Model Predictive Control: For An Artical Pancreas

Model Predictive Control
for an artical pancreas

B.Sc. Thesis
Matias Srensen og Simon Kristiansen
Technical University of Denmark

Kongens Lyngby 2007
Technical University of Denmark

Informatics and Mathematical Modelling
Building 321, DK-2800 Kongens Lyngby, Denmark
Phone +45 45253351, Fax +45 45882673
reception@imm.dtu.dk
www.imm.dtu.dk
Supervisor: John Bagterp Jrgensen, IMM
Summary
This thesis deals with linear Model Predictive Control, MPC, with the goal of
making a controller for an articial pancreas. A diabetic is simulated by a mathematical model, and based on this model the MPC will compute the optimal
insulin input, taking constraints, disturbances and noise into account. Below is
a brief description of each chapter.
Chapter 2 describes linearization of dierential equations in continuous time.
These are converted into a linear state-space model with discrete time representation, which is a requirement for linear MPC.
In Chapter 3, the basic idea for MPC is reviewed. By starting of with MPC on
basic form, the control model is extended step by step. Constraints are imposed
on the input and on the input rate of movement, which makes sure the input
appears in a controlled volume and speed.
Soft constraints are used for the output, to ensure the output are held inside
the wanted boundaries, but, if needed, the boundaries can be violated. In some
cases it would be impossible to stay within the constraints, and this would make
the problem infeasible, if soft constraints wasn't used.
Feedforward and feedback is also described.
These two approaches will make
the system more robust, and gives a better reaction speed on changes in the
process, such as disturbances and noise.
Chapter 4 is the implementation of the MPC problem in
divided into three dierent phases:
ATLAB. This is
Design, simulation and evaluation.
user calls a le, where the wanted test scenario is initialized.
The
This includes
properties such as constraints and meal disturbances. After all the needed parameters are specied, the MPC controller is designed, and the simulation is
completed. Finally the simulation is evalutated by various plots.
This implementation is used on a modied version of Bergmans minimal model
in Chapter 5. This model consists of ve dierential equations, which simulates
a type 1 diabetic patient. The knowledge from the previous chapters is used to
transform the model into a linear state space model with discrete time representation, and then use this with MPC. This chapter also discuss the selection
ii
of noise and weight matrices. Furthermore simulations are done, to nally test
the functionality of the controller.
The thesis concludes that, it is possible that MPC can be used for the purpose
of an insulin pump, but severe testing, and a better model would be needed.
Resum
Formlet med dette projekt er at undersge om Model Predictive Control, MPC,
kan bruges som kontrolanordning til en insulinpumpe, hvilket kan blive brugt
til at udvikle en kunstig bugspytkirtel til mennesker som ikke producerer insulin
selv.
De basale ider for liner MPC vil blive gennemget og implementeret, og
anvendt p en matematisk model for en type 1 diabetiker.
Dette inkluderer
begrnsninger p insulin input og input-hastigheden, hvilket srger for der ikke

bliver injiceret for meget insulin, og/eller dette foregr for hurtigt.
P systemets output, patientens blodsukker niveau, bruges der blde begrnsninger, hvilket betyder at grnserne kan blive krydset, for at sikre at kontrol
problemet ikke bliver ulseligt. Dette betyder at der er en risiko for at patientens blodsukker niveau ikke ligger p et sundt niveau i en kortere periode, men
som det vil blive vist, sker dette kun i ekstreme tilflde.
Kontrolleren hndterer forstyrelser i systemet vha.
feedforward og feedback.
Feedback fr kontrol algoritmen til at evaluere det mlte output og dermed

basere det kommende insulin input p dette, mens feedforward kan komme
fremtidige forstyrrelser, ssom indtagelse af mad og drikke, i forkbet. Sdan
hndtering af forstyrrelser gr at systemet bliver mere robust og hurtigt reagerende.
I simuleringerne bliver patienten udsat for forskellige scenarier med forskellige
strrelser mltider.
Her bliver kontrollerens ydelse testet, og det viser sig at
den hndterer de este scnearier godt, dog skal det haves i mente at den ktive
patient er baseret p en noget mangelfuld model.
iv
Contents
Summary
Resum
iii
Introduction
Linearization
2.1
Continuous-Discrete Time Conversion
. . . . . . . . . . . . . . .
2.2
Summary
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1
. . . . . . . . . . . . . . . . . . . . . . . . .
3.2
Unconstrained MPC
3.1.1
Regularization
. . . . . . . . . . . . . . . . . . . . . . . .
12
3.1.2
Disturbance . . . . . . . . . . . . . . . . . . . . . . . . . .
14
3.1.3
Feed-forward/feedback . . . . . . . . . . . . . . . . . . . .
17
Constrained MPC
. . . . . . . . . . . . . . . . . . . . . . . . . .
19
3.2.1
Input constraints . . . . . . . . . . . . . . . . . . . . . . .
19
3.2.2
Constraints on input rate of movement . . . . . . . . . . .
20
3.2.3
Output constraints . . . . . . . . . . . . . . . . . . . . . .
22
vi
CONTENTS
3.2.4
. . . . . . . . . . . . . . . . . . .
23
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
26
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
28
3.3
Kalman lter
3.4
Summary
Implementation
29
4.1
ScenaX.m
4.2
MPCControl.m .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.1
MPCDesign.m
4.2.2
MPCSimulate.m
4.2.3
MPCPlot
. . . . . . . . . . . . . . . . . . . . . . . . .
30
30
30
. . . . . . . . . . . . . . . . . . . . . . . .
31
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
34
Case study - A minimal model
35
5.1
Linearization
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
38
5.2
MPC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
39
5.2.1
Weight matrices
39
5.2.2
Horizon and sampling time
5.2.3
Noise
5.3
Soft output constraints
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . .
42
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
43
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
44
5.3.1
Scenario I . . . . . . . . . . . . . . . . . . . . . . . . . . .
44
5.3.2
Scenario II
. . . . . . . . . . . . . . . . . . . . . . . . . .
54
5.3.3
Scenario III . . . . . . . . . . . . . . . . . . . . . . . . . .
56
Simulations
Conclusion
59
CONTENTS
vii
A Impulse-response method
61
B Matlab programs
63
B.1
Scena1.m
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
63
B.2
Scena2.m
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
64
B.3
Scena3.m
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
65
B.4
MPCControl.m . . . . . . . . . . . . . . . . . . . . . . . . . . . .
66
B.5
MPCDesign.m
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
66
B.6
DesignKalman.m . . . . . . . . . . . . . . . . . . . . . . . . . . .
68
B.7
DesignDiscreteMatrices.m . . . . . . . . . . . . . . . . . . . . . .
68
B.8
DesignParameters.m
. . . . . . . . . . . . . . . . . . . . . . . . .
69
B.9
DesignConstraints.m . . . . . . . . . . . . . . . . . . . . . . . . .
69
B.10 DesignMPCMatrices.m . . . . . . . . . . . . . . . . . . . . . . . .
70
B.11 MPCSimulate.m
. . . . . . . . . . . . . . . . . . . . . . . . . . .
72
B.12 MPCCompute.m
. . . . . . . . . . . . . . . . . . . . . . . . . . .
73
B.13 BergmanMinimalModel.m . . . . . . . . . . . . . . . . . . . . . .
74
B.14 MPCPlot.m . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
75
B.15 InvestSampling.m . . . . . . . . . . . . . . . . . . . . . . . . . . .
77
B.16 InvestWeights.m
78
Bibliography
. . . . . . . . . . . . . . . . . . . . . . . . . . .
81
viii
CONTENTS
Chapter
1
Introduction
In year 2000 there were around 171 mio.
people with diabetes in the world.
WHO has estimated that this number will be 366 mio. in 2030, i.e. diabetes is
a rising problem.
In the USA alone, there is used 130 billion USD a year on diabetes, which is
10% of the health care budget.
There are two types of diabetes, type 1 and type 2. Type 1 is called the insulin
dependent diabetes and occur when the cells
in the pancreas, which producess insulin, are
destroyed.
The human body needs insulin to move glucose from the food into cells,
throughout
the body, so the energy in the food can be

used.
Figure 1.1 illustrates how the pancreas works
in a healthy body.
When there isn't produced any insulin, the
glucose is accumulated in the blood, which
can cause serious damage.
Therefore peo-
ple with diabetes needs to have an external

source of insulin, which means injections into
Figure
1.1:
The
diabetes-
the veins.
glucose-regulation
These injections are done with a needle several times at day, with the correct proportioning of insulin for the individual.
If the patient gets too much
insulin he can go into hypoglycemia, and too little can cause hyperglycemia.
Hyperglycemia can lead to blindness, kidney failure and other long terms complication, while hypoglycemia can lead to loss of consciousness and coma.
device which could automate the injection of insulin would obviously be a big
advantage for the patient.
Introduction
A comparatively new device for type 1 diabetes is an insulin pump, which makes
the injection for the diseased person.
The pump consists essentially of three components: A sensor, which measures
the blood glucose concentration in the body, a controller that estimates the
needed insulin quantity, and a pump which makes the injection. See Figure 1.2.
This project will deal with a control algorithm for an insulin pump, for which
linear Model Predictive Control will be used.
MPC is a controller build on a model for the specic case.
The controller
calculates the optimal quantity of insulin based on measurements of the blood

sugar on the subcutaneous layer. In theory, MPC could be used to make a device, which could act as an articial pancreas, reducing the impact the disease
has on the patients life.
Figure 1.2: The insulin pump
Chapter
2
Linearization
This chapter will show how to convert a model consisting of rst order dierential equations, into a linear model with discrete time representation, and thereby
making it possible to use this model with linear MPC.
Such a system of dierential equations in continuous time is denoted by;
dx(t)
dt
= x(t)
= f (x(t), u(t), d(t))
(2.1)
The rst step is to identify an equilibrium point, a steady-state,

i.e.
f (xs , us , ds ) = 0.
(xs , us , ds ),
Making a Taylor expansion around this point yields an
approximation to the system (2.1);
x(t)
= f (x(t), u(t), d(t))

f
f
(x(t) xs ) +
(u(t) us )
' f (xs , us , ds ) +
x (xs ,us ,ds )
u (xs ,us ,ds )

f
+
(d(t) ds )
d (xs ,us ,ds )
(u(t) us ) + E
(d(t) ds ) ,
= A (x(t) xs ) + B
(2.2)
where;
,
A=
x (xs ,us ,ds )
x(t)
B=
,
u (xs ,us ,ds )
is transcribed so it containes
x(t)
dx(t)
dt
d (x(t)
(2.3)
xs :
xs )
dt
E=
,
d (xs ,us ,ds )
dx(t)
dt
d xs
dt
(2.4)
Linearization
Introducing the deviation variables,
D(t) = d(t) ds ,
X(t) = x(t) xs , U (t) = u(t) us
and
gives;
xs )
(u(t) us ) + E
(d(t) ds )
= A (x(t) xs ) + B
dt
dX(t)
(t) + ED(t)
= AX(t)
+ BU
dt
(t) + ED(t)
X(t)
= AX(t)
+ BU
d (x(t)
(2.5)
This is a system of linear time invariant dierential equations.

The solution to this system is given in [7];
X(t) = eA(tt0 ) X(t0 ) +
(s) + ED(s)
ds
eA(ts) BU
(2.6)
t0
Where
denotes the matrix exponential function.
2.1 Continuous-Discrete Time Conversion

Having the linearized system, the next step is to convert it from continuous
time to discrete time. Considering equation (2.6), and let
xk , uk
and
dk
still be
deviation variables;
xk+1 = e
k+1 tk )
A(t
tk+1
x(tk ) +
+ Ed(s)
eA(tk+1 s) Bu(s)
ds
tk
eATs x(tk ) +
Ts
eA(tk+1 s) Bu(s)
+ Ed(s)
ds
Ts = tk+1 tk
eATs x(tk ) +
Ts
eA Bu(
) + Ed(
)
= tk+1 s
eATs xk +
Ts

eA B
d uk +
Ts

eA E
d dk
A xk + B uk + E dk
where the matrices
A = eATs ,
A, B
and
Z
B=
0
Ts
(2.7)
are;
d,
eA B
Z
E=
0
Ts
d
eA E
(2.8)
2.1 Continuous-Discrete Time Conversion
Ts
is the sampling time.
A practical way to nd

A
0
B
I
A, B
and
E,
E
= eM Ts ,
I
M=
is to solve the equation (see [1]);
A B
0 0
E
0

(2.9)
tk be the time dened as tk = t0 + kTs and let x(tk ) = xk , u(tk ) = uk

d(tk ) = dk , for t tk < tk+1 , be the states at time tk .
Let
and
Then the evolution of the states is governed by the dierence equation, as above;
x(tk+1 ) = A x(tk ) + B u(tk ) + E d(tk ) xk+1 = A xk + B uk + E dk

(2.10)
Over a time-interval, from 1 to
k,
the values of
are;
x1 = A xs + B us + E ds
x2 = A x1 + B u1 + E d1
= A (A xs + B us + E ds ) + B u1 + E d1
= A2 xs + ABus + AEds + B u1 + E d1
x3 = A x2 + B u2 + E d2

= A A2 xs + ABus + AEds + B u1 + E d1 + B u2 + E d2
= A3 xs + A2 Bus + A2 Eds + ABu1 + AEd1 + Bu2 + Ed2
.
.
.
xk = Ak xs + (Ak1 B)us + (Ak1 E)ds + + Buk1 + Edk1

= Ak x s +
k1
X
k1
X

Ak1j B uj +
Ak1j E dj
j=0
(2.11)
j=0
By dening the output vector
as the measurable states, the output in
k -time
is;
zk = C xk
= C A k xs +
k1
X
k1
X

C Ak1j B uj +
Ak1j E dj
j=0
= C A k xs +
k1
X
j=0
Hkj uj +
j=0
Hk
is the
k 'th

Hk =
Hkj,d dj ,
(2.12)
j=0
Markov parameter.
0,
CAk1 B,
k1
X
k=0
,
k1
Hk
and
Hk,d
are denoted below.

Hk,d =
0,
k=0
CAk1 E, k 1
(2.13)
Linearization
These parameters will be of use in later sections.

The system can be set up as a linear discrete-time state-space model of the
form;
Linearized discrete time state-space model
xk+1 = A xk + B uk + E dk
zk = C xk
(2.14)
2.2 Summary
2.2 Summary
In this chapter it has been shown how to convert a system of rst order dierential equations in continuous time, into a linear model in discrete time. This
process is summarized below.
Linearization of model,
- Given a system of dierential equations,
x(t)
= f (x(t), u(t), d(t)),
(xs , us , ds ).

= f
, B
u
indentify a steady-state point

- Calculate

= f
E
d
A =
f
x
(xs ,us ,ds )
(xs ,us ,ds )
and
(xs ,us ,ds )
- The corresponding linear model is
(t) + ED(t)
X(t)
= AX(t)
+ BU
,
with X, U and D as deviation variables.
Converting to discrete time,

- Calculate
A = eATs ,
B=
- Model in discrete time is
R Ts
0
d,
eA B
and
E=
xk+1 = A xk + B uk + E dk .,
R Ts
0
d .
eA E
where
x, u and
are deviation variables.
- The output is
zk = C xk ,
with
indicating the measureable states.
Linearization
Chapter

Linear Model Predictive Control will now be introduced.
Figur 3.1 illustrates the basic idea for the MPC.
Figure 3.1: The basic MPC

Where
is the actual output and
manipulates the input,

as possible.
u,
is the measured output.
The controller
to achieve an output as close to the setpoint,
r,
This basic control model will be expanded step by step in this
chapter, for instance by adding disturbance and dierent sorts of constraints. It

will also be shown how these extended control problems can be formulated as a
Quadratic Programming (QP) problem, since QP problems are easily solved by
known algorithms.
3.1 Unconstrained MPC

In this section focus will be on the basic control model, the unconstrained MPC.
The goal of the controller is, as mentioned, to make the dierence between the
output,
zk ,
and the reference,
rk ,
as small as possible.
This can be done by
using a least squares problem. The weighted 2-norm is used:
1X
||zk rk ||2Qy
2
k=0
(3.1)
10
Since
z0
1
2
2 ||z0 r0 ||Qz is discarded. This gives the
can't be inuenced, the term
rst control problem;
(3.2)
Basic control problem
min z =
s.t.
1
2
N
X
||zk rk ||2Qz
k=1
k = 0, 1, . . . , N 1
k = 0, 1, . . . , N
xk+1 = Axk + Buk ,

zk = Cxk ,
From (2.12) it's known that;
zk = CAk x0 +
k1
X
Hkj uj +
k1
X
Hkj,d dj
j=0
j=0
= zx0 + zuj + zdj

For now the term
zdj
(3.3)
is discarded, such that;
zk = zx0 + zuj = CAk x0 +
k1
X
Hkj uj
(3.4)
j=0
This can be written as;
z0 = Cx0

CA
H1
z1
H
z CA2
2

z3 = CA3 x0 + H3
.
. .
.
. .
.
. .
CAN
HN
zN
|
| {z } | {z }
Z
(3.5)
0
H1
H2
0
0
H1
.
.
.
.
.
.
..
HN 1
HN 2
{z
Z = x0 + U
R
By introducing
r1
r2

R= .
..
rN
0
0
0
u0
u
1
u2
.
.
.
.
.
.
uN 1
H1
} | {z }
(3.6)
(3.7)
as a vector containing the setpoints
(3.8)
11
the objective function can be written as;
z =
1X
1
2
||zk rk ||2Qz = ||Z R||Qz
2
2
(3.9)
k=1
Where the weight matrix
Qz
Qz
0
Qz = .
..
0
Qz
.
.
.
..
is given by;
0
0
.
.
.
(3.10)
Qz
Plugging equation (3.7) into (3.9) gives;
1
2
||Z R||Qz
2
1
2
= ||x0 + U R||Qz
2
1
2
= ||U (R x0 )||Qz
2
1
2
= ||U b||Qz ,
b = R x0
2
z =
(3.11)
(3.12)
(3.13)
(3.14)
To make this problem easier to solve, it is convenient to express it as a QP

problem;
1
2
||U b||Qz
2
1
= (U b)T Qz (U b),
by denition
2
1
1
= U T T Qz U (T Qz b)T U + bT Qz b
2
2
1
= U T Hz U + gzT U + z
2
z =
where
Hz , gz
and
of weighted norm
(3.15)
are given by;
Hz = T Qz
(3.16)
gz = T Qz b
= T Qz (R x0 )
= T Qz x0 T Qz R
= Mx 0 x 0 + MR R
1
z = bT Qz b
2
(3.17)
(3.18)
12
This is the QP problem equivalent to problem (3.2);
(3.19)
QP formulation of problem (3.2)
min z =
U
1 T
U Hz U + gzT U
2
Hz = T Qz
gz = Mx0 x0 + MR R
is discarded since it doesn't inuence the solution to the problem. Note that
the gradient is dynamic and needs to be updated for every timestep, as opposed
to the Hessian, which is static.
3.1.1
Regularization
Regularization is done by introducing a new term,

where
u , in the objective function,
uk = uk uk1 ;
(3.20)
Control problem, with regularization
min = z + u =
s.t.
1
2
N
X
||zk rk ||2Qz +
k=1
xk+1 = Axk + Buk ,

zk = Cxk ,
1
2
N
1
X
||uk ||2S
k=0
k = 0, 1, . . . , N 1
k = 0, 1, . . . , N
This new term minimizes the dierence between two consecutive steps in
which gives more smooth input, ie. tries to ensure that steps in
u,
are either
continously decreasing or continously increasing.

Again this should be formulated as a QP problem. Compared to problem (3.2),
the only dierence is a new term in the objective function, so to formulate as a
13
QP problem, this new term has to be rewritten;
u =
N 1
1 X
||uk ||2S
2
k=0
1
2
1
2
N
1
X
||uk uk1 ||2S
k=0
N
1
X
(uk uk1 )T S(uk uk1 )
k=0
u0
u1
u2
2 .
..
uN 1
2S
S
..
..

S
0

T
0
+
. u1
.
.
0
| {z }
..
S
2S
2S
S
{z
HS
u0
u1
u2
u0
u1
u2
.
.
S .
S
uN 1
}
. + 2 u1 Su1
.
.
uN 1
Mu1
This
1
1
T
= U T HS U + (Mu1 u1 ) U + u1 Su1
2
2
shows, that introducing u extends the QP problem
(3.21)
by the following
terms;
Hu = HS
(3.22)
gu = Mu1 u1
(3.23)
1
2 u1 Su1 is discarded, because of the lack of inuence
on the solution to the problem. The new QP problem is;
Like with
z ,
the term
min =
U
1 T
U HU + g T U
2
H = Hz + Hu = T Qz + HS
g = gz + gu = Mx0 x0 + MR R + Mu1 u1
(3.24)
14
3.1.2
Disturbance
The MPC problem will now be extended by adding disturbance.
Virtually
every system has some disturbance. Figure 3.2 illustrates how the disturbance
inuences the MPC problem.
Figure 3.2: MPC with disturbance
The disturbance in this model is relatively easily handled by including the last
term,
zdj ,
from equation (3.3), in the problem;
(3.25)
Control problem with regularization and disturbance
min = z + u =
N
N 1
1X
1 X
||zk rk ||2Qz +
||uk ||2S
2
2
k=1
s.t.
k=0
xk+1 = Axk + Buk + Edk ,

zk = Cxk ,
As mentioned in Chapter 2.1,

analogous to
zuj .
Hi
k = 0, 1, . . . , N 1
k = 0, 1, . . . , N
is analogous to
Hi,d ,
so
zdj
must then be
Therefore it's easy to see that including the term
new term, analogous to
U ,
Z = x0 + U + d D
in
Z.
This new term is named
zdj
yields a
d D ;
(3.26)
15
Where;
H1,d
H
2,d
H
d =
.3,d
.
.
HN,d
0
H1,d
H2,d
0
0
H1,d
0
0
0
.
.
.
.
.
.
..
.
.
.
HN 1,d
HN 2,d
The introduction of this new term in
d0
d1
D= .
..
dN 1
H1,d
means, that
(3.27)
has to be rewritten as a
QP;
z =
1X
||zk rk ||2Qz
2
k=1
=
=
=
=
=
=
=
1
||Z R||2Qz
2
1
||U (R x0 d D)||2Qz ,
by using (3.26)
2
1
||U b||2Qz ,
b = R x0 d D
2
1 T T
(U bT )Qz (U b)
2

1
U T T Qz U bT Qz U U T T Qz b + bT Qz b
2
1 T T
1
U Qz U (T Qz b)T U + bT Qz b
2
2
1 T
U Hz U + gzT U + z
(3.28)
2
Where;
Hz = T Qz
(3.29)
gz = Qz b
= T Qz (R x0 d D)
= Mx 0 x 0 + MR R + MD D
1
z = bT Qz b
2
(3.30)
(3.31)
16
The QP problem is summarized below. Compared to (3.24), the problem has

been expanded by a new term in the gradient.
min =
U
1 T
U HU + g T U
2
H = T Qz + HS
g = Mx0 x0 + MR R + MD D + Mu1 u1
(3.32)
3.1.3
17
Feed-forward/feedback
Two important aspects in a MPC controller is feedforward and feedback. They

both react on changes in the process, and can therefore make the controller
better in terms of reaction speed and robustness.
The feedback approach is illustrated on Figure 3.3.
Figure 3.3: MPC with feedback

Feedback is often used to control the dynamic behavior of the system. When
providing a system with feedback, the measured output,
y,
and thereby the po-
tential occurring disturbance, is fed back to the controller.

This ensures that the measured output is approximately the same as the wanted
value, the reference,
r,
which means the system will be more robust. The draw-
back with feedback is slow reaction speed, since output has to be measured
before it can be used.
The feedforward approach is illustrated on Figure 3.4.
Feedforward is a loop
in the system, which takes some known disturbance and forward it to the controller.
Figure 3.4: MPC with feedforward

When a system exhibits feedforward behavior, it responds to disturbances which
are predened and therefore known.
This means, that the system can respond more quickly to the disturbance and
the measured output will be more robust.
18
But meanwhile there is a relatively big chance that the output isn't comparable
to the wanted values, the references,
r,
since it cannot do much about novel
disturbance.
When combining these two appoaches, the system will be aware of both the
known and the unmeasured disturbance, so the system in theory will be trustworthy and fast. Figur 3.5 illustrates a controller where both these approaches
are used.
Figure 3.5: The MPC-model with feed-forward/feedback
3.2 Constrained MPC
19
3.2 Constrained MPC

The MPC will be extended such that it contains constraints.
The MPC problem has to take the limits of the physical system into consideration. It's infrequent that a system has no boundaries. Constraints are imposed
on the input quantity, the input rate of movement and on the output. For all
constraints it's assumed that, for every timestep
3.2.1
k , the boundaries are the same.
Input constraints
The input constraints are limitations to the maximum and minimum input volume.
(3.33)
MPC with input constraint
min =
N
N 1
1X
1 X
||zk rk ||2Qz +
||uk ||2S
2
2
k=1
s.t.
k=0

zk = Cxk ,
k = 0, 1, . . . , N 1
k = 0, 1, . . . , N
umin uk umax
k = 0, 1, . . . , N 1
To formulate this as a QP problem this new constraint is put in vector form;
umin uk umax

umin
u0
umin u1

. .
.. ..
umin
| {z }
uN 1
min
U
umax
umax

.
..
(3.34)
umax
| {z }
max
U
This yields a constrained QP problem;

(3.35)
1 T
U HU + g T U
2
min U U
max
U
min =
U
s.t.
The Hessian and gradient for this problem are the same as in problem (3.32).
20
3.2.2
Constraints on input rate of movement
It is also possible to have constraints on how fast the input constraints can
change.
The input rate of movement is the change from
therefore called
uk .
to
k+1
and is
The control problem is extended again;
(3.36)
MPC with constraints on input and input rate
min =
s.t.
1
2
N
X
||zk rk ||2Qz +
k=1
1
2
N
1
X
||uk ||2S
k=0

zk = Cxk ,
k = 0, 1, . . . , N 1
k = 0, 1, . . . , N
umin uk umax ,
umin uk umax ,
k = 0, 1, . . . , N 1
k = 0, 1, . . . , N 1
This new constraint for
uk
can be written as;
umin uk umax
umin
u0 u1
umax
umin u1 u0 umax
.
.
.
.
.
..
.
.
umin
uN uN 1

umin + u1
I
umin I I
.
..
.

.
.
umax
..
u0
umax + u1
u1 umax
.
.
.
.
umin
Since the rst row contains
u1
umin
{z }
Umin
I
|
{z
umax
this can be written as;
umin + u1 u0 umax + u1

I I
umin
umin
I I

.
..
..
..
.
.
|
uN 1
u1
u2
.
..
(3.37)
umax
umax

.
.
.
uN 1
I
}
umax
{z }
Umax
(3.38)
3.2 Constrained MPC
An input constraint for
21
u0
is already given in (3.34), combining with (3.37)
gives;
u0 umax + u1
u0 umax
u0 umin + u1
u0 umin
u0 min(umax , umax + u1 )
u0 max(umin , umin + u1 )
(3.39)
So the input constraints becomes;

u0
max(umin , umin + u1 )
u1
u
min
.
.
.
..
.
uN 1
umin
{z
min(umax , umax + u1 )

umax

.
.

Umin
umax
{z
Umax
}
(3.40)
The QP problem becomes;
min =
U
s.t.
(3.41)
1 T
U HU + g T U
2
Umin U Umax
Umin U Umax
Note that
timestep.
Umin
and
Umax
are dynamic and needs to be re-calculated for every
22
3.2.3
Output constraints
Output constraints are analogous to the input constraints, i.e.
limitations to
the maximum and minimum output.
MPC with output and input constraints
min =
s.t.
1
2
N
X
||zk rk ||2Qz +
k=0
1
2
N
1
X
(3.42)
||uk ||2S
k=0

zk = Cxk ,
umin uk umax ,
umin uk umax ,
zmin zk zmax ,
k
k
k
k
k
= 0, 1, . . . , N 1
= 0, 1, . . . , N
= 0, 1, . . . , N 1
= 0, 1, . . . , N 1
= 1, 2, . . . , N
The output at k = 0 cannot be aected, so the constraint here is disregarded.

This leads to:
zmin zk zmax
zmin
zmax
z1
zmin z2 zmax
. . .
.. .. ..
zmin
| {z }
Zmin
Using
zN
| {z }
Z
(3.43)
zmax
| {z }
Zmax
Z = x0 + U + d D
from equation (3.26) yields;
Zmin Z Zmax
Zmin x0 + U + d D Zmax
Zmin x0 d D U Zmax x0 d D
|
{z
}
|
{z
}
min
Z
max
Z
(3.44)
3.2 Constrained MPC
23
The problem transforms into;
(3.45)
min =
U
s.t.
3.2.4
1 T
U HU + g T U
2
Umin U Umax
Umin U Umax
Zmin U Zmax
Soft output constraints
It can occour that the control problem will be infeasible. This is mainly due to
the use of constraints greatly complicating the problem. In some cases it can
simply be impossible to stay within the boundaries.
A solution to this problem is to soften the constraints, meaning the boundaries
can be violated occasionally, if needed. In this case, soft output constraints are
used.
The easiest way to softening output constraints is to introduce a new
slack variable,
(3.46)
MPC with input and soft output constraints
min = z + + u
=
k=1
s.t.
N 1
1 X
1
+
||zk rk ||2Qz + ||k ||2S + sT
||uk ||2S
k
2
2
2
N
X
1

zk = Cxk ,
umin uk umax ,
umin uk umax ,
< zk k zmax ,
zmin zk + k < ,
0 k < ,
k=0
k
k
k
k
k
k
k
= 0, 1, . . . , N 1
= 0, 1, . . . , N
= 0, 1, . . . , N 1
= 0, 1, . . . , N 1
= 1, 2, . . . , N
= 1, 2, . . . , N
= 1, 2, . . . , N
If it's not neccesary to violate output constraints, the solution to the problem
24
will simply have
= 0,
and will be equivalent to the solution with hard output
constraints. If it's not possible to use hard constraints,
is selected as small as
possible, so the output constraints are only violated by a minimum.

The approach to formulate (3.46) as a QP problem is a little dierent than
for the previous problems. First
is introduced;
1
2

= .
..
(3.47)
N
The new term in the objective function,
is formulated as QP;
1 T
1
||k ||2S + sT
T
k = S + s
2
2
(3.48)
where;

s
..
s = .
S =
..
(3.49)
S
For this problem the Hessian is
and the gradient is
should be formulated in terms of the vector
,
U
s .
The QP problem

U
U=
(3.50)
Which means the objective function for the QP should be
1 T
2 U HU
.
+ gT U
The
Hessian and gradient for this problem are;

= H
H
0

0
,
S

g =
g
s
U and should

Umin
U
Umax
| {z } |{z} | {z }
The constraints for
min
U
(3.51)
then be written as;
(3.52)
max
U
It's obvious that as in the previous QP problem, (3.45), the constraint
Umin U Umax
should still be imposed.
Formulated in terms of
yields;
Umin

U
Umax
0
(3.53)
3.2 Constrained MPC
The upper bound on
zk
25
needs to be transcribed;
< zk k zmax
< Z Zmax
< x0 + D D + U Zmax ,
using (3.26)
< U Zmax ( x0 + D D)
< U Zmax

U
< I
Zmax
(3.54)
Equivalently, this can be done for the lower bound, which gives;

U
Zmin I
<
(3.55)
Combining equations (3.53), (3.54), and (3.55) in matrix formulation yields;
0
Umin
Umax
I U Zmax
I
Zmin
| {z } | {z }
| {z }
bmin
(3.56)
bmax
So the QP problem is;
1 T
U H U + gT U
2
min U
U
max
U
bmin AU bmax
min =
U
s.t.
(3.57)
26
3.3 Kalman lter

The basic ideas behind the Kalman lter will now be introduced. Kalman Filter
is used to minimize the impact of noise in the problem. For further and more
detailed information, see [8] and [9].
To make the simulations more realistic, noice is assumed on both process and
output. This noise is denoted
Both
wk
and
vk ,
respectively;
xk+1 = Axk + Buk + wk
(3.58)
yk = Cxk + vk
(3.59)
wk
and
vk
are assumed to be white noise and normal distributed, with
mean zero and covariance
Qw
and
Rv ,
respectively.
wk N (0, Qw )
(3.60)
vk N (0, Rv )
(3.61)
An estimate of
is given by:
x
k+1|k = E {xk+1 |yk , yk1 , . . . , y0 }
= E {xk+1 |Y}
= E {Axk + Buk + wk |Y}
= A E {xk |Y} + Buk + E {wk |Y}
= A
xk|k + Buk + w
k|k
Since the mean of
wk
(3.62)
is equal to 0, the noise estimate becomes:
x
k+1|k = A
xk|k + Buk
This means that
Pk|k ,
xk|k
is normal distibuted with mean
(3.63)
x
k|k
and error covariance
i.e:
xk|k N x
k|k , Pk|k
The system can be augmented by adding an integrated disturbance,
(3.64)
k ,
to
achieve oset-free performance;
xk+1 = Axk + B (uk + k ) + wk
(3.65)
k+1 = Ik +
(3.66)
yk = Cxk + vk
(3.67)
Transformed into matrix form gives;

xk+1
A
=
k+1
0

y= C

xk
B
w
+
uk + k
k
0

xk
+ vk
0
k
B
I
(3.68)
(3.69)
3.3 Kalman lter
where
27
is normal distributed with zero mean and covariance
disturbance noise and normal distributed with
Q,k
and
is
N (0, Q ).
Now the Kalman Filter is used to estimate the state
x
k+1|k .
First phase of the Kalman lter is the time update. It is used to produce an
estimate of the current state, using the estimate of the previous state;
x
k|k1 = A
xk1|k1 + Buk
(3.70)
Pk|k1 = APk1|k1 A + Qw
(3.71)
The second phase of the lter, the measurement update, renes measurement
information from the current timestep to get a new, more accurate, estimate;
ek = yk C x
k|k1
(3.72)
Sk = CPk|k1 C + Rv
Kk Sk = Pk|k1 C T ,
(3.73)
(optimal Kalman gain, solve for
Kk )
(3.74)
x
k|k = x
k|k1 + Kk ek
(3.75)
Pk|k = (I Kk C) Pk|k1
(3.76)
An illustration of the Kalman Filter is shown on gure 3.6.
Figure 3.6: The Kalman Filter cycle
28
3.4 Summary
This chapter derived the MPC problem (3.46) shown on Figure 3.7.
Figure 3.7: The MPC-model

The goal of the MPC controller is to make the output,
ence,
r,
time-step.
output,
bances,
z,
as close to the refer-
as possible. This is done by calculating the optimal input,
y.
d.
u,
for each
Feedback is used, such that the controller analyzes the measured

When feedforward is used, the controller analyzes the future disturFeedback and feedforward can well be used simultaneously.
The described MPC algorithm also takes care of input constraints, constraints
on input rate of movement, and soft output constraints. The soft output constraints keeps us from ending up with an infeasible problem, but can also violate
the physical limitations for the output, which may cause the system to malfunction and should therefore be used with concern.
Also the use of the Kalman Filter is described. The lter makes an estimate of
x,
when there is an addition of noise on the output measurements and/or the
process.
Formulating the control problem as a QP problem makes it relatively easy solvable.
Chapter
4
Implementation
This chapter will describe the implementation of the MPC problem (3.46) in
ATLAB.
The process tree is seen on Figure 4.1.
Figure 4.1: MPC tree
The individual
typewriter font
ATLAB les will now be described in detail.
nominates
ATLAB entries.
Names in
30
Implementation
4.1 ScenaX.m
SceneX.m
is the le which is called by the user.
The
species the name of
the scene. The idea is that, all parameters the user should have inuence on,
can be adjusted in this le. Table 4.1 shows the basic variables which this le
initializes.
Parameters(s)
N
umin , umax
umin , umax
zmin , zmax
Q, S , S , s
ATLAB name(s)
umin, umax
udmin, udmax
zmin, zmax
Q, S, Spsi, spsi
Describtion
Size of prediction horizon.
Bounds on
u.
Bound on rate of change in

Bounds on
u.
z.
Weight matrices.
Table 4.1: Basic MPC variables initialized in
SceneX.m
Additionally, this le introduces the following:
The timevector
t,
the sampling time
A vector
which species timesteps in terms of the starttime
Ts,
and the ending time
tf.
t0,
mealsv, which denes at which points in time a meal is consumed.
sigmaw and sigmav, which is used to scale the simulated noise, and the
seed parameter, which denes what set of random numbers to be used.
feed
determines if feedforward should be used. This will be described in
details later.
4.2 MPCControl.m
This is a wrapper script, which simply calls the three scripts needed for the
dierent phases of MPC: Design, simulation and evaluation.
4.2.1
MPCDesign.m
In this design le, everything is prepared for the actual simulation. This mainly
consists of setting up the many matrices found in Chapter 3.
4.2
MPCControl.m
First the function
31
ParameterDesign.m
is called, which assigns the basic model
parameters, shown in Table 5.1, to a vector
p, to make handling of these easier.
Several new variables are initialized;
Parameters(s)
xs , us , ds , zs
x0 , u0
R, D
x, u, y , U
A, B , C , D
ATLAB name(s)
big
xs, us, ds, zs
x0, u0
R, D
x, u, y, ubar
Ac, Bc, Cc, Ec
Describtion
Large number which acts as innity.
Steady-state point for the system.
Starting points for the simulation.
Vectors of lenght N.
Initialized to zero.
System matrices in continous time.
Table 4.2: Basic MPC variables initialized in MPCControl.m
MPCDesign
also performs the following operations;
The constraints for

The vectors
and
u
v
and
are transformed into deviation constraints.
are generated by the
randn command and scaled, so

seed option is used so it is
they have the desired standard deviation. The

possible to reproduce the same noise vector.
DesignConstraints.m assigns the constraints for u, u, z ,

T
vectors, i.e. Umin = umin
umin umin , etc.
and
to
See equations (3.38), (3.40), (3.43), and (3.52).
DesignDiscreteMatrices.m
converts the system matrices
into the equivalent matrices in discrete time,
A, B , C , E .
, C , E
A, B
This is done by
equation (2.9).
DesignKalman.m.
The Kalman Filter matrices is designed in
See section
3.3.
Finally
xp
These are all created by basic
, , d , , Mx0 , Mu1 , MR ,
H
DesignMPCMatrices.m designs the matrices
MD , , s .
ATLAB matrix operations.
is formed, which is the augmented system with integrated dis-
turbance, corresponding to equation (3.68).
4.2.2
MPCSimulate.m
Now the actual simulation is started. This consists of a loop which runs over the
timesteps dened in
t.
First noise is simulated on model
and measurement
y,
32
Implementation
by adding values of vectors
and
v.
Meals are simulated by adjusting
Dm (t),
and are assumed to span only one timestep, that is from tk to tk+1 . For instance,
a meal at
t = 420
Dm (420),
see Figure 4.2.
D(t) [mg/dL/min]
min (7 hours) is simulated by adding the size of the meal to
6.5
7.5
t [hour]
8.5
Figure 4.2: Modelling meals, change in

On Figure 4.3 is shown how the implementation treats
Dm
when feedforward is
used and when it isn't. See Section 3.1.3 for further details of these.
4.2
MPCControl.m
33
(a) Feedforward in use,
D is updated with meals
(b) Feedforward not in use,

times
D is kept to all-zeros at all
Figure 4.3: Implementation of feedforward
Now comes the actual control step, where the optimal

timestep, this is done in
First
y , u1 , R
MPCCompute.m.
and
D are transformed
x is updated with
current prediction for
is calculated for this
Below is the outline of
MPCCompute.m;
into deviation variables and the

the Kalman gain
Kfx.
The gradient
Matrix system (3.56) is formed.
Starting guess for
ATLAB's quadprog
qpsolver.m, where qpsolver.m has shown
itself to greatly outperform quadprog.
is used as uk . If the solver fails, uk1 is used as uk .
The rst element of U
is updated.
is made by forming
U
init = [uk uk+1 uk+N 1 uk+N 1
U
T
k k+1 k+N 1 k+N 1 ] .
Then
is found by solving the QP problem (3.57) by
or John Bagterp Jrgensens
xp
is updated.
34
Implementation
The found
is returned as a physical variable (as opposed to deviation
variable).
The program now returns to
MPCSimulate.m, where xk+1
the Bergman model using the
ode15s.
4.2.3
is found by evaluating
ATLAB Ordinary Dierential Equation solver
MPCPlot
After the simulation-loop is nished the results are plotted in convenient ways.
The results include the state vector
disturbances
Dm .
x,
the calculated input
and the meal
Plots are made with x-axis in hours, and constraints are
shown where applicable.
Chapter
Case study - A minimal

model
In this case study, the Bergman minimal model will be used with the implemented MPC.
Based on measurements of the glucose level in the subcutaneous layer, the controller will calculate the optimal amount of insulin to inject into the patient.
To model the patient, a modied version of Bergmans minimal model is used.
As the name implies, this model is small, and is mainly used to test the MPC.
It would be unrealistic to think of it as an accurate model of a real patient.
The model consists of ve dierential equations, (see [2] for further details):
dG
= P1 (G + Gb ) Xr G + Dm (t)
dt
dXr
dt
dI
dt
= P2 Xr + P3 (I Ib )
U (t)
VI
G Gsc
=
Rutln
5
= nI +
dGsc
dt
dDm
dt
= Dm (t)
(5.1)
(5.2)
(5.3)
(5.4)
(5.5)
36
A very brief description of the ve states;
(mg/dL): Blood plasma glucose concentration above basal value.
Xr
I
(mU/L): Insulin in the remote compartment.
(mU/L): Plasma insulin concentration above basal value.
Gsc (mg/dL): Glucose concentration on the subcutaneous layer.

approximates G, and is the one which are measurable.
Dm
This state
(mg/dL/min): Meal glucose disturbance.
and the input is the manipulated insulin infusion rate (U (t), mU/min).
time variable
The
is measured in minutes. The standard parameters for the model
can be found in Table 5.1.
Gb , Xbr , Ib , Gbsc
and
Dm
denote the basal values for
the system.
The parameters are assumed to be optimal, so they will not be inspected in this
project.
37
Name
Value
P1
P2
P3
n
VI
Rutln
Gb
Xbr
Ib
Gbsc
Dbm
0.028735 min1
0.028355 min1
5.035 105 mU/L
5/54 min
12 L
0.7400mg/dL/min
0.05
81.3 mg/dL
0
15mU/L
Gb 5 Rutln
0
Table 5.1: Bergman model parameter values
Boundaries on the blood sugar level is needed to avoid the person going into
hyperglycemia or hypoglycemia. These limits corresponds to output constraints.
The input constraints restricts how much, and how fast, the insulin can be
injected, and ensures the system obeys physiological and physical limits.
The chosen constraints are shown below, respectively;
mg/dL
z 180
mg/dL
(5.6)
mU/min
u 100
mU/min
(5.7)
16.7
mU/min
u 16.7
60
mU/min
(5.8)
Neither these will be investigated during this project.

To use the equations in MPC, let
and
be the inputs
U (t)
and
D(t),
and let
be the system of equations (5.1-5.5);
G
X
r
x= I
Gsc
D
G
X
r
x = I

Gsc
Dm
(5.9)
A steady-state point for the system is given by;
Gb
X
br
xs = Ib ,
Gbsc
Dbm
us = n Ib VI ,
ds = 0
(5.10)
38
This model will now be used with MPC.

First step is to linearize the model.
5.1 Linearization
By using the procedure shown in Chapter 2, it's clear that this system of dierential equations can be set up like this;
+ BU
+ ED
X = AX
X, U
D = d ds .
where
In this case,
and
and
(5.11)
are deviation variables,
is the state vector,
X = x xs , U = u us
and
is the input variable for insulin injection,
is the input variable for meal consumption. The matrices
A, B
and
are the partial derivatives of the model:
A=
x
,
(xs ,us ,ds )
B=
u
,
(xs ,us ,ds )
E=
d
(5.12)
(xs ,us ,ds )
which for this system gives;
P1 Xb Gb 0
0
P2 P3
A =
0
0
n
0.2
0
0
0
0
0
h
iT
= 0 0 1 0 0
B
VI

= 1 0 0 0 T
E
0
0
0
0.2
0
1
0
(5.13)
(5.14)
(5.15)
Converting to discrete time yields (referring to Section 2.1);
Xk+1 = A Xk + BUk + EDk
(5.16)
By using parameter values from Table 5.1, the matrices
A, B
and
can be
evaluated by equation (2.9). These matrices are not evaluated here since they
depend on the used sampling time
Ts .
The model is now linearized and converted to discrete time, and ready to be
used with the MPC controller.
5.2 MPC
39
5.2 MPC
The goal for the controller is to keep the glucose level, for the diabetic, at a
healthy level at all times.
The sensor in the insulin pump is able to measure the glucose on the sub-
Gsc . Since
state which is
this is the only

C = 0 0 0 1 0 . The setpoint for
value, Gbsc .
cutaneous layer, that is, the state

measureable, the vector
is given by
this state is selected as it's basal
The constraints used are giving in equations (5.7-5.8), and it should be noted
that the output constraints are implemented as soft constraints.
5.2.1
Weight matrices
There are four weigths used in the objective function:
Qz , S , s
and
S.
These
can be chosen to tune the controller.
and
are weights for the slack variable
k ,
which are property of the soft
output constraints, and are kept constant as identity matrices.
Qz
and
matrices are weights for the output and regularization term, respec-
tively. Tuning can be done by keeping one of these constant while varying the
other, since it's the ratio of these that actually matters in the objective function.
Qz
is selected as identity matrix, and S is varied, see Figures 5.1 and 5.2. The
patient is given a meal at
t = 1 hour.
S , there is
For the very low values for

that
close to no regularization, which means
varies greatly between timesteps. If the value of
controller responds too slow.
is selceted to
102 ,
is set too high, the
since this values gives some
regularization and is still low enough to provide a decent response from the
controller in regards of speed.
It would also be possible to tune the Kalman Filter, in this thesis the weights
are selected as;
Qw = 102
and
Q = 102 .
The seleted weights are summarized below;
Qz = 1,
S = 102 ,
S = 1,
s = 1,
Qw se = 102 ,
Q = 102
(5.17)
40
4000
u(t) [mU/min]
u(t) [mU/min]
5000
5000
10000
1.2
1.4
1.6
1.8
2000
0
2000
4000
1.2
1.4
t [hour]
90
80
100
90
80
70
1.2
1.4
1.6
1.8
1.2
1.4
t [hour]
(a)
S = 105
(b)
u(t) [mU/min]
u(t) [mU/min]
1.8
1.8
S = 104
1000
1000
0
1000
1
1.2
1.4
1.6
1.8
500
0
500
1.2
1.4
t [hour]
z(t)
Constrains
z(t) [mg/dL]
80
1.2
1.4
1.6
1.8
z(t)
Constrains
120
100
100
80
1.2
1.4
t [hour]
(c)
1.6
t [hour]
120
z(t) [mg/dL]
1.6
t [hour]
2000
1.6
1.8
1.8
t [hour]
S = 103
(d)
400
S = 102
200
u(t) [mU/min]
u(t) [mU/min]
z(t)
Constrains
110
70
200
0
150
100
50
0
50
200
1
1.2
1.4
1.6
1.8
1.2
1.4
t [hour]
1.6
t [hour]
z(t)
Constrains
z(t)
Constrains
160
z(t) [mg/dL]
140
z(t) [mg/dL]
1.8
120
z(t)
Constrains
100
z(t) [mg/dL]
z(t) [mg/dL]
110
2000
1.6
t [hour]
120
100
80
140
120
100
80
60
60
1
1.2
1.4
1.6
1.8
1.2
1.4
t [hour]
(e)
S = 101
1.6
t [hour]
(f)
S=1
Figure 5.1: Weight matrix S varied, without feedforward
1.8
5.2 MPC
41
3000
2000
u(t) [mU/min]
u(t) [mU/min]
5000
1000
0
1000
5000
0.5
1.5
t [hour]
2.5
2000
0.5
80
1.5
t [hour]
2.5
80
0.5
400
500
0
1.5
t [hour]
2.5
2.5
S = 104
200
0
80
1.5
t [hour]
(c)
1.5
t [hour]
z(t)
Constrains
120
100
400
0.5
z(t) [mg/dL]
z(t) [mg/dL]
2.5
z(t)
Constrains
120
2.5
100
80
0.5
S = 103
1.5
t [hour]
(d)
400
2.5
2.5
S = 102
200
u(t) [mU/min]
u(t) [mU/min]
200
500
200
0
1.5
t [hour]
150
100
50
0
50
0.5
2.5
z(t)
Constrains
1.5
t [hour]
z(t)
Constrains
160
z(t) [mg/dL]
140
z(t) [mg/dL]
1.5
t [hour]
(b)
600
120
100
80
60
0.5
S = 105
u(t) [mU/min]
u(t) [mU/min]
90
1000
200
0.5
z(t)
Constrains
100
1500
0.5
2.5
70
1
(a)
1000
0.5
110
90
70
0.5
1.5
t [hour]
120
z(t)
Constrains
100
z(t) [mg/dL]
z(t) [mg/dL]
110
140
120
100
80
1.5
t [hour]
(e)
S = 101
2.5
60
0.5
1.5
t [hour]
(f)
S=1
Figure 5.2: Weight matrix S varied, with feedforward
2.5
42
5.2.2
Horizon and sampling time
An appropriate horizon
N,
and a sampling time
Ts ,
are now selected. This is
done by investigating how a stepchange in the meal-disturbance,
u(t) will eect the non-linearized system.
Dm ,
and in
See Appendix B.15 for the implemen-
tation of this.
Figure
5.3
illustrates
what
is
meant
by
stepchange.
The eect of this stepchange on the output
is seen on Figure 5.4. This Figure shows that,

for the stepchange in
Dm ,
the system reaches
a new steady-state after approximately three

hours (180 minutes), which is slightly faster
than for the stepchange in
u.
A general rule-of-thumb says, that one should
4
t [hour]
use a sampling time of about 180 min/12 samples = 15 min. Since computational speed is
Figure 5.3: Stepchange
not an issue for this problem , the used sampling time is
Ts = 8
min.
The horizon should be at least three hours, so
4000
N = 25
is chosen.
80
3500
75
3000
70
z(t) [mg/dL]
z(t) [mg/dL]
2500
2000
65
60
1500
55
1000
50
500
0
3
t [hour]
(a) Stepchange in
Dm
45
4
t [hour]
(b) Stepchange in
Figure 5.4: Eect of stepchange
5.2 MPC
5.2.3
43
Noise
As described in Section 3.3, noise can be simulated on output measurement

and/or the model itself;
xk+1 = Axk + Buk + Edk + wk x

k+1 = Axk + Buk + Edk
yk = Cxk + vk yk = Cxk
where
and
are both assumed to be Gaussian distributed;
w N (0, Qw )
v N (0, Rv )
These vectors of noise are created in
ATLAB by use of the
randn
function,
which is able to create random numbers with mean zero and standard deviation
one.
To obtain the desired distribution for the noise, the noise is scaled with
periments are required to determine the size of
parameter for
randn,
and
v .
Ex-
By use of the seed
it's possible to reproduce the same random vectors of
noise, and thereby produce simulations with the same noise vectors scaled differently.
By experiments,

w = 0.7
and
103
v
0.4
are chosen to be;
0.8

0 ,
v = 0.8
(5.18)
These have shown themselfs to give simulations with a reasonable level of noise.
44
5.3 Simulations
The performance of the controller will now be tested with three dierent scenarios. In the rst test scenario our virtual patient is given three regular meals.
This is the main test scenario, and here the eect of feedforward will be investigated, as well as measurement noise and process noise. In the last two scenarios
the size of the meals are varied. Scenario II models a patient eating big meals,
and in Scenario III the patient ingests alot of meals throughout the day.
The main objective for the controller is to keep the patient within healthy blood
sugar limits at all time.
5.3.1
Scenario I
Tabel 5.2 shows at what hours the patient ingests meals;

Time (hours)
D(t)
(mg/dL/min)
12
18
10
Table 5.2: Scenario I meal disturbances
This is the basic simulation scenario.

feedforward.
This is now tested with and without
5.3 Simulations
5.3.1.1
45
Feedforward
The feedforward approach was described in Section 3.1.3.
In this section the
controller will be equiped with this approach, which simulates a controller where
the patient informs of future incoming meals.
The simulations are done without any process or measurement noise. Figures
5.5 and 5.6 illustrates the evolution of the states and the input/output, respectively, without feedforward, while Figures 5.7 and 5.8 illustrates the case with
feedforward.
200
G(t), Gsc(t) [mg/dL]
Dm(t) [mg/dL/min]
15
10
100
50
9 12 15 18 21 24
t [hour]
9 12 15 18 21 24
t [hour]
9 12 15 18 21 24
t [hour]
50
40
0.02
I(t) [mU/L]
Xr(t) [mU/L]
0.03
0.01
0
0.01
G
Gsc
150
30
20
10
9 12 15 18 21 24
t [hour]
Figure 5.5: The evolution of states without feedforward
46
u(t) [mU/min]
100
u(t)
Constraints
80
60
40
20
0
0
z(t) [mg/dL]
200
10
12
14
t [hour]
16
18
20
22
24
10
12
14
t [hour]
16
18
20
22
24
z(t)
Setpoint
150
100
50
0
Figure 5.6: Insulin input and output without feedforward
200
Dm(t) [mg/dL/min]
15
10
100
50
9 12 15 18 21 24
t [hour]
9 12 15 18 21 24
t [hour]
9 12 15 18 21 24
t [hour]
50
40
0.02
I(t) [mU/L]
Xr(t) [mU/L]
0.03
0.01
0
0.01
G
Gsc
150
30
20
10
9 12 15 18 21 24
t [hour]
Figure 5.7: The evolution of states with feedforward
5.3 Simulations
47
u(t) [mU/min]
100
u(t)
Constraints
80
60
40
20
0
0
z(t) [mg/dL]
200
10
12
14
t [hour]
16
18
20
22
24
10
12
14
t [hour]
16
18
20
22
24
z(t)
Setpoint
150
100
50
0
Figure 5.8: Insulin input and output with feedforward
Using feedforward, the controller knows when there is going to be an increase

in the glucose value, because of meal consumption. That means that before the
consumptions are made, the controller injects some insulin. The glucose level is
held on a lower value than without feedforward and therefore it is more likely
for the patient to maintain a heathy blood glucose level.
From now, on all simulations are done with feedforward.
48
5.3.1.2
Measurement noise
As mentioned in Section 5.2.3, the system can be simulated with measurement

noise and/or process noise. In this section only the measurement noise, noise
on
y,
will be used .
A way to think of measurement noise, is how well the sensor, that measure the
glucose level in the subcutaneous layer, works.
Figure 5.9 illustrates the evolution of the ve states for the model, Figure 5.10
the input and the outputs, while 5.11 illustrates the dierence between the measured output and the real output.
160
Dm(t) [mg/dL/min]
15
10
120
100
9 12 15 18 21 24
t [hour]
9 12 15 18 21 24
t [hour]
50
40
0.02
I(t) [mU/L]
Xr(t) [mU/L]
80
60
9 12 15 18 21 24
t [hour]
0.03
0.01
0
0.01
G
Gsc
140
30
20
10
9 12 15 18 21 24
t [hour]
Figure 5.9: The evolution of states with measurement noise
5.3 Simulations
49
u(t) [mU/min]
100
u(t)
Constraints
80
60
40
20
0
0
z(t) [mg/dL]
200
10
12
14
t [hour]
16
18
20
22
24
10
12
14
t [hour]
16
18
20
22
24
z(t)
Setpoint
150
100
50
0
Figure 5.10: Insulin input and output with measurement noise
150
z
y
140
130
120
110
100
90
80
70
60
10
15
20
25
t [hour]
Figure 5.11: The measured and real output with measurement noise
50
Compared to Figures 5.7 and 5.8, Figures 5.9 and 5.10 does show some oscillations, which means the control problem is harder, though, the controller still
performs nicely, as the output is kept within the given blood glucose limits. On
Figure 5.11 the measured output
is compared to the real output
z.
It shows
that these two accompanies each other pretty good, even though oscillations are
present here too.
5.3.1.3
Process noise
This section will handle the process noise, that is, noise on
x.
Process noise could simulate physiological aspects the model doesn't consider.
200
Dm(t) [mg/dL/min]
15
10
100
50
9 12 15 18 21 24
t [hour]
9 12 15 18 21 24
t [hour]
9 12 15 18 21 24
t [hour]
50
40
0.02
I(t) [mU/L]
Xr(t) [mU/L]
0.03
0.01
0
0.01
G
Gsc
150
30
20
10
9 12 15 18 21 24
t [hour]
Figure 5.12: The evolution of states with process noise
5.3 Simulations
51
u(t) [mU/min]
100
u(t)
Constraints
80
60
40
20
0
0
z(t) [mg/dL]
200
10
12
14
t [hour]
16
18
20
22
24
10
12
14
t [hour]
16
18
20
22
24
z(t)
Setpoint
150
100
50
0
Figure 5.13: Insulin input and output with process noise
160
z
y
140
120
100
80
60
40
10
15
20
25
t [hour]
Figure 5.14: The measured and real output with process noise
52
Again the controller handles the noise nicely.
5.3.1.4
Measurement and process noise
Virtually every system has both measurement noise and process noise, here they
are both applied to the scenario. This is expected to be the most realistic simulation, as these types of noise would be present in the physical system.
200
Dm(t) [mg/dL/min]
15
10
150
100
50
9 12 15 18 21 24
t [hour]
9 12 15 18 21 24
t [hour]
9 12 15 18 21 24
t [hour]
50
40
0.02
I(t) [mU/L]
Xr(t) [mU/L]
0.03
0.01
0
0.01
G
Gsc
30
20
10
9 12 15 18 21 24
t [hour]
Figure 5.15: The evolution of states with process and measurement noise
5.3 Simulations
53
u(t) [mU/min]
100
u(t)
Constraints
80
60
40
20
0
0
z(t) [mg/dL]
200
10
12
14
t [hour]
16
18
20
22
24
10
12
14
t [hour]
16
18
20
22
24
z(t)
Setpoint
150
100
50
0
Figure 5.16: Insulin input and output with process and measurement noise
150
z
y
140
130
120
110
100
90
80
70
60
50
10
15
20
25
t [hour]
Figure 5.17: The measured and real output with process and measurement noise
54
5.3.2
Scenario II
In the next two scenarios, the simulations will be done with both measurement
and process noise. Here the diabetic eats ve big meals.
Time (hours)
D(t)
(mg/dL/min)
12
10
15
18
13
22
Table 5.3: Scenario II meals disturbances

Figures 5.18 and 5.19 illustrates the states and the input/output respectively.
10
100
50
9 12 15 18 21 24
t [hour]
0.04
50
0.03
40
0.02
0.01
0
0.01
G
Gsc
150
I(t) [mU/L]
Xr(t) [mU/L]
200
Dm(t) [mg/dL/min]
15
9 12 15 18 21 24
t [hour]
9 12 15 18 21 24
t [hour]
30
20
10
9 12 15 18 21 24
t [hour]
5.3 Simulations
55
u(t) [mU/min]
100
u(t)
Constraints
80
60
40
20
0
0
z(t) [mg/dL]
200
10
12
14
t [hour]
16
18
20
22
24
10
12
14
t [hour]
16
18
20
22
24
z(t)
Setpoint
150
100
50
0
Figure 5.19: Insulin input and output
Even with these big meals, the controller manages to keep the blood sugar at a
healthy quantity. Figure 5.19 does show that
and lower bound, and at
t 20
comes close to both the upper
the lower bound is violated for a very short
period of time. This is the eect of soft output constraints, and is to be expected
when the ingested meals are big.
56
5.3.3
Scenario III
In this scenario the diabetic eats meals all day long.

Time (hours)
D(t)
(mg/dL/min)
10
12
13
14
15
16
18
13
20
21
22
Table 5.4: Scenario III meals disturbances

Again, Figures 5.20 and 5.21, illustrates the states and the input/output
10
100
50
9 12 15 18 21 24
t [hour]
0.04
50
0.03
40
0.02
0.01
0
0.01
G
Gsc
150
I(t) [mU/L]
Xr(t) [mU/L]
200
Dm(t) [mg/dL/min]
15
9 12 15 18 21 24
t [hour]
9 12 15 18 21 24
t [hour]
30
20
10
9 12 15 18 21 24
t [hour]
5.3 Simulations
57
u(t) [mU/min]
100
u(t)
Constraints
80
60
40
20
0
0
z(t) [mg/dL]
200
10
12
14
t [hour]
16
18
20
22
24
10
12
14
t [hour]
16
18
20
22
24
z(t)
Setpoint
150
100
50
0
Figure 5.21: Insulin input and output
Again the controller performs good, but, like with Scenario II, the lower boundary on the output is violated briey.
58
Chapter
6
Conclusion
What the simulations show, is that the MPC controller, based on Bergmans
minimal model, does a good job keeping the patient inside the blood glucose
limits.
This is the case for both Scenario I, where meals of ordinary size is
simulated, Scenario II, where the patient is assumed to eat relatively big meals,
and Scenario III, where big meals combined with many small meals are ingested.
The simulations also show, that the controller handles noise, on both the measurements and on the process, well. This is important since a practical implementation of the system should denitely be expected to contain noise.
Feedforward showed itself to be an advantage, as the controller is able to take
precaution of future meals, and thereby making it more likely for the patient
to stay within the healthy blood glucose limits. This can be seen by comparing
Figures 5.6 and 5.8.
Though it should be kept in mind, that if a controller
on a real diabetic should use feedforward, it would require the patient to tell
the device what hours a day he would eat, and how big the meals would be.
Therefore feedforward may not be very useful in practice.
It should be kept in mind that the model used for simulations is a minimal
model, so what can be concluded is that there is a change the MPC would
work for the insulin pump case, but the model used is too inadequate to draw
any denite conclusion. Another reason why MPC wouldn't be used is seen on
Figure 5.19 for Scenario II at
t 20,
where the the eect of soft constraints
reveals itself. The amount of blood glucose slightly violates it's lower boundary,
which is potentially disastrous for the patient.
On the other hand, soft con-
straints are needed to make sure the control problem stay feasible.
MPC is an interesting tool to use as controller in an insulin pump, but the
model of the patient must be improved to make it really useful.
60
Conclusion
Appendix
Impulse-response method
The MPC method solves the control problem using a model. Instead of using a
linear model, the impuls-response method and the step response could be used
[1].
The idea behind these methods, is that the process can apply an impuls at any
input, and then measure the response. Thereby the name impuls-response.
Related to the state-space model from the previous chapters:
xk+1 = A xk + B uk
(A.1)
yk = C xk
(A.2)
where
E dk
yk
is a vector of measured outputs.
is neglected from the equation, due the fact that we only want make an
impuls at one input statement, and in this case it is
x0 = 0 and
that u is an
Now assuming that

input at,
vector
t0 ,
such
integer,
u0 6= 0
and
uk = 0
for
k 1.
I.e the
will look like;

u = u0
T
(A.3)
By using these assymptions the following outputs,
xk ,
uk .
then applying that there only is a change in the
zk ,
at the dierent states,
is;
x0 = 0
y0 = H0 u0 = 0
(A.4)
x1 = B u0
y1 = H1 u0 = C Bu0
(A.5)
x2 = A Bu0
y2 = H2 u0 = C ABu0
(A.6)
.
.
.
.
.
.
(A.7)
xk = A
Where
k1
Bu0
Hk = C Ak1 B
yk = Hk u0 = C A
k1
Bu0
(A.8)
is called the k'th Markov parameter, as mentioned in
the previous chapters, and the impuls response sequence is indicated by
Hk
for
62
Impulse-response method
State
Output
Impulse response
Step response
sequence
sequence
sequence
sequence
x0 = 0
x1 = Bu0
x2 = ABu0
y0 = 0
y1 = C Bu0
y2 = C ABu0
H0 = 0
H1 = C B
H2 = C AB
S0 = 0
S1 = C B
S2 = C AB + C B
.
.
.
.
.
.
.
.
.
.
.
.
xk = Ak1 Bu0
yk = C Ak1 Bu0
Hk = C Ak1 B
Sk =
Pk1
i=0
C Ai B
Table A.1: Relation between the dierent sequences
k = 0 . . . N.
Step responce is then the summerize of all the values of
Hi in the input sequence.
I.e. the step response can be dened as;
Sk =
k1
X
Hi
(A.9)
i=0
The relation between the sequences is outlined in Table A;
Using these terms, there can be achived an output using a single input at a
given time
instead of
ts .
uk .
It is also possible to use the other inputs variable,
is replaced with
E,
and
u0
with
d0 .
dk ,
as inputs
The state and the Impulse response
sequence would be;
x0 = 0
H0 = 0
(A.10)
x1 = E d 0
H1 = C B
(A.11)
x2 = A Eu0
H2 = C AB
(A.12)
.
.
.
.
.
.
(A.13)
xk = A
k1
Ed0
Hk = C A
And it's easily to nd the sequences for the output,
k1
y,
(A.14)
and the step response, S,
from this.
The impuls-responce method can be used, when the MPC does not have a
model to build the controller on. Then the method is used to make the Markov
parameter and then solve the problem in an indirect way.
Appendix
B
Matlab programs
This appendix contains the complete listing of the
ATLAB scripts used.
B.1 Scena1.m
1
%S c e n a 1 .m
%M a t i a s
%T e c h n i c a l
%S p r i n g
%B . S c .
Srensen
s042300
University
and
of
Simon
Kristiansen
s042264
Denmark
2007
Thesis
%F i l e
Model
Predictive
Control
for
an
Artificial
Pancreas
description :
10
11
%S c e n a r i o
12
13
I .
Three
normal
meals
are
simulated .
14
15
clear
all
16
close
all
17
18
%Time
19
t0
0.0;
%[ min ]
Start
20
Ts =
8;
%[ min ]
Sampling
21
tf
6024;
%[ min ]
End
22
=
=
=
( t 0 : Ts : t f ) ' ;
%Time
23
24
%S i z e
25
N =
of
horizon
25;
26
27
%C o n s t r a i n t s
28
umin
0;
umax
29
udmin =
16.7;
udmax =
16.7;
30
zmin
60;
zmax
180;
=
=
=
=
100;
31
32
%M e a l s
33
mealsv
( disturbance )
34
m e a l s v ( 8 6 0 / Ts )
35
m e a l s v ( 1 2 6 0 / Ts )
zeros (1 , length ( t ) ) ;
=
=
5;
7;
vector
time
64
36
Matlab programs
m e a l s v ( 1 8 6 0 / Ts )
10;
1 e 3
0.4
37
38
%N o i s e
39
sigmaw
40
%sigmaw
41
42
43
sigmav
%s i g m a v
seed
[0.7
[0
0.8
0];
0];
0.8;
0;
3507;
%C h o i c e
for
randn
44
45
%Use
46
feed
feedforward ?
=
0;
%1 =
yes ,
0 = no
47
48
%W e i g h t
49
Q =
eye ( 1 ) ;
matrices
e y e ( 1 ) 1 e 2;
50
S =
51
Seta
eye ( 1 ) ;
52
seta
eye ( 1 ) ;
53
54
%S t a r t
55
MPCControl
MPC
B.2 Scena2.m
1
%S c e n a 2 .m
%M a t i a s
%T e c h n i c a l
%S p r i n g
%B . S c .
Srensen
s042300
University
and
of
Simon
Kristiansen
s042264
Denmark
2007
Thesis
%F i l e
Model
Predictive
Control
for
an
Artificial
Pancreas
description :
10
11
%S c e n a r i o
12
13
II .
Five
big
meals
are
simulated .
14
15
clear
all
16
close
all
17
18
%Time
19
t0
0.0;
%[ min ]
Start
20
Ts =
8;
%[ min ]
Sampling
21
tf
6024;
%[ min ]
End
22
=
=
( t 0 : Ts : t f ) ' ;
%Time
23
24
%S i z e
25
N =
of
horizon
25;
26
27
28
umin
0;
umax
29
udmin =
16.7;
udmax =
16.7;
30
zmin
60;
zmax
180;
=
=
=
=
100;
31
32
%M e a l s
33
mealsv
34
m e a l s v ( 8 6 0 / Ts )
35
m e a l s v ( 1 2 6 0 / Ts )
36
m e a l s v ( r o u n d ( 1 5 6 0 / Ts ) )
37
m e a l s v ( 1 8 6 0 / Ts )
13;
38
m e a l s v ( 2 2 6 0 / Ts )
6;
39
( disturbance )
=
=
=
7;
10;
=
5;
vector
time
B.3 Scena3.m
40
%N o i s e
41
sigmaw
[0.7
42
sigmav
0.8;
43
seed
65
1 e 3
3507;
0.4
%C h o i c e
0.8
for
0];
randn
44
45
%Use
46
feed
feedforward ?
=
1;
%1 =
yes ,
0 = no
47
48
%W e i g h t
49
Q =
eye ( 1 ) ;
matrices
e y e ( 1 ) 1 e 2;
50
S =
51
Seta
eye ( 1 ) ;
52
seta
eye ( 1 ) ;
53
54
%S t a r t
55
MPCControl
MPC
B.3 Scena3.m
1
%S c e n a 3 .m
%M a t i a s
%T e c h n i c a l
%S p r i n g
%B . S c .
Srensen
s042300
University
and
of
Simon
Kristiansen
s042264
Denmark
2007
Thesis
%F i l e
Model
Predictive
Control
for
an
Artificial
Pancreas
description :
10
11
%S c e n a r i o
12
13
III .
Many
meals
throughout
the
24
14
15
clear
all
16
close
all
17
18
%Time
19
t0
0.0;
%[ min ]
Start
20
Ts =
8;
%[ min ]
Sampling
21
tf
6024;
%[ min ]
End
22
=
=
( t 0 : Ts : t f ) ' ;
%Time
vector
23
24
%S i z e
25
N =
of
horizon
25;
26
27
28
umin
29
30
( bequette
article )
0;
umax
udmin =
16.7;
udmax =
16.7;
zmin
60;
zmax
180;
=
=
100;
31
32
%M e a l s
33
mealsv
( disturbance )
34
m e a l s v ( r o u n d ( 8 6 0 / Ts ) )
35
m e a l s v ( r o u n d ( 1 0 6 0 / Ts ) )
5;
36
m e a l s v ( r o u n d ( 1 2 6 0 / Ts ) )
5;
37
m e a l s v ( r o u n d ( 1 3 6 0 / Ts ) )
7;
38
m e a l s v ( r o u n d ( 1 4 6 0 / Ts ) )
4;
39
m e a l s v ( r o u n d ( 1 5 6 0 / Ts ) )
5;
40
m e a l s v ( r o u n d ( 1 6 6 0 / Ts ) )
6;
41
m e a l s v ( r o u n d ( 1 8 6 0 / Ts ) )
13;
42
m e a l s v ( r o u n d ( 2 0 6 0 / Ts ) )
5;
43
m e a l s v ( r o u n d ( 2 1 6 0 / Ts ) )
3;
=
7;
time
hour
period .
66
44
Matlab programs
m e a l s v ( r o u n d ( 2 2 6 0 / Ts ) )
4;
45
46
%N o i s e
47
sigmaw
[0.7
48
sigmav
0.8;
49
seed
3507;
1 e 3
0.4
%C h o i c e
0.8
for
0];
randn
50
51
%Use
52
feed
feedforward ?
=
1;
%1 =
yes ,
0 = no
53
54
%W e i g h t
55
Q =
eye ( 1 ) ;
matrices
e y e ( 1 ) 1 e 2;
56
S =
57
Seta
eye ( 1 ) ;
58
seta
eye ( 1 ) ;
59
60
%S t a r t
61
MPCControl
MPC
B.4 MPCControl.m
1
%MPCControl .m
%M a t i a s
%T e c h n i c a l
%S p r i n g
%B . S c .
Srensen
s042300
University
of
and
Simon
Kristiansen
s042264
Denmark
2007
Thesis
%F i l e
Model
Predictive
Control
for
an
Artificial
Pancreas
description :
10
11
%Wrapper
12
13
file
for
the
three
MPC
phases .
14
15
MPCDesign
16
MPCSimulate
17
MPCPlot
B.5 MPCDesign.m
1
%MPCDesign .m
%M a t i a s
%T e c h n i c a l
%S p r i n g
%B . S c .
Srensen
s042300
University
of
and
Simon
Kristiansen
s042264
Denmark
2007
Thesis
%F i l e
Model
Predictive
Control
for
an
Artificial
Pancreas
description :
10
11
%D e s i g n
12
%v a r i o u s
13
14
phase ,
design
initializes
15
16
%R e t r i e v e
17
p =
18
P1 = p ( 1 ) ;
the
MPC
matrices
by
calling
files .
parameters
DesignParameters ( ) ;
B.5 MPCDesign.m
19
P2 = p ( 2 ) ;
20
P3 = p ( 3 ) ;
21
n = p(4);
22
VI = p ( 5 ) ;
23
Rutln
= p(6);
24
alpha
= p(7);
25
Gb = p ( 8 ) ;
26
Xb = p ( 9 ) ;
27
Ib
28
Gbsc = p ( 1 1 ) ;
29
Db = p ( 1 2 ) ;
67
= p(10);
30
31
big
10^10;
%A c t s
as
infinity
32
33
%S t e a d y s t a t e
34
xs
[ Gb ;
35
us
= VI n I b ;
36
ds
= Db ;
37
zs
= Gbsc ;
values
Xb ;
Ib ;
( basal
Gbsc ;
values )
Db ] ;
38
39
%S t a r t i n g
40
x0
xs ;
41
u0 =
us ;
points
42
43
%I n i t i a l i z e
44
x =
zeros (5 ,
45
u =
zeros ( length ( t ) , 1 ) ;
46
ubar
47
y =
variables
length ( t ) ) ;
x(: ,
z e r o s ( 2 N , 1 ) ;
1)
u(1)
x0 ;
= u0 ;
ubar ( 1 )
= u0 ;
zeros ( length ( t ) , 1 ) ;
48
49
%D i s t u r b a n c e
50
D =
z e r o s (N,
1);
51
52
%S e t
53
R =
point
r e p m a t ( Gbsc ,
N,
1);
54
55
%G e n e r a t e
noise
56
randn ( ' seed ' ,
57
w =
58
for
59
60
seed ) ;
randn ( 5 , l e n g t h ( t ) ) ;
k =
1:5
w( k , : )
= w( k , : )
sigmaw ( k ) ;
end
61
randn ( ' seed ' ,
62
v =
seed ) ;
randn ( l e n g t h ( t ) , 1 ) sigmav ;
63
64
65
%T r a n s f o r m
66
umin = umin
into
67
zmin
%S e t
up
zmin
deviation
variables
us ;
umax = umax
zs ;
zmax = zmax
us ;
zs ;
68
69
70
[ Umin ,
71
constraints
Umax ,
Udmin ,
Udmax ,
Zmin ,
D e s i g n C o n s t r a i n t s ( umin ,
umax ,
Zmax ,
72
73
%M a t r i c e s
74
Ac =
in
continous
P1Xb
Gb
P2
time
1;
P3
0;
. . .
0;
. . .
0;
. . .
75
76
77
0.2
0.2
78
79
Bc =
[0;
80
Ec =
[1;
81
Cc =
[0
0;
0;
0
1 / VI ;
0;
1
0;
0;
0];
a l p h a
];
0];
82
83
%M a t r i c e s
in
discrete
time
ubarmin ,
udmax ,
. . .
a l p h a
];
udmin ,
ubarmax ]
zmin ,
zmax ,
. . .
big ,
N) ;
68
84
Matlab programs
[ A, B , E , C ]
D e s i g n D i s c r e t e M a t r i c e s ( Ac , Bc , Ec , Cc , Ts ) ;
85
86
87
%Kalman
[ Kfx ,
Filter
A,
B,
design
C,
E]
DesignKalman (A,
B,
C,
E);
88
89
90
%MPC
matrices
[ Hbar , Gamma, Gammad , P h i , Mx0 , Mum1,MR,MD, Lambda , b a r s e t a ]
91
. . .
D e s i g n M P C M a t r i c e s ( A , B , E , C , Q, S , N , S e t a , s e t a ) ;
92
93
[ x0x s ;
xp =
0];
B.6 DesignKalman.m
1
%D e s i g n C o n s t r a i n t s .m
%M a t i a s
%T e c h n i c a l
%S p r i n g
%B . S c .
Srensen
s042300
University
of
and
Simon
Kristiansen
s042264
Denmark
2007
Thesis
%F i l e
Model
Predictive
Control
for
an
Artificial
Pancreas
description :
10
11
%D e s i g n s
12
%A , B , E
13
14
matrices
and
needed
such
they
for
the
contain
Kalman
the
Filter ,
integrated
and
extends
noise .
15
16
function
[ Kfx ,
Ae ,
Be ,
Ce ,
Ee ]
DesignKalman (A,
B,
C,
E)
17
18
n =
s i z e (A , 1 ) ;
19
m =
s i z e (C , 1 ) ;
20
21
%Kalman
22
Qw =
design
e y e ( n , n ) 1 e 2;
e y e (m,m) 1 e 2 ;
23
Qxi
24
Rv =
e y e (m,m ) ;
25
Ad =
e y e (m,m ) ;
26
Ae =
[A B;
27
Be =
[B;
z e r o s (m, n )
Ad ] ;
z e r o s (m,m ) ] ;
28
Ee =
[E;
29
Ce =
[C
z e r o s (m,m ) ] ;
30
Qe =
[ Qw
z e r o s (m,m ) ] ;
z e r o s ( n ,m ) ;
z e r o s (m,
n)
Qxi ] ;
31
32
P =
33
Re = Ce P Ce '+Rv ;
d a r e ( Ae ' ,
34
Kfx = P ( Ce '
35
%k e y b o a r d
Ce ' ,
/
Qe ,
Rv ) ;
Re ) ;
B.7 DesignDiscreteMatrices.m
1
%M a t i a s
%T e c h n i c a l
%S p r i n g
%B . S c .
Srensen
s042300
University
of
and
Simon
Kristiansen
s042264
Denmark
2007
Thesis
Model
Predictive
Control
for
an
Artificial
Pancreas
B.8 DesignParameters.m
%F i l e
69
description :
10
11
%C o n v e r t s
12
13
matrices
A, B , E
and
into
discrete
time .
14
15
function
[ Ad , Bd , Ed , Cd ]
D e s i g n D i s c r e t e M a t r i c e s ( Ac ,
Bc ,
Ec ,
Cc ,
Ts )
16
17
M1 =
18
M2 = expm (M1 Ts ) ;
[ Ac
Bc
Ec ;
zeros (2 ,
7)];
19
Ad = M2 ( 1 : s i z e ( Ac , 1 ) ,
20
Bd = M2 ( 1 : s i z e ( Ac , 1 ) ,
s i z e ( Ac , 2 ) + 1 : s i z e ( Ac , 2 ) + 1 ) ;
21
Ed = M2 ( 1 : s i z e ( Ec , 1 ) ,
s i z e ( Ac , 2 ) + s i z e ( Bc , 2 ) + 1 : e n d ) ;
22
Cd = Cc ;
1 : s i z e ( Ac , 2 ) ) ;
B.8 DesignParameters.m
1
%D e s i g n P a r a m e t e r s .m
%M a t i a s
%T e c h n i c a l
%S p r i n g
%B . S c .
Srensen
s042300
University
of
and
Simon
Kristiansen
s042264
Denmark
2007
Thesis
%F i l e
Model
Predictive
Control
for
an
Artificial
Pancreas
description :
10
11
%C o n t a i n s
12
13
basic
parameters
for
the
Bergman
model .
14
15
function
p =
DesignParameters ( )
16
17
%Model
18
P1 =
0.028735;
parameters
%[ min ^ ( 1 ) ]
19
P2 =
0.028344;
%[ min ^ ( 1 ) ]
20
P3 =
5 . 0 3 5 e 5;
%[mU/L ]
21
Gb =
81.3;
%[mg/ dL ]
22
Ib
15;
%[mU/L ]
23
VI =
12;
%[L ]
24
n =
25
Rutln
26
Gbsc = Gb5 R u t l n ;
27
alpha
28
Xb =
0;
%[mU/L ]
29
Db =
0;
%[mg/ dL / min ]
%[ min ^ ( 1 ) ]
5/54;
=
0.74;
%[mg/ dL / min ]
%[mg/ dL ]
0.05;
%[]
30
31
%P a r a m e t e r
32
p =
[ P1 ;
vektor
P2 ;
P3 ;
n;
VI ;
Rutln ;
alpha ;
Gb ;
Xb ;
Ib ;
Gbsc ;
Db ] ;
B.9 DesignConstraints.m
1
%M a t i a s
%T e c h n i c a l
%S p r i n g
%B . S c .
Srensen
s042300
University
of
and
Simon
Kristiansen
s042264
Denmark
2007
Thesis
Model
Predictive
Control
for
an
Artificial
Pancreas
70
Matlab programs
%F i l e
description :
10
11
%A s s i g n s
12
13
constraints
for
u,
Delta_u ,
and
ubar
into
vectors ;
14
15
function
[ Umin ,
16
Umax ,
Udmin ,
Udmax ,
D e s i g n C o n s t r a i n t s ( umin ,
Zmin ,
umax ,
Zmax ,
udmax ,
ubarmin ,
udmin ,
ubarmax ]
zmin ,
zmax ,
. . .
big ,
17
18
19
Umin =
r e p m a t ( umin ,
N,
1);
20
Umax =
r e p m a t ( umax ,
N,
1);
21
22
Udmin =
r e p m a t ( udmin ,
N 1 ,
1);
23
Udmax =
r e p m a t ( udmax ,
N 1 ,
1);
24
25
Zmin =
r e p m a t ( zmin , N , 1 ) ;
26
Zmax =
r e p m a t ( zmax , N , 1 ) ;
27
28
ubarmin
[ Umin ;
z e r o s (N , 1 ) ] ;
29
ubarmax =
[ Umax ;
o n e s ( N, 1 ) b i g ] ;
B.10 DesignMPCMatrices.m
1
%D e s i g n M P C M a t r i c e s .m
%M a t i a s
%T e c h n i c a l
%S p r i n g
%B . S c .
Srensen
s042300
University
and
of
Simon
Kristiansen
s042264
Denmark
2007
Thesis
%F i l e
Model
Predictive
Control
for
an
Artificial
Pancreas
description :
10
11
%D e s i g n s
various
12
%C r e d i t :
John
13
14
matrices
Bagterp
needed
for
the
MPC
algorithm .
Jrgensen
15
16
function
17
[ barH , Gamma, Gammad , P h i , Mx0 , Mum1,MR,MD, Lambda ,

D e s i g n M P C M a t r i c e s ( A , B , E , Cz , Qz , S , N , S e t a , s e t a )
18
19
nx =
s i z e (A , 2 ) ;
20
nu =
s i z e (B , 2 ) ;
21
nd =
s i z e (E , 2 ) ;
22
nz
s i z e ( Cz , 1 ) ;
23
24
% Form Gamma,
25
Gamma =
26
Gammad =
27
Phi
Gammad ,
Phi
z e r o s (N nz , N nu ) ;
z e r o s (N nz , N nd ) ;
z e r o s (N nz , nx ) ;
28
29
T = Cz ;
30
kz
31
for
0;
k = 1 :N
32
Gamma( k z + 1 : k z+nz , 1 : nu )
33
Gammad( k z + 1 : k z+nz , 1 : nd )
34
T = TA ;
35
P h i ( k z + 1 : k z+nz , 1 : nx )
36
kz
37
end
k z+n z ;
= TB ;
= TE ;
= T;
barseta ]
. . .
N)
B.10 DesignMPCMatrices.m
71
38
39
for
k = 2 :N
40
Gamma ( ( k 1) n z + 1 : end , ( k 1) nu + 1 : k nu
41
Gammad ( ( k 1) n z + 1 : end , ( k 1) nd + 1 : k nd )
42
= Gamma ( 1 : ( N+1k ) nz , 1 : nu ) ;
= Gammad ( 1 : ( N+1k ) nz , 1 : nd ) ;
end
43
44
% Form QZ
45
QZ =
z e r o s (N nz , N n z ) ;
46
kz
0;
47
for
k = 1 :N
48
QZ( k z + 1 : k z+nz , k z + 1 : k z+n z )
49
kz
50
= Qz ;
k z+n z ;
end
51
52
% Form
53
HS =
HS
z e r o s (N nu , N nu ) ;
54
55
if
N == 1
56
57
HS = S ;
else
58
k =0;
59
HS ( 1 : nu , 1 : nu )
60
HS(1+ nu : nu+nu , 1 : nu )
2 S ;
=
S ;
61
62
k = 1 :N2
for
63
ku = k nu ;
64
HS ( kunu + 1 : ku , ku + 1 : ku+nu )
65
HS ( ku + 1 : ku+nu , ku + 1 : ku+nu )
S ;
2 S ;
HS ( ku+nu + 1 : ku+2 nu , ku + 1 : ku+nu )
66
67
S ;
end
68
k=N 1 ;
69
70
ku = k nu ;
71
HS ( kunu + 1 : ku , ku + 1 : ku+nu )
72
HS ( ku + 1 : ku+nu , ku + 1 : ku+nu )
= S;
73
S ;
end
74
75
76
% Form Mum1
77
Mum1 = [ S ;
z e r o s ( ( N 1) nu , nu ) ] ;
78
79
barSeta
e y e (N) S e t a ;
80
barseta
o n e s ( N, 1 ) s e t a ;
81
82
% Form H , Mx0 ,MR,MD
83
T = Gamma' QZ ;
84
H = TGamma + HS ;
85
H =
86
Mx0 = T P h i ;
(H+H ' ) / 2 ;
T ;
87
MR =
88
MD = TGammad ;
89
90
% barH ,
barg
91
92
93
barH =
z e r o s (N, N ) ;
[H
95
% Form
Lambda
96
Lambda =
97
T = [ e y e ( nu , nu )
z ;
barSeta ] ;
94
z e r o s ( ( N 1) nu , N nu ) ;
e y e ( nu , nu ) ] ;
98
99
for
101
k = 1 :N1
Lambda ( ( k 1) nu + 1 : k nu , ( k 1) nu + 1 : ( k +1) nu )
100
end
T;
72
Matlab programs
B.11 MPCSimulate.m
1
%MPCPlot .m
%M a t i a s
%T e c h n i c a l
%S p r i n g
%B . S c .
Srensen
s042300
University
and
of
Simon
Kristiansen
s042264
Denmark
2007
Thesis
%F i l e
Model
Predictive
Control
for
an
Artificial
Pancreas
description :
10
11
%S i m u l a t i o n
phase ,
12
%i n s u l i n
and
13
14
u,
loops
uses
over
this
as
each
input
timestep
to
the
and
finds
Bergman
the
optimal
model .
15
16
17
18
%I t e r a t e
for
k = 1 : l e n g t h ( t ) 1
display (k)
19
20
21
%M e a s u r e m e n t
22
y(k)
Noise
noise
= x(4 ,k)
+ v(k );
23
24
%Model
25
x(1 ,k)
= x(1 ,k)
noise
+ w( 1 , k ) ;
26
x(2 ,k)
= x(2 ,k)
+ w( 2 , k ) ;
27
x(3 ,k)
= x(3 ,k)
+ w( 3 , k ) ;
28
x(4 ,k)
= x(4 ,k)
+ w( 4 , k ) ;
29
30
%Meal
31
x(5 ,k)
Disturbance
disturbance
= x(5 ,k)
at
this
point
in
time
mealsv ( k ) ;
32
33
34
%W i t h o u t
if
feed
35
36
37
feedforward
== 0
% Do
%With
elseif
38
feed
if
== 1
l e n g t h ( t )1 k ;
n =
39
nothing
Feedforward
N <= n ;
40
%A s s i g n
41
D ( 1 : N)
42
future
=
%S i m u l a t i o n
44
D =
45
is
ending ,
assign
zeros
to
z e r o s (N , 1 ) ;
D( 1 : n )
46
m e a l s v ( k : k+n 1 ) ;
end
else
48
display ( ' Fejl
49
return ;
50
to
else
43
47
meals
m e a l s v ( k : k+N 1 ) ;
feed
parameter ' )
end
51
52
53
%Dummy
if
54
55
to
handle
u0
un = u0 ;
else
un = u ( k 1 ) ;
56
57
statement
k == 1
end
58
59
60
%MPC
step
[ unow , xp , i n f o ]
= MPCCompute ( y ( k ) , un , R , D, Zmin , Zmax ,
. . .
61
z s , z s , u s , d s , xp , u b a r ,
62
Hbar , Mx0 , Mum1,MR,MD, b a r s e t a ,
63
Phi ,
Gammad ,
Gamma,
. . .
Lambda ,
. . .
big ,
. . .
B.12 MPCCompute.m
73
64
umin ,
65
ubarmin ,
umax ,
66
Udmin ,
67
Kfx ,
68
A,
udmin ,
ubarmax
Udmax ,
udmax ,
. . .
. . .
. . .
. . .
B,
E,
C,
N) ;
69
70
71
%Check
72
if
for
info
error
== 1
73
display ( ' Error
74
return
75
in
QP
solver ! ');
end
76
77
%S i m u l a t e
78
[ T,X]
step
o d e 1 5 s ( @BergmanMinimalModel ,
[ t (k)
t (k +1)] ,
x(: ,
k) ,
79
. . .
[] ,
p,
unow ) ;
80
81
% Remember
82
x(: ,
k +1) = X( end ,
83
u(k)
= unow ;
84
variables
:);
end
B.12 MPCCompute.m
1
%MPCCompute .m
%M a t i a s
%T e c h n i c a l
%S p r i n g
%B . S c .
Srensen
s042300
University
and
of
Simon
Kristiansen
s042264
Denmark
2007
Thesis
%F i l e
Model
Predictive
Control
for
an
Artificial
Pancreas
description :
10
11
%C a l c u l a t e s
12
13
the
insulin
input
for
this
timestep ,
u0 .
14
15
function
[ u0 , xp , i n f o ]
= MPCCompute ( y , um1 , R , D, Zmin , Zmax ,
. . .
16
y s , z s , u s , d s , xp , u b a r ,
17
Hbar , Mx0 , Mum1,MR,MD, s b a r e t a ,
18
Phi ,
19
umin ,
GammaD,
umax ,
20
ubarmin ,
21
dUmin ,
22
Kfx ,
23
A,
24
25
nu =
l e n g t h ( um1 ) ;
26
nd =
l e n g t h ( ds ) ;
27
nz
length ( zs ) ;
28
%Form
29
dy = y
deviation
variables
ys ;
30
dum1 = um1u s ;
31
dR = Rr e p m a t ( z s ,
N,
1);
32
dD = Dr e p m a t ( d s ,
N,
1);
33
dd = dD ( 1 : nd ,
1);
34
35
%Kalman
36
37
x0
38
= dy
Filter
C xp ;
= xp + Kfx e ;
B,
Gamma,
dumin ,
ubarmax
dUmax ,
. . .
E,
C,
N)
. . .
. . .
Lambda ,
dumax ,
. . .
. . .
big ,
. . .
. . .
74
Matlab programs
39
%U p d a t e
40
g = Mx0 x 0
41
gbar
gradient
+ MRdR + MDdD + Mum1 dum1 ;
[g;
sbareta ] ;
42
43
%C r e a t e
bounds
for
ubar
44
u b a r m i n ( 1 : nu , 1 )
= max ( umin , dumin+dum1 ) ;
45
ubarmax ( 1 : nu , 1 )
= min ( umax , dumax+dum1 ) ;
46
47
48
%C r e a t e
c
bounds
Phi x0
for
+ GammaDdD ;
49
Zbarmin
50
Zbarmax = Zmax
= Zmin
c ;
c ;
51
52
%S e t u p
matrix
53
bmin =
[ dUmin ;
system
b i g o n e s (N,
for
boundaries
54
bmax =
[ dUmax ;
Zbarmax ;
55
Abar =
[ Lambda
z e r o s ( N 1 ,
1);
Zbarmin ] ;
b i g o n e s (N , 1 ) ] ;
N) ;
Gamma
e y e ( N , N ) ;
Gamma
e y e (N, N ) ] ;
56
57
58
%C r e a t e
startguess
ubarinit
u b a r ( nu + 1 :N nu , 1 ) ;
u b a r ( ( N 1) nu + 1 :N nu , 1 ) ;
u b a r ( nu N+n z + 1 : end , 1 ) ;
59
. . .
u b a r ( nu N+(N 1) n z + 1 : e n d )
];
60
61
62
% QPSOLVER
63
64
[ u b a r , i n f o ]= q p s o l v e r ( Hbar , g b a r , u b a r m i n , ubarmax , Abar , bmin , bmax , u b a r i n i t ) ;
65
66
if
info
67
== 0
du0 =
68
u b a r ( 1 : nu , 1 ) ;
else
69
du0 = dum1 ;
70
end
%
73
74
% QUADPROG
75
76
77
78
79
80
81
82
83
du0 =
84
info
85
86
87
88
89
71
72
Abar2 = [ Abar ; Abar ] ;

b b a r = [ bmax ; bmin ] ;
[ u b a r , f e v a l , EXITFLAG ]
. . .
q u a d p r o g ( Hbar , g b a r , Abar2 , b b a r , [ ] , [ ] , u b a r m i n , ubarmax , u b a r i n i t ) ;

if
EXITFLAG == 1
u b a r ( 1 : nu , 1 ) ;
0;
else
du0 = dum1 ;
info
1;
end
90
91
92
% Update
93
xp = A x 0
Kalman
Filter
+ B du0 + E dd ;
94
95
% Form
96
u0 = du0 +
physical
variable
us ;
B.13 BergmanMinimalModel.m
1
%B e r g m a n M i n i m a l M o d e l .m
B.14 MPCPlot.m
%M a t i a s
%T e c h n i c a l
%S p r i n g
%B . S c .
Srensen
75
s042300
University
of
and
Simon
Kristiansen
s042264
Denmark
2007
Thesis
%F i l e
Model
Predictive
Control
for
an
Artificial
Pancreas
description :
10
11
%I m p l e m e n t a t i o n
12
%d i f f e r e n t i a l
13
14
of
the
Bergman
' ' minimal
model ' ' ,
which
consists
of
equations .
15
16
function
xdot
BergmanMinimalModel ( t , x , p , u )
17
18
xdot
zeros (5 ,1);
19
20
G = x(1 ,1);
21
X = x(2 ,1);
22
23
Gsc = x ( 4 , 1 ) ;
= x(3 ,1);
24
D = x(5 ,1);
25
26
P1 = p ( 1 ) ;
27
P2 = p ( 2 ) ;
28
P3 = p ( 3 ) ;
29
n = p(4);
30
VI = p ( 5 ) ;
31
Rutln
= p(6);
32
alpha
= p(7);
33
Gb = p ( 8 ) ;
34
Xb = p ( 9 ) ;
35
Ib
36
Gbsc = p ( 1 1 ) ;
= p(10);
37
Db = p ( 1 2 ) ;
38
39
xdot ( 1 , 1 )
40
xdot ( 2 , 1 )
41
xdot ( 3 , 1 )
42
xdot ( 4 , 1 )
43
xdot ( 5 , 1 )
P1 (GGb ) XG +
P2 X + P3 ( I I b ) ;
n I + u / VI ;
(GGsc ) / 5 R u t l n ;
a l p h a D ;
D;
B.14 MPCPlot.m
1
%MPCPlot .m
%M a t i a s
%T e c h n i c a l
%S p r i n g
%B . S c .
Srensen
s042300
University
of
and
Simon
Kristiansen
s042264
Denmark
2007
Thesis
%F i l e
Model
Predictive
Control
for
an
Artificial
Pancreas
description :
10
11
%E v a l u a t i o n
12
13
phase ,
plots
14
15
16
17
%t
t
should
=
t /60;
be
in
hours
the
simulations .
five
76
18
Matlab programs
%P l o t
the
five
states
19
figure ;
20
subplot (221);
21
plot (t , x ( 5 , : ) ) ;
22
y l a b e l ( 'D_{m} ( t )
23
xlabel ( ' t
24
25
V =
axis ;
s e t ( gca ,
[ mg/ dL / min ] ' ) ;
[ hour ] ' ) ;
axis ([0
t f /60
' XTick ' ,
V( 3 )
V( 4 ) ] )
0:3:24)
26
27
subplot (222);
28
plot (t , x (1 ,:) , ' b ' , t , x (4 ,:) , ' r ' ) ;
29
l e g e n d ( ' G' , 'G_{ s c } ' , ' L o c a t i o n ' ,
30
y l a b e l ( 'G( t ) ,
31
32
33
xlabel ( ' t
V =
' NorthWest ' ) ;
[ mg/ dL ] ' ) ;
[ hour ] ' ) ;
axis ;
s e t ( gca ,
G_{ s c } ( t )
axis ([0
t f /60
' XTick ' ,
V( 3 )
V( 4 ) ] )
0:3:24)
34
35
subplot (223);
36
plot (t , x ( 2 , : ) ) ;
37
y l a b e l ( 'X_{ r } ( t )
38
xlabel ( ' t
39
40
V =
axis ;
s e t ( gca ,
[mU/L ] ' ) ;
[ hour ] ' )
axis ([0
t f /60
' XTick ' ,
V( 3 )
V( 4 ) ] )
0:3:24)
41
42
subplot (224);
43
plot (t , x ( 3 , : ) ) ;
44
ylabel ( ' I ( t )
45
xlabel ( ' t
46
47
V =
axis ;
s e t ( gca ,
[mU/L ] ' ) ;
[ hour ] ' )
axis ([0
t f /60
' XTick ' ,
V( 3 )
V( 4 ) ] )
0:3:24)
48
49
%P l o t
input
and
output
50
figure
51
subplot (211);
52
plot (t ,
53
ylabel ( 'u( t )
54
xlabel ( ' t
55
hold
56
stairs (t ,
r e p m a t ( umin+u s ,
length ( t )) ,
'k ' ,
' linewidth ' ,
2);
57
stairs (t ,
r e p m a t ( umax+u s ,
length ( t )) ,
'k ' ,
' linewidth ' ,
2);
58
legend ( ' u( t ) ' , ' Constraints ' , ' Location ' ,
59
60
V =
u,
'b ' )
[mU/ min ] ' ) ;
[ hour ] ' )
on
axis ;
s e t ( gca ,
axis ([0
t f /60
' XTick ' ,
V( 3 ) 1 0
' NorthWest ' ) ;
V( 4 ) 1 . 1 ] )
0:2:24)
61
62
subplot (212);
63
plot (t ,
64
hold
x(4 ,
:) ,
65
plot (t ,
66
plot (t ,
r e p m a t ( z m i n+z s ,
length ( t )) ,
'k ' ,
' linewidth ' ,
2);
67
plot (t ,
r e p m a t ( zmax+z s ,
length ( t )) ,
'k ' ,
' linewidth ' ,
2);
68
xlabel ( ' t
ylabel ( ' z( t )
70
legend ( ' z ( t ) ' ,
72
V =
axis ;
s e t ( gca ,
length ( t )) ,
'r ' ,
' linewidth ' ,
[ hour ] ' )
69
71
'b ' ) ;
on
[ mg/ dL ] ' ) ;
' Setpoint ' , ' Location ' ,
axis ([0
' XTick ' ,
t f /60
V( 3 ) 0 . 9
0:2:24)
73
74
75
%F o r
investigation
76
st
0.5;
77
% en
2.5;
78
figure
79
subplot (211);
80
plot ( t ,
81
ylabel ( 'u( t )
82
xlabel ( ' t
u,
'b ' )
[mU/ min ] ' ) ;
[ hour ] ' )
' NorthWest ' ) ;
V( 4 ) 1 . 1 ] )
1);
B.15 InvestSampling.m
hold
77
83
84
% V =
axis ;
on
axis ([ st
en
85
% %s e t ( g c a ,
' XTick ' ,
0:1:24)
86
87
88
plot ( t ,
89
hold
90
plot ( t ,
91
xlabel ( ' t
92
ylabel ( ' z( t )
93
legend ( ' z ( t ) ' , ' Constrains ' , ' Location ' ,
94
% V =
V( 3 ) 1 0
V( 4 ) 1 . 1 ] )
subplot (212);
x(4 ,
:) ,
'b ' ) ;
on
length ( t )) ,
'r ' ,
' linewidth ' ,
1);
[ hour ] ' )
axis ;
[ mg/ dL ] ' ) ;
axis ([ st
V( 3 ) 0 . 9
en
' NorthEast ' ) ;
V( 4 ) 1 . 1 ] )
95
96
97
98
99
100
% y
vs
figure
plot (t ,
hold
x(4 ,
:) , 'b')
on
1 ) , y ( 1 : end 1 ) , ' r
101
p l o t ( t ( 1 : end
102
legend ( ' z ' , ' y ' )
103
xlabel ( ' t
')
[ hour ] ' )
B.15 InvestSampling.m
1
%I n v e s t S a m p l i n g .m
%M a t i a s
%T e c h n i c a l
%S p r i n g
%B . S c .
Srensen
s042300
University
of
and
Simon
Kristiansen
s042264
Denmark
2007
Thesis
%F i l e
Model
Predictive
Control
for
an
Artificial
Pancreas
description :
10
11
%S c e n a r i o
12
%and
13
%r e s u l t i n g
14
15
the
where
sampling
graphs .
stepchange
time
See
16
17
clear
all
18
close
all
19
20
p =
21
P1 = p ( 1 ) ;
DesignParameters ( ) ;
22
P2 = p ( 2 ) ;
23
P3 = p ( 3 ) ;
24
n = p(4);
25
VI = p ( 5 ) ;
26
Rutln
= p(6);
27
alpha
= p(7);
28
Gb = p ( 8 ) ;
29
Xb = p ( 9 ) ;
30
Ib
31
Gbsc = p ( 1 1 ) ;
= p(10);
32
Db = p ( 1 2 ) ;
33
34
t0
35
Ts =
36
tf
37
38
=
=
=
0.0;
1;
6024;
t 0 : Ts : t f ;
is
in
either
investigated
thesis
for
or
by
is
looking
specified ,
at
the
details .
78
Matlab programs
39
xs
40
us
= n I b VI ;
[ Gb ;
0;
Ib ;
42
x0
xs ;
43
u0 =
us ;
Gbsc ;
0];
41
44
45
N =
length ( t ) ;
47
u =
z e r o s ( 1 ,N ) ;
48
u ( 1 : end )
49
u ( 6 0 : end )
46
=
us ;
us
10;
50
51
x =
52
x(: ,
z e r o s ( 5 ,N ) ;
1)
x0 ;
53
54
55
mealsv
zeros ( length ( t ))4;
%m e a l s v ( 1 : 6 0 )
0;
56
57
for
k =
58
1:
N1
display (k );
59
60
x(5 ,k)
= x(5 ,k)
[ T,
mealsv ( k ) ;
61
62
X]
o d e 1 5 s ( @BergmanMinimalModel , [ t ( k )
t (k +1)] , x ( : , k ) , [ ] , p , u(k ) ) ;
63
64
65
x(: ,
k +1) = X( end ,
:) ';
end
66
67
figure
68
en
69
860;
/60;
70
71
p l o t ( t ( 1 : en ) ,
72
ylabel ( ' z( t )
73
xlabel ( ' t
74
hold
75
s e t ( gca ,
x(4 ,
1 : en ) ,
'b ' ) ;
[ mg/ dL ] ' ) ;
[ hour ] ' ) ;
on
' XTick ' ,
0 : 1 : en / 6 0 )
76
77
figure
78
s t a i r s ( t ( 1 : en ) ,
79
xlabel ( ' t
80
y l a b e l ( 'D( t )
81
s e t ( gca ,
' XTick ' ,
0:1:24)
82
s e t ( gca ,
' YTick ' ,
[ ] ) ;
83
V =
m e a l s v ( 1 : en ) )
[ hour ] ' ) ;
axis ;
[ mg/ dL / min ] ' ) ;
a x i s ( [ V( 1 )
V( 2 )
V( 3 )
0.5
V( 4 ) + 0 . 5 ] )
V( 3 )
0.5
V( 4 ) + 0 . 5 ] )
84
85
figure
86
s t a i r s ( t ( 1 : en ) ,
87
xlabel ( ' t
88
s e t ( gca ,
' XTick ' ,
0:1:24)
89
s e t ( gca ,
' YTick ' ,
[ ] ) ;
90
V =
u ( 1 : en ) )
[ hour ] ' ) ;
axis ;
a x i s ( [ V( 1 )
V( 2 )
1
%I n v e s t W e i g h t s .m
%M a t i a s
%T e c h n i c a l
%S p r i n g
%B . S c .
Srensen
s042300
University
of
and
Simon
Kristiansen
s042264
Denmark
2007
Thesis
Model
Predictive
Control
for
an
Artificial
Pancreas
%F i l e
79
description :
10
11
%W e i g h t
12
13
matrix
varied .
14
15
clear
all
16
close
all
17
18
%Time
19
t0
0.0;
%[ min ]
Start
20
Ts =
3;
%[ min ]
Sampling
21
tf
6024;
%[ min ]
End
22
=
=
( t 0 : Ts : t f ) ' ;
%Time
time
vector
23
24
%S i z e
25
N =
of
horizon
25;
26
27
28
big
( bequette
article )
10^10;
29
% umin
0;
umax
30
% udmin =
16.7;
udmax =
100;
16.7;
31
% zmin
60;
zmax
180;
33
umin
umax
big ;
udmin =
udmax =
big ;
35
zmin
b i g
b i g
b i g
34
zmax
big ;
32
=
=
36
37
%M e a l s
38
mealsv
( disturbance )
39
m e a l s v ( 1 6 0 / Ts )
=
10;
40
41
%N o i s e
42
noisefact
43
seed
0;
%S t a n d a r d
3501;
%C h o i c e
deviation
for
of
noise
randn
44
45
%Use
46
feed
feedforward ?
=
1;
%1 =
yes ,
0 = no
47
48
%W e i g h t
matrices
49
Q =
eye ( 1 ) ;
50
S =
e y e ( 1 ) 1 e 1;
51
Seta
eye ( 1 ) ;
52
seta
eye ( 1 ) ;
53
54
%S t a r t
55
MPCControl ( t ,
56
MPC
Ts ,
tf ,
mealsv ,
umin ,
umax ,
noisefact ,
Q,
udmin ,
S,
udmax ,
Seta ,
seta ,
zmin ,
N,
zmax ,
seed ,
. . .
feed )
80
Matlab programs
Bibliography
[1] Jan M. Maciejowski:
Predictive Control with Constraints, Prentice Hall,
2001
Model Predictive Control of

Blood Glucose in Type 1 Diabetics Using Subcutaneous Glucose Measurements, Rensselaer Polytechnic Institute, 2002
[2] Sandra M. Lynch, B. Wayne Bequette:
et al.: Hypoglycemia Prediction and Detection Using

Optimal Estimation, Diabetes Technology & Therapeutics, vol. 7, num.
[3] Cesar C. Palerm

1, 2005
A Critical Assessment of Algorithms and Challenges in the Development of a Closed-Loop Articial Pancreas, Dia-
[4] B. Wayne Bequette:
betes Technology & Therapeutics, vol. 7, num. 1, 2005
et al.: Modeling Insulin Action for Development of a ClosedLoop Articial Pancreas, Diabetes Technology & Therapeutics, vol. 7,
[5] H.M. Steil
num. 1, 2005
[6] Nakhle H. Asmar:
Partial Dierential Equations, second edition, Pren-
tice Hall, 2004

[7] John B. Jrgensen, Sten B. Jrgensen:
Model Predictive Control, Tech-
nical University of Denmark, 1998

[8] Greg Welch, Gary Bishop:
An Introduction to the Kalman Filter, Uni-
versity of North Carolina,2006

James B. Rawlings:
Disturbance Models for
Oset-Free Model-Predictive Control, AIChE Journal Vol.49 No.2, 2003
[9] Gabriele Pannocchia,

Model Predictive Control: For An Artical Pancreas

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Model Predictive Control: For An Artical Pancreas

Uploaded by

Copyright:

Available Formats

Model Predictive Control

for an artical pancreas

Matias Srensen og Simon Kristiansen

Technical University of Denmark

Technical University of Denmark

These two approaches will make

Design, simulation and evaluation.

user calls a le, where the wanted test scenario is initialized.

begrnsninger p insulin input og input-hastigheden, hvilket srger for der ikke

Feedback fr kontrol algoritmen til at evaluere det mlte output og dermed

Her bliver kontrollerens ydelse testet, og det viser sig at

Continuous-Discrete Time Conversion

Model Predictive Control

Constraints on input rate of movement . . . . . . . . . . .

Case study - A minimal model

Horizon and sampling time

Soft output constraints

In year 2000 there were around 171 mio.

people with diabetes in the world.

the body, so the energy in the food can be

ple with diabetes needs to have an external

If the patient gets too much

calculates the optimal quantity of insulin based on measurements of the blood

Figure 1.2: The insulin pump

= f (x(t), u(t), d(t))

The rst step is to identify an equilibrium point, a steady-state,

Making a Taylor expansion around this point yields an

approximation to the system (2.1);

= f (x(t), u(t), d(t))

Introducing the deviation variables,

X(t) = x(t) xs , U (t) = u(t) us

This is a system of linear time invariant dierential equations.

X(t) = eA(tt0 ) X(t0 ) +

denotes the matrix exponential function.

2.1 Continuous-Discrete Time Conversion

2.1 Continuous-Discrete Time Conversion

is the sampling time.

A practical way to nd

is to solve the equation (see [1]);

tk be the time dened as tk = t0 + kTs and let x(tk ) = xk , u(tk ) = uk

x(tk+1 ) = A x(tk ) + B u(tk ) + E d(tk ) xk+1 = A xk + B uk + E dk

xk = Ak xs + (Ak1 B)us + (Ak1 E)ds + + Buk1 + Edk1

By dening the output vector

as the measurable states, the output in

are denoted below.

These parameters will be of use in later sections.

Linearized discrete time state-space model

= f (x(t), u(t), d(t)),

indentify a steady-state point

(xs ,us ,ds )

(xs ,us ,ds )

(xs ,us ,ds )

- The corresponding linear model is

with X, U and D as deviation variables.

Converting to discrete time,

- Model in discrete time is

are deviation variables.

indicating the measureable states.

Model Predictive Control

Figure 3.1: The basic MPC

is the actual output and

manipulates the input,

is the measured output.

for an artical pancreas

user calls a le, where the wanted test scenario is initialized.

Case study - A minimal model

The rst step is to identify an equilibrium point, a steady-state,

This is a system of linear time invariant dierential equations.

A practical way to nd

tk be the time dened as tk = t0 + kTs and let x(tk ) = xk , u(tk ) = uk

By dening the output vector

chapter, for instance by adding disturbance and dierent sorts of constraints. It

can't be inuenced, the term

rst control problem;

inuences the MPC problem.