
βββββ
π 0 0 0 . . .π½ π 0 0 . . .
βππ½ π½ π 0 . . ....
......
.... . .
βββββ
οΈΈ οΈ·οΈ· οΈΈπ΄πΊ
π’(π); (10)
π’(0) =β1β π2[1,βπ, π2, . . . ]π ; π½ = 1β π2
2) GOMPC: generalised functions and MPC: The predic-tions using d.o.f. based on generalised functions [13] are:
πΆ =
ββββββ
ππ...
ππ+ππβ1
...
ββββββ =
βββπ’(0)ππ’(1)π
...
βββ π = π»πΊπ (11)
where π is the ππΊ dimension decision variable when oneuses the first ππΊ column of π»πΊ. The only difference betweenLaguerre MPC and Kautz or generalised MPC is that of π»πΊ
matrix. For further details readers are referred to [11] and [13].Algorithm 2.2: GOMPC
πβ = πππ minπ
ππ [
ββπ=0
π΄ππΊπ’(0)ππ’(0)π (π΄π
πΊ)π ]π
π .π‘. ππ₯π +ππ»πΊ π β€ π (12)
Define πβπ = π»πΊπβπ and implement π’π = βπΎπ₯π + πππ π
βπ.
where πππ is the π-th standard basis vector.Remark 2.2: It is straightforward to show, with conven-
tional arguments, that all algorithms (i.e. LOMPC, KOMPC,GOMPC) using terminal constraints within MPC problem for-mulation provides recursive feasibility and Lyapunov stability.
III. BEST ALTERNATIVE PARAMETERISATION SELECTION
In generalised parameterisation there are two clear choiceswithin the future input predictions. First one can choose theorder of the dynamics, that is the number of poles ππ in πΊπ(π§),and second is the parameter selection(s), that is the actualvalues of ππ. This section discusses the systematic selectionof the parameterisation dynamics by asking what impact thischoice has on feasibility and performance?
A. The best choice for order of the prediction dynamics
The prediction dynamics for the 3rd order prediction dy-namics from (8) (which can easily be extended to ππ‘β order)can be defined as:
π’(π + 1) =
βββββββ
π 0 0 . . .π π 0 . . .
βππ (1β ππ) π . . .ππ2 βπ(1β ππ) (1β ππ) . . .
......
.... . .
βββββββ
οΈΈ οΈ·οΈ· οΈΈπ΄πΊ
π’(π)
π’(0) = πΎ[1, 1, 1,βπ, ππ,βπππ, π2ππ, . . . ]π , (13)
πΎ =β(1β π2)(1β π2)(1β π2),
The dynamic structure of the prediction in (13) is quite genericwith distinct eigenvalues. To fulfill the algebraic relationsin (11), π΄πΊ used in (13) must be a square matrix with amaximum number of dimensions the same as ππ. Moreovera key observation from the prediction matrix structure is thatπππ(π΄πΊ) = ππ is an upper bound on dynamic dimensions. Infact, one could select πππ(π΄πΊ) β€ ππ.
B. Multiobjective solution to select best parameterisation
The main objective of this paper is to formulate a system-atic method for handling the compromise between feasibility,performance and computational burden. This section showshow one can produce trade-off curves between feasible volumeπ£, performance loss π½ and computational burden (implicitlylinked to ππ) of the parameterised MPC problem. Such trade-off curves allow us to formulate a multi-objective optimi-sation problem as a function of parameterisation parametersπΌ = [π1, . . . , ππ] β (0, 1)π:
π½(ππ, π) := minπΌ
π½,maxπΌ
π£
π .π‘. ππ₯+ππ»π β€ π,
π£ =π£ππ(ππ»)
π£ππ(ππππ‘),
π½ =1
π
β π½π»(π₯)β π½πππ‘(π₯)
π½πππ‘(π₯), (14)
πΌ = [π1, . . . , ππ], 0 < ππ < 1, π β€ ππ,
π’(π + 1) = π΄πΊπ’(π),π» =
(π’π (0) π’π (1) . . .)π
,
where ππππ‘ := {(π₯, π)β£ππ₯ + ππ β€ π} represents the MCASwith ππ = 20 (used as an approximate for the global maximumMCAS) for comparison, ππ» = {(π₯, π)β£ππ₯ + ππ»π β€ π} isthe polytope sliced by the matrix π» , π£ππ(.) is the volume,π½πππ‘(π₯) = ππ£π. {π½π(π₯, π)β£(π₯, π) β ππππ‘} and π½π»(π₯) =ππ£π. {π½π(π₯,π»π)β£(π₯, π) β ππ»} for convex function π½π(π₯, π).
1) Volume approximation: Computing the volume of a highdimensional polytope is a complex task, and can, in the worstcase be exponential in the size of the data π and π or ππ» .Consequently, this paper approximates the volume. First selecta large number of equi-spaced (by solid angle) or randomdirections in the state space i.e. π₯ = (π₯1, . . . , π₯π) and then,for each direction, the distance from the origin to the boundaryof MCAS is determined by solving a linear programming (LP)and clearly the larger the distance, hereafter denoted as radius,the better the feasibility. Finally radii are normalised againstthe radii obtained for ππππ‘. Although this might be consideredsomewhat arbitrary, it seems a pragmatic way of indicatingrelative volumes of different feasible regions compared to thebest feasible shape with a large ππ. One might argue thata precise volume measurement would, in some sense, be anequally arbitrary comparison.
2) Performance approximation: Similarly in case of calcu-lating the performance loss, it is converted from integral to
248
sampled sum i.e.
π½ :β 1
π
β π½π»(π₯)
π½πππ‘(π₯)β 1 (15)
That is, one computes the predicted performance for all thegiven state directions and compares each of these to theβglobalβ optimum. Equation (15) normalises these so that thebest achievable performance would correspond to a π½ measureof unity.
By deploying explicit numeric measures of volume andperformance, the multi-objective optimisation problem is ableto generate trade-off curves between the d.o.f. ππ, average radiiand average performance loss of the parameterised problemwith different parameterisation settings.
C. Pragmatic selection
In practice it is known from parameterisation insight thatthe optimal ππ is highly nonlinear in terms of its dependenceupon the current state π₯π; more over the solution is to someextent unpredictable. The following gives a simple approachto identify a pragmatic selection of pole location(s) and orderselection of parameterisation dynamics.
1) Order selection: The generalised functions with higherorder orthonormal functions have more flexibility to improvefeasibility while retaining good performance with a limitednumber of d.o.f. [13]. So πππ(π΄πΊ) = ππ is a sub-optimalchoice to select order parameterisation dynamics.
2) Parameter selection: It is observed from simulationresults that there is a relation between the closed loop polesand good locations of parameterisation dynamics. The polelocations of generalised functions can be selected to be equalto poles of the closed loop system using an optimal gain πΎ;this may be sub-optimal but is efficient.
D. Summary
In summary, there are two choices: (i) the order of param-eterisation dynamics (πππ(π΄πΊ)) and (ii) parameter value(s)or pole location(s) ([π1, . . . , ππ]). One can use multi-objectiveoptimisation to find πππ(π΄πΊ) and [π1, . . . , ππ]. Another keyobservation is that the maximum πππ(π΄πΊ) β€ ππ and closedloop eigenvalues or pole locations with an optimal gain πΎ area pragmatic choice to tune the parameterisation dynamics.
IV. NUMERICAL EXAMPLES
In this section three numerical examples are presented toillustrate the efficacy of the proposed approach. The aim isto compare two aspects: (i) feasible regions; (ii) performanceloss for the generalised parameterisation dynamics with differ-ent parameter(s) or pole location(s) selection by varying thed.o.f. or ππ. The trade-off curves are shown between averageradius and performance loss as a function of the differentparameterisation parameter(s) and d.o.f. or ππ. The parallelcoordinates are also shown to represent an alternative variationof the parameter(s) and the effects on average performance lossand average feasibility gain. The multi-objective optimisationis done using the NSG-II Matlab toolbox.
A. Examples
Three numerical examples are simulated using weight-ings π = πΆππΆ, π = 2 and are simulated by varyingππ = [2, . . . , 6] and with different parameter values. Trade-off curves between average radius and average performanceloss are plotted by solving the multi-objective optimisationwith varying parameter values. Similarly the effect on averageperformance loss and average radius are shown by varyingππ. Conversely, Table 1 shows the average performance lossand average radius using a pragmatic approach to parameterselection.
1) Example 1:
π΄ =[1.2
]; π΅ =
[1]; πΆ =
[1];
β 0.2 β€ π’π β€ 0.2; β0.2 β€ Ξπ’π β€ 0.2; β5 β€ πΆπ₯π β€ 5;
2) Example 2:
π΄ =
[0.6 β0.41 1.4
]; π΅ =
[0.20.05
]; πΆ =
[1 β2.2
]β 0.8 β€ π’π β€ 1.5; β0.4 β€ Ξπ’π β€ 0.4; β5 β€ πΆπ₯π β€ 5;
3) Example 3:
π΄ =
β‘β£ 1 0.1 0.3
0 1 0.10.4 0 1
β€β¦ ; π΅ =
β‘β£011
β€β¦ ;
πΆ =[1 0 0
]β 2 β€ π’π β€ 2; β1 β€ Ξπ’π β€ 1; β5 β€ πΆπ₯π β€ 5;
B. Optimal selection based on multiobjective solution
The multi-objective procedure was run for different sampledirections, which were chosen uniformly on the unit circle.The parameters ππ representing d.o.f. were chosen in the range{2, . . . , 6}. The results for Examples 1, 2, 3 are shown inFigures 1-2, 3-5 and 6-7 respectively. In all plots, colour circlesrepresent the parameter values πΌ.
Figures 1, 3 and 6 show the parallel coordinates and trade-off between average performance loss and average radiusby varying ππ with different parameter values πΌ. Parallelcoordinates are used to plot all axes representing differentobjectives parallel to each other. Only the Laguerre functionparameterisation simulation results are shown in the figures,but there are similar results for higher order parameterisationsi.e. Kautz and 3rd order functions. These trade-off curvesmay be used as a selection criteria to choose best functionparameterisation πΌ.
Figures 2, 4 and 7 show the trade-off curves betweenaverage performance loss and average radius for ππ = 4with different parameterisation dynamics. It is observed that3ππ order function parameterisation improves both the feasibleregion without too much detriment to performance. An averageperformance loss and feasible region gain for example 2 isshown in figure 5 by varying ππ β {2, . . . , 6} for a Laguerreparameterisation. One can see that best parameter selectionmay improve the feasible region and performance with alimited number of d.o.f. ππ.
249
a Avg. radius Avg. perf. lossβ1
β0.5
0
0.5
1
Coordinate
Coo
rdin
ate
Val
ue
n
c=2
nc=3
nc=4
nc=5
nc=6
0.5 0.6 0.7 0.8 0.9 1β0.8
β0.6
β0.4
β0.2
0
Average radius
Ave
rage
per
form
ance
loss
n
c=2
nc=3
nc=4
nc=5
nc=6
PragmaticSelection
Fig. 1. Parallel coordinates and trade-off curves for Example 1
0.65 0.7 0.75 0.8 0.85 0.9 0.95 1β0.7
β0.6
β0.5
β0.4
β0.3
β0.2
β0.1
0
0.1
Average radius
Ave
rage
per
form
ance
loss
LaguerreKautz3rd OrderPragmaticSelection
Fig. 2. Trade-off as a function of the parameter πΌ : ππ = 4 for Example 1
C. Sub optimal selection based on closed loop eigenvalues
Table I, II and III show the pragmatic selection of param-eterisation dynamics using closed loop eigenvalues with anoptimal gain πΎ. It is clear from these that pragmatic choicesare sub optimal. The selection of parameterisation is based onclosed loop eigenvalues so parameterisation order is selectedbased on example dimensions.
Laguerre parameterisation is the only choice for example1. With a sub optimal choice of ππ = 2 achieves 70 %of feasible set and 22 % of performance loss. Similarly inexample 2, there are two parameterisation choices Laguerreand Kautz functions. The Laguerre function achieved 90 %of feasible set with 6 % performance loss. On the otherhandthe Kautz function achieved 95 % of feasible set with 1 %performance loss with ππ = 2. In example 3, there are threeparameterisation choices i.e. Laguerre, Kautz and 3rd orderfunctions. All achieved more than 90 % global feasible setand maximum of 8 % of performance loss.
Remark 4.1: It is clear from figures 1-4, 6 and 7 that thepragmatic choice has solution in vicinity of an optimal paretosolution.
Remark 4.2: The optimisation structure of an approximateglobal optimal control is different to the parameterised one,so in few cases performance may even improve.
a Avg. radius Avg. perf. lossβ0.5
0
0.5
1
Coordinate
Coo
rdin
ate
Val
ue
n
c=2
nc=3
nc=4
nc=5
nc=6
0.7 0.75 0.8 0.85 0.9 0.95 1β0.2
β0.1
0
0.1
0.2
Average radius
Ave
rage
per
form
ance
loss
n
c=2
nc=3
nc=4
nc=5
nc=6
PragmaticSelection
Fig. 3. Parallel coordinates and trade-off curves for Example 2
0.92 0.93 0.94 0.95 0.96 0.97 0.98β0.1
β0.08
β0.06
β0.04
β0.02
0
0.02
0.04
Average radius
Ave
rage
per
form
ance
loss
LaguerreKautz3rd Order
PragmaticSelection
Fig. 4. Trade-off as a function of the parameter πΌ : ππ = 4 for Example 2
V. CONCLUSION
The main contribution of this paper was to present asystematic approach to selecting the best alternative param-eterisation for MPC. Two techniques were discussed basedon a multi-objective optimisation and a pragmatic approachusing closed loop poles. Examples demonstrate that in manycases nominal closed loop poles are a good sub optimal choiceto tune the parameterisation dynamics. However an optimalselection may be made using a multi-objective optimisationtechnique. In future work there is a need to further investigatein parallel issues such as: will these choices lead to an efficientonline optimisation QP structure and/or low complexity multi-parametric solution? Finally, it is an interesting future area toconsider these choices within a robust scenario.
TABLE IPRAGMATIC SELECTION FOR πΌ
Example 1Parameterisation ππ Avg. perf. loss Avg. radius
Laguerre 2 22% 0.6904a=0.52 3 22% 0.8011
4 22% 0.87205 22% 0.92376 22% 0.9578
250
2 4 60.7
0.75
0.8
0.85
0.9
0.95
1
Number of d.o.f. (nc)
Ave
rage
rad
ius
a=0.02a=0.06a=0.11a=0.20a=0.41a=0.45a=0.46
2 4 6β0.18
β0.16
β0.14
β0.12
β0.1
β0.08
β0.06
β0.04
β0.02
0
Number of d.o.f. (nc)
Ave
rage
per
form
ance
loss
a=0.02a=0.06a=0.11a=0.20a=0.41a=0.45a=0.46
Fig. 5. Average radius and average performance loss for Example 2
a Avg. radius Ang. perf. lossβ0.5
0
0.5
1
Coordinate
Coo
rdin
ate
Val
ue
n
c=2
nc=3
nc=4
nc=5
nc=6
0.86 0.88 0.9 0.92 0.94 0.96 0.98 1β0.2
β0.1
0
0.1
0.2
Average radius
Ave
rage
per
form
ance
loss
n
c=2
nc=3
nc=4
nc=5
nc=6
PragmaticSelection
Fig. 6. Parallel coordinates and trade-off curves for Example 3
REFERENCES
[1] E. Camacho and C. Bordons, Model predictive control, Springer, 2003.[2] D. Q. Mayne, J. B. Rawlings, C. Rao and P. Scokaret, βConstrained model
predictive controlβ, Automatica, vol. 36, 2000, pp 789-814.[3] J. A. Rossiter, Model-based predictive control, A practical approach, CRC
Press, London; 2003.[4] E. Gilbert and K. Tan, βLinear sysetms with state and control constraints:
The theory and application of maximal output admissible setβ, IEEETrans. on Automatic Control, vol. 36, 1991, pp 1008-1020.
[5] P. Grieder, Efficient computation of feedback controllers for constrainedsystems, PhD Thesis, ETH, 2004
[6] L. Imsland and J. A. Rossiter, βTime varying terminal controlβ, Proc.16th IFAc World Congress, Prague, Czech Republic, 2005.
[7] L. Imsland, J. A. Rossiter, B. Pluymers and J. Suykens, βRobust triplemode MPCβ, IJC, vol. 81, 2005, pp 679-689.
TABLE IIPRAGMATIC SELECTION FOR πΌ
Example 2Parameterisation ππ Avg. perf. loss Avg. radius
Laguerre 2 6 % 0.9098a=0.72 3 3 % 0.9548
4 2% 0.96835 2 % 0.97996 2 % 0.9896
Kautz 2 1 % 0.9468(a,b)=0.72Β±0.37j 3 0 % 0.9584
4 0.5 % 0.98295 2 % 0.98856 2 % 0.9907
0.92 0.93 0.94 0.95 0.96 0.97 0.98 0.99 1β0.1
β0.08
β0.06
β0.04
β0.02
0
0.02
Average radius
Ave
rage
per
form
ance
loss
LaguerreKautz3rd OrderPragmaticSelection
Fig. 7. Trade-off as a function of the parameter πΌ : ππ = 4 for Example 3
TABLE IIIPRAGMATIC SELECTION FOR πΌ
Example 3Parameterisation ππ Avg. perf. loss Avg. radius
Laguerre 2 0.2 % 0.9113a=0.36 3 0.2% 0.9494
4 0.2% 0.96585 0.2% 0.97906 0.2% 0.98732 8 % 0.9580
a=0.74 3 6 % 0.97314 2 % 0.98655 0.7 % 0.99066 0.6 % 0.9936
Kautz 2 8 % 0.9580(a,b)=0.74Β±0.02j 3 8 % 0.9600
4 0.4% 0.97645 0.4% 0.98316 0.4% 0.9954
Generalised 2 2 % 0.9113(a,b,c)=0.36,0.74Β±0.02j 3 0 % 0.9494
4 0 % 0.96895 0 % 0.99306 0 % 0.9934
[8] J. A. Rossiter, B. Kouvaritakis, M. Bacic, βInterpolation based computa-tionally efficient predictive controlβ, IJC, vol. 77, 2004, pp 290-301.
[9] J. A. Rossiter and L. Wang, βExploiting Laguerre functions to improvethe feasibility/performance compromise in MPCβ, Proc. CDC., Cancun,Mexico, 2008.
[10] J. A. Rossiter, L. Wang and G. Valencia-Palomo, βEfficient algorithmsfor trading off feasibility and performance in predictive controlβ, IJC,vol. 83, 2010, pp 789-797.
[11] B. Khan, J. A. Rossiter and G. Valencia-Palomo, βExploiting Kautz func-tions to improve feasibility in MPCβ, Proc. 18th IFAC World Congress,Milan, Itlay, 2011.
[12] B. Khan and J. A. Rossiter, βTriple Mode MPC or Laguerre MPC: acomparisonβ, Proc. ACC, San Francisco, California, USA, 2011.
[13] B. Khan and J. A. Rossiter, βGeneralised Parameterisation for MPCβ,Proc. The 13th IASTED Inter. Conf. On Intelligent Systems and Control(ISC), Cambridge, UK, 2011.
[14] B. Khan and J. A. Rossiter, βComputational Efficiency of LaguerreMPC using Active set Methodsβ, Proc. The 13th IASTED Inter. Conf.On Intelligent Systems and Control (ISC), Cambridge, UK, 2011.
[15] G. Valencia-Palomo, J. A. Rossiter, C. N. Colin, R. Gondhalekar andB. Khan, βAlternative parameterisations for predictive control: how andwhy?β, Proc. ACC, San Francisco, USA, 2011.
[16] P. Scokaert and J. B. Rawlings, βConstrained linear quadratic regula-tionβ, IEEE Trans. on Automatic Control, vol. 43, 1998, pp 1163-1168.
[17] L. Wang, Model predictive control design and implementation usingMATLAB, Springer, 2009.
251