Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
Research ArticleTrajectory Tracking and Stabilization of a Quadrotor UsingModel Predictive Control of Laguerre Functions
Mapopa Chipofya1 Deok Jin Lee2 and Kil To Chong1
1 Department of Electronic Engineering Chonbuk National University Room 421 Engineering Building No 7 567 Baekje-daeroDeokjin-gu Jeonju-si Jeollabuk-do 561-756 Republic of Korea
2Department of Mechanical Engineering Kunsan National University Republic of Korea
Correspondence should be addressed to Kil To Chong kitchongjbnuackr
Received 12 September 2014 Accepted 29 October 2014
Academic Editor Sakthivel Rathinasamy
Copyright copy 2015 Mapopa Chipofya et al This is an open access article distributed under the Creative Commons AttributionLicense which permits unrestricted use distribution and reproduction in any medium provided the original work is properlycited
This paper presents a solution to stability and trajectory tracking of a quadrotor system using amodel predictive controller designedusing a type of orthonormal functions called Laguerre functions A linear model of the quadrotor is derived and used To checkthe performance of the controller we compare it with a linear quadratic regulator and a more traditional linear state space MPCSimulations for trajectory tracking and stability are performed in MATLAB and results provided in this paper
1 Introduction
A quadrotor is a helicopter which has four equally spacedrotors usually arranged at the corners of a square bodyWhen these rotors spin they push air downwards and inthe process create a thrust force that keeps the quadrotoraloft Unlike the conventional helicopter (with two rotors)which requires a swashplate mechanism in order to havemore degrees of freedom such mechanism is not needed inquadrotor systems since the two additional rotors provide thesame level of control as that with conventional helicoptersfixed with swashplate mechanism
The task to control a quadrotor is a fundamentallydifficult and interesting problemWith six degrees of freedom(three translational and three rotational) and only fourindependent inputs (rotor speeds) quadrotors are severelyunderactuated In order to achieve six degrees of freedomrotational and translationalmotion are coupledThe resultingdynamics are highly nonlinear especially after accounting forthe complicated aerodynamic effects Finally unlike groundvehicles quadrotors have very little friction to prevent theirmotion so they must provide their own damping in orderto stop moving and remain stable Put together these factorscreate a very challenging control problem
2 Literature Review
Amongst the many control techniques used in quadrotorcontrol PID LQR and recently MPC have been widely used
While PID is the most popular choice for controllingseveral types of processes it turns out that tuning becomes abig challenge especially forMIMOsystems like the quadrotorSeveral techniques have been used to divide the quadrotorcontrol problem to several SISO systems While this hasworked in some cases it has to be pointed out that this takesaway the most natural way of controlling the system anddesigning many SISO controllers may make maintenancemore difficult
LQR is a relatively modern control method that is verypowerful yet limited to applications where linear systemmodels are available To use this method to control non-linear systems a linear model has to be obtained from thecorresponding nonlinearmodel To evaluate the performanceof the LQR controller we need to compare it with resultsobtained from other control techniques This means design-ing the same system using another control technique whichcould be cumbersome Because of this LQR is only verypopular to use in inherently linear systems or where results ofcontrol obtained using other control techniques are availablefor comparison
Hindawi Publishing CorporationAbstract and Applied AnalysisVolume 2015 Article ID 916864 11 pageshttpdxdoiorg1011552015916864
2 Abstract and Applied Analysis
On the other hand MPC can handle both linear andnonlinear systems [1] Comparisons between the nonlinearsystem and its corresponding linearmodel can easily bemadewith very little modification Compared to PID tuning MPCis easier even for complex MIMO systems
However in trying to track the reference trajectory inan optimal way and at the same time obey the constraintsimposed MPC makes much more calculations than eitherPID or LQRThis computation burden makes MPC less suit-able for ldquofastrdquo processes It requires very powerful processorsto be used in hardware implementation Not surprisingly alot of effort is being put in to reduce the computations so thatMPC can be faster and easily be implemented even on lowcost processors
In [2] Wang proposed a method of designing MPC usingorthonormal functions This method makes less computa-tions than the traditional MPC With reduced number ofcomputations this technique can be used where rapid systemdynamics are requiredWe will use this method in our task tocontrol a quadrotor using MPC
Other MPC applications to quadrotor can be found in[3 4] In [4] the control structure consists of a controllerbased on MPC to track the reference trajectory and a secondone based on a nonlinear 119867
infincontrol technique to stabilize
the rotational movements Similarly in [3] anMPC controlleris designed to control position and a second feedforwardcontroller is used to perform stabilization of the quadrotorsystem
In both [3 4] the control loop uses two controllers thecontroller based onMPC which is used for tracking the posi-tion reference trajectory and a second controller (which isdesigned using other techniques) that is used for stabilizationof the rotational movements In this paper MPC is used tocontrol both the position and rotational movements This isgood for uniformity and ease of maintenance
3 Kinematics and Dynamics of a Quadrotor
A quadrotor is completely described by twelve states Thefirst six states describe translational motion of the quadrotorThese states are 119909 119910 and 119911 which represent position ininertial frame and 119906 V and 119908 which denote velocity vectorsalong the body frame Similarly the remaining six variablesdescribe rotational motionsThese are 120601 120579 and 120595 commonlyknown as Euler angles and 119901 119902 and 119903 which denote angularvelocities
A free body diagram of a quadrotor is shown in Figure 1The force 119865
lowastand moment 119872
lowast(where lowast = 119891 119887 119903 and 119897)
produced by the rotors can be described by their angularspeed (120596) as 119865
lowast= 119896
11205962 and 119872
lowast= 119896
21205962 respectively
Similarly torques 120591120601 120591120579 and 120591
120595and the thrust force (119865) are
related to angular speed of rotors as
(
119865
120591120601
120591120579
120591120595
) = 119872(
(
1205962
119891
1205962
119903
1205962
119887
1205962
119897
)
)
(1)
where119872 = [
1198961
1198961
11989611198961
0 minus1198971198961
0 1198971198961
1198971198961
0 minus11989711989610
minus11989621198962minus11989621198962
] and 119897 is the length of arm of the
quadrotorEquations describing the kinematics of the quadrotor are
given by
(
119910
) = 119877 (120601 120579 120595)(
119906
V119908
)
(
120601
120579
) = (
1 119904120601119905120579 119888120601119905120579
0 119888120601 minus119904120601
0119904120601
119888120579
119888120601
119888120579
)(
119901
119902
119903
)
(2)
where 119877(120601 120579 120595) = [
119888120579119888120595119904120601119904120579119888120595minus119888120601119904120595119888120601119904120579119888120595+119904120601119904120595
119888120579119904120595119904120601119904120579119904120595+119888120601119888120595119888120601119904120579119904120595minus119904120601119888120595
minus119904120579
119904120601119888120579
119888120601119888120579
] 119888 stands for
cos 119904 stand for sin and 119905 stands for tan Similarly dynamicequations can be derived using Newtons laws of motionwhich are listed in
(
V
) = (
119903V minus 119902119908
119901119908 minus 119903119902
119902119906 minus 119901V) + (
minus119892 sin 120579119892 cos 120579 sin120601119892 cos 120579 cos120601
) +1
119898(
0
0
minus119865
)
(
119902
119903
) =
(((
(
119869119910minus 119869
119911
119869119909
119902119903
119869119911minus 119869
119909
119869119910
119901119903
119869119909minus 119869
119910
119869119911
119901119902
)))
)
+((
(
1
119869119909
120591120601
1
119869119910
120591120579
1
119869119911
120591120595
))
)
(3)
where 119869119909 119869119910 and 119869
119911are the respectivemoment of inertia about
119883 119884 and 119885 axes of the quadrotor For the simulation of thecontroller of quadrotor we have considered 119865 120591
120601 120591120579 and 120591
120595
as the inputs to the system
31 Linearization In order to design a linear controller it isvital that (2) to (3) be simplified to linear ones To do that rollpitch and heading angles are assumed to operate within verysmall angles [5] which leads to the following
(
120601
120579
) =((
(
1
119869119909
120591120601
1
119869119910
120591120579
1
119869119911
120591120595
))
)
= 119892 minus119865
119898
= minus119892120579
V = 119892120601
(4)
4 Controller Design
In order to control attitude (120601 120579 and 120595) altitude (119911) andposition (119909 119910) we propose a control architecture given in
Abstract and Applied Analysis 3
X
Y
Z
(a)
Fl Ff
Fb Fr
120591l
120591b
120591f
120591r(r 120595)
(q 120579)
(p 120601)
u
120596
(b)
Figure 1 Inertial frame (a) and forces and torques acting on the quadrotor (b)
Positioncontroller
Altitudecontroller
Attitudecontroller
Quadrotordynamics
MPC
zref
xref
yref
120595ref
120579ref
120601ref
[x y][120579 120601 120595]
120591120579
120591120601
120591120595
F
z
Figure 2 Block diagram showing the structure of the controllers
Figure 2 where we have utilized MPC to design the threecontrollers The attitude controller generates 120591
120601 120591
120579 and 120591
120595
actuator signals whereas the altitude controller generatesrequired thrust for the system The position controller con-trols position in the 119909 and 119910 directions and generates controlsignals 120579 and 120601 which when combined with the 120595 signal actas reference signals to the attitude controller
In the next section we look at the design of each of thethree controllers
41 Discretization In this section discrete-time state-spaceequations necessary for altitude position and attitude con-trol are derived
411 Altitude Controller The linear equations responsible foraltitude control are
= 119908
= 119892 minus119865
119898
(5)
These two equations can be combined to give the followingsecond order ODE
= 119892 minus119865
119898 (6)
This second order ODE can be simplified to
= 119866 (7)
where 119866 = 119892 minus 119865119898 or 119865 = 119898(119892 minus 119866) Equation (7) can bewritten in state-space form as in the following equation
[
] = [
0 1
0 0] [
119911
] + [
0
1]119866 (8)
Using the forward Euler method and choosing a samplinginterval Δ119879 we can express (8) in discrete form as in
[119911 (119896 + 1)
V119911(119896 + 1)
] = [1 Δ119879
0 1] [
119911 (119896)
V119911(119896)
] + [0
Δ119879]119866 (119896) (9)
The states 119911(119896) and V119911(119896) represent the position and velocity
respectively and we could choose a proper output matrix 119862depending on which state we are going to measure In thispaper we are interested in the position and therefore we willuse the output matrix 119862 = [1 0]
412 Position Controller The equations responsible for posi-tion and velocity control in 119909 direction are
= 119906
= minus119892120579
(10)
These two equations can be combined to give the secondorder ODE
= minus119892120579 (11)
4 Abstract and Applied Analysis
Similarly the equations responsible for position and velocitycontrol in 119910 direction are
119910 = V
V = 119892120601
(12)
which give rise to the second order ODE
119910 = 119892120601 (13)
Equations (11) and (13) can be combined and form a state-space equation such as
[[[
[
119910
119910
]]]
]
=
[[[
[
0 1 0 0
0 0 0 0
0 0 0 1
0 0 0 0
]]]
]
[[[
[
119909
119910
119910
]]]
]
+
[[[
[
0 0
minus119892 0
0 0
0 119892
]]]
]
[120579
120601] (14)
As stated before we can use the forward Euler method toexpress (14) in discrete form Consider
[[[
[
119909 (119896 + 1)
V119909(119896 + 1)
119910 (119896 + 1)
V119910(119896 + 1)
]]]
]
=
[[[
[
1 Δ119879 0 0
0 1 0 0
0 0 1 Δ119879
0 0 0 1
]]]
]
[[[
[
119909 (119896)
V119909(119896)
119910 (119896)
V119910(119896)
]]]
]
+
[[[
[
0 0
minus119892Δ119879 0
0 0
0 119892Δ119879
]]]
]
[120579 (119896)
120601 (119896)]
(15)
The states 119909(119896) and 119910(119896) represent the position while V119909(119896)
and V119910(119896) represent velocity Since we are interested in the
position we will use the output matrix 119862 = [1 0 0 0
0 0 1 0] but
we could have used a different 2 times 4 output matrix if wewere interested in the velocity But it is important to notethat we can not use a 3 times 4 or 4 times 4 as there is no guaranteethat we will be able to control each of the measured outputsindependently with zero steady state errors This is generallythe case for a system with119898 inputs 119902 outputs and 119902 gt 119898 [6]
413 Attitude Controller In attitude control we are interestedin controlling the roll 120601 pitch 120579 and yaw (heading) 120595 suchthat we generate appropriate torque signals responsible forsteering the quadrotor in the desired direction with therequired attitude The equations responsible for roll controlare
120601 = 119901
=1
119869119909
120591120601
(16)
These two equations can be combined to give the secondorder ODE
120601 =1
119869119909
120591120601 (17)
Similarly the equations responsible for pitch are given by120579 = 119902
119902 =1
119869119910
120591120579
(18)
which give rise to the second order ODE
120579 =1
119869119910
120591120579 (19)
Also the equations responsible for heading are expressed as
= 119903
119903 =1
119869119911
120591120595
(20)
which like in roll and pitch give rise to the second orderODE
=1
119869119911
120591120579 (21)
Equations (17) (19) and (21) can be combined and written instate-space form as in
[[[[[[[
[
120601
120601
120579
120579
]]]]]]]
]
=
[[[[[[[
[
0 1 0 0 0 0
0 0 0 0 0 0
0 0 0 1 0 0
0 0 0 0 0 0
0 0 0 0 0 1
0 0 0 0 0 0
]]]]]]]
]
[[[[[[[
[
120601
120601
120579
120579
120595
]]]]]]]
]
+
[[[[[[[
[
0 0 0
119869minus1
1199090 0
0 0 0
0 119869minus1
1199100
0 0 0
0 0 119869minus1
119911
]]]]]]]
]
[
[
120591120601
120591120579
120591120595
]
]
(22)
where 120591120601 120591
120579 and 120591
120595are the torques required to give the
required roll pitch and yaw respectivelyAgain using the forward Eulermethod we express (22) in
discrete form as
[[[[[[[
[
120601 (119896 + 1)
V120601(119896 + 1)
120579 (119896 + 1)
V120579(119896 + 1)
120595 (119896 + 1)
V120595(119896 + 1)
]]]]]]]
]
=
[[[[[[[
[
1 Δ119879 0 0 0 0
0 1 0 0 0 0
0 0 1 Δ119879 0 0
0 0 0 1 0 0
0 0 0 0 1 Δ119879
0 0 0 0 0 1
]]]]]]]
]
[[[[[[[
[
120601 (119896)
V120601(119896)
120579 (119896)
V120579(119896)
120595 (119896)
V120595(119896)
]]]]]]]
]
+
[[[[[[[
[
0 0 0
Δ119879119869minus1
1199090 0
0 0 0
0 Δ119879119869minus1
1199100
0 0 0
0 0 Δ119879119869minus1
119911
]]]]]]]
]
[
[
120591120601(119896)
120591120579(119896)
120591120595(119896)
]
]
(23)
Since we are interested in measuring 120601(119896) 120579(119896) and 120595(119896) wewill use the output matrix 119862 = [
1 0 0 0 0 0
0 0 1 0 0 0
0 0 0 0 1 0
]So far we have formulated the necessary equations used to
build the three controllers Figure 2 shows in block diagramform what we have done so far But that is not the wholestory In the next section we will look at exactly how modelpredictive control is applied to the quadrotor systemusing thediscrete state-space equations formulated in this section
Abstract and Applied Analysis 5
Plant
Observer
MPC
u(t) y(t)
x(t)
Figure 3 Block diagram of SS-MPC
5 Model Predictive Control
One of the many tenets deployed in control of complexsystems like the quadrotor using MPC is to use modelsthat reduce the computation burden which is a necessitywhen it comes to online implementation It is also agreedupon that linear models perform better in this respect thanthe nonlinear models they represent In this and subse-quent sections the performance of a qudrotor controllerdesigned using Laguerre-based MPC (LMPC) is comparedwith that designed using themore common linear state-spaceapproach presented in [7] We will call the linear state-spacemethod SS-MPC
51 Linear State-Space MPC (SS-MPC) Linear state-spaceMPC is modeled using a linear state-space relation andplant constraints are modeled using linear equalities andinequalities When combined with convex quadratic costfunction this implies that the desired control action canbe obtained at each sample interval via the solution of acorresponding quadratic program This is attractive becausethe quadratic programs can be solved efficiently online [7]The general block diagram representing SSMPC is shown inFigure 3
511 Plant Theplant ismodeled by using the following state-space system
119909119896+1
= 119860119909119896+ 119861119906
119896+ 119908
119896
119910119896= 119862119909
119896+ 119863119906
119896+ V
119896
(24)
where 119908119896is the state noise and V
119896is the measurement noise
Both 119908119896and V
119896are assumed to be Gaussian distributed with
zero mean respective covariances of 119882 and 119881 and crosscovariance 119885 This is represented mathematically as
[119908119896
V119896
] sim 119873([0
0] [
119882 119885
119885119879
119881]) (25)
512 Observer Using Gaussian assumptions stated above itis possible to make optimal predictions of state and outputusing the Kalman filter as
119909119896+1|119896
= 119860119909119896|119896minus1
+ 119861119906 + 119870 (119910119896minus 119910
119896|119896minus1)
119910119896|119896minus1
= 119862119909119896|119896minus1
+ 119863119906119896
(26)
The closed loop gain 119870 is found by solving the discrete-timealgebraic Riccati equation (DARE)
119870 = (119860119875119862119879
+ 119885) (119862119875119862119879
+ 119881)minus1
119875 = 119882 + 119860119875119860119879
minus (119860119875119862119879
+ 119885) (119862119875119862119879
+ 119881)minus1
(119885119879
+ 119862119875119860119879
)
(27)
513 Optimal Estimation of State and Output The generalequation representing the optimal estimate of state at instant119895 and written in terms of the initial state 119909
119896+1|119896and future
control inputs 119906119896+119894|119896
is given by
119909119896+119895|119896
= 119860119895minus1
119909119896+1|119896
+
119895minus1
sum
119899=1
119860119895minus119899minus1
119861119906119896+119899|119896
(28)
Similarly the output equation is given by
119910119896+119895|119896
= 119862119860119895minus1
119909119896+1|119896
+ 119862(
119895minus1
sum
119899=1
119860119895minus119899minus1
119861119906119896+119899|119896
) + 119863119906119896+119895|119896
(29)
The vector containing all output vectors can be written as
119884119896=[[
[
119910119896+1|119896
119910119896+119873|119896
]]
]
(30)
while that containing all control actions can be written as
119880119896=[[
[
119906119896+1|119896
119906119896+119873|119896
]]
]
(31)
Therefore we can write 119884119896as
119884119896= 119865119909
119896+1|119896+ Φ119880
119896 (32)
where
119865 =
[[[[[[
[
119862
119862119860
1198621198602
119862119860119873minus1
]]]]]]
]
Φ =
[[[[[[
[
119863
119862119861 119863
119862119860119861 119862119861 119863
119862119860119873minus2
119861 sdot sdot sdot sdot sdot sdot 119862119861 119863
]]]]]]
]
(33)
6 Abstract and Applied Analysis
514 Cost Function The prime goal in MPC is to rejectdisturbances whilst tracking a reference signal But at eachinstant it has to ensure that the control signal is withinreasonable range that is practical for actuation An objectivefunction that combines these two tasks is
119869 (119909119896+1|119896
119880119896) =
1
2
119873
sum
119899=1
1003817100381710038171003817119910119896+119899|119896 minus 119903119896+119899
1003817100381710038171003817
2
119876+1003817100381710038171003817119906119896+119899|119896 minus 119906
119896+119899minus1|119896
1003817100381710038171003817
2
119878
(34)
where both119876 and 119878 are assumed to be symmetric and positivedefinite
52 Laguerre-Based MPC (LMPC) Model predictive controldesigned using Laguerre functions is developed and summa-rized in [2] The author presented this design technique forboth single-variable and multivariable systems Fortunatelythe derivations for the single-variable case and the multi-variable case are very similar and knowing one would giveenough insight into the derivation of the other For brevityonly the single-variable case is presented and it is hoped thatthis is enough to set forth the LMPC design technique
Consider a plant with 119901 inputs 119902 outputs and 119899 states asdescribed by (35) where the subscript119898 stands for model
119909119898(119896 + 1) = 119860
119898119909119898(119896) + 119861
119898119906 (119896) + 119861
119889120596 (119896)
119910 (119896) = 119862119898119909 (119896)
(35)
where 119906(119896) is the input variable 119910(119896) is the process outputand 119909
119898(119896) is the state vector 120596(119896) is the input disturbance
and is assumed to be a sequence of integrated white noiseThe plant described by (35) can be expressed in aug-
mented state-space form [2 6] as
119909 (119896 + 1) = 119860119909 (119896) + 119861Δ119906 (119896) + 119861120598120598 (119896)
119910 (119896) = 119862119909 (119896)
(36)
where 119860 = [119860119898
0119899times119902
119862119898119860119898119868119902times119902
] 119861 = [119861119898
119862119898119861119898
] 119861120598= [
119861119889
119862119898119861119889
] 119862 =
[0119902times119899
119868119902times119902
] and the state 119909(119896) = [Δ119909119898(119896) 119910(119896)]
119879 120598(119896) is azero mean white noise sequence related to the disturbance bythe difference equation 120598(119896) = 120596(119896) minus 120596(119896 minus 1) [6]
521 Design Framework In designing MPC using Laguerrefunctions the control trajectory Δ119880 is expressed using a setof orthonormal functions called Laguerre functions Since thestate and output vectors can also be described in terms ofΔ119880 it follows that they too can be expressed using Laguerrefunctions
At sampling instant 119896119894 the state variable 119909(119896
119894) is available
through measurement The future control trajectory is givenby
Δ119880 = [Δ119906 (119896119894) Δ119906 (119896
119894+ 1) Δ119906 (119896
119894+ 119873
119888minus 1)] (37)
where119873119888is the control horizon
Laguerre functions can approximate the incrementalterms contained inΔ119880The 119911-transfroms of the discrete-timeLaguerre functions are written as
Γ1(119911) =
radic1 minus 1198862
1 minus 119886119911minus1
Γ2(119911) =
radic1 minus 1198862
1 minus 119886119911minus1(
119911minus1
minus 119886
1 minus 119886119911minus1)
Γ119873(119911) =
radic1 minus 1198862
1 minus 119886119911minus1(
119911minus1
minus 119886
1 minus 119886119911minus1)
119873minus1
(38)
In the discrete-time Laguerre network of (38) 119886 is the poleof the discrete-time Laguerre network and 0 le 119886 lt 1 forstability of the network The parameter 119886 is called the scalingfactor and 119873 = 1 2 3 is the number of Laguerre termsused in the network Note that
Γ119896(119911) = Γ
119896minus1(119911) (
119911minus1
minus 119886
1 minus 119886119911minus1) (39)
with the first term as Γ1(119911) = radic1 minus 1198862(1 minus 119886119911
minus1
)Let 119897
119895(119896) be the inverse 119911-transform of the 119895th term in
the discrete Laguerre network and 119871(119896) the vector containingall inverse 119911-transform terms Then taking advantage of (39)successive inverse 119911-transform vectors are obtained throughthe difference equation
119871 (119896 + 1) = 119860119897119871 (119896) (40)
Thematrix119860119897has size119873times119873 and is a function of parameters
119886 and 120573 = 1 minus 1198862 The initial condition is given by
119871 (0)⊤
= radic120573 [1 minus119886 1198862
sdot sdot sdot (minus1)119873minus1
119886119873minus1
] (41)
In the case where119873 = 5 matrix 119860119897and the initial condition
119871(0) are
[[[[[
[
119886 0 0 0 0
120573 119886 0 0 0
minus119886120573 120573 119886 0 0
1198862
120573 minus119886120573 120573 119886 0
minus1198863
120573 1198862
120573 minus119886120573 120573 119886
]]]]]
]
(42)
119871 (0) = radic120573
[[[[[
[
1
minus119886
1198862
minus1198863
1198864
]]]]]
]
(43)
respectivelyAt sampling instant 119896
119894 the control trajectory is described
using Laguerre functions as in
Δ119906 (119896119894+ 119896) =
119873
sum
119895=1
119888119895(119896119894) 119897119895(119896119894) (44)
Abstract and Applied Analysis 7
where 119896 = 0 1 2 119873119901with 119873
119901as the prediction horizon
119873 terms used in the expansion and 119897119895(119896119894) is the inverse 119911-
transform of the 119895th term in the discrete Laguerre networkEquation (44) can be rewritten as
Δ119906 (119896119894+ 119896) = 119871 (119896)
⊤
120578 (45)
where 119871(119896) = [1198971(119896) 119897
2(119896) sdot sdot sdot 119897
119873(119896)]
119879 120578 = [11988811198882sdot sdot sdot 119888
119873]119879
and the coefficients 1198881 1198882 119888
119873are obtained from system
dataAt an arbitrary future instant 119898 the state is described
using Laguerre functions as
119909 (119896119894+ 119898 | 119896
119894) = 119860
119898
119909 (119896119894) +
119898minus1
sum
119894=0
119860119898minus119894minus1
119861119871 (119894)119879
120578 (46)
Similarly the output is described as
119910 (119896119894+ 119898 | 119896
119894) = 119862119860
119898
119909 (119896119894) +
119898minus1
sum
119894=0
119862119860119898minus119894minus1
119861119871 (119894)119879
120578 (47)
522 Cost Function The cost function is used to choose theoptimal control trajectory Δ119880 to bring the predicted outputas close as possible to the set-point The cost function
119869 =
119873119901
sum
119898=1
119909 (119896119894+ 119898 | 119896
119894)119879
119876119909 (119896119894+ 119898 | 119896
119894) + 120578
119879
119877119871120578 (48)
minimizes the error between the set-point signal and theoutput in the shortest possible time by carefully tuning theweighting matrices 119876 ge 0 and 119877
119871gt 0
523 Cost Function Minimization The state variable in (46)can be rewritten as
119909 (119896119894+ 119898 | 119896
119894) = 119860
119898
119909 (119896119894) + 120601 (119898)
119879
120578 (49)
where 120601(119898)119879
= sum119898minus1
119894=0119860119898minus119894minus1
119861119871(119894)119879
By substituting (49) into (48) and performing the partialderivative 120597119869120597120578 = 0 the Laguerre coefficients vector is foundto be
120578 = minusΩminus1
Ψ119909 (119896119894) (50)
withΩ = sum119873119901
119898=1120601(119898)119876120601(119898)
119879
+ 119877119871and Ψ = 120601(119898)119876119860
119898
524 Receding Horizon Control In receding horizon control(RHC) only the first term in Δ119880 that is Δ119906(119896
119894) is imple-
mented at instant 119896119894 The rest of the sequence is ignored
Only the most recent measurement is taken to form the statevector for calculation of the control signal This procedure isrepeated in real-time to give the receding horizon control law[6 8]
525 Stability In stability analysis we make use of the tech-nique of exponential data weighting originated by Andersonand Moore [9] and applied to MPC in [6] More specifically
we will concentrate on the discrete exponential factor 119890120582Δ119905where Δ119905 is the sampling interval and the discrete weightsform a geometric sequence 120572119895 119895 = 0 1 2
The proposed cost function is similar to the one usedin linear quadratic regulator (LQR) systems but with theinclusion of discrete weights
119869119908=
119873119901
sum
119895=1
120572minus2119895
119909 (119896119894+ 119895 | 119896
119894)119879
119876119909 (119896119894+ 119895 | 119896
119894)
+
119873119901
sum
119895=0
120572minus2119895
Δ119906 (119896119894+ 119895)
119879
119877Δ119906 (119896119894+ 119895)
(51)
For 120572 gt 1 the exponential weights 120572minus2119895 119895 = 1 2 3 119873119901
put more emphasis on the current state 119909(119896119894+ 119895 | 119896
119894) and less
emphasis on subsequent future statesThe exponentially weighted cost function can be
expressed more compactly as
119869119908=
119873119901
sum
119895=1
119909 (119896119894+ 119895 | 119896
119894)119879
119876119909 (119896119894+ 119895 | 119896
119894)
+
119873119901
sum
119895=0
Δ (119896119894+ 119895)
119879
119877Δ (119896119894+ 119895)
(52)
with state equation119909 (119896 + 1) = 119860
120572119909 (119896) + 119861
120572Δ (119896) (53)
and 119860120572= 119860120572 and 119861
120572= 119861120572
With 119876 ge 0 119877 gt 0 and 119873119901
rarr infin minimizing thecost function 119869
119908is equivalent to the DLQR problem which is
solved using algebraic Riccati equation (54)Thepair (119860120572 119861
120572)
is assumed to be controllable and (119860120572 119862) observable with
119876 = 119862119879
119862 Then there is a stabilizing state feedback controlgain matrix119870
119908
119870119908= (119877 + 120572
minus2
119861119879
119875119908119861)
minus1
120572minus2
119861119879
119875119908119860
119860119879
120572[119875
119908minus 119875
119908
119861
120572(119877 +
119861119879
120572119875119908
119861
120572)
minus1
119861119879
120572119875119908]119860
120572+ 119876 minus 119875
119908= 0
(54)that gives a stable closed loop system with all its eigenvaluesinside the unit circle and the the closed loop system beingdescribed by
119909 (119896119894+ 119895 + 1 | 119896
119894) = 120572
minus1
(119860 minus 119861119870119908) 119909 (119896
119894+ 119895 | 119896
119894) (55)
From (55) the transformed system has all its eigenvaluesinside the unit circle by taking119873
119901rarr infin So
120572minus1 1003816100381610038161003816120582max (119860 minus 119861119870
119908)1003816100381610038161003816 lt 1 (56)
which gives1003816100381610038161003816120582max (119860 minus 119861119870
119908)1003816100381610038161003816 lt 120572 (57)
Thus by choosing 120572 gt 1 there is a great chance that the systemwill be stable Several simulations on the quadrotor systemindicate that a choice of 120572 slightly greater than unity makesthe system stable almost all the time
8 Abstract and Applied Analysis
Table 1
Parameter ValueMass119898 07 kgLength 119897 0275m119869119909
01063 kgsdotm2
119869119910
01063 kgsdotm2
119869119911
02122 kgsdotm2
Table 2
Control horizon119873119888
119873 Required 119886
10 5 0606515 5 0716520 5 0778825 5 0818750 5 09048
6 Simulation Parameters and Results
In order to illustrate the controllerrsquos stability we will checkif the eigenvalues of each of the three individual controllersappear inside a unit circle Andwewill use a trajectory in a 3Dplane to check the overall controllerrsquos ability to track a desiredtrajectory
Table 1 shows the parameters of the quadrotor used insimulation
61 Stability Analysis In this section we look at the locationof eigenvalues of LMPC and check if they fall within the unitcircle for stability of the system We also compare them tothose obtained from the optimal DLQR system To ensurestability as explained in Section 525 we use 120572 = 12
By looking at Figure 4 we note that all the eigenvaluesappear inside the unit circle as required Also the eigenvaluesobtained from LMPC closely match with those of the optimalDLQRThis not only shows the system is stable but also showsthe controllers perform optimally
62 Trajectory Tracking In this section we look at theperformance of LMPC when used to track a 3D trajectoryAlso LMPC is compared to SS-MPC and individual statetrajectories are observed The starting point of the trajectoryis (119909 119910 119911) = (0 0 0) and final point is (119909 119910 119911) = (5 5 10)Parameter 119886 = 08187 while119873 = 5 See Figure 5
Clearly both LMPC and SS-MPC track the desired tra-jectory very well See Figures 5 6 and 7 However LMPCperforms better than SS-MPC in that it requires only fiveparameters (119873 = 5) to capture the control trajectorycompared to SS-MPCrsquosminimumof 25 (119873
119888= 25) parameters
(see Table 2 for relationship of119873 and119873119888with 119886 = 08187)
63 Control Horizon in Relation to 119886 and 119873 The controlhorizon (119873
119888) is related to the parameters 119886 and119873 [2] by
119886 asymp 119890minus119873119873
119888 (58)
minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1
minus08
minus06
minus04
minus02
0
02
04
06
08
1
Real
Imag
inar
yLMPCDLQR
Figure 4 Eigenvalues for DLQR and LMPC in altitude control
0 1 2 3 4 5 6
02
46
0
2
4
6
8
10
xy
z
DesiredSS-MPCLMPC
Figure 5 Tracking a 3D trajectory
631 Constant 119873 Using (58) and choosing a constantparameter119873 = 5 we get Table 2
Thus we can increase the control horizon by increasingthe parameter 119886 without necessarily changing the order119873
632 Constant Control Horizon 119873119888 Assuming we want to
achieve a control horizon of 30 we choose one parameterusually 119873 because it is an integer and directly defines theorder of the discrete Laguerre networkTherefater we simplyuse (58) to find the corresponding scaling factor 119886 as shownin Table 3
Abstract and Applied Analysis 9
0 500 1000 1500 2000 2500 3000 3500 40000
1
2
3
4
5
6
Time (ms)
x p
ositi
on
DesiredSS-MPCLMPC
(a)
0 500 1000 1500 2000 2500 3000 3500 4000Time (ms)
0
1
2
3
4
5
6
y p
ositi
on
DesiredSS-MPCLMPC
(b)
Figure 6 Tracking 119909 (a) and 119910 in (b)
2
4
6
8
10
12
00
500 1000 1500 2000 2500 3000 3500 4000Time (ms)
z p
ositi
on (m
)
DesiredSS-MPCLMPC
(a)
0
01
02
03
04
05
0 500 1000 1500 2000 2500 3000 3500 4000Time (ms)
Hea
ding
angl
e120595
DesiredSS-MPCLMPC
(b)
Figure 7 Altitude (119911) (a) and heading angle (120595) (b)
Table 3
Target control horizon119873119888
Chosen119873 Required 119886
30 1 0967230 3 0904830 5 0846530 7 0791930 9 07408
Figures 8 9 10 11 12 and 13 compare two LMPC designsboth of which yield a control horizon of 30 One designuses 119873 = 1 and the other 119873 = 9 Thus we can choose aLaguerre network with lower number of terms119873 (that gives
lower computation burden) but with a larger scaling factor119886 to achieve the same control horizon The computationalcost is lower if a smaller number of parameters are usedFor 119873 = 1 only one Laguerre term is used to capture thecontrol trajectory while119873 = 9 requires nine Laguerre termsto capture the same control trajectory
7 Conclusion
This paper has presented a solution to stability and trajectorytracking of quadrotor systems using a model predictivecontroller designed by using a special type of orthonor-mal functions called Laguerre functions A quadratic cost
10 Abstract and Applied Analysis
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
x (m
)
DesiredLMPC
Figure 8 119909 trajectory119873 = 1
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
x (m
)
DesiredLMPC
Figure 9 119909 trajectory119873 = 9
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
y (m
)
DesiredLMPC
Figure 10 119910 trajectory119873 = 1
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
y (m
)
DesiredLMPC
Figure 11 119910 trajectory119873 = 9
0
2
4
6
8
10A
ltitu
de (m
)
Altitude
0 05 1 15 2 25 3 35 4Time (s)
DesiredLMPC
Figure 12 119911 trajectory119873 = 1
0
2
4
6
8
10
Alti
tude
(m)
Altitude
0 05 1 15 2 25 3 35 4Time (s)
DesiredLMPC
Figure 13 119911 trajectory119873 = 9
Abstract and Applied Analysis 11
function similar to that used in linear quadratic regulator(LQR) has been used In stability analysis LMPC has beencompared to the optimal DLQR system whereas in trajectorytracking LMPC has been compared to a popular linear state-space MPC desgin technique in which plant constraints aremodeled using linear equalities and inequalities The resultsfrom the simulations indicate that the controller performsvery well and is considered feasible With this perceivedfeasibility LMPCoffers the added advantage that it can handlesystems where rapid sampling and more complicated processdynamics are required [5 10] By selecting appropriate valuesof 119886 and119873 LMPC reduces the number of parameters requiredfor accurate prediction when using the traditional (MPC)approach This is a big advantage in that with the reducednumber of parameters online implementation might bepossible where the traditional MPC would have failed
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
This work was supported by the National Research Foun-dation of Korea (NRF) grant funded by the Koreangovernment (MEST) (no 2013R1A2A2A01068127 and no2013R1A1A2A10009458)
References
[1] J M Maciejowski Predictive Control with Constraints PrenticeHall New York NY USA 1st edition 2002
[2] L Wang ldquoDiscrete time model predictive control design usingLaguerre functionsrdquo in Proceedings of the American ControlConference pp 2430ndash2435 Arlington Va USA June 2001
[3] K Alexis G Nikolakopoulos and A Tzes ldquoExperimentalmodel predictive attitude tracking control of a quadrotorhelicopter subject to wind-gustsrdquo in Proceedings of the 18thMediterranean Conference on Control amp Automation (MED rsquo10)pp 1461ndash1466 IEEE Marrakech Morocco June 2010
[4] G V RaffoM G Ortega and F R Rubio ldquoMPCwith nonlinear119867infin
control for path tracking of a quad-rotor helicopterrdquo inProceedings of the 17th World Congress International Federationof Automatic Control (IFAC rsquo08) July 2008
[5] A M Singh D J Lee D P Hong and K T Chong ldquoSucces-sive loop closure based controller design for an autonomousquadrotor vehiclerdquo Applied Mechanics and Materials vol 483pp 361ndash367 2014
[6] L Wang Model Predictive Control Design and ImplementationUsing MATLAB Advances in Industrial Control Springer NewYork NY USA 1st edition 2009
[7] A G Wills ldquoTechnical report EE04025-Notes on linear modelpredictive controlrdquo December 2004
[8] Y Gao C G Lee and K T Chong ldquoReceding horizon trackingcontrol for wheeled mobile robots with time-delayrdquo Journal ofMechanical Science and Technology vol 22 no 12 pp 2403ndash2416 2008
[9] B D O Anderson and J B Moore Linear Optimal ControlPrentice-Hall 1971
[10] J A Rossiter and L Wang ldquoExploiting Laguerre functions toimprove the feasibilityperformance compromise in MPCrdquo inProceedings of the 47th IEEE Conference onDecision and Control(CDC rsquo08) pp 4737ndash4742 December 2008
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
2 Abstract and Applied Analysis
On the other hand MPC can handle both linear andnonlinear systems [1] Comparisons between the nonlinearsystem and its corresponding linearmodel can easily bemadewith very little modification Compared to PID tuning MPCis easier even for complex MIMO systems
However in trying to track the reference trajectory inan optimal way and at the same time obey the constraintsimposed MPC makes much more calculations than eitherPID or LQRThis computation burden makes MPC less suit-able for ldquofastrdquo processes It requires very powerful processorsto be used in hardware implementation Not surprisingly alot of effort is being put in to reduce the computations so thatMPC can be faster and easily be implemented even on lowcost processors
In [2] Wang proposed a method of designing MPC usingorthonormal functions This method makes less computa-tions than the traditional MPC With reduced number ofcomputations this technique can be used where rapid systemdynamics are requiredWe will use this method in our task tocontrol a quadrotor using MPC
Other MPC applications to quadrotor can be found in[3 4] In [4] the control structure consists of a controllerbased on MPC to track the reference trajectory and a secondone based on a nonlinear 119867
infincontrol technique to stabilize
the rotational movements Similarly in [3] anMPC controlleris designed to control position and a second feedforwardcontroller is used to perform stabilization of the quadrotorsystem
In both [3 4] the control loop uses two controllers thecontroller based onMPC which is used for tracking the posi-tion reference trajectory and a second controller (which isdesigned using other techniques) that is used for stabilizationof the rotational movements In this paper MPC is used tocontrol both the position and rotational movements This isgood for uniformity and ease of maintenance
3 Kinematics and Dynamics of a Quadrotor
A quadrotor is completely described by twelve states Thefirst six states describe translational motion of the quadrotorThese states are 119909 119910 and 119911 which represent position ininertial frame and 119906 V and 119908 which denote velocity vectorsalong the body frame Similarly the remaining six variablesdescribe rotational motionsThese are 120601 120579 and 120595 commonlyknown as Euler angles and 119901 119902 and 119903 which denote angularvelocities
A free body diagram of a quadrotor is shown in Figure 1The force 119865
lowastand moment 119872
lowast(where lowast = 119891 119887 119903 and 119897)
produced by the rotors can be described by their angularspeed (120596) as 119865
lowast= 119896
11205962 and 119872
lowast= 119896
21205962 respectively
Similarly torques 120591120601 120591120579 and 120591
120595and the thrust force (119865) are
related to angular speed of rotors as
(
119865
120591120601
120591120579
120591120595
) = 119872(
(
1205962
119891
1205962
119903
1205962
119887
1205962
119897
)
)
(1)
where119872 = [
1198961
1198961
11989611198961
0 minus1198971198961
0 1198971198961
1198971198961
0 minus11989711989610
minus11989621198962minus11989621198962
] and 119897 is the length of arm of the
quadrotorEquations describing the kinematics of the quadrotor are
given by
(
119910
) = 119877 (120601 120579 120595)(
119906
V119908
)
(
120601
120579
) = (
1 119904120601119905120579 119888120601119905120579
0 119888120601 minus119904120601
0119904120601
119888120579
119888120601
119888120579
)(
119901
119902
119903
)
(2)
where 119877(120601 120579 120595) = [
119888120579119888120595119904120601119904120579119888120595minus119888120601119904120595119888120601119904120579119888120595+119904120601119904120595
119888120579119904120595119904120601119904120579119904120595+119888120601119888120595119888120601119904120579119904120595minus119904120601119888120595
minus119904120579
119904120601119888120579
119888120601119888120579
] 119888 stands for
cos 119904 stand for sin and 119905 stands for tan Similarly dynamicequations can be derived using Newtons laws of motionwhich are listed in
(
V
) = (
119903V minus 119902119908
119901119908 minus 119903119902
119902119906 minus 119901V) + (
minus119892 sin 120579119892 cos 120579 sin120601119892 cos 120579 cos120601
) +1
119898(
0
0
minus119865
)
(
119902
119903
) =
(((
(
119869119910minus 119869
119911
119869119909
119902119903
119869119911minus 119869
119909
119869119910
119901119903
119869119909minus 119869
119910
119869119911
119901119902
)))
)
+((
(
1
119869119909
120591120601
1
119869119910
120591120579
1
119869119911
120591120595
))
)
(3)
where 119869119909 119869119910 and 119869
119911are the respectivemoment of inertia about
119883 119884 and 119885 axes of the quadrotor For the simulation of thecontroller of quadrotor we have considered 119865 120591
120601 120591120579 and 120591
120595
as the inputs to the system
31 Linearization In order to design a linear controller it isvital that (2) to (3) be simplified to linear ones To do that rollpitch and heading angles are assumed to operate within verysmall angles [5] which leads to the following
(
120601
120579
) =((
(
1
119869119909
120591120601
1
119869119910
120591120579
1
119869119911
120591120595
))
)
= 119892 minus119865
119898
= minus119892120579
V = 119892120601
(4)
4 Controller Design
In order to control attitude (120601 120579 and 120595) altitude (119911) andposition (119909 119910) we propose a control architecture given in
Abstract and Applied Analysis 3
X
Y
Z
(a)
Fl Ff
Fb Fr
120591l
120591b
120591f
120591r(r 120595)
(q 120579)
(p 120601)
u
120596
(b)
Figure 1 Inertial frame (a) and forces and torques acting on the quadrotor (b)
Positioncontroller
Altitudecontroller
Attitudecontroller
Quadrotordynamics
MPC
zref
xref
yref
120595ref
120579ref
120601ref
[x y][120579 120601 120595]
120591120579
120591120601
120591120595
F
z
Figure 2 Block diagram showing the structure of the controllers
Figure 2 where we have utilized MPC to design the threecontrollers The attitude controller generates 120591
120601 120591
120579 and 120591
120595
actuator signals whereas the altitude controller generatesrequired thrust for the system The position controller con-trols position in the 119909 and 119910 directions and generates controlsignals 120579 and 120601 which when combined with the 120595 signal actas reference signals to the attitude controller
In the next section we look at the design of each of thethree controllers
41 Discretization In this section discrete-time state-spaceequations necessary for altitude position and attitude con-trol are derived
411 Altitude Controller The linear equations responsible foraltitude control are
= 119908
= 119892 minus119865
119898
(5)
These two equations can be combined to give the followingsecond order ODE
= 119892 minus119865
119898 (6)
This second order ODE can be simplified to
= 119866 (7)
where 119866 = 119892 minus 119865119898 or 119865 = 119898(119892 minus 119866) Equation (7) can bewritten in state-space form as in the following equation
[
] = [
0 1
0 0] [
119911
] + [
0
1]119866 (8)
Using the forward Euler method and choosing a samplinginterval Δ119879 we can express (8) in discrete form as in
[119911 (119896 + 1)
V119911(119896 + 1)
] = [1 Δ119879
0 1] [
119911 (119896)
V119911(119896)
] + [0
Δ119879]119866 (119896) (9)
The states 119911(119896) and V119911(119896) represent the position and velocity
respectively and we could choose a proper output matrix 119862depending on which state we are going to measure In thispaper we are interested in the position and therefore we willuse the output matrix 119862 = [1 0]
412 Position Controller The equations responsible for posi-tion and velocity control in 119909 direction are
= 119906
= minus119892120579
(10)
These two equations can be combined to give the secondorder ODE
= minus119892120579 (11)
4 Abstract and Applied Analysis
Similarly the equations responsible for position and velocitycontrol in 119910 direction are
119910 = V
V = 119892120601
(12)
which give rise to the second order ODE
119910 = 119892120601 (13)
Equations (11) and (13) can be combined and form a state-space equation such as
[[[
[
119910
119910
]]]
]
=
[[[
[
0 1 0 0
0 0 0 0
0 0 0 1
0 0 0 0
]]]
]
[[[
[
119909
119910
119910
]]]
]
+
[[[
[
0 0
minus119892 0
0 0
0 119892
]]]
]
[120579
120601] (14)
As stated before we can use the forward Euler method toexpress (14) in discrete form Consider
[[[
[
119909 (119896 + 1)
V119909(119896 + 1)
119910 (119896 + 1)
V119910(119896 + 1)
]]]
]
=
[[[
[
1 Δ119879 0 0
0 1 0 0
0 0 1 Δ119879
0 0 0 1
]]]
]
[[[
[
119909 (119896)
V119909(119896)
119910 (119896)
V119910(119896)
]]]
]
+
[[[
[
0 0
minus119892Δ119879 0
0 0
0 119892Δ119879
]]]
]
[120579 (119896)
120601 (119896)]
(15)
The states 119909(119896) and 119910(119896) represent the position while V119909(119896)
and V119910(119896) represent velocity Since we are interested in the
position we will use the output matrix 119862 = [1 0 0 0
0 0 1 0] but
we could have used a different 2 times 4 output matrix if wewere interested in the velocity But it is important to notethat we can not use a 3 times 4 or 4 times 4 as there is no guaranteethat we will be able to control each of the measured outputsindependently with zero steady state errors This is generallythe case for a system with119898 inputs 119902 outputs and 119902 gt 119898 [6]
413 Attitude Controller In attitude control we are interestedin controlling the roll 120601 pitch 120579 and yaw (heading) 120595 suchthat we generate appropriate torque signals responsible forsteering the quadrotor in the desired direction with therequired attitude The equations responsible for roll controlare
120601 = 119901
=1
119869119909
120591120601
(16)
These two equations can be combined to give the secondorder ODE
120601 =1
119869119909
120591120601 (17)
Similarly the equations responsible for pitch are given by120579 = 119902
119902 =1
119869119910
120591120579
(18)
which give rise to the second order ODE
120579 =1
119869119910
120591120579 (19)
Also the equations responsible for heading are expressed as
= 119903
119903 =1
119869119911
120591120595
(20)
which like in roll and pitch give rise to the second orderODE
=1
119869119911
120591120579 (21)
Equations (17) (19) and (21) can be combined and written instate-space form as in
[[[[[[[
[
120601
120601
120579
120579
]]]]]]]
]
=
[[[[[[[
[
0 1 0 0 0 0
0 0 0 0 0 0
0 0 0 1 0 0
0 0 0 0 0 0
0 0 0 0 0 1
0 0 0 0 0 0
]]]]]]]
]
[[[[[[[
[
120601
120601
120579
120579
120595
]]]]]]]
]
+
[[[[[[[
[
0 0 0
119869minus1
1199090 0
0 0 0
0 119869minus1
1199100
0 0 0
0 0 119869minus1
119911
]]]]]]]
]
[
[
120591120601
120591120579
120591120595
]
]
(22)
where 120591120601 120591
120579 and 120591
120595are the torques required to give the
required roll pitch and yaw respectivelyAgain using the forward Eulermethod we express (22) in
discrete form as
[[[[[[[
[
120601 (119896 + 1)
V120601(119896 + 1)
120579 (119896 + 1)
V120579(119896 + 1)
120595 (119896 + 1)
V120595(119896 + 1)
]]]]]]]
]
=
[[[[[[[
[
1 Δ119879 0 0 0 0
0 1 0 0 0 0
0 0 1 Δ119879 0 0
0 0 0 1 0 0
0 0 0 0 1 Δ119879
0 0 0 0 0 1
]]]]]]]
]
[[[[[[[
[
120601 (119896)
V120601(119896)
120579 (119896)
V120579(119896)
120595 (119896)
V120595(119896)
]]]]]]]
]
+
[[[[[[[
[
0 0 0
Δ119879119869minus1
1199090 0
0 0 0
0 Δ119879119869minus1
1199100
0 0 0
0 0 Δ119879119869minus1
119911
]]]]]]]
]
[
[
120591120601(119896)
120591120579(119896)
120591120595(119896)
]
]
(23)
Since we are interested in measuring 120601(119896) 120579(119896) and 120595(119896) wewill use the output matrix 119862 = [
1 0 0 0 0 0
0 0 1 0 0 0
0 0 0 0 1 0
]So far we have formulated the necessary equations used to
build the three controllers Figure 2 shows in block diagramform what we have done so far But that is not the wholestory In the next section we will look at exactly how modelpredictive control is applied to the quadrotor systemusing thediscrete state-space equations formulated in this section
Abstract and Applied Analysis 5
Plant
Observer
MPC
u(t) y(t)
x(t)
Figure 3 Block diagram of SS-MPC
5 Model Predictive Control
One of the many tenets deployed in control of complexsystems like the quadrotor using MPC is to use modelsthat reduce the computation burden which is a necessitywhen it comes to online implementation It is also agreedupon that linear models perform better in this respect thanthe nonlinear models they represent In this and subse-quent sections the performance of a qudrotor controllerdesigned using Laguerre-based MPC (LMPC) is comparedwith that designed using themore common linear state-spaceapproach presented in [7] We will call the linear state-spacemethod SS-MPC
51 Linear State-Space MPC (SS-MPC) Linear state-spaceMPC is modeled using a linear state-space relation andplant constraints are modeled using linear equalities andinequalities When combined with convex quadratic costfunction this implies that the desired control action canbe obtained at each sample interval via the solution of acorresponding quadratic program This is attractive becausethe quadratic programs can be solved efficiently online [7]The general block diagram representing SSMPC is shown inFigure 3
511 Plant Theplant ismodeled by using the following state-space system
119909119896+1
= 119860119909119896+ 119861119906
119896+ 119908
119896
119910119896= 119862119909
119896+ 119863119906
119896+ V
119896
(24)
where 119908119896is the state noise and V
119896is the measurement noise
Both 119908119896and V
119896are assumed to be Gaussian distributed with
zero mean respective covariances of 119882 and 119881 and crosscovariance 119885 This is represented mathematically as
[119908119896
V119896
] sim 119873([0
0] [
119882 119885
119885119879
119881]) (25)
512 Observer Using Gaussian assumptions stated above itis possible to make optimal predictions of state and outputusing the Kalman filter as
119909119896+1|119896
= 119860119909119896|119896minus1
+ 119861119906 + 119870 (119910119896minus 119910
119896|119896minus1)
119910119896|119896minus1
= 119862119909119896|119896minus1
+ 119863119906119896
(26)
The closed loop gain 119870 is found by solving the discrete-timealgebraic Riccati equation (DARE)
119870 = (119860119875119862119879
+ 119885) (119862119875119862119879
+ 119881)minus1
119875 = 119882 + 119860119875119860119879
minus (119860119875119862119879
+ 119885) (119862119875119862119879
+ 119881)minus1
(119885119879
+ 119862119875119860119879
)
(27)
513 Optimal Estimation of State and Output The generalequation representing the optimal estimate of state at instant119895 and written in terms of the initial state 119909
119896+1|119896and future
control inputs 119906119896+119894|119896
is given by
119909119896+119895|119896
= 119860119895minus1
119909119896+1|119896
+
119895minus1
sum
119899=1
119860119895minus119899minus1
119861119906119896+119899|119896
(28)
Similarly the output equation is given by
119910119896+119895|119896
= 119862119860119895minus1
119909119896+1|119896
+ 119862(
119895minus1
sum
119899=1
119860119895minus119899minus1
119861119906119896+119899|119896
) + 119863119906119896+119895|119896
(29)
The vector containing all output vectors can be written as
119884119896=[[
[
119910119896+1|119896
119910119896+119873|119896
]]
]
(30)
while that containing all control actions can be written as
119880119896=[[
[
119906119896+1|119896
119906119896+119873|119896
]]
]
(31)
Therefore we can write 119884119896as
119884119896= 119865119909
119896+1|119896+ Φ119880
119896 (32)
where
119865 =
[[[[[[
[
119862
119862119860
1198621198602
119862119860119873minus1
]]]]]]
]
Φ =
[[[[[[
[
119863
119862119861 119863
119862119860119861 119862119861 119863
119862119860119873minus2
119861 sdot sdot sdot sdot sdot sdot 119862119861 119863
]]]]]]
]
(33)
6 Abstract and Applied Analysis
514 Cost Function The prime goal in MPC is to rejectdisturbances whilst tracking a reference signal But at eachinstant it has to ensure that the control signal is withinreasonable range that is practical for actuation An objectivefunction that combines these two tasks is
119869 (119909119896+1|119896
119880119896) =
1
2
119873
sum
119899=1
1003817100381710038171003817119910119896+119899|119896 minus 119903119896+119899
1003817100381710038171003817
2
119876+1003817100381710038171003817119906119896+119899|119896 minus 119906
119896+119899minus1|119896
1003817100381710038171003817
2
119878
(34)
where both119876 and 119878 are assumed to be symmetric and positivedefinite
52 Laguerre-Based MPC (LMPC) Model predictive controldesigned using Laguerre functions is developed and summa-rized in [2] The author presented this design technique forboth single-variable and multivariable systems Fortunatelythe derivations for the single-variable case and the multi-variable case are very similar and knowing one would giveenough insight into the derivation of the other For brevityonly the single-variable case is presented and it is hoped thatthis is enough to set forth the LMPC design technique
Consider a plant with 119901 inputs 119902 outputs and 119899 states asdescribed by (35) where the subscript119898 stands for model
119909119898(119896 + 1) = 119860
119898119909119898(119896) + 119861
119898119906 (119896) + 119861
119889120596 (119896)
119910 (119896) = 119862119898119909 (119896)
(35)
where 119906(119896) is the input variable 119910(119896) is the process outputand 119909
119898(119896) is the state vector 120596(119896) is the input disturbance
and is assumed to be a sequence of integrated white noiseThe plant described by (35) can be expressed in aug-
mented state-space form [2 6] as
119909 (119896 + 1) = 119860119909 (119896) + 119861Δ119906 (119896) + 119861120598120598 (119896)
119910 (119896) = 119862119909 (119896)
(36)
where 119860 = [119860119898
0119899times119902
119862119898119860119898119868119902times119902
] 119861 = [119861119898
119862119898119861119898
] 119861120598= [
119861119889
119862119898119861119889
] 119862 =
[0119902times119899
119868119902times119902
] and the state 119909(119896) = [Δ119909119898(119896) 119910(119896)]
119879 120598(119896) is azero mean white noise sequence related to the disturbance bythe difference equation 120598(119896) = 120596(119896) minus 120596(119896 minus 1) [6]
521 Design Framework In designing MPC using Laguerrefunctions the control trajectory Δ119880 is expressed using a setof orthonormal functions called Laguerre functions Since thestate and output vectors can also be described in terms ofΔ119880 it follows that they too can be expressed using Laguerrefunctions
At sampling instant 119896119894 the state variable 119909(119896
119894) is available
through measurement The future control trajectory is givenby
Δ119880 = [Δ119906 (119896119894) Δ119906 (119896
119894+ 1) Δ119906 (119896
119894+ 119873
119888minus 1)] (37)
where119873119888is the control horizon
Laguerre functions can approximate the incrementalterms contained inΔ119880The 119911-transfroms of the discrete-timeLaguerre functions are written as
Γ1(119911) =
radic1 minus 1198862
1 minus 119886119911minus1
Γ2(119911) =
radic1 minus 1198862
1 minus 119886119911minus1(
119911minus1
minus 119886
1 minus 119886119911minus1)
Γ119873(119911) =
radic1 minus 1198862
1 minus 119886119911minus1(
119911minus1
minus 119886
1 minus 119886119911minus1)
119873minus1
(38)
In the discrete-time Laguerre network of (38) 119886 is the poleof the discrete-time Laguerre network and 0 le 119886 lt 1 forstability of the network The parameter 119886 is called the scalingfactor and 119873 = 1 2 3 is the number of Laguerre termsused in the network Note that
Γ119896(119911) = Γ
119896minus1(119911) (
119911minus1
minus 119886
1 minus 119886119911minus1) (39)
with the first term as Γ1(119911) = radic1 minus 1198862(1 minus 119886119911
minus1
)Let 119897
119895(119896) be the inverse 119911-transform of the 119895th term in
the discrete Laguerre network and 119871(119896) the vector containingall inverse 119911-transform terms Then taking advantage of (39)successive inverse 119911-transform vectors are obtained throughthe difference equation
119871 (119896 + 1) = 119860119897119871 (119896) (40)
Thematrix119860119897has size119873times119873 and is a function of parameters
119886 and 120573 = 1 minus 1198862 The initial condition is given by
119871 (0)⊤
= radic120573 [1 minus119886 1198862
sdot sdot sdot (minus1)119873minus1
119886119873minus1
] (41)
In the case where119873 = 5 matrix 119860119897and the initial condition
119871(0) are
[[[[[
[
119886 0 0 0 0
120573 119886 0 0 0
minus119886120573 120573 119886 0 0
1198862
120573 minus119886120573 120573 119886 0
minus1198863
120573 1198862
120573 minus119886120573 120573 119886
]]]]]
]
(42)
119871 (0) = radic120573
[[[[[
[
1
minus119886
1198862
minus1198863
1198864
]]]]]
]
(43)
respectivelyAt sampling instant 119896
119894 the control trajectory is described
using Laguerre functions as in
Δ119906 (119896119894+ 119896) =
119873
sum
119895=1
119888119895(119896119894) 119897119895(119896119894) (44)
Abstract and Applied Analysis 7
where 119896 = 0 1 2 119873119901with 119873
119901as the prediction horizon
119873 terms used in the expansion and 119897119895(119896119894) is the inverse 119911-
transform of the 119895th term in the discrete Laguerre networkEquation (44) can be rewritten as
Δ119906 (119896119894+ 119896) = 119871 (119896)
⊤
120578 (45)
where 119871(119896) = [1198971(119896) 119897
2(119896) sdot sdot sdot 119897
119873(119896)]
119879 120578 = [11988811198882sdot sdot sdot 119888
119873]119879
and the coefficients 1198881 1198882 119888
119873are obtained from system
dataAt an arbitrary future instant 119898 the state is described
using Laguerre functions as
119909 (119896119894+ 119898 | 119896
119894) = 119860
119898
119909 (119896119894) +
119898minus1
sum
119894=0
119860119898minus119894minus1
119861119871 (119894)119879
120578 (46)
Similarly the output is described as
119910 (119896119894+ 119898 | 119896
119894) = 119862119860
119898
119909 (119896119894) +
119898minus1
sum
119894=0
119862119860119898minus119894minus1
119861119871 (119894)119879
120578 (47)
522 Cost Function The cost function is used to choose theoptimal control trajectory Δ119880 to bring the predicted outputas close as possible to the set-point The cost function
119869 =
119873119901
sum
119898=1
119909 (119896119894+ 119898 | 119896
119894)119879
119876119909 (119896119894+ 119898 | 119896
119894) + 120578
119879
119877119871120578 (48)
minimizes the error between the set-point signal and theoutput in the shortest possible time by carefully tuning theweighting matrices 119876 ge 0 and 119877
119871gt 0
523 Cost Function Minimization The state variable in (46)can be rewritten as
119909 (119896119894+ 119898 | 119896
119894) = 119860
119898
119909 (119896119894) + 120601 (119898)
119879
120578 (49)
where 120601(119898)119879
= sum119898minus1
119894=0119860119898minus119894minus1
119861119871(119894)119879
By substituting (49) into (48) and performing the partialderivative 120597119869120597120578 = 0 the Laguerre coefficients vector is foundto be
120578 = minusΩminus1
Ψ119909 (119896119894) (50)
withΩ = sum119873119901
119898=1120601(119898)119876120601(119898)
119879
+ 119877119871and Ψ = 120601(119898)119876119860
119898
524 Receding Horizon Control In receding horizon control(RHC) only the first term in Δ119880 that is Δ119906(119896
119894) is imple-
mented at instant 119896119894 The rest of the sequence is ignored
Only the most recent measurement is taken to form the statevector for calculation of the control signal This procedure isrepeated in real-time to give the receding horizon control law[6 8]
525 Stability In stability analysis we make use of the tech-nique of exponential data weighting originated by Andersonand Moore [9] and applied to MPC in [6] More specifically
we will concentrate on the discrete exponential factor 119890120582Δ119905where Δ119905 is the sampling interval and the discrete weightsform a geometric sequence 120572119895 119895 = 0 1 2
The proposed cost function is similar to the one usedin linear quadratic regulator (LQR) systems but with theinclusion of discrete weights
119869119908=
119873119901
sum
119895=1
120572minus2119895
119909 (119896119894+ 119895 | 119896
119894)119879
119876119909 (119896119894+ 119895 | 119896
119894)
+
119873119901
sum
119895=0
120572minus2119895
Δ119906 (119896119894+ 119895)
119879
119877Δ119906 (119896119894+ 119895)
(51)
For 120572 gt 1 the exponential weights 120572minus2119895 119895 = 1 2 3 119873119901
put more emphasis on the current state 119909(119896119894+ 119895 | 119896
119894) and less
emphasis on subsequent future statesThe exponentially weighted cost function can be
expressed more compactly as
119869119908=
119873119901
sum
119895=1
119909 (119896119894+ 119895 | 119896
119894)119879
119876119909 (119896119894+ 119895 | 119896
119894)
+
119873119901
sum
119895=0
Δ (119896119894+ 119895)
119879
119877Δ (119896119894+ 119895)
(52)
with state equation119909 (119896 + 1) = 119860
120572119909 (119896) + 119861
120572Δ (119896) (53)
and 119860120572= 119860120572 and 119861
120572= 119861120572
With 119876 ge 0 119877 gt 0 and 119873119901
rarr infin minimizing thecost function 119869
119908is equivalent to the DLQR problem which is
solved using algebraic Riccati equation (54)Thepair (119860120572 119861
120572)
is assumed to be controllable and (119860120572 119862) observable with
119876 = 119862119879
119862 Then there is a stabilizing state feedback controlgain matrix119870
119908
119870119908= (119877 + 120572
minus2
119861119879
119875119908119861)
minus1
120572minus2
119861119879
119875119908119860
119860119879
120572[119875
119908minus 119875
119908
119861
120572(119877 +
119861119879
120572119875119908
119861
120572)
minus1
119861119879
120572119875119908]119860
120572+ 119876 minus 119875
119908= 0
(54)that gives a stable closed loop system with all its eigenvaluesinside the unit circle and the the closed loop system beingdescribed by
119909 (119896119894+ 119895 + 1 | 119896
119894) = 120572
minus1
(119860 minus 119861119870119908) 119909 (119896
119894+ 119895 | 119896
119894) (55)
From (55) the transformed system has all its eigenvaluesinside the unit circle by taking119873
119901rarr infin So
120572minus1 1003816100381610038161003816120582max (119860 minus 119861119870
119908)1003816100381610038161003816 lt 1 (56)
which gives1003816100381610038161003816120582max (119860 minus 119861119870
119908)1003816100381610038161003816 lt 120572 (57)
Thus by choosing 120572 gt 1 there is a great chance that the systemwill be stable Several simulations on the quadrotor systemindicate that a choice of 120572 slightly greater than unity makesthe system stable almost all the time
8 Abstract and Applied Analysis
Table 1
Parameter ValueMass119898 07 kgLength 119897 0275m119869119909
01063 kgsdotm2
119869119910
01063 kgsdotm2
119869119911
02122 kgsdotm2
Table 2
Control horizon119873119888
119873 Required 119886
10 5 0606515 5 0716520 5 0778825 5 0818750 5 09048
6 Simulation Parameters and Results
In order to illustrate the controllerrsquos stability we will checkif the eigenvalues of each of the three individual controllersappear inside a unit circle Andwewill use a trajectory in a 3Dplane to check the overall controllerrsquos ability to track a desiredtrajectory
Table 1 shows the parameters of the quadrotor used insimulation
61 Stability Analysis In this section we look at the locationof eigenvalues of LMPC and check if they fall within the unitcircle for stability of the system We also compare them tothose obtained from the optimal DLQR system To ensurestability as explained in Section 525 we use 120572 = 12
By looking at Figure 4 we note that all the eigenvaluesappear inside the unit circle as required Also the eigenvaluesobtained from LMPC closely match with those of the optimalDLQRThis not only shows the system is stable but also showsthe controllers perform optimally
62 Trajectory Tracking In this section we look at theperformance of LMPC when used to track a 3D trajectoryAlso LMPC is compared to SS-MPC and individual statetrajectories are observed The starting point of the trajectoryis (119909 119910 119911) = (0 0 0) and final point is (119909 119910 119911) = (5 5 10)Parameter 119886 = 08187 while119873 = 5 See Figure 5
Clearly both LMPC and SS-MPC track the desired tra-jectory very well See Figures 5 6 and 7 However LMPCperforms better than SS-MPC in that it requires only fiveparameters (119873 = 5) to capture the control trajectorycompared to SS-MPCrsquosminimumof 25 (119873
119888= 25) parameters
(see Table 2 for relationship of119873 and119873119888with 119886 = 08187)
63 Control Horizon in Relation to 119886 and 119873 The controlhorizon (119873
119888) is related to the parameters 119886 and119873 [2] by
119886 asymp 119890minus119873119873
119888 (58)
minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1
minus08
minus06
minus04
minus02
0
02
04
06
08
1
Real
Imag
inar
yLMPCDLQR
Figure 4 Eigenvalues for DLQR and LMPC in altitude control
0 1 2 3 4 5 6
02
46
0
2
4
6
8
10
xy
z
DesiredSS-MPCLMPC
Figure 5 Tracking a 3D trajectory
631 Constant 119873 Using (58) and choosing a constantparameter119873 = 5 we get Table 2
Thus we can increase the control horizon by increasingthe parameter 119886 without necessarily changing the order119873
632 Constant Control Horizon 119873119888 Assuming we want to
achieve a control horizon of 30 we choose one parameterusually 119873 because it is an integer and directly defines theorder of the discrete Laguerre networkTherefater we simplyuse (58) to find the corresponding scaling factor 119886 as shownin Table 3
Abstract and Applied Analysis 9
0 500 1000 1500 2000 2500 3000 3500 40000
1
2
3
4
5
6
Time (ms)
x p
ositi
on
DesiredSS-MPCLMPC
(a)
0 500 1000 1500 2000 2500 3000 3500 4000Time (ms)
0
1
2
3
4
5
6
y p
ositi
on
DesiredSS-MPCLMPC
(b)
Figure 6 Tracking 119909 (a) and 119910 in (b)
2
4
6
8
10
12
00
500 1000 1500 2000 2500 3000 3500 4000Time (ms)
z p
ositi
on (m
)
DesiredSS-MPCLMPC
(a)
0
01
02
03
04
05
0 500 1000 1500 2000 2500 3000 3500 4000Time (ms)
Hea
ding
angl
e120595
DesiredSS-MPCLMPC
(b)
Figure 7 Altitude (119911) (a) and heading angle (120595) (b)
Table 3
Target control horizon119873119888
Chosen119873 Required 119886
30 1 0967230 3 0904830 5 0846530 7 0791930 9 07408
Figures 8 9 10 11 12 and 13 compare two LMPC designsboth of which yield a control horizon of 30 One designuses 119873 = 1 and the other 119873 = 9 Thus we can choose aLaguerre network with lower number of terms119873 (that gives
lower computation burden) but with a larger scaling factor119886 to achieve the same control horizon The computationalcost is lower if a smaller number of parameters are usedFor 119873 = 1 only one Laguerre term is used to capture thecontrol trajectory while119873 = 9 requires nine Laguerre termsto capture the same control trajectory
7 Conclusion
This paper has presented a solution to stability and trajectorytracking of quadrotor systems using a model predictivecontroller designed by using a special type of orthonor-mal functions called Laguerre functions A quadratic cost
10 Abstract and Applied Analysis
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
x (m
)
DesiredLMPC
Figure 8 119909 trajectory119873 = 1
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
x (m
)
DesiredLMPC
Figure 9 119909 trajectory119873 = 9
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
y (m
)
DesiredLMPC
Figure 10 119910 trajectory119873 = 1
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
y (m
)
DesiredLMPC
Figure 11 119910 trajectory119873 = 9
0
2
4
6
8
10A
ltitu
de (m
)
Altitude
0 05 1 15 2 25 3 35 4Time (s)
DesiredLMPC
Figure 12 119911 trajectory119873 = 1
0
2
4
6
8
10
Alti
tude
(m)
Altitude
0 05 1 15 2 25 3 35 4Time (s)
DesiredLMPC
Figure 13 119911 trajectory119873 = 9
Abstract and Applied Analysis 11
function similar to that used in linear quadratic regulator(LQR) has been used In stability analysis LMPC has beencompared to the optimal DLQR system whereas in trajectorytracking LMPC has been compared to a popular linear state-space MPC desgin technique in which plant constraints aremodeled using linear equalities and inequalities The resultsfrom the simulations indicate that the controller performsvery well and is considered feasible With this perceivedfeasibility LMPCoffers the added advantage that it can handlesystems where rapid sampling and more complicated processdynamics are required [5 10] By selecting appropriate valuesof 119886 and119873 LMPC reduces the number of parameters requiredfor accurate prediction when using the traditional (MPC)approach This is a big advantage in that with the reducednumber of parameters online implementation might bepossible where the traditional MPC would have failed
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
This work was supported by the National Research Foun-dation of Korea (NRF) grant funded by the Koreangovernment (MEST) (no 2013R1A2A2A01068127 and no2013R1A1A2A10009458)
References
[1] J M Maciejowski Predictive Control with Constraints PrenticeHall New York NY USA 1st edition 2002
[2] L Wang ldquoDiscrete time model predictive control design usingLaguerre functionsrdquo in Proceedings of the American ControlConference pp 2430ndash2435 Arlington Va USA June 2001
[3] K Alexis G Nikolakopoulos and A Tzes ldquoExperimentalmodel predictive attitude tracking control of a quadrotorhelicopter subject to wind-gustsrdquo in Proceedings of the 18thMediterranean Conference on Control amp Automation (MED rsquo10)pp 1461ndash1466 IEEE Marrakech Morocco June 2010
[4] G V RaffoM G Ortega and F R Rubio ldquoMPCwith nonlinear119867infin
control for path tracking of a quad-rotor helicopterrdquo inProceedings of the 17th World Congress International Federationof Automatic Control (IFAC rsquo08) July 2008
[5] A M Singh D J Lee D P Hong and K T Chong ldquoSucces-sive loop closure based controller design for an autonomousquadrotor vehiclerdquo Applied Mechanics and Materials vol 483pp 361ndash367 2014
[6] L Wang Model Predictive Control Design and ImplementationUsing MATLAB Advances in Industrial Control Springer NewYork NY USA 1st edition 2009
[7] A G Wills ldquoTechnical report EE04025-Notes on linear modelpredictive controlrdquo December 2004
[8] Y Gao C G Lee and K T Chong ldquoReceding horizon trackingcontrol for wheeled mobile robots with time-delayrdquo Journal ofMechanical Science and Technology vol 22 no 12 pp 2403ndash2416 2008
[9] B D O Anderson and J B Moore Linear Optimal ControlPrentice-Hall 1971
[10] J A Rossiter and L Wang ldquoExploiting Laguerre functions toimprove the feasibilityperformance compromise in MPCrdquo inProceedings of the 47th IEEE Conference onDecision and Control(CDC rsquo08) pp 4737ndash4742 December 2008
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
Abstract and Applied Analysis 3
X
Y
Z
(a)
Fl Ff
Fb Fr
120591l
120591b
120591f
120591r(r 120595)
(q 120579)
(p 120601)
u
120596
(b)
Figure 1 Inertial frame (a) and forces and torques acting on the quadrotor (b)
Positioncontroller
Altitudecontroller
Attitudecontroller
Quadrotordynamics
MPC
zref
xref
yref
120595ref
120579ref
120601ref
[x y][120579 120601 120595]
120591120579
120591120601
120591120595
F
z
Figure 2 Block diagram showing the structure of the controllers
Figure 2 where we have utilized MPC to design the threecontrollers The attitude controller generates 120591
120601 120591
120579 and 120591
120595
actuator signals whereas the altitude controller generatesrequired thrust for the system The position controller con-trols position in the 119909 and 119910 directions and generates controlsignals 120579 and 120601 which when combined with the 120595 signal actas reference signals to the attitude controller
In the next section we look at the design of each of thethree controllers
41 Discretization In this section discrete-time state-spaceequations necessary for altitude position and attitude con-trol are derived
411 Altitude Controller The linear equations responsible foraltitude control are
= 119908
= 119892 minus119865
119898
(5)
These two equations can be combined to give the followingsecond order ODE
= 119892 minus119865
119898 (6)
This second order ODE can be simplified to
= 119866 (7)
where 119866 = 119892 minus 119865119898 or 119865 = 119898(119892 minus 119866) Equation (7) can bewritten in state-space form as in the following equation
[
] = [
0 1
0 0] [
119911
] + [
0
1]119866 (8)
Using the forward Euler method and choosing a samplinginterval Δ119879 we can express (8) in discrete form as in
[119911 (119896 + 1)
V119911(119896 + 1)
] = [1 Δ119879
0 1] [
119911 (119896)
V119911(119896)
] + [0
Δ119879]119866 (119896) (9)
The states 119911(119896) and V119911(119896) represent the position and velocity
respectively and we could choose a proper output matrix 119862depending on which state we are going to measure In thispaper we are interested in the position and therefore we willuse the output matrix 119862 = [1 0]
412 Position Controller The equations responsible for posi-tion and velocity control in 119909 direction are
= 119906
= minus119892120579
(10)
These two equations can be combined to give the secondorder ODE
= minus119892120579 (11)
4 Abstract and Applied Analysis
Similarly the equations responsible for position and velocitycontrol in 119910 direction are
119910 = V
V = 119892120601
(12)
which give rise to the second order ODE
119910 = 119892120601 (13)
Equations (11) and (13) can be combined and form a state-space equation such as
[[[
[
119910
119910
]]]
]
=
[[[
[
0 1 0 0
0 0 0 0
0 0 0 1
0 0 0 0
]]]
]
[[[
[
119909
119910
119910
]]]
]
+
[[[
[
0 0
minus119892 0
0 0
0 119892
]]]
]
[120579
120601] (14)
As stated before we can use the forward Euler method toexpress (14) in discrete form Consider
[[[
[
119909 (119896 + 1)
V119909(119896 + 1)
119910 (119896 + 1)
V119910(119896 + 1)
]]]
]
=
[[[
[
1 Δ119879 0 0
0 1 0 0
0 0 1 Δ119879
0 0 0 1
]]]
]
[[[
[
119909 (119896)
V119909(119896)
119910 (119896)
V119910(119896)
]]]
]
+
[[[
[
0 0
minus119892Δ119879 0
0 0
0 119892Δ119879
]]]
]
[120579 (119896)
120601 (119896)]
(15)
The states 119909(119896) and 119910(119896) represent the position while V119909(119896)
and V119910(119896) represent velocity Since we are interested in the
position we will use the output matrix 119862 = [1 0 0 0
0 0 1 0] but
we could have used a different 2 times 4 output matrix if wewere interested in the velocity But it is important to notethat we can not use a 3 times 4 or 4 times 4 as there is no guaranteethat we will be able to control each of the measured outputsindependently with zero steady state errors This is generallythe case for a system with119898 inputs 119902 outputs and 119902 gt 119898 [6]
413 Attitude Controller In attitude control we are interestedin controlling the roll 120601 pitch 120579 and yaw (heading) 120595 suchthat we generate appropriate torque signals responsible forsteering the quadrotor in the desired direction with therequired attitude The equations responsible for roll controlare
120601 = 119901
=1
119869119909
120591120601
(16)
These two equations can be combined to give the secondorder ODE
120601 =1
119869119909
120591120601 (17)
Similarly the equations responsible for pitch are given by120579 = 119902
119902 =1
119869119910
120591120579
(18)
which give rise to the second order ODE
120579 =1
119869119910
120591120579 (19)
Also the equations responsible for heading are expressed as
= 119903
119903 =1
119869119911
120591120595
(20)
which like in roll and pitch give rise to the second orderODE
=1
119869119911
120591120579 (21)
Equations (17) (19) and (21) can be combined and written instate-space form as in
[[[[[[[
[
120601
120601
120579
120579
]]]]]]]
]
=
[[[[[[[
[
0 1 0 0 0 0
0 0 0 0 0 0
0 0 0 1 0 0
0 0 0 0 0 0
0 0 0 0 0 1
0 0 0 0 0 0
]]]]]]]
]
[[[[[[[
[
120601
120601
120579
120579
120595
]]]]]]]
]
+
[[[[[[[
[
0 0 0
119869minus1
1199090 0
0 0 0
0 119869minus1
1199100
0 0 0
0 0 119869minus1
119911
]]]]]]]
]
[
[
120591120601
120591120579
120591120595
]
]
(22)
where 120591120601 120591
120579 and 120591
120595are the torques required to give the
required roll pitch and yaw respectivelyAgain using the forward Eulermethod we express (22) in
discrete form as
[[[[[[[
[
120601 (119896 + 1)
V120601(119896 + 1)
120579 (119896 + 1)
V120579(119896 + 1)
120595 (119896 + 1)
V120595(119896 + 1)
]]]]]]]
]
=
[[[[[[[
[
1 Δ119879 0 0 0 0
0 1 0 0 0 0
0 0 1 Δ119879 0 0
0 0 0 1 0 0
0 0 0 0 1 Δ119879
0 0 0 0 0 1
]]]]]]]
]
[[[[[[[
[
120601 (119896)
V120601(119896)
120579 (119896)
V120579(119896)
120595 (119896)
V120595(119896)
]]]]]]]
]
+
[[[[[[[
[
0 0 0
Δ119879119869minus1
1199090 0
0 0 0
0 Δ119879119869minus1
1199100
0 0 0
0 0 Δ119879119869minus1
119911
]]]]]]]
]
[
[
120591120601(119896)
120591120579(119896)
120591120595(119896)
]
]
(23)
Since we are interested in measuring 120601(119896) 120579(119896) and 120595(119896) wewill use the output matrix 119862 = [
1 0 0 0 0 0
0 0 1 0 0 0
0 0 0 0 1 0
]So far we have formulated the necessary equations used to
build the three controllers Figure 2 shows in block diagramform what we have done so far But that is not the wholestory In the next section we will look at exactly how modelpredictive control is applied to the quadrotor systemusing thediscrete state-space equations formulated in this section
Abstract and Applied Analysis 5
Plant
Observer
MPC
u(t) y(t)
x(t)
Figure 3 Block diagram of SS-MPC
5 Model Predictive Control
One of the many tenets deployed in control of complexsystems like the quadrotor using MPC is to use modelsthat reduce the computation burden which is a necessitywhen it comes to online implementation It is also agreedupon that linear models perform better in this respect thanthe nonlinear models they represent In this and subse-quent sections the performance of a qudrotor controllerdesigned using Laguerre-based MPC (LMPC) is comparedwith that designed using themore common linear state-spaceapproach presented in [7] We will call the linear state-spacemethod SS-MPC
51 Linear State-Space MPC (SS-MPC) Linear state-spaceMPC is modeled using a linear state-space relation andplant constraints are modeled using linear equalities andinequalities When combined with convex quadratic costfunction this implies that the desired control action canbe obtained at each sample interval via the solution of acorresponding quadratic program This is attractive becausethe quadratic programs can be solved efficiently online [7]The general block diagram representing SSMPC is shown inFigure 3
511 Plant Theplant ismodeled by using the following state-space system
119909119896+1
= 119860119909119896+ 119861119906
119896+ 119908
119896
119910119896= 119862119909
119896+ 119863119906
119896+ V
119896
(24)
where 119908119896is the state noise and V
119896is the measurement noise
Both 119908119896and V
119896are assumed to be Gaussian distributed with
zero mean respective covariances of 119882 and 119881 and crosscovariance 119885 This is represented mathematically as
[119908119896
V119896
] sim 119873([0
0] [
119882 119885
119885119879
119881]) (25)
512 Observer Using Gaussian assumptions stated above itis possible to make optimal predictions of state and outputusing the Kalman filter as
119909119896+1|119896
= 119860119909119896|119896minus1
+ 119861119906 + 119870 (119910119896minus 119910
119896|119896minus1)
119910119896|119896minus1
= 119862119909119896|119896minus1
+ 119863119906119896
(26)
The closed loop gain 119870 is found by solving the discrete-timealgebraic Riccati equation (DARE)
119870 = (119860119875119862119879
+ 119885) (119862119875119862119879
+ 119881)minus1
119875 = 119882 + 119860119875119860119879
minus (119860119875119862119879
+ 119885) (119862119875119862119879
+ 119881)minus1
(119885119879
+ 119862119875119860119879
)
(27)
513 Optimal Estimation of State and Output The generalequation representing the optimal estimate of state at instant119895 and written in terms of the initial state 119909
119896+1|119896and future
control inputs 119906119896+119894|119896
is given by
119909119896+119895|119896
= 119860119895minus1
119909119896+1|119896
+
119895minus1
sum
119899=1
119860119895minus119899minus1
119861119906119896+119899|119896
(28)
Similarly the output equation is given by
119910119896+119895|119896
= 119862119860119895minus1
119909119896+1|119896
+ 119862(
119895minus1
sum
119899=1
119860119895minus119899minus1
119861119906119896+119899|119896
) + 119863119906119896+119895|119896
(29)
The vector containing all output vectors can be written as
119884119896=[[
[
119910119896+1|119896
119910119896+119873|119896
]]
]
(30)
while that containing all control actions can be written as
119880119896=[[
[
119906119896+1|119896
119906119896+119873|119896
]]
]
(31)
Therefore we can write 119884119896as
119884119896= 119865119909
119896+1|119896+ Φ119880
119896 (32)
where
119865 =
[[[[[[
[
119862
119862119860
1198621198602
119862119860119873minus1
]]]]]]
]
Φ =
[[[[[[
[
119863
119862119861 119863
119862119860119861 119862119861 119863
119862119860119873minus2
119861 sdot sdot sdot sdot sdot sdot 119862119861 119863
]]]]]]
]
(33)
6 Abstract and Applied Analysis
514 Cost Function The prime goal in MPC is to rejectdisturbances whilst tracking a reference signal But at eachinstant it has to ensure that the control signal is withinreasonable range that is practical for actuation An objectivefunction that combines these two tasks is
119869 (119909119896+1|119896
119880119896) =
1
2
119873
sum
119899=1
1003817100381710038171003817119910119896+119899|119896 minus 119903119896+119899
1003817100381710038171003817
2
119876+1003817100381710038171003817119906119896+119899|119896 minus 119906
119896+119899minus1|119896
1003817100381710038171003817
2
119878
(34)
where both119876 and 119878 are assumed to be symmetric and positivedefinite
52 Laguerre-Based MPC (LMPC) Model predictive controldesigned using Laguerre functions is developed and summa-rized in [2] The author presented this design technique forboth single-variable and multivariable systems Fortunatelythe derivations for the single-variable case and the multi-variable case are very similar and knowing one would giveenough insight into the derivation of the other For brevityonly the single-variable case is presented and it is hoped thatthis is enough to set forth the LMPC design technique
Consider a plant with 119901 inputs 119902 outputs and 119899 states asdescribed by (35) where the subscript119898 stands for model
119909119898(119896 + 1) = 119860
119898119909119898(119896) + 119861
119898119906 (119896) + 119861
119889120596 (119896)
119910 (119896) = 119862119898119909 (119896)
(35)
where 119906(119896) is the input variable 119910(119896) is the process outputand 119909
119898(119896) is the state vector 120596(119896) is the input disturbance
and is assumed to be a sequence of integrated white noiseThe plant described by (35) can be expressed in aug-
mented state-space form [2 6] as
119909 (119896 + 1) = 119860119909 (119896) + 119861Δ119906 (119896) + 119861120598120598 (119896)
119910 (119896) = 119862119909 (119896)
(36)
where 119860 = [119860119898
0119899times119902
119862119898119860119898119868119902times119902
] 119861 = [119861119898
119862119898119861119898
] 119861120598= [
119861119889
119862119898119861119889
] 119862 =
[0119902times119899
119868119902times119902
] and the state 119909(119896) = [Δ119909119898(119896) 119910(119896)]
119879 120598(119896) is azero mean white noise sequence related to the disturbance bythe difference equation 120598(119896) = 120596(119896) minus 120596(119896 minus 1) [6]
521 Design Framework In designing MPC using Laguerrefunctions the control trajectory Δ119880 is expressed using a setof orthonormal functions called Laguerre functions Since thestate and output vectors can also be described in terms ofΔ119880 it follows that they too can be expressed using Laguerrefunctions
At sampling instant 119896119894 the state variable 119909(119896
119894) is available
through measurement The future control trajectory is givenby
Δ119880 = [Δ119906 (119896119894) Δ119906 (119896
119894+ 1) Δ119906 (119896
119894+ 119873
119888minus 1)] (37)
where119873119888is the control horizon
Laguerre functions can approximate the incrementalterms contained inΔ119880The 119911-transfroms of the discrete-timeLaguerre functions are written as
Γ1(119911) =
radic1 minus 1198862
1 minus 119886119911minus1
Γ2(119911) =
radic1 minus 1198862
1 minus 119886119911minus1(
119911minus1
minus 119886
1 minus 119886119911minus1)
Γ119873(119911) =
radic1 minus 1198862
1 minus 119886119911minus1(
119911minus1
minus 119886
1 minus 119886119911minus1)
119873minus1
(38)
In the discrete-time Laguerre network of (38) 119886 is the poleof the discrete-time Laguerre network and 0 le 119886 lt 1 forstability of the network The parameter 119886 is called the scalingfactor and 119873 = 1 2 3 is the number of Laguerre termsused in the network Note that
Γ119896(119911) = Γ
119896minus1(119911) (
119911minus1
minus 119886
1 minus 119886119911minus1) (39)
with the first term as Γ1(119911) = radic1 minus 1198862(1 minus 119886119911
minus1
)Let 119897
119895(119896) be the inverse 119911-transform of the 119895th term in
the discrete Laguerre network and 119871(119896) the vector containingall inverse 119911-transform terms Then taking advantage of (39)successive inverse 119911-transform vectors are obtained throughthe difference equation
119871 (119896 + 1) = 119860119897119871 (119896) (40)
Thematrix119860119897has size119873times119873 and is a function of parameters
119886 and 120573 = 1 minus 1198862 The initial condition is given by
119871 (0)⊤
= radic120573 [1 minus119886 1198862
sdot sdot sdot (minus1)119873minus1
119886119873minus1
] (41)
In the case where119873 = 5 matrix 119860119897and the initial condition
119871(0) are
[[[[[
[
119886 0 0 0 0
120573 119886 0 0 0
minus119886120573 120573 119886 0 0
1198862
120573 minus119886120573 120573 119886 0
minus1198863
120573 1198862
120573 minus119886120573 120573 119886
]]]]]
]
(42)
119871 (0) = radic120573
[[[[[
[
1
minus119886
1198862
minus1198863
1198864
]]]]]
]
(43)
respectivelyAt sampling instant 119896
119894 the control trajectory is described
using Laguerre functions as in
Δ119906 (119896119894+ 119896) =
119873
sum
119895=1
119888119895(119896119894) 119897119895(119896119894) (44)
Abstract and Applied Analysis 7
where 119896 = 0 1 2 119873119901with 119873
119901as the prediction horizon
119873 terms used in the expansion and 119897119895(119896119894) is the inverse 119911-
transform of the 119895th term in the discrete Laguerre networkEquation (44) can be rewritten as
Δ119906 (119896119894+ 119896) = 119871 (119896)
⊤
120578 (45)
where 119871(119896) = [1198971(119896) 119897
2(119896) sdot sdot sdot 119897
119873(119896)]
119879 120578 = [11988811198882sdot sdot sdot 119888
119873]119879
and the coefficients 1198881 1198882 119888
119873are obtained from system
dataAt an arbitrary future instant 119898 the state is described
using Laguerre functions as
119909 (119896119894+ 119898 | 119896
119894) = 119860
119898
119909 (119896119894) +
119898minus1
sum
119894=0
119860119898minus119894minus1
119861119871 (119894)119879
120578 (46)
Similarly the output is described as
119910 (119896119894+ 119898 | 119896
119894) = 119862119860
119898
119909 (119896119894) +
119898minus1
sum
119894=0
119862119860119898minus119894minus1
119861119871 (119894)119879
120578 (47)
522 Cost Function The cost function is used to choose theoptimal control trajectory Δ119880 to bring the predicted outputas close as possible to the set-point The cost function
119869 =
119873119901
sum
119898=1
119909 (119896119894+ 119898 | 119896
119894)119879
119876119909 (119896119894+ 119898 | 119896
119894) + 120578
119879
119877119871120578 (48)
minimizes the error between the set-point signal and theoutput in the shortest possible time by carefully tuning theweighting matrices 119876 ge 0 and 119877
119871gt 0
523 Cost Function Minimization The state variable in (46)can be rewritten as
119909 (119896119894+ 119898 | 119896
119894) = 119860
119898
119909 (119896119894) + 120601 (119898)
119879
120578 (49)
where 120601(119898)119879
= sum119898minus1
119894=0119860119898minus119894minus1
119861119871(119894)119879
By substituting (49) into (48) and performing the partialderivative 120597119869120597120578 = 0 the Laguerre coefficients vector is foundto be
120578 = minusΩminus1
Ψ119909 (119896119894) (50)
withΩ = sum119873119901
119898=1120601(119898)119876120601(119898)
119879
+ 119877119871and Ψ = 120601(119898)119876119860
119898
524 Receding Horizon Control In receding horizon control(RHC) only the first term in Δ119880 that is Δ119906(119896
119894) is imple-
mented at instant 119896119894 The rest of the sequence is ignored
Only the most recent measurement is taken to form the statevector for calculation of the control signal This procedure isrepeated in real-time to give the receding horizon control law[6 8]
525 Stability In stability analysis we make use of the tech-nique of exponential data weighting originated by Andersonand Moore [9] and applied to MPC in [6] More specifically
we will concentrate on the discrete exponential factor 119890120582Δ119905where Δ119905 is the sampling interval and the discrete weightsform a geometric sequence 120572119895 119895 = 0 1 2
The proposed cost function is similar to the one usedin linear quadratic regulator (LQR) systems but with theinclusion of discrete weights
119869119908=
119873119901
sum
119895=1
120572minus2119895
119909 (119896119894+ 119895 | 119896
119894)119879
119876119909 (119896119894+ 119895 | 119896
119894)
+
119873119901
sum
119895=0
120572minus2119895
Δ119906 (119896119894+ 119895)
119879
119877Δ119906 (119896119894+ 119895)
(51)
For 120572 gt 1 the exponential weights 120572minus2119895 119895 = 1 2 3 119873119901
put more emphasis on the current state 119909(119896119894+ 119895 | 119896
119894) and less
emphasis on subsequent future statesThe exponentially weighted cost function can be
expressed more compactly as
119869119908=
119873119901
sum
119895=1
119909 (119896119894+ 119895 | 119896
119894)119879
119876119909 (119896119894+ 119895 | 119896
119894)
+
119873119901
sum
119895=0
Δ (119896119894+ 119895)
119879
119877Δ (119896119894+ 119895)
(52)
with state equation119909 (119896 + 1) = 119860
120572119909 (119896) + 119861
120572Δ (119896) (53)
and 119860120572= 119860120572 and 119861
120572= 119861120572
With 119876 ge 0 119877 gt 0 and 119873119901
rarr infin minimizing thecost function 119869
119908is equivalent to the DLQR problem which is
solved using algebraic Riccati equation (54)Thepair (119860120572 119861
120572)
is assumed to be controllable and (119860120572 119862) observable with
119876 = 119862119879
119862 Then there is a stabilizing state feedback controlgain matrix119870
119908
119870119908= (119877 + 120572
minus2
119861119879
119875119908119861)
minus1
120572minus2
119861119879
119875119908119860
119860119879
120572[119875
119908minus 119875
119908
119861
120572(119877 +
119861119879
120572119875119908
119861
120572)
minus1
119861119879
120572119875119908]119860
120572+ 119876 minus 119875
119908= 0
(54)that gives a stable closed loop system with all its eigenvaluesinside the unit circle and the the closed loop system beingdescribed by
119909 (119896119894+ 119895 + 1 | 119896
119894) = 120572
minus1
(119860 minus 119861119870119908) 119909 (119896
119894+ 119895 | 119896
119894) (55)
From (55) the transformed system has all its eigenvaluesinside the unit circle by taking119873
119901rarr infin So
120572minus1 1003816100381610038161003816120582max (119860 minus 119861119870
119908)1003816100381610038161003816 lt 1 (56)
which gives1003816100381610038161003816120582max (119860 minus 119861119870
119908)1003816100381610038161003816 lt 120572 (57)
Thus by choosing 120572 gt 1 there is a great chance that the systemwill be stable Several simulations on the quadrotor systemindicate that a choice of 120572 slightly greater than unity makesthe system stable almost all the time
8 Abstract and Applied Analysis
Table 1
Parameter ValueMass119898 07 kgLength 119897 0275m119869119909
01063 kgsdotm2
119869119910
01063 kgsdotm2
119869119911
02122 kgsdotm2
Table 2
Control horizon119873119888
119873 Required 119886
10 5 0606515 5 0716520 5 0778825 5 0818750 5 09048
6 Simulation Parameters and Results
In order to illustrate the controllerrsquos stability we will checkif the eigenvalues of each of the three individual controllersappear inside a unit circle Andwewill use a trajectory in a 3Dplane to check the overall controllerrsquos ability to track a desiredtrajectory
Table 1 shows the parameters of the quadrotor used insimulation
61 Stability Analysis In this section we look at the locationof eigenvalues of LMPC and check if they fall within the unitcircle for stability of the system We also compare them tothose obtained from the optimal DLQR system To ensurestability as explained in Section 525 we use 120572 = 12
By looking at Figure 4 we note that all the eigenvaluesappear inside the unit circle as required Also the eigenvaluesobtained from LMPC closely match with those of the optimalDLQRThis not only shows the system is stable but also showsthe controllers perform optimally
62 Trajectory Tracking In this section we look at theperformance of LMPC when used to track a 3D trajectoryAlso LMPC is compared to SS-MPC and individual statetrajectories are observed The starting point of the trajectoryis (119909 119910 119911) = (0 0 0) and final point is (119909 119910 119911) = (5 5 10)Parameter 119886 = 08187 while119873 = 5 See Figure 5
Clearly both LMPC and SS-MPC track the desired tra-jectory very well See Figures 5 6 and 7 However LMPCperforms better than SS-MPC in that it requires only fiveparameters (119873 = 5) to capture the control trajectorycompared to SS-MPCrsquosminimumof 25 (119873
119888= 25) parameters
(see Table 2 for relationship of119873 and119873119888with 119886 = 08187)
63 Control Horizon in Relation to 119886 and 119873 The controlhorizon (119873
119888) is related to the parameters 119886 and119873 [2] by
119886 asymp 119890minus119873119873
119888 (58)
minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1
minus08
minus06
minus04
minus02
0
02
04
06
08
1
Real
Imag
inar
yLMPCDLQR
Figure 4 Eigenvalues for DLQR and LMPC in altitude control
0 1 2 3 4 5 6
02
46
0
2
4
6
8
10
xy
z
DesiredSS-MPCLMPC
Figure 5 Tracking a 3D trajectory
631 Constant 119873 Using (58) and choosing a constantparameter119873 = 5 we get Table 2
Thus we can increase the control horizon by increasingthe parameter 119886 without necessarily changing the order119873
632 Constant Control Horizon 119873119888 Assuming we want to
achieve a control horizon of 30 we choose one parameterusually 119873 because it is an integer and directly defines theorder of the discrete Laguerre networkTherefater we simplyuse (58) to find the corresponding scaling factor 119886 as shownin Table 3
Abstract and Applied Analysis 9
0 500 1000 1500 2000 2500 3000 3500 40000
1
2
3
4
5
6
Time (ms)
x p
ositi
on
DesiredSS-MPCLMPC
(a)
0 500 1000 1500 2000 2500 3000 3500 4000Time (ms)
0
1
2
3
4
5
6
y p
ositi
on
DesiredSS-MPCLMPC
(b)
Figure 6 Tracking 119909 (a) and 119910 in (b)
2
4
6
8
10
12
00
500 1000 1500 2000 2500 3000 3500 4000Time (ms)
z p
ositi
on (m
)
DesiredSS-MPCLMPC
(a)
0
01
02
03
04
05
0 500 1000 1500 2000 2500 3000 3500 4000Time (ms)
Hea
ding
angl
e120595
DesiredSS-MPCLMPC
(b)
Figure 7 Altitude (119911) (a) and heading angle (120595) (b)
Table 3
Target control horizon119873119888
Chosen119873 Required 119886
30 1 0967230 3 0904830 5 0846530 7 0791930 9 07408
Figures 8 9 10 11 12 and 13 compare two LMPC designsboth of which yield a control horizon of 30 One designuses 119873 = 1 and the other 119873 = 9 Thus we can choose aLaguerre network with lower number of terms119873 (that gives
lower computation burden) but with a larger scaling factor119886 to achieve the same control horizon The computationalcost is lower if a smaller number of parameters are usedFor 119873 = 1 only one Laguerre term is used to capture thecontrol trajectory while119873 = 9 requires nine Laguerre termsto capture the same control trajectory
7 Conclusion
This paper has presented a solution to stability and trajectorytracking of quadrotor systems using a model predictivecontroller designed by using a special type of orthonor-mal functions called Laguerre functions A quadratic cost
10 Abstract and Applied Analysis
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
x (m
)
DesiredLMPC
Figure 8 119909 trajectory119873 = 1
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
x (m
)
DesiredLMPC
Figure 9 119909 trajectory119873 = 9
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
y (m
)
DesiredLMPC
Figure 10 119910 trajectory119873 = 1
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
y (m
)
DesiredLMPC
Figure 11 119910 trajectory119873 = 9
0
2
4
6
8
10A
ltitu
de (m
)
Altitude
0 05 1 15 2 25 3 35 4Time (s)
DesiredLMPC
Figure 12 119911 trajectory119873 = 1
0
2
4
6
8
10
Alti
tude
(m)
Altitude
0 05 1 15 2 25 3 35 4Time (s)
DesiredLMPC
Figure 13 119911 trajectory119873 = 9
Abstract and Applied Analysis 11
function similar to that used in linear quadratic regulator(LQR) has been used In stability analysis LMPC has beencompared to the optimal DLQR system whereas in trajectorytracking LMPC has been compared to a popular linear state-space MPC desgin technique in which plant constraints aremodeled using linear equalities and inequalities The resultsfrom the simulations indicate that the controller performsvery well and is considered feasible With this perceivedfeasibility LMPCoffers the added advantage that it can handlesystems where rapid sampling and more complicated processdynamics are required [5 10] By selecting appropriate valuesof 119886 and119873 LMPC reduces the number of parameters requiredfor accurate prediction when using the traditional (MPC)approach This is a big advantage in that with the reducednumber of parameters online implementation might bepossible where the traditional MPC would have failed
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
This work was supported by the National Research Foun-dation of Korea (NRF) grant funded by the Koreangovernment (MEST) (no 2013R1A2A2A01068127 and no2013R1A1A2A10009458)
References
[1] J M Maciejowski Predictive Control with Constraints PrenticeHall New York NY USA 1st edition 2002
[2] L Wang ldquoDiscrete time model predictive control design usingLaguerre functionsrdquo in Proceedings of the American ControlConference pp 2430ndash2435 Arlington Va USA June 2001
[3] K Alexis G Nikolakopoulos and A Tzes ldquoExperimentalmodel predictive attitude tracking control of a quadrotorhelicopter subject to wind-gustsrdquo in Proceedings of the 18thMediterranean Conference on Control amp Automation (MED rsquo10)pp 1461ndash1466 IEEE Marrakech Morocco June 2010
[4] G V RaffoM G Ortega and F R Rubio ldquoMPCwith nonlinear119867infin
control for path tracking of a quad-rotor helicopterrdquo inProceedings of the 17th World Congress International Federationof Automatic Control (IFAC rsquo08) July 2008
[5] A M Singh D J Lee D P Hong and K T Chong ldquoSucces-sive loop closure based controller design for an autonomousquadrotor vehiclerdquo Applied Mechanics and Materials vol 483pp 361ndash367 2014
[6] L Wang Model Predictive Control Design and ImplementationUsing MATLAB Advances in Industrial Control Springer NewYork NY USA 1st edition 2009
[7] A G Wills ldquoTechnical report EE04025-Notes on linear modelpredictive controlrdquo December 2004
[8] Y Gao C G Lee and K T Chong ldquoReceding horizon trackingcontrol for wheeled mobile robots with time-delayrdquo Journal ofMechanical Science and Technology vol 22 no 12 pp 2403ndash2416 2008
[9] B D O Anderson and J B Moore Linear Optimal ControlPrentice-Hall 1971
[10] J A Rossiter and L Wang ldquoExploiting Laguerre functions toimprove the feasibilityperformance compromise in MPCrdquo inProceedings of the 47th IEEE Conference onDecision and Control(CDC rsquo08) pp 4737ndash4742 December 2008
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
4 Abstract and Applied Analysis
Similarly the equations responsible for position and velocitycontrol in 119910 direction are
119910 = V
V = 119892120601
(12)
which give rise to the second order ODE
119910 = 119892120601 (13)
Equations (11) and (13) can be combined and form a state-space equation such as
[[[
[
119910
119910
]]]
]
=
[[[
[
0 1 0 0
0 0 0 0
0 0 0 1
0 0 0 0
]]]
]
[[[
[
119909
119910
119910
]]]
]
+
[[[
[
0 0
minus119892 0
0 0
0 119892
]]]
]
[120579
120601] (14)
As stated before we can use the forward Euler method toexpress (14) in discrete form Consider
[[[
[
119909 (119896 + 1)
V119909(119896 + 1)
119910 (119896 + 1)
V119910(119896 + 1)
]]]
]
=
[[[
[
1 Δ119879 0 0
0 1 0 0
0 0 1 Δ119879
0 0 0 1
]]]
]
[[[
[
119909 (119896)
V119909(119896)
119910 (119896)
V119910(119896)
]]]
]
+
[[[
[
0 0
minus119892Δ119879 0
0 0
0 119892Δ119879
]]]
]
[120579 (119896)
120601 (119896)]
(15)
The states 119909(119896) and 119910(119896) represent the position while V119909(119896)
and V119910(119896) represent velocity Since we are interested in the
position we will use the output matrix 119862 = [1 0 0 0
0 0 1 0] but
we could have used a different 2 times 4 output matrix if wewere interested in the velocity But it is important to notethat we can not use a 3 times 4 or 4 times 4 as there is no guaranteethat we will be able to control each of the measured outputsindependently with zero steady state errors This is generallythe case for a system with119898 inputs 119902 outputs and 119902 gt 119898 [6]
413 Attitude Controller In attitude control we are interestedin controlling the roll 120601 pitch 120579 and yaw (heading) 120595 suchthat we generate appropriate torque signals responsible forsteering the quadrotor in the desired direction with therequired attitude The equations responsible for roll controlare
120601 = 119901
=1
119869119909
120591120601
(16)
These two equations can be combined to give the secondorder ODE
120601 =1
119869119909
120591120601 (17)
Similarly the equations responsible for pitch are given by120579 = 119902
119902 =1
119869119910
120591120579
(18)
which give rise to the second order ODE
120579 =1
119869119910
120591120579 (19)
Also the equations responsible for heading are expressed as
= 119903
119903 =1
119869119911
120591120595
(20)
which like in roll and pitch give rise to the second orderODE
=1
119869119911
120591120579 (21)
Equations (17) (19) and (21) can be combined and written instate-space form as in
[[[[[[[
[
120601
120601
120579
120579
]]]]]]]
]
=
[[[[[[[
[
0 1 0 0 0 0
0 0 0 0 0 0
0 0 0 1 0 0
0 0 0 0 0 0
0 0 0 0 0 1
0 0 0 0 0 0
]]]]]]]
]
[[[[[[[
[
120601
120601
120579
120579
120595
]]]]]]]
]
+
[[[[[[[
[
0 0 0
119869minus1
1199090 0
0 0 0
0 119869minus1
1199100
0 0 0
0 0 119869minus1
119911
]]]]]]]
]
[
[
120591120601
120591120579
120591120595
]
]
(22)
where 120591120601 120591
120579 and 120591
120595are the torques required to give the
required roll pitch and yaw respectivelyAgain using the forward Eulermethod we express (22) in
discrete form as
[[[[[[[
[
120601 (119896 + 1)
V120601(119896 + 1)
120579 (119896 + 1)
V120579(119896 + 1)
120595 (119896 + 1)
V120595(119896 + 1)
]]]]]]]
]
=
[[[[[[[
[
1 Δ119879 0 0 0 0
0 1 0 0 0 0
0 0 1 Δ119879 0 0
0 0 0 1 0 0
0 0 0 0 1 Δ119879
0 0 0 0 0 1
]]]]]]]
]
[[[[[[[
[
120601 (119896)
V120601(119896)
120579 (119896)
V120579(119896)
120595 (119896)
V120595(119896)
]]]]]]]
]
+
[[[[[[[
[
0 0 0
Δ119879119869minus1
1199090 0
0 0 0
0 Δ119879119869minus1
1199100
0 0 0
0 0 Δ119879119869minus1
119911
]]]]]]]
]
[
[
120591120601(119896)
120591120579(119896)
120591120595(119896)
]
]
(23)
Since we are interested in measuring 120601(119896) 120579(119896) and 120595(119896) wewill use the output matrix 119862 = [
1 0 0 0 0 0
0 0 1 0 0 0
0 0 0 0 1 0
]So far we have formulated the necessary equations used to
build the three controllers Figure 2 shows in block diagramform what we have done so far But that is not the wholestory In the next section we will look at exactly how modelpredictive control is applied to the quadrotor systemusing thediscrete state-space equations formulated in this section
Abstract and Applied Analysis 5
Plant
Observer
MPC
u(t) y(t)
x(t)
Figure 3 Block diagram of SS-MPC
5 Model Predictive Control
One of the many tenets deployed in control of complexsystems like the quadrotor using MPC is to use modelsthat reduce the computation burden which is a necessitywhen it comes to online implementation It is also agreedupon that linear models perform better in this respect thanthe nonlinear models they represent In this and subse-quent sections the performance of a qudrotor controllerdesigned using Laguerre-based MPC (LMPC) is comparedwith that designed using themore common linear state-spaceapproach presented in [7] We will call the linear state-spacemethod SS-MPC
51 Linear State-Space MPC (SS-MPC) Linear state-spaceMPC is modeled using a linear state-space relation andplant constraints are modeled using linear equalities andinequalities When combined with convex quadratic costfunction this implies that the desired control action canbe obtained at each sample interval via the solution of acorresponding quadratic program This is attractive becausethe quadratic programs can be solved efficiently online [7]The general block diagram representing SSMPC is shown inFigure 3
511 Plant Theplant ismodeled by using the following state-space system
119909119896+1
= 119860119909119896+ 119861119906
119896+ 119908
119896
119910119896= 119862119909
119896+ 119863119906
119896+ V
119896
(24)
where 119908119896is the state noise and V
119896is the measurement noise
Both 119908119896and V
119896are assumed to be Gaussian distributed with
zero mean respective covariances of 119882 and 119881 and crosscovariance 119885 This is represented mathematically as
[119908119896
V119896
] sim 119873([0
0] [
119882 119885
119885119879
119881]) (25)
512 Observer Using Gaussian assumptions stated above itis possible to make optimal predictions of state and outputusing the Kalman filter as
119909119896+1|119896
= 119860119909119896|119896minus1
+ 119861119906 + 119870 (119910119896minus 119910
119896|119896minus1)
119910119896|119896minus1
= 119862119909119896|119896minus1
+ 119863119906119896
(26)
The closed loop gain 119870 is found by solving the discrete-timealgebraic Riccati equation (DARE)
119870 = (119860119875119862119879
+ 119885) (119862119875119862119879
+ 119881)minus1
119875 = 119882 + 119860119875119860119879
minus (119860119875119862119879
+ 119885) (119862119875119862119879
+ 119881)minus1
(119885119879
+ 119862119875119860119879
)
(27)
513 Optimal Estimation of State and Output The generalequation representing the optimal estimate of state at instant119895 and written in terms of the initial state 119909
119896+1|119896and future
control inputs 119906119896+119894|119896
is given by
119909119896+119895|119896
= 119860119895minus1
119909119896+1|119896
+
119895minus1
sum
119899=1
119860119895minus119899minus1
119861119906119896+119899|119896
(28)
Similarly the output equation is given by
119910119896+119895|119896
= 119862119860119895minus1
119909119896+1|119896
+ 119862(
119895minus1
sum
119899=1
119860119895minus119899minus1
119861119906119896+119899|119896
) + 119863119906119896+119895|119896
(29)
The vector containing all output vectors can be written as
119884119896=[[
[
119910119896+1|119896
119910119896+119873|119896
]]
]
(30)
while that containing all control actions can be written as
119880119896=[[
[
119906119896+1|119896
119906119896+119873|119896
]]
]
(31)
Therefore we can write 119884119896as
119884119896= 119865119909
119896+1|119896+ Φ119880
119896 (32)
where
119865 =
[[[[[[
[
119862
119862119860
1198621198602
119862119860119873minus1
]]]]]]
]
Φ =
[[[[[[
[
119863
119862119861 119863
119862119860119861 119862119861 119863
119862119860119873minus2
119861 sdot sdot sdot sdot sdot sdot 119862119861 119863
]]]]]]
]
(33)
6 Abstract and Applied Analysis
514 Cost Function The prime goal in MPC is to rejectdisturbances whilst tracking a reference signal But at eachinstant it has to ensure that the control signal is withinreasonable range that is practical for actuation An objectivefunction that combines these two tasks is
119869 (119909119896+1|119896
119880119896) =
1
2
119873
sum
119899=1
1003817100381710038171003817119910119896+119899|119896 minus 119903119896+119899
1003817100381710038171003817
2
119876+1003817100381710038171003817119906119896+119899|119896 minus 119906
119896+119899minus1|119896
1003817100381710038171003817
2
119878
(34)
where both119876 and 119878 are assumed to be symmetric and positivedefinite
52 Laguerre-Based MPC (LMPC) Model predictive controldesigned using Laguerre functions is developed and summa-rized in [2] The author presented this design technique forboth single-variable and multivariable systems Fortunatelythe derivations for the single-variable case and the multi-variable case are very similar and knowing one would giveenough insight into the derivation of the other For brevityonly the single-variable case is presented and it is hoped thatthis is enough to set forth the LMPC design technique
Consider a plant with 119901 inputs 119902 outputs and 119899 states asdescribed by (35) where the subscript119898 stands for model
119909119898(119896 + 1) = 119860
119898119909119898(119896) + 119861
119898119906 (119896) + 119861
119889120596 (119896)
119910 (119896) = 119862119898119909 (119896)
(35)
where 119906(119896) is the input variable 119910(119896) is the process outputand 119909
119898(119896) is the state vector 120596(119896) is the input disturbance
and is assumed to be a sequence of integrated white noiseThe plant described by (35) can be expressed in aug-
mented state-space form [2 6] as
119909 (119896 + 1) = 119860119909 (119896) + 119861Δ119906 (119896) + 119861120598120598 (119896)
119910 (119896) = 119862119909 (119896)
(36)
where 119860 = [119860119898
0119899times119902
119862119898119860119898119868119902times119902
] 119861 = [119861119898
119862119898119861119898
] 119861120598= [
119861119889
119862119898119861119889
] 119862 =
[0119902times119899
119868119902times119902
] and the state 119909(119896) = [Δ119909119898(119896) 119910(119896)]
119879 120598(119896) is azero mean white noise sequence related to the disturbance bythe difference equation 120598(119896) = 120596(119896) minus 120596(119896 minus 1) [6]
521 Design Framework In designing MPC using Laguerrefunctions the control trajectory Δ119880 is expressed using a setof orthonormal functions called Laguerre functions Since thestate and output vectors can also be described in terms ofΔ119880 it follows that they too can be expressed using Laguerrefunctions
At sampling instant 119896119894 the state variable 119909(119896
119894) is available
through measurement The future control trajectory is givenby
Δ119880 = [Δ119906 (119896119894) Δ119906 (119896
119894+ 1) Δ119906 (119896
119894+ 119873
119888minus 1)] (37)
where119873119888is the control horizon
Laguerre functions can approximate the incrementalterms contained inΔ119880The 119911-transfroms of the discrete-timeLaguerre functions are written as
Γ1(119911) =
radic1 minus 1198862
1 minus 119886119911minus1
Γ2(119911) =
radic1 minus 1198862
1 minus 119886119911minus1(
119911minus1
minus 119886
1 minus 119886119911minus1)
Γ119873(119911) =
radic1 minus 1198862
1 minus 119886119911minus1(
119911minus1
minus 119886
1 minus 119886119911minus1)
119873minus1
(38)
In the discrete-time Laguerre network of (38) 119886 is the poleof the discrete-time Laguerre network and 0 le 119886 lt 1 forstability of the network The parameter 119886 is called the scalingfactor and 119873 = 1 2 3 is the number of Laguerre termsused in the network Note that
Γ119896(119911) = Γ
119896minus1(119911) (
119911minus1
minus 119886
1 minus 119886119911minus1) (39)
with the first term as Γ1(119911) = radic1 minus 1198862(1 minus 119886119911
minus1
)Let 119897
119895(119896) be the inverse 119911-transform of the 119895th term in
the discrete Laguerre network and 119871(119896) the vector containingall inverse 119911-transform terms Then taking advantage of (39)successive inverse 119911-transform vectors are obtained throughthe difference equation
119871 (119896 + 1) = 119860119897119871 (119896) (40)
Thematrix119860119897has size119873times119873 and is a function of parameters
119886 and 120573 = 1 minus 1198862 The initial condition is given by
119871 (0)⊤
= radic120573 [1 minus119886 1198862
sdot sdot sdot (minus1)119873minus1
119886119873minus1
] (41)
In the case where119873 = 5 matrix 119860119897and the initial condition
119871(0) are
[[[[[
[
119886 0 0 0 0
120573 119886 0 0 0
minus119886120573 120573 119886 0 0
1198862
120573 minus119886120573 120573 119886 0
minus1198863
120573 1198862
120573 minus119886120573 120573 119886
]]]]]
]
(42)
119871 (0) = radic120573
[[[[[
[
1
minus119886
1198862
minus1198863
1198864
]]]]]
]
(43)
respectivelyAt sampling instant 119896
119894 the control trajectory is described
using Laguerre functions as in
Δ119906 (119896119894+ 119896) =
119873
sum
119895=1
119888119895(119896119894) 119897119895(119896119894) (44)
Abstract and Applied Analysis 7
where 119896 = 0 1 2 119873119901with 119873
119901as the prediction horizon
119873 terms used in the expansion and 119897119895(119896119894) is the inverse 119911-
transform of the 119895th term in the discrete Laguerre networkEquation (44) can be rewritten as
Δ119906 (119896119894+ 119896) = 119871 (119896)
⊤
120578 (45)
where 119871(119896) = [1198971(119896) 119897
2(119896) sdot sdot sdot 119897
119873(119896)]
119879 120578 = [11988811198882sdot sdot sdot 119888
119873]119879
and the coefficients 1198881 1198882 119888
119873are obtained from system
dataAt an arbitrary future instant 119898 the state is described
using Laguerre functions as
119909 (119896119894+ 119898 | 119896
119894) = 119860
119898
119909 (119896119894) +
119898minus1
sum
119894=0
119860119898minus119894minus1
119861119871 (119894)119879
120578 (46)
Similarly the output is described as
119910 (119896119894+ 119898 | 119896
119894) = 119862119860
119898
119909 (119896119894) +
119898minus1
sum
119894=0
119862119860119898minus119894minus1
119861119871 (119894)119879
120578 (47)
522 Cost Function The cost function is used to choose theoptimal control trajectory Δ119880 to bring the predicted outputas close as possible to the set-point The cost function
119869 =
119873119901
sum
119898=1
119909 (119896119894+ 119898 | 119896
119894)119879
119876119909 (119896119894+ 119898 | 119896
119894) + 120578
119879
119877119871120578 (48)
minimizes the error between the set-point signal and theoutput in the shortest possible time by carefully tuning theweighting matrices 119876 ge 0 and 119877
119871gt 0
523 Cost Function Minimization The state variable in (46)can be rewritten as
119909 (119896119894+ 119898 | 119896
119894) = 119860
119898
119909 (119896119894) + 120601 (119898)
119879
120578 (49)
where 120601(119898)119879
= sum119898minus1
119894=0119860119898minus119894minus1
119861119871(119894)119879
By substituting (49) into (48) and performing the partialderivative 120597119869120597120578 = 0 the Laguerre coefficients vector is foundto be
120578 = minusΩminus1
Ψ119909 (119896119894) (50)
withΩ = sum119873119901
119898=1120601(119898)119876120601(119898)
119879
+ 119877119871and Ψ = 120601(119898)119876119860
119898
524 Receding Horizon Control In receding horizon control(RHC) only the first term in Δ119880 that is Δ119906(119896
119894) is imple-
mented at instant 119896119894 The rest of the sequence is ignored
Only the most recent measurement is taken to form the statevector for calculation of the control signal This procedure isrepeated in real-time to give the receding horizon control law[6 8]
525 Stability In stability analysis we make use of the tech-nique of exponential data weighting originated by Andersonand Moore [9] and applied to MPC in [6] More specifically
we will concentrate on the discrete exponential factor 119890120582Δ119905where Δ119905 is the sampling interval and the discrete weightsform a geometric sequence 120572119895 119895 = 0 1 2
The proposed cost function is similar to the one usedin linear quadratic regulator (LQR) systems but with theinclusion of discrete weights
119869119908=
119873119901
sum
119895=1
120572minus2119895
119909 (119896119894+ 119895 | 119896
119894)119879
119876119909 (119896119894+ 119895 | 119896
119894)
+
119873119901
sum
119895=0
120572minus2119895
Δ119906 (119896119894+ 119895)
119879
119877Δ119906 (119896119894+ 119895)
(51)
For 120572 gt 1 the exponential weights 120572minus2119895 119895 = 1 2 3 119873119901
put more emphasis on the current state 119909(119896119894+ 119895 | 119896
119894) and less
emphasis on subsequent future statesThe exponentially weighted cost function can be
expressed more compactly as
119869119908=
119873119901
sum
119895=1
119909 (119896119894+ 119895 | 119896
119894)119879
119876119909 (119896119894+ 119895 | 119896
119894)
+
119873119901
sum
119895=0
Δ (119896119894+ 119895)
119879
119877Δ (119896119894+ 119895)
(52)
with state equation119909 (119896 + 1) = 119860
120572119909 (119896) + 119861
120572Δ (119896) (53)
and 119860120572= 119860120572 and 119861
120572= 119861120572
With 119876 ge 0 119877 gt 0 and 119873119901
rarr infin minimizing thecost function 119869
119908is equivalent to the DLQR problem which is
solved using algebraic Riccati equation (54)Thepair (119860120572 119861
120572)
is assumed to be controllable and (119860120572 119862) observable with
119876 = 119862119879
119862 Then there is a stabilizing state feedback controlgain matrix119870
119908
119870119908= (119877 + 120572
minus2
119861119879
119875119908119861)
minus1
120572minus2
119861119879
119875119908119860
119860119879
120572[119875
119908minus 119875
119908
119861
120572(119877 +
119861119879
120572119875119908
119861
120572)
minus1
119861119879
120572119875119908]119860
120572+ 119876 minus 119875
119908= 0
(54)that gives a stable closed loop system with all its eigenvaluesinside the unit circle and the the closed loop system beingdescribed by
119909 (119896119894+ 119895 + 1 | 119896
119894) = 120572
minus1
(119860 minus 119861119870119908) 119909 (119896
119894+ 119895 | 119896
119894) (55)
From (55) the transformed system has all its eigenvaluesinside the unit circle by taking119873
119901rarr infin So
120572minus1 1003816100381610038161003816120582max (119860 minus 119861119870
119908)1003816100381610038161003816 lt 1 (56)
which gives1003816100381610038161003816120582max (119860 minus 119861119870
119908)1003816100381610038161003816 lt 120572 (57)
Thus by choosing 120572 gt 1 there is a great chance that the systemwill be stable Several simulations on the quadrotor systemindicate that a choice of 120572 slightly greater than unity makesthe system stable almost all the time
8 Abstract and Applied Analysis
Table 1
Parameter ValueMass119898 07 kgLength 119897 0275m119869119909
01063 kgsdotm2
119869119910
01063 kgsdotm2
119869119911
02122 kgsdotm2
Table 2
Control horizon119873119888
119873 Required 119886
10 5 0606515 5 0716520 5 0778825 5 0818750 5 09048
6 Simulation Parameters and Results
In order to illustrate the controllerrsquos stability we will checkif the eigenvalues of each of the three individual controllersappear inside a unit circle Andwewill use a trajectory in a 3Dplane to check the overall controllerrsquos ability to track a desiredtrajectory
Table 1 shows the parameters of the quadrotor used insimulation
61 Stability Analysis In this section we look at the locationof eigenvalues of LMPC and check if they fall within the unitcircle for stability of the system We also compare them tothose obtained from the optimal DLQR system To ensurestability as explained in Section 525 we use 120572 = 12
By looking at Figure 4 we note that all the eigenvaluesappear inside the unit circle as required Also the eigenvaluesobtained from LMPC closely match with those of the optimalDLQRThis not only shows the system is stable but also showsthe controllers perform optimally
62 Trajectory Tracking In this section we look at theperformance of LMPC when used to track a 3D trajectoryAlso LMPC is compared to SS-MPC and individual statetrajectories are observed The starting point of the trajectoryis (119909 119910 119911) = (0 0 0) and final point is (119909 119910 119911) = (5 5 10)Parameter 119886 = 08187 while119873 = 5 See Figure 5
Clearly both LMPC and SS-MPC track the desired tra-jectory very well See Figures 5 6 and 7 However LMPCperforms better than SS-MPC in that it requires only fiveparameters (119873 = 5) to capture the control trajectorycompared to SS-MPCrsquosminimumof 25 (119873
119888= 25) parameters
(see Table 2 for relationship of119873 and119873119888with 119886 = 08187)
63 Control Horizon in Relation to 119886 and 119873 The controlhorizon (119873
119888) is related to the parameters 119886 and119873 [2] by
119886 asymp 119890minus119873119873
119888 (58)
minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1
minus08
minus06
minus04
minus02
0
02
04
06
08
1
Real
Imag
inar
yLMPCDLQR
Figure 4 Eigenvalues for DLQR and LMPC in altitude control
0 1 2 3 4 5 6
02
46
0
2
4
6
8
10
xy
z
DesiredSS-MPCLMPC
Figure 5 Tracking a 3D trajectory
631 Constant 119873 Using (58) and choosing a constantparameter119873 = 5 we get Table 2
Thus we can increase the control horizon by increasingthe parameter 119886 without necessarily changing the order119873
632 Constant Control Horizon 119873119888 Assuming we want to
achieve a control horizon of 30 we choose one parameterusually 119873 because it is an integer and directly defines theorder of the discrete Laguerre networkTherefater we simplyuse (58) to find the corresponding scaling factor 119886 as shownin Table 3
Abstract and Applied Analysis 9
0 500 1000 1500 2000 2500 3000 3500 40000
1
2
3
4
5
6
Time (ms)
x p
ositi
on
DesiredSS-MPCLMPC
(a)
0 500 1000 1500 2000 2500 3000 3500 4000Time (ms)
0
1
2
3
4
5
6
y p
ositi
on
DesiredSS-MPCLMPC
(b)
Figure 6 Tracking 119909 (a) and 119910 in (b)
2
4
6
8
10
12
00
500 1000 1500 2000 2500 3000 3500 4000Time (ms)
z p
ositi
on (m
)
DesiredSS-MPCLMPC
(a)
0
01
02
03
04
05
0 500 1000 1500 2000 2500 3000 3500 4000Time (ms)
Hea
ding
angl
e120595
DesiredSS-MPCLMPC
(b)
Figure 7 Altitude (119911) (a) and heading angle (120595) (b)
Table 3
Target control horizon119873119888
Chosen119873 Required 119886
30 1 0967230 3 0904830 5 0846530 7 0791930 9 07408
Figures 8 9 10 11 12 and 13 compare two LMPC designsboth of which yield a control horizon of 30 One designuses 119873 = 1 and the other 119873 = 9 Thus we can choose aLaguerre network with lower number of terms119873 (that gives
lower computation burden) but with a larger scaling factor119886 to achieve the same control horizon The computationalcost is lower if a smaller number of parameters are usedFor 119873 = 1 only one Laguerre term is used to capture thecontrol trajectory while119873 = 9 requires nine Laguerre termsto capture the same control trajectory
7 Conclusion
This paper has presented a solution to stability and trajectorytracking of quadrotor systems using a model predictivecontroller designed by using a special type of orthonor-mal functions called Laguerre functions A quadratic cost
10 Abstract and Applied Analysis
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
x (m
)
DesiredLMPC
Figure 8 119909 trajectory119873 = 1
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
x (m
)
DesiredLMPC
Figure 9 119909 trajectory119873 = 9
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
y (m
)
DesiredLMPC
Figure 10 119910 trajectory119873 = 1
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
y (m
)
DesiredLMPC
Figure 11 119910 trajectory119873 = 9
0
2
4
6
8
10A
ltitu
de (m
)
Altitude
0 05 1 15 2 25 3 35 4Time (s)
DesiredLMPC
Figure 12 119911 trajectory119873 = 1
0
2
4
6
8
10
Alti
tude
(m)
Altitude
0 05 1 15 2 25 3 35 4Time (s)
DesiredLMPC
Figure 13 119911 trajectory119873 = 9
Abstract and Applied Analysis 11
function similar to that used in linear quadratic regulator(LQR) has been used In stability analysis LMPC has beencompared to the optimal DLQR system whereas in trajectorytracking LMPC has been compared to a popular linear state-space MPC desgin technique in which plant constraints aremodeled using linear equalities and inequalities The resultsfrom the simulations indicate that the controller performsvery well and is considered feasible With this perceivedfeasibility LMPCoffers the added advantage that it can handlesystems where rapid sampling and more complicated processdynamics are required [5 10] By selecting appropriate valuesof 119886 and119873 LMPC reduces the number of parameters requiredfor accurate prediction when using the traditional (MPC)approach This is a big advantage in that with the reducednumber of parameters online implementation might bepossible where the traditional MPC would have failed
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
This work was supported by the National Research Foun-dation of Korea (NRF) grant funded by the Koreangovernment (MEST) (no 2013R1A2A2A01068127 and no2013R1A1A2A10009458)
References
[1] J M Maciejowski Predictive Control with Constraints PrenticeHall New York NY USA 1st edition 2002
[2] L Wang ldquoDiscrete time model predictive control design usingLaguerre functionsrdquo in Proceedings of the American ControlConference pp 2430ndash2435 Arlington Va USA June 2001
[3] K Alexis G Nikolakopoulos and A Tzes ldquoExperimentalmodel predictive attitude tracking control of a quadrotorhelicopter subject to wind-gustsrdquo in Proceedings of the 18thMediterranean Conference on Control amp Automation (MED rsquo10)pp 1461ndash1466 IEEE Marrakech Morocco June 2010
[4] G V RaffoM G Ortega and F R Rubio ldquoMPCwith nonlinear119867infin
control for path tracking of a quad-rotor helicopterrdquo inProceedings of the 17th World Congress International Federationof Automatic Control (IFAC rsquo08) July 2008
[5] A M Singh D J Lee D P Hong and K T Chong ldquoSucces-sive loop closure based controller design for an autonomousquadrotor vehiclerdquo Applied Mechanics and Materials vol 483pp 361ndash367 2014
[6] L Wang Model Predictive Control Design and ImplementationUsing MATLAB Advances in Industrial Control Springer NewYork NY USA 1st edition 2009
[7] A G Wills ldquoTechnical report EE04025-Notes on linear modelpredictive controlrdquo December 2004
[8] Y Gao C G Lee and K T Chong ldquoReceding horizon trackingcontrol for wheeled mobile robots with time-delayrdquo Journal ofMechanical Science and Technology vol 22 no 12 pp 2403ndash2416 2008
[9] B D O Anderson and J B Moore Linear Optimal ControlPrentice-Hall 1971
[10] J A Rossiter and L Wang ldquoExploiting Laguerre functions toimprove the feasibilityperformance compromise in MPCrdquo inProceedings of the 47th IEEE Conference onDecision and Control(CDC rsquo08) pp 4737ndash4742 December 2008
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
Abstract and Applied Analysis 5
Plant
Observer
MPC
u(t) y(t)
x(t)
Figure 3 Block diagram of SS-MPC
5 Model Predictive Control
One of the many tenets deployed in control of complexsystems like the quadrotor using MPC is to use modelsthat reduce the computation burden which is a necessitywhen it comes to online implementation It is also agreedupon that linear models perform better in this respect thanthe nonlinear models they represent In this and subse-quent sections the performance of a qudrotor controllerdesigned using Laguerre-based MPC (LMPC) is comparedwith that designed using themore common linear state-spaceapproach presented in [7] We will call the linear state-spacemethod SS-MPC
51 Linear State-Space MPC (SS-MPC) Linear state-spaceMPC is modeled using a linear state-space relation andplant constraints are modeled using linear equalities andinequalities When combined with convex quadratic costfunction this implies that the desired control action canbe obtained at each sample interval via the solution of acorresponding quadratic program This is attractive becausethe quadratic programs can be solved efficiently online [7]The general block diagram representing SSMPC is shown inFigure 3
511 Plant Theplant ismodeled by using the following state-space system
119909119896+1
= 119860119909119896+ 119861119906
119896+ 119908
119896
119910119896= 119862119909
119896+ 119863119906
119896+ V
119896
(24)
where 119908119896is the state noise and V
119896is the measurement noise
Both 119908119896and V
119896are assumed to be Gaussian distributed with
zero mean respective covariances of 119882 and 119881 and crosscovariance 119885 This is represented mathematically as
[119908119896
V119896
] sim 119873([0
0] [
119882 119885
119885119879
119881]) (25)
512 Observer Using Gaussian assumptions stated above itis possible to make optimal predictions of state and outputusing the Kalman filter as
119909119896+1|119896
= 119860119909119896|119896minus1
+ 119861119906 + 119870 (119910119896minus 119910
119896|119896minus1)
119910119896|119896minus1
= 119862119909119896|119896minus1
+ 119863119906119896
(26)
The closed loop gain 119870 is found by solving the discrete-timealgebraic Riccati equation (DARE)
119870 = (119860119875119862119879
+ 119885) (119862119875119862119879
+ 119881)minus1
119875 = 119882 + 119860119875119860119879
minus (119860119875119862119879
+ 119885) (119862119875119862119879
+ 119881)minus1
(119885119879
+ 119862119875119860119879
)
(27)
513 Optimal Estimation of State and Output The generalequation representing the optimal estimate of state at instant119895 and written in terms of the initial state 119909
119896+1|119896and future
control inputs 119906119896+119894|119896
is given by
119909119896+119895|119896
= 119860119895minus1
119909119896+1|119896
+
119895minus1
sum
119899=1
119860119895minus119899minus1
119861119906119896+119899|119896
(28)
Similarly the output equation is given by
119910119896+119895|119896
= 119862119860119895minus1
119909119896+1|119896
+ 119862(
119895minus1
sum
119899=1
119860119895minus119899minus1
119861119906119896+119899|119896
) + 119863119906119896+119895|119896
(29)
The vector containing all output vectors can be written as
119884119896=[[
[
119910119896+1|119896
119910119896+119873|119896
]]
]
(30)
while that containing all control actions can be written as
119880119896=[[
[
119906119896+1|119896
119906119896+119873|119896
]]
]
(31)
Therefore we can write 119884119896as
119884119896= 119865119909
119896+1|119896+ Φ119880
119896 (32)
where
119865 =
[[[[[[
[
119862
119862119860
1198621198602
119862119860119873minus1
]]]]]]
]
Φ =
[[[[[[
[
119863
119862119861 119863
119862119860119861 119862119861 119863
119862119860119873minus2
119861 sdot sdot sdot sdot sdot sdot 119862119861 119863
]]]]]]
]
(33)
6 Abstract and Applied Analysis
514 Cost Function The prime goal in MPC is to rejectdisturbances whilst tracking a reference signal But at eachinstant it has to ensure that the control signal is withinreasonable range that is practical for actuation An objectivefunction that combines these two tasks is
119869 (119909119896+1|119896
119880119896) =
1
2
119873
sum
119899=1
1003817100381710038171003817119910119896+119899|119896 minus 119903119896+119899
1003817100381710038171003817
2
119876+1003817100381710038171003817119906119896+119899|119896 minus 119906
119896+119899minus1|119896
1003817100381710038171003817
2
119878
(34)
where both119876 and 119878 are assumed to be symmetric and positivedefinite
52 Laguerre-Based MPC (LMPC) Model predictive controldesigned using Laguerre functions is developed and summa-rized in [2] The author presented this design technique forboth single-variable and multivariable systems Fortunatelythe derivations for the single-variable case and the multi-variable case are very similar and knowing one would giveenough insight into the derivation of the other For brevityonly the single-variable case is presented and it is hoped thatthis is enough to set forth the LMPC design technique
Consider a plant with 119901 inputs 119902 outputs and 119899 states asdescribed by (35) where the subscript119898 stands for model
119909119898(119896 + 1) = 119860
119898119909119898(119896) + 119861
119898119906 (119896) + 119861
119889120596 (119896)
119910 (119896) = 119862119898119909 (119896)
(35)
where 119906(119896) is the input variable 119910(119896) is the process outputand 119909
119898(119896) is the state vector 120596(119896) is the input disturbance
and is assumed to be a sequence of integrated white noiseThe plant described by (35) can be expressed in aug-
mented state-space form [2 6] as
119909 (119896 + 1) = 119860119909 (119896) + 119861Δ119906 (119896) + 119861120598120598 (119896)
119910 (119896) = 119862119909 (119896)
(36)
where 119860 = [119860119898
0119899times119902
119862119898119860119898119868119902times119902
] 119861 = [119861119898
119862119898119861119898
] 119861120598= [
119861119889
119862119898119861119889
] 119862 =
[0119902times119899
119868119902times119902
] and the state 119909(119896) = [Δ119909119898(119896) 119910(119896)]
119879 120598(119896) is azero mean white noise sequence related to the disturbance bythe difference equation 120598(119896) = 120596(119896) minus 120596(119896 minus 1) [6]
521 Design Framework In designing MPC using Laguerrefunctions the control trajectory Δ119880 is expressed using a setof orthonormal functions called Laguerre functions Since thestate and output vectors can also be described in terms ofΔ119880 it follows that they too can be expressed using Laguerrefunctions
At sampling instant 119896119894 the state variable 119909(119896
119894) is available
through measurement The future control trajectory is givenby
Δ119880 = [Δ119906 (119896119894) Δ119906 (119896
119894+ 1) Δ119906 (119896
119894+ 119873
119888minus 1)] (37)
where119873119888is the control horizon
Laguerre functions can approximate the incrementalterms contained inΔ119880The 119911-transfroms of the discrete-timeLaguerre functions are written as
Γ1(119911) =
radic1 minus 1198862
1 minus 119886119911minus1
Γ2(119911) =
radic1 minus 1198862
1 minus 119886119911minus1(
119911minus1
minus 119886
1 minus 119886119911minus1)
Γ119873(119911) =
radic1 minus 1198862
1 minus 119886119911minus1(
119911minus1
minus 119886
1 minus 119886119911minus1)
119873minus1
(38)
In the discrete-time Laguerre network of (38) 119886 is the poleof the discrete-time Laguerre network and 0 le 119886 lt 1 forstability of the network The parameter 119886 is called the scalingfactor and 119873 = 1 2 3 is the number of Laguerre termsused in the network Note that
Γ119896(119911) = Γ
119896minus1(119911) (
119911minus1
minus 119886
1 minus 119886119911minus1) (39)
with the first term as Γ1(119911) = radic1 minus 1198862(1 minus 119886119911
minus1
)Let 119897
119895(119896) be the inverse 119911-transform of the 119895th term in
the discrete Laguerre network and 119871(119896) the vector containingall inverse 119911-transform terms Then taking advantage of (39)successive inverse 119911-transform vectors are obtained throughthe difference equation
119871 (119896 + 1) = 119860119897119871 (119896) (40)
Thematrix119860119897has size119873times119873 and is a function of parameters
119886 and 120573 = 1 minus 1198862 The initial condition is given by
119871 (0)⊤
= radic120573 [1 minus119886 1198862
sdot sdot sdot (minus1)119873minus1
119886119873minus1
] (41)
In the case where119873 = 5 matrix 119860119897and the initial condition
119871(0) are
[[[[[
[
119886 0 0 0 0
120573 119886 0 0 0
minus119886120573 120573 119886 0 0
1198862
120573 minus119886120573 120573 119886 0
minus1198863
120573 1198862
120573 minus119886120573 120573 119886
]]]]]
]
(42)
119871 (0) = radic120573
[[[[[
[
1
minus119886
1198862
minus1198863
1198864
]]]]]
]
(43)
respectivelyAt sampling instant 119896
119894 the control trajectory is described
using Laguerre functions as in
Δ119906 (119896119894+ 119896) =
119873
sum
119895=1
119888119895(119896119894) 119897119895(119896119894) (44)
Abstract and Applied Analysis 7
where 119896 = 0 1 2 119873119901with 119873
119901as the prediction horizon
119873 terms used in the expansion and 119897119895(119896119894) is the inverse 119911-
transform of the 119895th term in the discrete Laguerre networkEquation (44) can be rewritten as
Δ119906 (119896119894+ 119896) = 119871 (119896)
⊤
120578 (45)
where 119871(119896) = [1198971(119896) 119897
2(119896) sdot sdot sdot 119897
119873(119896)]
119879 120578 = [11988811198882sdot sdot sdot 119888
119873]119879
and the coefficients 1198881 1198882 119888
119873are obtained from system
dataAt an arbitrary future instant 119898 the state is described
using Laguerre functions as
119909 (119896119894+ 119898 | 119896
119894) = 119860
119898
119909 (119896119894) +
119898minus1
sum
119894=0
119860119898minus119894minus1
119861119871 (119894)119879
120578 (46)
Similarly the output is described as
119910 (119896119894+ 119898 | 119896
119894) = 119862119860
119898
119909 (119896119894) +
119898minus1
sum
119894=0
119862119860119898minus119894minus1
119861119871 (119894)119879
120578 (47)
522 Cost Function The cost function is used to choose theoptimal control trajectory Δ119880 to bring the predicted outputas close as possible to the set-point The cost function
119869 =
119873119901
sum
119898=1
119909 (119896119894+ 119898 | 119896
119894)119879
119876119909 (119896119894+ 119898 | 119896
119894) + 120578
119879
119877119871120578 (48)
minimizes the error between the set-point signal and theoutput in the shortest possible time by carefully tuning theweighting matrices 119876 ge 0 and 119877
119871gt 0
523 Cost Function Minimization The state variable in (46)can be rewritten as
119909 (119896119894+ 119898 | 119896
119894) = 119860
119898
119909 (119896119894) + 120601 (119898)
119879
120578 (49)
where 120601(119898)119879
= sum119898minus1
119894=0119860119898minus119894minus1
119861119871(119894)119879
By substituting (49) into (48) and performing the partialderivative 120597119869120597120578 = 0 the Laguerre coefficients vector is foundto be
120578 = minusΩminus1
Ψ119909 (119896119894) (50)
withΩ = sum119873119901
119898=1120601(119898)119876120601(119898)
119879
+ 119877119871and Ψ = 120601(119898)119876119860
119898
524 Receding Horizon Control In receding horizon control(RHC) only the first term in Δ119880 that is Δ119906(119896
119894) is imple-
mented at instant 119896119894 The rest of the sequence is ignored
Only the most recent measurement is taken to form the statevector for calculation of the control signal This procedure isrepeated in real-time to give the receding horizon control law[6 8]
525 Stability In stability analysis we make use of the tech-nique of exponential data weighting originated by Andersonand Moore [9] and applied to MPC in [6] More specifically
we will concentrate on the discrete exponential factor 119890120582Δ119905where Δ119905 is the sampling interval and the discrete weightsform a geometric sequence 120572119895 119895 = 0 1 2
The proposed cost function is similar to the one usedin linear quadratic regulator (LQR) systems but with theinclusion of discrete weights
119869119908=
119873119901
sum
119895=1
120572minus2119895
119909 (119896119894+ 119895 | 119896
119894)119879
119876119909 (119896119894+ 119895 | 119896
119894)
+
119873119901
sum
119895=0
120572minus2119895
Δ119906 (119896119894+ 119895)
119879
119877Δ119906 (119896119894+ 119895)
(51)
For 120572 gt 1 the exponential weights 120572minus2119895 119895 = 1 2 3 119873119901
put more emphasis on the current state 119909(119896119894+ 119895 | 119896
119894) and less
emphasis on subsequent future statesThe exponentially weighted cost function can be
expressed more compactly as
119869119908=
119873119901
sum
119895=1
119909 (119896119894+ 119895 | 119896
119894)119879
119876119909 (119896119894+ 119895 | 119896
119894)
+
119873119901
sum
119895=0
Δ (119896119894+ 119895)
119879
119877Δ (119896119894+ 119895)
(52)
with state equation119909 (119896 + 1) = 119860
120572119909 (119896) + 119861
120572Δ (119896) (53)
and 119860120572= 119860120572 and 119861
120572= 119861120572
With 119876 ge 0 119877 gt 0 and 119873119901
rarr infin minimizing thecost function 119869
119908is equivalent to the DLQR problem which is
solved using algebraic Riccati equation (54)Thepair (119860120572 119861
120572)
is assumed to be controllable and (119860120572 119862) observable with
119876 = 119862119879
119862 Then there is a stabilizing state feedback controlgain matrix119870
119908
119870119908= (119877 + 120572
minus2
119861119879
119875119908119861)
minus1
120572minus2
119861119879
119875119908119860
119860119879
120572[119875
119908minus 119875
119908
119861
120572(119877 +
119861119879
120572119875119908
119861
120572)
minus1
119861119879
120572119875119908]119860
120572+ 119876 minus 119875
119908= 0
(54)that gives a stable closed loop system with all its eigenvaluesinside the unit circle and the the closed loop system beingdescribed by
119909 (119896119894+ 119895 + 1 | 119896
119894) = 120572
minus1
(119860 minus 119861119870119908) 119909 (119896
119894+ 119895 | 119896
119894) (55)
From (55) the transformed system has all its eigenvaluesinside the unit circle by taking119873
119901rarr infin So
120572minus1 1003816100381610038161003816120582max (119860 minus 119861119870
119908)1003816100381610038161003816 lt 1 (56)
which gives1003816100381610038161003816120582max (119860 minus 119861119870
119908)1003816100381610038161003816 lt 120572 (57)
Thus by choosing 120572 gt 1 there is a great chance that the systemwill be stable Several simulations on the quadrotor systemindicate that a choice of 120572 slightly greater than unity makesthe system stable almost all the time
8 Abstract and Applied Analysis
Table 1
Parameter ValueMass119898 07 kgLength 119897 0275m119869119909
01063 kgsdotm2
119869119910
01063 kgsdotm2
119869119911
02122 kgsdotm2
Table 2
Control horizon119873119888
119873 Required 119886
10 5 0606515 5 0716520 5 0778825 5 0818750 5 09048
6 Simulation Parameters and Results
In order to illustrate the controllerrsquos stability we will checkif the eigenvalues of each of the three individual controllersappear inside a unit circle Andwewill use a trajectory in a 3Dplane to check the overall controllerrsquos ability to track a desiredtrajectory
Table 1 shows the parameters of the quadrotor used insimulation
61 Stability Analysis In this section we look at the locationof eigenvalues of LMPC and check if they fall within the unitcircle for stability of the system We also compare them tothose obtained from the optimal DLQR system To ensurestability as explained in Section 525 we use 120572 = 12
By looking at Figure 4 we note that all the eigenvaluesappear inside the unit circle as required Also the eigenvaluesobtained from LMPC closely match with those of the optimalDLQRThis not only shows the system is stable but also showsthe controllers perform optimally
62 Trajectory Tracking In this section we look at theperformance of LMPC when used to track a 3D trajectoryAlso LMPC is compared to SS-MPC and individual statetrajectories are observed The starting point of the trajectoryis (119909 119910 119911) = (0 0 0) and final point is (119909 119910 119911) = (5 5 10)Parameter 119886 = 08187 while119873 = 5 See Figure 5
Clearly both LMPC and SS-MPC track the desired tra-jectory very well See Figures 5 6 and 7 However LMPCperforms better than SS-MPC in that it requires only fiveparameters (119873 = 5) to capture the control trajectorycompared to SS-MPCrsquosminimumof 25 (119873
119888= 25) parameters
(see Table 2 for relationship of119873 and119873119888with 119886 = 08187)
63 Control Horizon in Relation to 119886 and 119873 The controlhorizon (119873
119888) is related to the parameters 119886 and119873 [2] by
119886 asymp 119890minus119873119873
119888 (58)
minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1
minus08
minus06
minus04
minus02
0
02
04
06
08
1
Real
Imag
inar
yLMPCDLQR
Figure 4 Eigenvalues for DLQR and LMPC in altitude control
0 1 2 3 4 5 6
02
46
0
2
4
6
8
10
xy
z
DesiredSS-MPCLMPC
Figure 5 Tracking a 3D trajectory
631 Constant 119873 Using (58) and choosing a constantparameter119873 = 5 we get Table 2
Thus we can increase the control horizon by increasingthe parameter 119886 without necessarily changing the order119873
632 Constant Control Horizon 119873119888 Assuming we want to
achieve a control horizon of 30 we choose one parameterusually 119873 because it is an integer and directly defines theorder of the discrete Laguerre networkTherefater we simplyuse (58) to find the corresponding scaling factor 119886 as shownin Table 3
Abstract and Applied Analysis 9
0 500 1000 1500 2000 2500 3000 3500 40000
1
2
3
4
5
6
Time (ms)
x p
ositi
on
DesiredSS-MPCLMPC
(a)
0 500 1000 1500 2000 2500 3000 3500 4000Time (ms)
0
1
2
3
4
5
6
y p
ositi
on
DesiredSS-MPCLMPC
(b)
Figure 6 Tracking 119909 (a) and 119910 in (b)
2
4
6
8
10
12
00
500 1000 1500 2000 2500 3000 3500 4000Time (ms)
z p
ositi
on (m
)
DesiredSS-MPCLMPC
(a)
0
01
02
03
04
05
0 500 1000 1500 2000 2500 3000 3500 4000Time (ms)
Hea
ding
angl
e120595
DesiredSS-MPCLMPC
(b)
Figure 7 Altitude (119911) (a) and heading angle (120595) (b)
Table 3
Target control horizon119873119888
Chosen119873 Required 119886
30 1 0967230 3 0904830 5 0846530 7 0791930 9 07408
Figures 8 9 10 11 12 and 13 compare two LMPC designsboth of which yield a control horizon of 30 One designuses 119873 = 1 and the other 119873 = 9 Thus we can choose aLaguerre network with lower number of terms119873 (that gives
lower computation burden) but with a larger scaling factor119886 to achieve the same control horizon The computationalcost is lower if a smaller number of parameters are usedFor 119873 = 1 only one Laguerre term is used to capture thecontrol trajectory while119873 = 9 requires nine Laguerre termsto capture the same control trajectory
7 Conclusion
This paper has presented a solution to stability and trajectorytracking of quadrotor systems using a model predictivecontroller designed by using a special type of orthonor-mal functions called Laguerre functions A quadratic cost
10 Abstract and Applied Analysis
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
x (m
)
DesiredLMPC
Figure 8 119909 trajectory119873 = 1
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
x (m
)
DesiredLMPC
Figure 9 119909 trajectory119873 = 9
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
y (m
)
DesiredLMPC
Figure 10 119910 trajectory119873 = 1
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
y (m
)
DesiredLMPC
Figure 11 119910 trajectory119873 = 9
0
2
4
6
8
10A
ltitu
de (m
)
Altitude
0 05 1 15 2 25 3 35 4Time (s)
DesiredLMPC
Figure 12 119911 trajectory119873 = 1
0
2
4
6
8
10
Alti
tude
(m)
Altitude
0 05 1 15 2 25 3 35 4Time (s)
DesiredLMPC
Figure 13 119911 trajectory119873 = 9
Abstract and Applied Analysis 11
function similar to that used in linear quadratic regulator(LQR) has been used In stability analysis LMPC has beencompared to the optimal DLQR system whereas in trajectorytracking LMPC has been compared to a popular linear state-space MPC desgin technique in which plant constraints aremodeled using linear equalities and inequalities The resultsfrom the simulations indicate that the controller performsvery well and is considered feasible With this perceivedfeasibility LMPCoffers the added advantage that it can handlesystems where rapid sampling and more complicated processdynamics are required [5 10] By selecting appropriate valuesof 119886 and119873 LMPC reduces the number of parameters requiredfor accurate prediction when using the traditional (MPC)approach This is a big advantage in that with the reducednumber of parameters online implementation might bepossible where the traditional MPC would have failed
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
This work was supported by the National Research Foun-dation of Korea (NRF) grant funded by the Koreangovernment (MEST) (no 2013R1A2A2A01068127 and no2013R1A1A2A10009458)
References
[1] J M Maciejowski Predictive Control with Constraints PrenticeHall New York NY USA 1st edition 2002
[2] L Wang ldquoDiscrete time model predictive control design usingLaguerre functionsrdquo in Proceedings of the American ControlConference pp 2430ndash2435 Arlington Va USA June 2001
[3] K Alexis G Nikolakopoulos and A Tzes ldquoExperimentalmodel predictive attitude tracking control of a quadrotorhelicopter subject to wind-gustsrdquo in Proceedings of the 18thMediterranean Conference on Control amp Automation (MED rsquo10)pp 1461ndash1466 IEEE Marrakech Morocco June 2010
[4] G V RaffoM G Ortega and F R Rubio ldquoMPCwith nonlinear119867infin
control for path tracking of a quad-rotor helicopterrdquo inProceedings of the 17th World Congress International Federationof Automatic Control (IFAC rsquo08) July 2008
[5] A M Singh D J Lee D P Hong and K T Chong ldquoSucces-sive loop closure based controller design for an autonomousquadrotor vehiclerdquo Applied Mechanics and Materials vol 483pp 361ndash367 2014
[6] L Wang Model Predictive Control Design and ImplementationUsing MATLAB Advances in Industrial Control Springer NewYork NY USA 1st edition 2009
[7] A G Wills ldquoTechnical report EE04025-Notes on linear modelpredictive controlrdquo December 2004
[8] Y Gao C G Lee and K T Chong ldquoReceding horizon trackingcontrol for wheeled mobile robots with time-delayrdquo Journal ofMechanical Science and Technology vol 22 no 12 pp 2403ndash2416 2008
[9] B D O Anderson and J B Moore Linear Optimal ControlPrentice-Hall 1971
[10] J A Rossiter and L Wang ldquoExploiting Laguerre functions toimprove the feasibilityperformance compromise in MPCrdquo inProceedings of the 47th IEEE Conference onDecision and Control(CDC rsquo08) pp 4737ndash4742 December 2008
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
6 Abstract and Applied Analysis
514 Cost Function The prime goal in MPC is to rejectdisturbances whilst tracking a reference signal But at eachinstant it has to ensure that the control signal is withinreasonable range that is practical for actuation An objectivefunction that combines these two tasks is
119869 (119909119896+1|119896
119880119896) =
1
2
119873
sum
119899=1
1003817100381710038171003817119910119896+119899|119896 minus 119903119896+119899
1003817100381710038171003817
2
119876+1003817100381710038171003817119906119896+119899|119896 minus 119906
119896+119899minus1|119896
1003817100381710038171003817
2
119878
(34)
where both119876 and 119878 are assumed to be symmetric and positivedefinite
52 Laguerre-Based MPC (LMPC) Model predictive controldesigned using Laguerre functions is developed and summa-rized in [2] The author presented this design technique forboth single-variable and multivariable systems Fortunatelythe derivations for the single-variable case and the multi-variable case are very similar and knowing one would giveenough insight into the derivation of the other For brevityonly the single-variable case is presented and it is hoped thatthis is enough to set forth the LMPC design technique
Consider a plant with 119901 inputs 119902 outputs and 119899 states asdescribed by (35) where the subscript119898 stands for model
119909119898(119896 + 1) = 119860
119898119909119898(119896) + 119861
119898119906 (119896) + 119861
119889120596 (119896)
119910 (119896) = 119862119898119909 (119896)
(35)
where 119906(119896) is the input variable 119910(119896) is the process outputand 119909
119898(119896) is the state vector 120596(119896) is the input disturbance
and is assumed to be a sequence of integrated white noiseThe plant described by (35) can be expressed in aug-
mented state-space form [2 6] as
119909 (119896 + 1) = 119860119909 (119896) + 119861Δ119906 (119896) + 119861120598120598 (119896)
119910 (119896) = 119862119909 (119896)
(36)
where 119860 = [119860119898
0119899times119902
119862119898119860119898119868119902times119902
] 119861 = [119861119898
119862119898119861119898
] 119861120598= [
119861119889
119862119898119861119889
] 119862 =
[0119902times119899
119868119902times119902
] and the state 119909(119896) = [Δ119909119898(119896) 119910(119896)]
119879 120598(119896) is azero mean white noise sequence related to the disturbance bythe difference equation 120598(119896) = 120596(119896) minus 120596(119896 minus 1) [6]
521 Design Framework In designing MPC using Laguerrefunctions the control trajectory Δ119880 is expressed using a setof orthonormal functions called Laguerre functions Since thestate and output vectors can also be described in terms ofΔ119880 it follows that they too can be expressed using Laguerrefunctions
At sampling instant 119896119894 the state variable 119909(119896
119894) is available
through measurement The future control trajectory is givenby
Δ119880 = [Δ119906 (119896119894) Δ119906 (119896
119894+ 1) Δ119906 (119896
119894+ 119873
119888minus 1)] (37)
where119873119888is the control horizon
Laguerre functions can approximate the incrementalterms contained inΔ119880The 119911-transfroms of the discrete-timeLaguerre functions are written as
Γ1(119911) =
radic1 minus 1198862
1 minus 119886119911minus1
Γ2(119911) =
radic1 minus 1198862
1 minus 119886119911minus1(
119911minus1
minus 119886
1 minus 119886119911minus1)
Γ119873(119911) =
radic1 minus 1198862
1 minus 119886119911minus1(
119911minus1
minus 119886
1 minus 119886119911minus1)
119873minus1
(38)
In the discrete-time Laguerre network of (38) 119886 is the poleof the discrete-time Laguerre network and 0 le 119886 lt 1 forstability of the network The parameter 119886 is called the scalingfactor and 119873 = 1 2 3 is the number of Laguerre termsused in the network Note that
Γ119896(119911) = Γ
119896minus1(119911) (
119911minus1
minus 119886
1 minus 119886119911minus1) (39)
with the first term as Γ1(119911) = radic1 minus 1198862(1 minus 119886119911
minus1
)Let 119897
119895(119896) be the inverse 119911-transform of the 119895th term in
the discrete Laguerre network and 119871(119896) the vector containingall inverse 119911-transform terms Then taking advantage of (39)successive inverse 119911-transform vectors are obtained throughthe difference equation
119871 (119896 + 1) = 119860119897119871 (119896) (40)
Thematrix119860119897has size119873times119873 and is a function of parameters
119886 and 120573 = 1 minus 1198862 The initial condition is given by
119871 (0)⊤
= radic120573 [1 minus119886 1198862
sdot sdot sdot (minus1)119873minus1
119886119873minus1
] (41)
In the case where119873 = 5 matrix 119860119897and the initial condition
119871(0) are
[[[[[
[
119886 0 0 0 0
120573 119886 0 0 0
minus119886120573 120573 119886 0 0
1198862
120573 minus119886120573 120573 119886 0
minus1198863
120573 1198862
120573 minus119886120573 120573 119886
]]]]]
]
(42)
119871 (0) = radic120573
[[[[[
[
1
minus119886
1198862
minus1198863
1198864
]]]]]
]
(43)
respectivelyAt sampling instant 119896
119894 the control trajectory is described
using Laguerre functions as in
Δ119906 (119896119894+ 119896) =
119873
sum
119895=1
119888119895(119896119894) 119897119895(119896119894) (44)
Abstract and Applied Analysis 7
where 119896 = 0 1 2 119873119901with 119873
119901as the prediction horizon
119873 terms used in the expansion and 119897119895(119896119894) is the inverse 119911-
transform of the 119895th term in the discrete Laguerre networkEquation (44) can be rewritten as
Δ119906 (119896119894+ 119896) = 119871 (119896)
⊤
120578 (45)
where 119871(119896) = [1198971(119896) 119897
2(119896) sdot sdot sdot 119897
119873(119896)]
119879 120578 = [11988811198882sdot sdot sdot 119888
119873]119879
and the coefficients 1198881 1198882 119888
119873are obtained from system
dataAt an arbitrary future instant 119898 the state is described
using Laguerre functions as
119909 (119896119894+ 119898 | 119896
119894) = 119860
119898
119909 (119896119894) +
119898minus1
sum
119894=0
119860119898minus119894minus1
119861119871 (119894)119879
120578 (46)
Similarly the output is described as
119910 (119896119894+ 119898 | 119896
119894) = 119862119860
119898
119909 (119896119894) +
119898minus1
sum
119894=0
119862119860119898minus119894minus1
119861119871 (119894)119879
120578 (47)
522 Cost Function The cost function is used to choose theoptimal control trajectory Δ119880 to bring the predicted outputas close as possible to the set-point The cost function
119869 =
119873119901
sum
119898=1
119909 (119896119894+ 119898 | 119896
119894)119879
119876119909 (119896119894+ 119898 | 119896
119894) + 120578
119879
119877119871120578 (48)
minimizes the error between the set-point signal and theoutput in the shortest possible time by carefully tuning theweighting matrices 119876 ge 0 and 119877
119871gt 0
523 Cost Function Minimization The state variable in (46)can be rewritten as
119909 (119896119894+ 119898 | 119896
119894) = 119860
119898
119909 (119896119894) + 120601 (119898)
119879
120578 (49)
where 120601(119898)119879
= sum119898minus1
119894=0119860119898minus119894minus1
119861119871(119894)119879
By substituting (49) into (48) and performing the partialderivative 120597119869120597120578 = 0 the Laguerre coefficients vector is foundto be
120578 = minusΩminus1
Ψ119909 (119896119894) (50)
withΩ = sum119873119901
119898=1120601(119898)119876120601(119898)
119879
+ 119877119871and Ψ = 120601(119898)119876119860
119898
524 Receding Horizon Control In receding horizon control(RHC) only the first term in Δ119880 that is Δ119906(119896
119894) is imple-
mented at instant 119896119894 The rest of the sequence is ignored
Only the most recent measurement is taken to form the statevector for calculation of the control signal This procedure isrepeated in real-time to give the receding horizon control law[6 8]
525 Stability In stability analysis we make use of the tech-nique of exponential data weighting originated by Andersonand Moore [9] and applied to MPC in [6] More specifically
we will concentrate on the discrete exponential factor 119890120582Δ119905where Δ119905 is the sampling interval and the discrete weightsform a geometric sequence 120572119895 119895 = 0 1 2
The proposed cost function is similar to the one usedin linear quadratic regulator (LQR) systems but with theinclusion of discrete weights
119869119908=
119873119901
sum
119895=1
120572minus2119895
119909 (119896119894+ 119895 | 119896
119894)119879
119876119909 (119896119894+ 119895 | 119896
119894)
+
119873119901
sum
119895=0
120572minus2119895
Δ119906 (119896119894+ 119895)
119879
119877Δ119906 (119896119894+ 119895)
(51)
For 120572 gt 1 the exponential weights 120572minus2119895 119895 = 1 2 3 119873119901
put more emphasis on the current state 119909(119896119894+ 119895 | 119896
119894) and less
emphasis on subsequent future statesThe exponentially weighted cost function can be
expressed more compactly as
119869119908=
119873119901
sum
119895=1
119909 (119896119894+ 119895 | 119896
119894)119879
119876119909 (119896119894+ 119895 | 119896
119894)
+
119873119901
sum
119895=0
Δ (119896119894+ 119895)
119879
119877Δ (119896119894+ 119895)
(52)
with state equation119909 (119896 + 1) = 119860
120572119909 (119896) + 119861
120572Δ (119896) (53)
and 119860120572= 119860120572 and 119861
120572= 119861120572
With 119876 ge 0 119877 gt 0 and 119873119901
rarr infin minimizing thecost function 119869
119908is equivalent to the DLQR problem which is
solved using algebraic Riccati equation (54)Thepair (119860120572 119861
120572)
is assumed to be controllable and (119860120572 119862) observable with
119876 = 119862119879
119862 Then there is a stabilizing state feedback controlgain matrix119870
119908
119870119908= (119877 + 120572
minus2
119861119879
119875119908119861)
minus1
120572minus2
119861119879
119875119908119860
119860119879
120572[119875
119908minus 119875
119908
119861
120572(119877 +
119861119879
120572119875119908
119861
120572)
minus1
119861119879
120572119875119908]119860
120572+ 119876 minus 119875
119908= 0
(54)that gives a stable closed loop system with all its eigenvaluesinside the unit circle and the the closed loop system beingdescribed by
119909 (119896119894+ 119895 + 1 | 119896
119894) = 120572
minus1
(119860 minus 119861119870119908) 119909 (119896
119894+ 119895 | 119896
119894) (55)
From (55) the transformed system has all its eigenvaluesinside the unit circle by taking119873
119901rarr infin So
120572minus1 1003816100381610038161003816120582max (119860 minus 119861119870
119908)1003816100381610038161003816 lt 1 (56)
which gives1003816100381610038161003816120582max (119860 minus 119861119870
119908)1003816100381610038161003816 lt 120572 (57)
Thus by choosing 120572 gt 1 there is a great chance that the systemwill be stable Several simulations on the quadrotor systemindicate that a choice of 120572 slightly greater than unity makesthe system stable almost all the time
8 Abstract and Applied Analysis
Table 1
Parameter ValueMass119898 07 kgLength 119897 0275m119869119909
01063 kgsdotm2
119869119910
01063 kgsdotm2
119869119911
02122 kgsdotm2
Table 2
Control horizon119873119888
119873 Required 119886
10 5 0606515 5 0716520 5 0778825 5 0818750 5 09048
6 Simulation Parameters and Results
In order to illustrate the controllerrsquos stability we will checkif the eigenvalues of each of the three individual controllersappear inside a unit circle Andwewill use a trajectory in a 3Dplane to check the overall controllerrsquos ability to track a desiredtrajectory
Table 1 shows the parameters of the quadrotor used insimulation
61 Stability Analysis In this section we look at the locationof eigenvalues of LMPC and check if they fall within the unitcircle for stability of the system We also compare them tothose obtained from the optimal DLQR system To ensurestability as explained in Section 525 we use 120572 = 12
By looking at Figure 4 we note that all the eigenvaluesappear inside the unit circle as required Also the eigenvaluesobtained from LMPC closely match with those of the optimalDLQRThis not only shows the system is stable but also showsthe controllers perform optimally
62 Trajectory Tracking In this section we look at theperformance of LMPC when used to track a 3D trajectoryAlso LMPC is compared to SS-MPC and individual statetrajectories are observed The starting point of the trajectoryis (119909 119910 119911) = (0 0 0) and final point is (119909 119910 119911) = (5 5 10)Parameter 119886 = 08187 while119873 = 5 See Figure 5
Clearly both LMPC and SS-MPC track the desired tra-jectory very well See Figures 5 6 and 7 However LMPCperforms better than SS-MPC in that it requires only fiveparameters (119873 = 5) to capture the control trajectorycompared to SS-MPCrsquosminimumof 25 (119873
119888= 25) parameters
(see Table 2 for relationship of119873 and119873119888with 119886 = 08187)
63 Control Horizon in Relation to 119886 and 119873 The controlhorizon (119873
119888) is related to the parameters 119886 and119873 [2] by
119886 asymp 119890minus119873119873
119888 (58)
minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1
minus08
minus06
minus04
minus02
0
02
04
06
08
1
Real
Imag
inar
yLMPCDLQR
Figure 4 Eigenvalues for DLQR and LMPC in altitude control
0 1 2 3 4 5 6
02
46
0
2
4
6
8
10
xy
z
DesiredSS-MPCLMPC
Figure 5 Tracking a 3D trajectory
631 Constant 119873 Using (58) and choosing a constantparameter119873 = 5 we get Table 2
Thus we can increase the control horizon by increasingthe parameter 119886 without necessarily changing the order119873
632 Constant Control Horizon 119873119888 Assuming we want to
achieve a control horizon of 30 we choose one parameterusually 119873 because it is an integer and directly defines theorder of the discrete Laguerre networkTherefater we simplyuse (58) to find the corresponding scaling factor 119886 as shownin Table 3
Abstract and Applied Analysis 9
0 500 1000 1500 2000 2500 3000 3500 40000
1
2
3
4
5
6
Time (ms)
x p
ositi
on
DesiredSS-MPCLMPC
(a)
0 500 1000 1500 2000 2500 3000 3500 4000Time (ms)
0
1
2
3
4
5
6
y p
ositi
on
DesiredSS-MPCLMPC
(b)
Figure 6 Tracking 119909 (a) and 119910 in (b)
2
4
6
8
10
12
00
500 1000 1500 2000 2500 3000 3500 4000Time (ms)
z p
ositi
on (m
)
DesiredSS-MPCLMPC
(a)
0
01
02
03
04
05
0 500 1000 1500 2000 2500 3000 3500 4000Time (ms)
Hea
ding
angl
e120595
DesiredSS-MPCLMPC
(b)
Figure 7 Altitude (119911) (a) and heading angle (120595) (b)
Table 3
Target control horizon119873119888
Chosen119873 Required 119886
30 1 0967230 3 0904830 5 0846530 7 0791930 9 07408
Figures 8 9 10 11 12 and 13 compare two LMPC designsboth of which yield a control horizon of 30 One designuses 119873 = 1 and the other 119873 = 9 Thus we can choose aLaguerre network with lower number of terms119873 (that gives
lower computation burden) but with a larger scaling factor119886 to achieve the same control horizon The computationalcost is lower if a smaller number of parameters are usedFor 119873 = 1 only one Laguerre term is used to capture thecontrol trajectory while119873 = 9 requires nine Laguerre termsto capture the same control trajectory
7 Conclusion
This paper has presented a solution to stability and trajectorytracking of quadrotor systems using a model predictivecontroller designed by using a special type of orthonor-mal functions called Laguerre functions A quadratic cost
10 Abstract and Applied Analysis
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
x (m
)
DesiredLMPC
Figure 8 119909 trajectory119873 = 1
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
x (m
)
DesiredLMPC
Figure 9 119909 trajectory119873 = 9
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
y (m
)
DesiredLMPC
Figure 10 119910 trajectory119873 = 1
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
y (m
)
DesiredLMPC
Figure 11 119910 trajectory119873 = 9
0
2
4
6
8
10A
ltitu
de (m
)
Altitude
0 05 1 15 2 25 3 35 4Time (s)
DesiredLMPC
Figure 12 119911 trajectory119873 = 1
0
2
4
6
8
10
Alti
tude
(m)
Altitude
0 05 1 15 2 25 3 35 4Time (s)
DesiredLMPC
Figure 13 119911 trajectory119873 = 9
Abstract and Applied Analysis 11
function similar to that used in linear quadratic regulator(LQR) has been used In stability analysis LMPC has beencompared to the optimal DLQR system whereas in trajectorytracking LMPC has been compared to a popular linear state-space MPC desgin technique in which plant constraints aremodeled using linear equalities and inequalities The resultsfrom the simulations indicate that the controller performsvery well and is considered feasible With this perceivedfeasibility LMPCoffers the added advantage that it can handlesystems where rapid sampling and more complicated processdynamics are required [5 10] By selecting appropriate valuesof 119886 and119873 LMPC reduces the number of parameters requiredfor accurate prediction when using the traditional (MPC)approach This is a big advantage in that with the reducednumber of parameters online implementation might bepossible where the traditional MPC would have failed
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
This work was supported by the National Research Foun-dation of Korea (NRF) grant funded by the Koreangovernment (MEST) (no 2013R1A2A2A01068127 and no2013R1A1A2A10009458)
References
[1] J M Maciejowski Predictive Control with Constraints PrenticeHall New York NY USA 1st edition 2002
[2] L Wang ldquoDiscrete time model predictive control design usingLaguerre functionsrdquo in Proceedings of the American ControlConference pp 2430ndash2435 Arlington Va USA June 2001
[3] K Alexis G Nikolakopoulos and A Tzes ldquoExperimentalmodel predictive attitude tracking control of a quadrotorhelicopter subject to wind-gustsrdquo in Proceedings of the 18thMediterranean Conference on Control amp Automation (MED rsquo10)pp 1461ndash1466 IEEE Marrakech Morocco June 2010
[4] G V RaffoM G Ortega and F R Rubio ldquoMPCwith nonlinear119867infin
control for path tracking of a quad-rotor helicopterrdquo inProceedings of the 17th World Congress International Federationof Automatic Control (IFAC rsquo08) July 2008
[5] A M Singh D J Lee D P Hong and K T Chong ldquoSucces-sive loop closure based controller design for an autonomousquadrotor vehiclerdquo Applied Mechanics and Materials vol 483pp 361ndash367 2014
[6] L Wang Model Predictive Control Design and ImplementationUsing MATLAB Advances in Industrial Control Springer NewYork NY USA 1st edition 2009
[7] A G Wills ldquoTechnical report EE04025-Notes on linear modelpredictive controlrdquo December 2004
[8] Y Gao C G Lee and K T Chong ldquoReceding horizon trackingcontrol for wheeled mobile robots with time-delayrdquo Journal ofMechanical Science and Technology vol 22 no 12 pp 2403ndash2416 2008
[9] B D O Anderson and J B Moore Linear Optimal ControlPrentice-Hall 1971
[10] J A Rossiter and L Wang ldquoExploiting Laguerre functions toimprove the feasibilityperformance compromise in MPCrdquo inProceedings of the 47th IEEE Conference onDecision and Control(CDC rsquo08) pp 4737ndash4742 December 2008
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
Abstract and Applied Analysis 7
where 119896 = 0 1 2 119873119901with 119873
119901as the prediction horizon
119873 terms used in the expansion and 119897119895(119896119894) is the inverse 119911-
transform of the 119895th term in the discrete Laguerre networkEquation (44) can be rewritten as
Δ119906 (119896119894+ 119896) = 119871 (119896)
⊤
120578 (45)
where 119871(119896) = [1198971(119896) 119897
2(119896) sdot sdot sdot 119897
119873(119896)]
119879 120578 = [11988811198882sdot sdot sdot 119888
119873]119879
and the coefficients 1198881 1198882 119888
119873are obtained from system
dataAt an arbitrary future instant 119898 the state is described
using Laguerre functions as
119909 (119896119894+ 119898 | 119896
119894) = 119860
119898
119909 (119896119894) +
119898minus1
sum
119894=0
119860119898minus119894minus1
119861119871 (119894)119879
120578 (46)
Similarly the output is described as
119910 (119896119894+ 119898 | 119896
119894) = 119862119860
119898
119909 (119896119894) +
119898minus1
sum
119894=0
119862119860119898minus119894minus1
119861119871 (119894)119879
120578 (47)
522 Cost Function The cost function is used to choose theoptimal control trajectory Δ119880 to bring the predicted outputas close as possible to the set-point The cost function
119869 =
119873119901
sum
119898=1
119909 (119896119894+ 119898 | 119896
119894)119879
119876119909 (119896119894+ 119898 | 119896
119894) + 120578
119879
119877119871120578 (48)
minimizes the error between the set-point signal and theoutput in the shortest possible time by carefully tuning theweighting matrices 119876 ge 0 and 119877
119871gt 0
523 Cost Function Minimization The state variable in (46)can be rewritten as
119909 (119896119894+ 119898 | 119896
119894) = 119860
119898
119909 (119896119894) + 120601 (119898)
119879
120578 (49)
where 120601(119898)119879
= sum119898minus1
119894=0119860119898minus119894minus1
119861119871(119894)119879
By substituting (49) into (48) and performing the partialderivative 120597119869120597120578 = 0 the Laguerre coefficients vector is foundto be
120578 = minusΩminus1
Ψ119909 (119896119894) (50)
withΩ = sum119873119901
119898=1120601(119898)119876120601(119898)
119879
+ 119877119871and Ψ = 120601(119898)119876119860
119898
524 Receding Horizon Control In receding horizon control(RHC) only the first term in Δ119880 that is Δ119906(119896
119894) is imple-
mented at instant 119896119894 The rest of the sequence is ignored
Only the most recent measurement is taken to form the statevector for calculation of the control signal This procedure isrepeated in real-time to give the receding horizon control law[6 8]
525 Stability In stability analysis we make use of the tech-nique of exponential data weighting originated by Andersonand Moore [9] and applied to MPC in [6] More specifically
we will concentrate on the discrete exponential factor 119890120582Δ119905where Δ119905 is the sampling interval and the discrete weightsform a geometric sequence 120572119895 119895 = 0 1 2
The proposed cost function is similar to the one usedin linear quadratic regulator (LQR) systems but with theinclusion of discrete weights
119869119908=
119873119901
sum
119895=1
120572minus2119895
119909 (119896119894+ 119895 | 119896
119894)119879
119876119909 (119896119894+ 119895 | 119896
119894)
+
119873119901
sum
119895=0
120572minus2119895
Δ119906 (119896119894+ 119895)
119879
119877Δ119906 (119896119894+ 119895)
(51)
For 120572 gt 1 the exponential weights 120572minus2119895 119895 = 1 2 3 119873119901
put more emphasis on the current state 119909(119896119894+ 119895 | 119896
119894) and less
emphasis on subsequent future statesThe exponentially weighted cost function can be
expressed more compactly as
119869119908=
119873119901
sum
119895=1
119909 (119896119894+ 119895 | 119896
119894)119879
119876119909 (119896119894+ 119895 | 119896
119894)
+
119873119901
sum
119895=0
Δ (119896119894+ 119895)
119879
119877Δ (119896119894+ 119895)
(52)
with state equation119909 (119896 + 1) = 119860
120572119909 (119896) + 119861
120572Δ (119896) (53)
and 119860120572= 119860120572 and 119861
120572= 119861120572
With 119876 ge 0 119877 gt 0 and 119873119901
rarr infin minimizing thecost function 119869
119908is equivalent to the DLQR problem which is
solved using algebraic Riccati equation (54)Thepair (119860120572 119861
120572)
is assumed to be controllable and (119860120572 119862) observable with
119876 = 119862119879
119862 Then there is a stabilizing state feedback controlgain matrix119870
119908
119870119908= (119877 + 120572
minus2
119861119879
119875119908119861)
minus1
120572minus2
119861119879
119875119908119860
119860119879
120572[119875
119908minus 119875
119908
119861
120572(119877 +
119861119879
120572119875119908
119861
120572)
minus1
119861119879
120572119875119908]119860
120572+ 119876 minus 119875
119908= 0
(54)that gives a stable closed loop system with all its eigenvaluesinside the unit circle and the the closed loop system beingdescribed by
119909 (119896119894+ 119895 + 1 | 119896
119894) = 120572
minus1
(119860 minus 119861119870119908) 119909 (119896
119894+ 119895 | 119896
119894) (55)
From (55) the transformed system has all its eigenvaluesinside the unit circle by taking119873
119901rarr infin So
120572minus1 1003816100381610038161003816120582max (119860 minus 119861119870
119908)1003816100381610038161003816 lt 1 (56)
which gives1003816100381610038161003816120582max (119860 minus 119861119870
119908)1003816100381610038161003816 lt 120572 (57)
Thus by choosing 120572 gt 1 there is a great chance that the systemwill be stable Several simulations on the quadrotor systemindicate that a choice of 120572 slightly greater than unity makesthe system stable almost all the time
8 Abstract and Applied Analysis
Table 1
Parameter ValueMass119898 07 kgLength 119897 0275m119869119909
01063 kgsdotm2
119869119910
01063 kgsdotm2
119869119911
02122 kgsdotm2
Table 2
Control horizon119873119888
119873 Required 119886
10 5 0606515 5 0716520 5 0778825 5 0818750 5 09048
6 Simulation Parameters and Results
In order to illustrate the controllerrsquos stability we will checkif the eigenvalues of each of the three individual controllersappear inside a unit circle Andwewill use a trajectory in a 3Dplane to check the overall controllerrsquos ability to track a desiredtrajectory
Table 1 shows the parameters of the quadrotor used insimulation
61 Stability Analysis In this section we look at the locationof eigenvalues of LMPC and check if they fall within the unitcircle for stability of the system We also compare them tothose obtained from the optimal DLQR system To ensurestability as explained in Section 525 we use 120572 = 12
By looking at Figure 4 we note that all the eigenvaluesappear inside the unit circle as required Also the eigenvaluesobtained from LMPC closely match with those of the optimalDLQRThis not only shows the system is stable but also showsthe controllers perform optimally
62 Trajectory Tracking In this section we look at theperformance of LMPC when used to track a 3D trajectoryAlso LMPC is compared to SS-MPC and individual statetrajectories are observed The starting point of the trajectoryis (119909 119910 119911) = (0 0 0) and final point is (119909 119910 119911) = (5 5 10)Parameter 119886 = 08187 while119873 = 5 See Figure 5
Clearly both LMPC and SS-MPC track the desired tra-jectory very well See Figures 5 6 and 7 However LMPCperforms better than SS-MPC in that it requires only fiveparameters (119873 = 5) to capture the control trajectorycompared to SS-MPCrsquosminimumof 25 (119873
119888= 25) parameters
(see Table 2 for relationship of119873 and119873119888with 119886 = 08187)
63 Control Horizon in Relation to 119886 and 119873 The controlhorizon (119873
119888) is related to the parameters 119886 and119873 [2] by
119886 asymp 119890minus119873119873
119888 (58)
minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1
minus08
minus06
minus04
minus02
0
02
04
06
08
1
Real
Imag
inar
yLMPCDLQR
Figure 4 Eigenvalues for DLQR and LMPC in altitude control
0 1 2 3 4 5 6
02
46
0
2
4
6
8
10
xy
z
DesiredSS-MPCLMPC
Figure 5 Tracking a 3D trajectory
631 Constant 119873 Using (58) and choosing a constantparameter119873 = 5 we get Table 2
Thus we can increase the control horizon by increasingthe parameter 119886 without necessarily changing the order119873
632 Constant Control Horizon 119873119888 Assuming we want to
achieve a control horizon of 30 we choose one parameterusually 119873 because it is an integer and directly defines theorder of the discrete Laguerre networkTherefater we simplyuse (58) to find the corresponding scaling factor 119886 as shownin Table 3
Abstract and Applied Analysis 9
0 500 1000 1500 2000 2500 3000 3500 40000
1
2
3
4
5
6
Time (ms)
x p
ositi
on
DesiredSS-MPCLMPC
(a)
0 500 1000 1500 2000 2500 3000 3500 4000Time (ms)
0
1
2
3
4
5
6
y p
ositi
on
DesiredSS-MPCLMPC
(b)
Figure 6 Tracking 119909 (a) and 119910 in (b)
2
4
6
8
10
12
00
500 1000 1500 2000 2500 3000 3500 4000Time (ms)
z p
ositi
on (m
)
DesiredSS-MPCLMPC
(a)
0
01
02
03
04
05
0 500 1000 1500 2000 2500 3000 3500 4000Time (ms)
Hea
ding
angl
e120595
DesiredSS-MPCLMPC
(b)
Figure 7 Altitude (119911) (a) and heading angle (120595) (b)
Table 3
Target control horizon119873119888
Chosen119873 Required 119886
30 1 0967230 3 0904830 5 0846530 7 0791930 9 07408
Figures 8 9 10 11 12 and 13 compare two LMPC designsboth of which yield a control horizon of 30 One designuses 119873 = 1 and the other 119873 = 9 Thus we can choose aLaguerre network with lower number of terms119873 (that gives
lower computation burden) but with a larger scaling factor119886 to achieve the same control horizon The computationalcost is lower if a smaller number of parameters are usedFor 119873 = 1 only one Laguerre term is used to capture thecontrol trajectory while119873 = 9 requires nine Laguerre termsto capture the same control trajectory
7 Conclusion
This paper has presented a solution to stability and trajectorytracking of quadrotor systems using a model predictivecontroller designed by using a special type of orthonor-mal functions called Laguerre functions A quadratic cost
10 Abstract and Applied Analysis
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
x (m
)
DesiredLMPC
Figure 8 119909 trajectory119873 = 1
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
x (m
)
DesiredLMPC
Figure 9 119909 trajectory119873 = 9
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
y (m
)
DesiredLMPC
Figure 10 119910 trajectory119873 = 1
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
y (m
)
DesiredLMPC
Figure 11 119910 trajectory119873 = 9
0
2
4
6
8
10A
ltitu
de (m
)
Altitude
0 05 1 15 2 25 3 35 4Time (s)
DesiredLMPC
Figure 12 119911 trajectory119873 = 1
0
2
4
6
8
10
Alti
tude
(m)
Altitude
0 05 1 15 2 25 3 35 4Time (s)
DesiredLMPC
Figure 13 119911 trajectory119873 = 9
Abstract and Applied Analysis 11
function similar to that used in linear quadratic regulator(LQR) has been used In stability analysis LMPC has beencompared to the optimal DLQR system whereas in trajectorytracking LMPC has been compared to a popular linear state-space MPC desgin technique in which plant constraints aremodeled using linear equalities and inequalities The resultsfrom the simulations indicate that the controller performsvery well and is considered feasible With this perceivedfeasibility LMPCoffers the added advantage that it can handlesystems where rapid sampling and more complicated processdynamics are required [5 10] By selecting appropriate valuesof 119886 and119873 LMPC reduces the number of parameters requiredfor accurate prediction when using the traditional (MPC)approach This is a big advantage in that with the reducednumber of parameters online implementation might bepossible where the traditional MPC would have failed
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
This work was supported by the National Research Foun-dation of Korea (NRF) grant funded by the Koreangovernment (MEST) (no 2013R1A2A2A01068127 and no2013R1A1A2A10009458)
References
[1] J M Maciejowski Predictive Control with Constraints PrenticeHall New York NY USA 1st edition 2002
[2] L Wang ldquoDiscrete time model predictive control design usingLaguerre functionsrdquo in Proceedings of the American ControlConference pp 2430ndash2435 Arlington Va USA June 2001
[3] K Alexis G Nikolakopoulos and A Tzes ldquoExperimentalmodel predictive attitude tracking control of a quadrotorhelicopter subject to wind-gustsrdquo in Proceedings of the 18thMediterranean Conference on Control amp Automation (MED rsquo10)pp 1461ndash1466 IEEE Marrakech Morocco June 2010
[4] G V RaffoM G Ortega and F R Rubio ldquoMPCwith nonlinear119867infin
control for path tracking of a quad-rotor helicopterrdquo inProceedings of the 17th World Congress International Federationof Automatic Control (IFAC rsquo08) July 2008
[5] A M Singh D J Lee D P Hong and K T Chong ldquoSucces-sive loop closure based controller design for an autonomousquadrotor vehiclerdquo Applied Mechanics and Materials vol 483pp 361ndash367 2014
[6] L Wang Model Predictive Control Design and ImplementationUsing MATLAB Advances in Industrial Control Springer NewYork NY USA 1st edition 2009
[7] A G Wills ldquoTechnical report EE04025-Notes on linear modelpredictive controlrdquo December 2004
[8] Y Gao C G Lee and K T Chong ldquoReceding horizon trackingcontrol for wheeled mobile robots with time-delayrdquo Journal ofMechanical Science and Technology vol 22 no 12 pp 2403ndash2416 2008
[9] B D O Anderson and J B Moore Linear Optimal ControlPrentice-Hall 1971
[10] J A Rossiter and L Wang ldquoExploiting Laguerre functions toimprove the feasibilityperformance compromise in MPCrdquo inProceedings of the 47th IEEE Conference onDecision and Control(CDC rsquo08) pp 4737ndash4742 December 2008
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
8 Abstract and Applied Analysis
Table 1
Parameter ValueMass119898 07 kgLength 119897 0275m119869119909
01063 kgsdotm2
119869119910
01063 kgsdotm2
119869119911
02122 kgsdotm2
Table 2
Control horizon119873119888
119873 Required 119886
10 5 0606515 5 0716520 5 0778825 5 0818750 5 09048
6 Simulation Parameters and Results
In order to illustrate the controllerrsquos stability we will checkif the eigenvalues of each of the three individual controllersappear inside a unit circle Andwewill use a trajectory in a 3Dplane to check the overall controllerrsquos ability to track a desiredtrajectory
Table 1 shows the parameters of the quadrotor used insimulation
61 Stability Analysis In this section we look at the locationof eigenvalues of LMPC and check if they fall within the unitcircle for stability of the system We also compare them tothose obtained from the optimal DLQR system To ensurestability as explained in Section 525 we use 120572 = 12
By looking at Figure 4 we note that all the eigenvaluesappear inside the unit circle as required Also the eigenvaluesobtained from LMPC closely match with those of the optimalDLQRThis not only shows the system is stable but also showsthe controllers perform optimally
62 Trajectory Tracking In this section we look at theperformance of LMPC when used to track a 3D trajectoryAlso LMPC is compared to SS-MPC and individual statetrajectories are observed The starting point of the trajectoryis (119909 119910 119911) = (0 0 0) and final point is (119909 119910 119911) = (5 5 10)Parameter 119886 = 08187 while119873 = 5 See Figure 5
Clearly both LMPC and SS-MPC track the desired tra-jectory very well See Figures 5 6 and 7 However LMPCperforms better than SS-MPC in that it requires only fiveparameters (119873 = 5) to capture the control trajectorycompared to SS-MPCrsquosminimumof 25 (119873
119888= 25) parameters
(see Table 2 for relationship of119873 and119873119888with 119886 = 08187)
63 Control Horizon in Relation to 119886 and 119873 The controlhorizon (119873
119888) is related to the parameters 119886 and119873 [2] by
119886 asymp 119890minus119873119873
119888 (58)
minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1
minus08
minus06
minus04
minus02
0
02
04
06
08
1
Real
Imag
inar
yLMPCDLQR
Figure 4 Eigenvalues for DLQR and LMPC in altitude control
0 1 2 3 4 5 6
02
46
0
2
4
6
8
10
xy
z
DesiredSS-MPCLMPC
Figure 5 Tracking a 3D trajectory
631 Constant 119873 Using (58) and choosing a constantparameter119873 = 5 we get Table 2
Thus we can increase the control horizon by increasingthe parameter 119886 without necessarily changing the order119873
632 Constant Control Horizon 119873119888 Assuming we want to
achieve a control horizon of 30 we choose one parameterusually 119873 because it is an integer and directly defines theorder of the discrete Laguerre networkTherefater we simplyuse (58) to find the corresponding scaling factor 119886 as shownin Table 3
Abstract and Applied Analysis 9
0 500 1000 1500 2000 2500 3000 3500 40000
1
2
3
4
5
6
Time (ms)
x p
ositi
on
DesiredSS-MPCLMPC
(a)
0 500 1000 1500 2000 2500 3000 3500 4000Time (ms)
0
1
2
3
4
5
6
y p
ositi
on
DesiredSS-MPCLMPC
(b)
Figure 6 Tracking 119909 (a) and 119910 in (b)
2
4
6
8
10
12
00
500 1000 1500 2000 2500 3000 3500 4000Time (ms)
z p
ositi
on (m
)
DesiredSS-MPCLMPC
(a)
0
01
02
03
04
05
0 500 1000 1500 2000 2500 3000 3500 4000Time (ms)
Hea
ding
angl
e120595
DesiredSS-MPCLMPC
(b)
Figure 7 Altitude (119911) (a) and heading angle (120595) (b)
Table 3
Target control horizon119873119888
Chosen119873 Required 119886
30 1 0967230 3 0904830 5 0846530 7 0791930 9 07408
Figures 8 9 10 11 12 and 13 compare two LMPC designsboth of which yield a control horizon of 30 One designuses 119873 = 1 and the other 119873 = 9 Thus we can choose aLaguerre network with lower number of terms119873 (that gives
lower computation burden) but with a larger scaling factor119886 to achieve the same control horizon The computationalcost is lower if a smaller number of parameters are usedFor 119873 = 1 only one Laguerre term is used to capture thecontrol trajectory while119873 = 9 requires nine Laguerre termsto capture the same control trajectory
7 Conclusion
This paper has presented a solution to stability and trajectorytracking of quadrotor systems using a model predictivecontroller designed by using a special type of orthonor-mal functions called Laguerre functions A quadratic cost
10 Abstract and Applied Analysis
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
x (m
)
DesiredLMPC
Figure 8 119909 trajectory119873 = 1
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
x (m
)
DesiredLMPC
Figure 9 119909 trajectory119873 = 9
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
y (m
)
DesiredLMPC
Figure 10 119910 trajectory119873 = 1
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
y (m
)
DesiredLMPC
Figure 11 119910 trajectory119873 = 9
0
2
4
6
8
10A
ltitu
de (m
)
Altitude
0 05 1 15 2 25 3 35 4Time (s)
DesiredLMPC
Figure 12 119911 trajectory119873 = 1
0
2
4
6
8
10
Alti
tude
(m)
Altitude
0 05 1 15 2 25 3 35 4Time (s)
DesiredLMPC
Figure 13 119911 trajectory119873 = 9
Abstract and Applied Analysis 11
function similar to that used in linear quadratic regulator(LQR) has been used In stability analysis LMPC has beencompared to the optimal DLQR system whereas in trajectorytracking LMPC has been compared to a popular linear state-space MPC desgin technique in which plant constraints aremodeled using linear equalities and inequalities The resultsfrom the simulations indicate that the controller performsvery well and is considered feasible With this perceivedfeasibility LMPCoffers the added advantage that it can handlesystems where rapid sampling and more complicated processdynamics are required [5 10] By selecting appropriate valuesof 119886 and119873 LMPC reduces the number of parameters requiredfor accurate prediction when using the traditional (MPC)approach This is a big advantage in that with the reducednumber of parameters online implementation might bepossible where the traditional MPC would have failed
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
This work was supported by the National Research Foun-dation of Korea (NRF) grant funded by the Koreangovernment (MEST) (no 2013R1A2A2A01068127 and no2013R1A1A2A10009458)
References
[1] J M Maciejowski Predictive Control with Constraints PrenticeHall New York NY USA 1st edition 2002
[2] L Wang ldquoDiscrete time model predictive control design usingLaguerre functionsrdquo in Proceedings of the American ControlConference pp 2430ndash2435 Arlington Va USA June 2001
[3] K Alexis G Nikolakopoulos and A Tzes ldquoExperimentalmodel predictive attitude tracking control of a quadrotorhelicopter subject to wind-gustsrdquo in Proceedings of the 18thMediterranean Conference on Control amp Automation (MED rsquo10)pp 1461ndash1466 IEEE Marrakech Morocco June 2010
[4] G V RaffoM G Ortega and F R Rubio ldquoMPCwith nonlinear119867infin
control for path tracking of a quad-rotor helicopterrdquo inProceedings of the 17th World Congress International Federationof Automatic Control (IFAC rsquo08) July 2008
[5] A M Singh D J Lee D P Hong and K T Chong ldquoSucces-sive loop closure based controller design for an autonomousquadrotor vehiclerdquo Applied Mechanics and Materials vol 483pp 361ndash367 2014
[6] L Wang Model Predictive Control Design and ImplementationUsing MATLAB Advances in Industrial Control Springer NewYork NY USA 1st edition 2009
[7] A G Wills ldquoTechnical report EE04025-Notes on linear modelpredictive controlrdquo December 2004
[8] Y Gao C G Lee and K T Chong ldquoReceding horizon trackingcontrol for wheeled mobile robots with time-delayrdquo Journal ofMechanical Science and Technology vol 22 no 12 pp 2403ndash2416 2008
[9] B D O Anderson and J B Moore Linear Optimal ControlPrentice-Hall 1971
[10] J A Rossiter and L Wang ldquoExploiting Laguerre functions toimprove the feasibilityperformance compromise in MPCrdquo inProceedings of the 47th IEEE Conference onDecision and Control(CDC rsquo08) pp 4737ndash4742 December 2008
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
Abstract and Applied Analysis 9
0 500 1000 1500 2000 2500 3000 3500 40000
1
2
3
4
5
6
Time (ms)
x p
ositi
on
DesiredSS-MPCLMPC
(a)
0 500 1000 1500 2000 2500 3000 3500 4000Time (ms)
0
1
2
3
4
5
6
y p
ositi
on
DesiredSS-MPCLMPC
(b)
Figure 6 Tracking 119909 (a) and 119910 in (b)
2
4
6
8
10
12
00
500 1000 1500 2000 2500 3000 3500 4000Time (ms)
z p
ositi
on (m
)
DesiredSS-MPCLMPC
(a)
0
01
02
03
04
05
0 500 1000 1500 2000 2500 3000 3500 4000Time (ms)
Hea
ding
angl
e120595
DesiredSS-MPCLMPC
(b)
Figure 7 Altitude (119911) (a) and heading angle (120595) (b)
Table 3
Target control horizon119873119888
Chosen119873 Required 119886
30 1 0967230 3 0904830 5 0846530 7 0791930 9 07408
Figures 8 9 10 11 12 and 13 compare two LMPC designsboth of which yield a control horizon of 30 One designuses 119873 = 1 and the other 119873 = 9 Thus we can choose aLaguerre network with lower number of terms119873 (that gives
lower computation burden) but with a larger scaling factor119886 to achieve the same control horizon The computationalcost is lower if a smaller number of parameters are usedFor 119873 = 1 only one Laguerre term is used to capture thecontrol trajectory while119873 = 9 requires nine Laguerre termsto capture the same control trajectory
7 Conclusion
This paper has presented a solution to stability and trajectorytracking of quadrotor systems using a model predictivecontroller designed by using a special type of orthonor-mal functions called Laguerre functions A quadratic cost
10 Abstract and Applied Analysis
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
x (m
)
DesiredLMPC
Figure 8 119909 trajectory119873 = 1
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
x (m
)
DesiredLMPC
Figure 9 119909 trajectory119873 = 9
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
y (m
)
DesiredLMPC
Figure 10 119910 trajectory119873 = 1
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
y (m
)
DesiredLMPC
Figure 11 119910 trajectory119873 = 9
0
2
4
6
8
10A
ltitu
de (m
)
Altitude
0 05 1 15 2 25 3 35 4Time (s)
DesiredLMPC
Figure 12 119911 trajectory119873 = 1
0
2
4
6
8
10
Alti
tude
(m)
Altitude
0 05 1 15 2 25 3 35 4Time (s)
DesiredLMPC
Figure 13 119911 trajectory119873 = 9
Abstract and Applied Analysis 11
function similar to that used in linear quadratic regulator(LQR) has been used In stability analysis LMPC has beencompared to the optimal DLQR system whereas in trajectorytracking LMPC has been compared to a popular linear state-space MPC desgin technique in which plant constraints aremodeled using linear equalities and inequalities The resultsfrom the simulations indicate that the controller performsvery well and is considered feasible With this perceivedfeasibility LMPCoffers the added advantage that it can handlesystems where rapid sampling and more complicated processdynamics are required [5 10] By selecting appropriate valuesof 119886 and119873 LMPC reduces the number of parameters requiredfor accurate prediction when using the traditional (MPC)approach This is a big advantage in that with the reducednumber of parameters online implementation might bepossible where the traditional MPC would have failed
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
This work was supported by the National Research Foun-dation of Korea (NRF) grant funded by the Koreangovernment (MEST) (no 2013R1A2A2A01068127 and no2013R1A1A2A10009458)
References
[1] J M Maciejowski Predictive Control with Constraints PrenticeHall New York NY USA 1st edition 2002
[2] L Wang ldquoDiscrete time model predictive control design usingLaguerre functionsrdquo in Proceedings of the American ControlConference pp 2430ndash2435 Arlington Va USA June 2001
[3] K Alexis G Nikolakopoulos and A Tzes ldquoExperimentalmodel predictive attitude tracking control of a quadrotorhelicopter subject to wind-gustsrdquo in Proceedings of the 18thMediterranean Conference on Control amp Automation (MED rsquo10)pp 1461ndash1466 IEEE Marrakech Morocco June 2010
[4] G V RaffoM G Ortega and F R Rubio ldquoMPCwith nonlinear119867infin
control for path tracking of a quad-rotor helicopterrdquo inProceedings of the 17th World Congress International Federationof Automatic Control (IFAC rsquo08) July 2008
[5] A M Singh D J Lee D P Hong and K T Chong ldquoSucces-sive loop closure based controller design for an autonomousquadrotor vehiclerdquo Applied Mechanics and Materials vol 483pp 361ndash367 2014
[6] L Wang Model Predictive Control Design and ImplementationUsing MATLAB Advances in Industrial Control Springer NewYork NY USA 1st edition 2009
[7] A G Wills ldquoTechnical report EE04025-Notes on linear modelpredictive controlrdquo December 2004
[8] Y Gao C G Lee and K T Chong ldquoReceding horizon trackingcontrol for wheeled mobile robots with time-delayrdquo Journal ofMechanical Science and Technology vol 22 no 12 pp 2403ndash2416 2008
[9] B D O Anderson and J B Moore Linear Optimal ControlPrentice-Hall 1971
[10] J A Rossiter and L Wang ldquoExploiting Laguerre functions toimprove the feasibilityperformance compromise in MPCrdquo inProceedings of the 47th IEEE Conference onDecision and Control(CDC rsquo08) pp 4737ndash4742 December 2008
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
10 Abstract and Applied Analysis
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
x (m
)
DesiredLMPC
Figure 8 119909 trajectory119873 = 1
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
x (m
)
DesiredLMPC
Figure 9 119909 trajectory119873 = 9
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
y (m
)
DesiredLMPC
Figure 10 119910 trajectory119873 = 1
0 05 1 15 2 25 3 35 40
1
2
3
4
5
6
Time (s)
y (m
)
DesiredLMPC
Figure 11 119910 trajectory119873 = 9
0
2
4
6
8
10A
ltitu
de (m
)
Altitude
0 05 1 15 2 25 3 35 4Time (s)
DesiredLMPC
Figure 12 119911 trajectory119873 = 1
0
2
4
6
8
10
Alti
tude
(m)
Altitude
0 05 1 15 2 25 3 35 4Time (s)
DesiredLMPC
Figure 13 119911 trajectory119873 = 9
Abstract and Applied Analysis 11
function similar to that used in linear quadratic regulator(LQR) has been used In stability analysis LMPC has beencompared to the optimal DLQR system whereas in trajectorytracking LMPC has been compared to a popular linear state-space MPC desgin technique in which plant constraints aremodeled using linear equalities and inequalities The resultsfrom the simulations indicate that the controller performsvery well and is considered feasible With this perceivedfeasibility LMPCoffers the added advantage that it can handlesystems where rapid sampling and more complicated processdynamics are required [5 10] By selecting appropriate valuesof 119886 and119873 LMPC reduces the number of parameters requiredfor accurate prediction when using the traditional (MPC)approach This is a big advantage in that with the reducednumber of parameters online implementation might bepossible where the traditional MPC would have failed
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
This work was supported by the National Research Foun-dation of Korea (NRF) grant funded by the Koreangovernment (MEST) (no 2013R1A2A2A01068127 and no2013R1A1A2A10009458)
References
[1] J M Maciejowski Predictive Control with Constraints PrenticeHall New York NY USA 1st edition 2002
[2] L Wang ldquoDiscrete time model predictive control design usingLaguerre functionsrdquo in Proceedings of the American ControlConference pp 2430ndash2435 Arlington Va USA June 2001
[3] K Alexis G Nikolakopoulos and A Tzes ldquoExperimentalmodel predictive attitude tracking control of a quadrotorhelicopter subject to wind-gustsrdquo in Proceedings of the 18thMediterranean Conference on Control amp Automation (MED rsquo10)pp 1461ndash1466 IEEE Marrakech Morocco June 2010
[4] G V RaffoM G Ortega and F R Rubio ldquoMPCwith nonlinear119867infin
control for path tracking of a quad-rotor helicopterrdquo inProceedings of the 17th World Congress International Federationof Automatic Control (IFAC rsquo08) July 2008
[5] A M Singh D J Lee D P Hong and K T Chong ldquoSucces-sive loop closure based controller design for an autonomousquadrotor vehiclerdquo Applied Mechanics and Materials vol 483pp 361ndash367 2014
[6] L Wang Model Predictive Control Design and ImplementationUsing MATLAB Advances in Industrial Control Springer NewYork NY USA 1st edition 2009
[7] A G Wills ldquoTechnical report EE04025-Notes on linear modelpredictive controlrdquo December 2004
[8] Y Gao C G Lee and K T Chong ldquoReceding horizon trackingcontrol for wheeled mobile robots with time-delayrdquo Journal ofMechanical Science and Technology vol 22 no 12 pp 2403ndash2416 2008
[9] B D O Anderson and J B Moore Linear Optimal ControlPrentice-Hall 1971
[10] J A Rossiter and L Wang ldquoExploiting Laguerre functions toimprove the feasibilityperformance compromise in MPCrdquo inProceedings of the 47th IEEE Conference onDecision and Control(CDC rsquo08) pp 4737ndash4742 December 2008
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
Abstract and Applied Analysis 11
function similar to that used in linear quadratic regulator(LQR) has been used In stability analysis LMPC has beencompared to the optimal DLQR system whereas in trajectorytracking LMPC has been compared to a popular linear state-space MPC desgin technique in which plant constraints aremodeled using linear equalities and inequalities The resultsfrom the simulations indicate that the controller performsvery well and is considered feasible With this perceivedfeasibility LMPCoffers the added advantage that it can handlesystems where rapid sampling and more complicated processdynamics are required [5 10] By selecting appropriate valuesof 119886 and119873 LMPC reduces the number of parameters requiredfor accurate prediction when using the traditional (MPC)approach This is a big advantage in that with the reducednumber of parameters online implementation might bepossible where the traditional MPC would have failed
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
This work was supported by the National Research Foun-dation of Korea (NRF) grant funded by the Koreangovernment (MEST) (no 2013R1A2A2A01068127 and no2013R1A1A2A10009458)
References
[1] J M Maciejowski Predictive Control with Constraints PrenticeHall New York NY USA 1st edition 2002
[2] L Wang ldquoDiscrete time model predictive control design usingLaguerre functionsrdquo in Proceedings of the American ControlConference pp 2430ndash2435 Arlington Va USA June 2001
[3] K Alexis G Nikolakopoulos and A Tzes ldquoExperimentalmodel predictive attitude tracking control of a quadrotorhelicopter subject to wind-gustsrdquo in Proceedings of the 18thMediterranean Conference on Control amp Automation (MED rsquo10)pp 1461ndash1466 IEEE Marrakech Morocco June 2010
[4] G V RaffoM G Ortega and F R Rubio ldquoMPCwith nonlinear119867infin
control for path tracking of a quad-rotor helicopterrdquo inProceedings of the 17th World Congress International Federationof Automatic Control (IFAC rsquo08) July 2008
[5] A M Singh D J Lee D P Hong and K T Chong ldquoSucces-sive loop closure based controller design for an autonomousquadrotor vehiclerdquo Applied Mechanics and Materials vol 483pp 361ndash367 2014
[6] L Wang Model Predictive Control Design and ImplementationUsing MATLAB Advances in Industrial Control Springer NewYork NY USA 1st edition 2009
[7] A G Wills ldquoTechnical report EE04025-Notes on linear modelpredictive controlrdquo December 2004
[8] Y Gao C G Lee and K T Chong ldquoReceding horizon trackingcontrol for wheeled mobile robots with time-delayrdquo Journal ofMechanical Science and Technology vol 22 no 12 pp 2403ndash2416 2008
[9] B D O Anderson and J B Moore Linear Optimal ControlPrentice-Hall 1971
[10] J A Rossiter and L Wang ldquoExploiting Laguerre functions toimprove the feasibilityperformance compromise in MPCrdquo inProceedings of the 47th IEEE Conference onDecision and Control(CDC rsquo08) pp 4737ndash4742 December 2008
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of