Anda di halaman 1dari 14

ooo!

L1098(95)00184-0

Auromorico.
Vol. 32, No. 4, pp. 519-532, 19%
Coovrieht
62 19% Elsevier Science Ltd
Printed&
&ea;Britain.
All rights reserved
OCKX-1098/% $15.00 + 0.00

Optimal Strategies for the Control of a Train*


PHIL HOWLETTt
analysis is used to find fuel-eficient driving strategies for a
long-haul freight train with only discrete levels of control. Key equations
are used to define strategies of optimal type and optimal switching points.

A Lagrangian

Key Words-Train

control; optimal control; discrete control; optimal switching; numerical algorithms.

Abstract-This paper describes a method for the calculation


of optimal control strategies in an important engineering
application. A train travels from one station to the next along
a track with non-constant gradient. The journey must be
completed within a given time, and it is desirable to minimise
the fuel consumption. We assume that only certain discrete
throttle settings are possible and that each setting determines
a constant rate of fuel supply. This assumption is based on
the control mechanism for a typical diesel-electric locomotive. For each fixed finite sequence of control settings, we
show that fuel consumption is minimised only if the settings
are changed when certain key equations are satisfied. The
strategy determined by these equations is called a strategy of
optimal type. Several realistic examples are given with the
results of associated numerical calculations. The examples
demonstrate the profound effect of even a small gradient on
a strategy of optimal type. We show that a strategy of
optimal type with alternate phases of coust and maximum
power can be used to approximate the idealised minimum
cost strategy.

rate of fuel supply. The power developed by the


locomotive is directly proportional to the rate of
fuel supply, and the corresponding acceleration
is non-negative. We shall assume that a single
brake control gives constant negative acceleration. In current operations dynamic braking is
used only to control the speed of the train. None
of the recovered energy is used to drive the
train.
In our model the train will be controlled using
a finite sequence of traction phases and a final
brake phase. During each phase, the control
setting is constant. A traction phase will be
classified as a power phase or, if the fuel supply
rate is zero, as a coast phase. For each control
strategy there is a uniquely defined speed profile
determined by the equations of motion. We shall
say that a strategy is feasible if the distance and
time constraints are satisfied.
We wish to find a feasible strategy that
minimises fuel consumption.

1. INTRODUCTION

1.1. An outline of the problem


A train travels from one station to the next along
a track with non-constant gradient. The journey
must be completed within a given time, and it is
desirable to minimise the fuel consumption.
We assume that only certain discrete control
settings are possible. This conforms to the most
common situation in railway locomotives. In
particular, the GM diesel-electric locomotives
(see the GM Locomotive service manual for the
GM JT26C-2S.S)
and many other long-haul
locomotives have eight traction control settings.
The JT26C-2SS
and other similar models are
currently used by Australian National to haul
freight trains some 4000 km across the Australian continent between Sydney and Perth. Each
traction control setting determines a constant

1.2. Applications of the work


The cost of fuel is the most significant
recurrent
expenditure
in many large rail
organisations. It is estimated that the cost for
purchase and lifetime maintenance of a locomotive is approximately equal to the total cost of
fuel consumed by the locomotive. Reduction of
fuel consumption using energy-efficient driving
strategies is therefore a primary concern, and
consequently
the development
of a suitable
driver advice system would be most beneficial.
The work in this paper is part of a long-term
project carried out by the Scheduling and
Control Group at the University of South
Australia. The aim of the project is to produce
an on-board advice unit that can be used by train
drivers
to ensure
energy-efficient
driving
strategies.
The first such device is now commercially
available under the name Metromiser. Metromiser is an on-board
computer
with an

* Received 14 July 1994; revised 16 March 1995; received


in final form 25 August 1995. This paper was not presented at
any IFAC meeting. This paper was recommended
for
publication in revised form by Associate Editor John N.
Tsitsiklis under the direction of Editor Tamer Basar.
Corresponding author Dr Phil Howlett. Tel. +61 8 302 3804;
Fax +61 8 302 3381; E-mail p.howlett@unisa.edu.au.
t Scheduling and Control Group, Centre for Industrial and
Applied Mathematics, University of South Australia, The
Levels, 5095, Australia.
519

P. Howlett
associated driver advice display unit designed for
use on metropolitan trains. Extensive in-service
trials with this unit (Benjamin et al., 1987) have
shown improved time-keeping and a reduction in
fuel consumption of between ten and twenty
percent. It has been clearly demonstrated that
the timing of control decisions is significantly
improved by Metromiser. This comment applies
particularly to the timing of the crucial switch
into the semi-final coast phase. Even for
relatively simple journeys it is not possible for a
driver to perform the necessary calculations or to
intuitively drive with near optimal control.
A more comprehensive device designed for
use on long-haul trains is currently
being
developed. The algorithms used in the new
device will be based on the work described in
this paper. In developing the new device, the
Group is collaborating
closely with railway
operators and system suppliers.

discrete control settings: power, coast and brake.


Consider the possible control options for the
driver. The strategy must begin with a power
phase and follow with alternate phases of coast
and power. The strategy must end with a
semi-final coast phase and a final brake phase,
although it is possible that one or other of these
latter two phases could become degenerate. We
assume that braking will be used only to stop the
train.
The driver can decide the number of phases
and the points at which the control settings will
be changed. These points are called the switching
points.

The nature of the problem is not changed by


allowing a greater range of discrete control
settings. The driver must first decide the precise
sequence of control settings and then determine
the optimal position of the switching points.
The driver must make these decisions in such
a way that fuel consumption is minimised.

1.3. Previous studies


The train control problem has been studied by
a number of authors. There are two widely used
control
models: a model with continuous
(Milroy, 1980; Kraft and Schneider,
1981;
Howlett, 1984, 1988a, b, 1990; Asnis et al., 1985;
Benjamin et al., 1987) in which the control
variable is the applied acceleration and the cost
of the strategy is the mechanical energy required
to drive the train; and a model with discrete
control (Benjamin et al., 1989; Cheng and
Howlett, 1990a, b, 1992, 1993; Howlett et al.,
1992, 1994a, b) in which the control variable is a
throttle setting and the cost of the strategy is the
total fuel consumption.
The current paper is one of a sequence of
the practical
papers that have investigated
implementation of energy-efficient driving strategies on metropolitan trains and on long-haul
freight trains. Because the implementation
of
these strategies on real trains is a fundamental
concern, it has been necessary to construct a
model that describes the real control mechanism.
This is the so-called fuel consumption model.
The model has been studied systematically on
level track, where a complete solution has been
obtained.
No previous
solution
has been
obtained for non-level track. For application of
this work to real freight trains, it is essential that
the effects of track gradient are properly
understood.

1.5. Formulation of a well-posed problem


Because the control strategy is restricted to a
finite sequence of discrete settings, the problem
described above is not well posed. In general,
there will be no admissible strategy for which the
minimum fuel consumption is achieved, and
hence no definitive conditions for optimality.
We overcome this difficulty in the following
way. Consider the set of all possible control
strategies. We divide the set into disjoint subsets,
and seek the best strategy from each subset.
Each subset is defined by a fixed finite sequence
of control settings. There are many different
strategies in each subset; all strategies use the
same sequence of control settings, but each one
is determined by different switching points.
To find a feasible strategy that minimises fuel
consumption within the given control subset, we
need to find the optimal locations for the
switching points subject to the distance and time
constraints. A strategy satisfying these requirements will be called a strategy of optimal type.
This problem is a standard finite-dimensional
constrained optimisation problem. Because we
are seeking the best location for a finite number
of points within a closed and bounded interval
the existence of such a strategy is guaranteed.
The original problem can now be regarded as
in comparison
of the various
a problem
strategies of optimal type.

1.4. A driver perspective on the problem


To understand the essence of the problem, it is
helpful to consider a simplified model.
Suppose the locomotive
has only three

1.6. A preview of the new results


We consider the train control problem on a
track with non-constant gradient. There has been
no practical solution to this problem. In this paper

521

Train control
we find key equations that determine necessary conditions for the strategies of optimal
type;
we show that the key equations can be used to
describe the nature of the strategies of optimal
type;
we show that the minimum fuel consumption
can be almost achieved using .a strategy of
optimal type and
we illustrate our results with several realistic
examples.
On level track it has been shown (Cheng and
Howlett, 1992) that the strategies of optimal type
are determined by key equations that provide
necessary and sufficient conditions. In this paper
the key equations of Cheng and Howlett are
generalised to give necessary conditions for a
strategy of optimal type on a track with
non-constant gradient. Although the existence of
a strategy of optimal type is clear, the existence
and uniqueness of solutions to the key equations
is no longer guaranteed. However, we do know
that the equations can often be solved and that
the solutions could be calculated in real time by
a small on-board computer. We believe that
these solutions can be used to provide practical
advice to drivers about energy-efficient driving
strategies.
We shall use the examples to emphasise that
there is no significant disadvantage in having
only a limited number of control settings. The
driver simply alternates between a level of
control above the desired value and a level
below the desired value.
The examples will also be used to show that an
inappropriate
choice of switching points can
disrupt schedules and increase fuel consumption.
The examples will demonstrate that even very
small changes in gradient can cause a dramatic
change in the position of the optimal switching
points. We shall also show that our methods can
be used to calculate energy-efficient
driving
strategies on steep track.
2. FORMULATION

OF THE

PROBLEM

2.1. Locomotive performance characteristics


It is useful to begin by considering the traction
and braking characteristics of the GM JT26C2% diesel-electric locomotive. The locomotive
has eight traction control settings. Each setting
determines a constant rate of fuel supply to the
diesel motors, and the observed acceleration is
inversely proportional
to the speed, provided
that the speed exceeds a known small value. This
is consistent with a constant power output in

each setting except at very low speeds. A typical


graph of observed acceleration against speed for
each traction control setting is shown in Fig.. 1.
Figure 2 shows a graph of power against rate of
fuel supply.
Dynamic braking is an important aspect of
control in a typical diesel-electric locomotive. A
minimal rate of fuel supply is normally required
to control the effective braking effort. Energy
regained by the locomotive from the dynamic
braking process is not stored for subsequent use,
and is normally lost as heat energy. The dynamic
braking performance of a typical locomotive is
shown on the graph in Fig. 3. The dynamic
braking control mechanism is designed so that
the negative
acceleration
is approximately
constant over the normal speed range. The
dynamic brakes are supplemented by air brakes,
which are used predominantly at low speed to
increase the braking capability to an acceptable
level. Because energy is lost during braking, it is
intuitively obvious that an optimal strategy will
use a relatively small braking time and that
braking will take place at relatively low speeds.
On the basis of the above argument we believe
that the precise form of the braking effort will
have little impact on our conclusions. Therefore
in our model we assume that the negative
acceleration during braking is constant.
2.2. The control strategies

To describe
the control
mechanism
we
introduce a control variable j. Each non-negative
value of the control variable determines
a
traction control, and the single negative value
determines a braking control. Let
Ce= (-1, 0, 1,2,. . . , m}

(1)

denote the set of all possible values for j. Let 6


be the fuel supply rate corresponding to the
control setting j, and assume that
O=f-,=fo<fi<f2<...<fm=1.

(2)

Consider a fixed sequence {j(k + I)}k=,,l,...,n of


control settings. We suppose that for k <n the
index j(k + 1) takes integer values between 0
and m from the set %. The semi-final phase is a
coast phase with j(n) = 0 and the final phase is a

Fig. 1. Tractive

effort against speed


locomotive.

for the GM JT26C-2SS

522

P. Howlett

Fig. 2. Power

against

fuel flow rate for the GM JT26C-2SS


locomotive.

brake phase with j(n + 1) = -1.

dt 1
dx=v

This sequence

defines a subset

(8)

u$=?+K,-p(v)+&).

y({j(k + l)]k=o.,.. .J

o=x~x,~x~I...~x,+,,

(4)

where x0 is the starting point, {x~}~=,,~,...,~


are the
switching points, x, + , is the stopping point and
j(k + 1) is the control setting on the interval
(xk, xk+,). We use &+, to denote the length of
the interval (xk, xk+,). The control strategy from
the given subset and with the above switching
points is denoted by
+ 1); (xk, ~~+,)l}k=~.,....,n).

(5)

2.3. The equations of motion


When j z-0, the power developed by the
locomotive is directly proportional to the rate of
fuel supply. When j ~0, there is a constant
negative acceleration applied to the train. If we
write Kj for the braking acceleration then Kj = 0
for j 2 0, and K-, = -K is a negative constant.
The equations of motion for a point mass train
are
dx
-$= n,
(6)

du HJ

-ET+

dt

Kj-

(9)

(3)

of control strategies, each with II + 1 distinct


control phases. Let us define

N[j(k

(6) and (7) are directly applicable


to a
point-mass train, it is important to remark that
similar equations
with a modified average
gradient acceleration can be used to describe a
train with distributed
mass. This issue is
discussed in detail elsewhere (Howlett et al.,
1994a). It is often convenient to rewrite the
equations of motion in the form

P(V) i-g(X),

(7)

where x is the distance along the track, u is the


speed of the train, p(v) is the resistive
acceleration
due to friction,
g(x) is the
gravitational
acceleration
due to the track
gradient and H is a constant. We assume that
p(O) > 0, that p(u) is strictly increasing and that
the graph y = VP(U) is strictly convex. Although

The solution of these equations


Section A.1 of the Appendix.

2.4. The cost function


For
the
control
strategy
S({[j(k + 1);
we
use rk+i to denote the
.n),
(xk! Xk+l)l)k=O,L.
time taken to traverse the interval (xk, xk+,).
The total fuel consumption is given by

J=

(10)

fi(k+l)rk+l.

k=O

2.5. A precise statement of the problem

The problem can now be stated precisely. Let


X be the length of the journey and let T be the
time allowed for the journey. For a fixed control
sequence {j(k + l)}k=O,I....,nr we wish to choose
the switching points {x~}~=,,~,.,.,, to define a
control strategy
S({[j(k + 1);

(xk,

Xk+I)l}k=O.I....,N)
E

W(k

3. Dynamic

braking

curves for
locomotive.

the

GM

l)}k=O.l....,n)t

(11)

with u(xo) = 0, u(x,,+,) = 0, cz=o rk+, = T and


in such a way that J =
c;=O tkt, =X
is
minimised. The correspondC:=Of,(k+l)~k+l
ing strategy is called a strategy of optimal type.
3. TRACK

GRADIENT
ANALYSIS
TERMINOLOGY

AND

Track gradient data is normally stored in


digital form, and hence the true gradient is
represented
as a piecewise-constant
gradient.
For both theoretical and computational reasons,
it is convenient to use the gradient data in this
form. It has been shown (Howlett et al., 1992)
that solutions to the equation of motion can be
computed to any desired accuracy using a
piecewise-constant
approximation
to the true
gradient.
Subdivide the interval [0, X] by setting
O=ho<h,<hz<...<h,,+,=X,

Fig.

is discussed in

(12)

JT26C-2%

and suppose that the track gradient is constant

523

Train control
on each subinterval (h,, h,+,). We consider a
fixed control sequence {j(k + l)}k=O,l,...,n and a
general strategy
+ 1); (x/C,&+l)lL=O,l,...,n)

S(C(k

effective speed at the point xk is denoted by


clr, = ?$k(&). For the subset Y({j(k + l)},d,J,...,~)
of control strategies, the following conditions are
necessary for a strategy of optimal type. There
are non-negative constants A and /.Lsuch that

E W(k + l)L=w,...,n). (13)


(19)

Equation (9) is now written in the following way.


When x E (xk, xk+,) tl (h,, h,+l), we have

$,

fG+v ;

dx

K,

u
=

I(k+l)

-P(u) + &

and, for each k = 1,2, . . . , n - 2,


(14)

F[k,r](u),

where g, is the acceleration due to the gradient


on the interval (h,, h,+l). We could, for example,
assume that g, =g(h,),
where g(x) is the
acceleration due to the known track gradient.
Ultimately we shall need to specify the locations
of the variable points {xk}k=l,Z,...,nin relation to
the fixed points {hr}r~0,1,2,...,p+l. We use the
notation {r(k)}k=O,l,...,n to denote the essentially
unknown sequence with r(0) = 0 and r(n + 1) =
p and with
(15)

for each k = 1,2, . . . , n.

4. THE MAIN RESULTS

4.1. The key equations


To explain the main results, it is necessary to
introduce some more terminology.
Use u, =
u(h,)
to denote
the speed at x = h, and
vk = u(xk) to denote the speed at x = xk. For
each k=1,2,...
, n and for r(k) <s 5 r(k + l),
define

For each r with r(k) 5 I < r(k + l), define


r(k+l)
93 [k,rl

n
s=r+l

(17)

Rm

and let Sk = %lk,r(k)]. When X E [Xk, Xk+,], the


effective speed ?$j(x) at the point x is defined by
vk(x)

v(X)

for

-=-

[hr(k+,p

xk+,]

and

Hfi(k+l)]

F[k.r(k,](vk)

Irk

9k

b + H.&k+l)l
-

Vk+l

F[k,r(k+l),(Vk+l).

(21)

These equations will be called the key equations,


and we shall show how they can be solved to
determine the strategies of optimal type. We
emphasise
that the key equations
provide
necessary conditions for the strategies of optimal
type, but in general may not provide sufficient
conditions. On the other hand, we believe that in
many specific situations where additional information is known the key equations will also be
sufficient. This is certainly the case if the train
speed increases during each power phase and
decreases during each coast phase. On the other
hand, there may be situations where maximum
power is applied but the speed decreases, or
situations where the coast control is used but
speed increases. In such situations we say that
the track is steep. For a track that contains
isolated steep sections, we assume that the train
will use maximum power on the steep inclines
and will coast on the steep declines. With this
additional assumption, the key equations now
have a unique solution, and can be used to find
the appropriate
switching points immediately
before and after each steep section. For this
reason, in a slight abuse of terminology, we shall
refer to the strategies obtained by solving the
key equations as strategies of optimal type. The
equations are derived in Section A.6 of the
Appendix.
4.2. An alternative form for the key equations
To solve the key equations, it is convenient to
rewrite them in an alternative form. We make
some additional definitions. For u > 0. let

by

---

[CL+

r(k+l)-I

2,

for x E [h,, h,+,]

&kkl+&)

and

qJ)

(18)
r(k) s r < r(k + 1). The

and define

= f+ P(U),

a function

(22)

{8&}&x) in the region

524

P. Howlett

x E [xk, ++,I by the formula {%~}&) = E,[u(x)]


for x E [hrtktl), Q+,] and by

r(k+l)-1
+

,;+,

&F&4 - ~,(Us+~)l +EI*(Ur(k+IJ


.s

(23)
for x E [h,, h,+l] and r(k) 5 r < r(k + 1). The
function {8+},(x) can be regarded
as the
effective energy density for the control strategy.
The key equations can now be rewritten as
follows. There are non-negative constants A and
p such that x, < X is the solution to
(24)
x,_ 1< x, is the solution to
(25)
and in general xk < & +,
@+.)k(X)

iS

the

SOhtiOn

= &tVk+,)

t0

(26)

for each k = 1,2, . . . , II - 2. On level track we


note that vk(x) = u and {%p}k(X)= E,(u) where
u = u(x). Since g,(,, = 0, it can be seen that our
key equations generalise the key equations for
level track (Cheng and Howlett, 1992).
4.3. The strategies of optimal type
It can be shown that the function {%p}k(r) is
well defined and continuous with a continuous
derivative at all points in the region x E
[xk, xk+,]. In particular, the derivative is given by
the formula

{up};

=&&y@,
F

(27)

when x E [h,, h,+,] and where u = u(x). From


this formula and the convexity of the graph
y = E,(u), it can be seen that the function
{CG;r}k(~)has a local turning point only when
u = We, where w, is the unique minimum turning
point for the graph y = E,(u). Note that the
critical speed wfi satisfies the equation
++$L)

= p.

(28)

We shall show that We defines an approximate


holding speed. These observations are crucial to
our understanding of the solutions to the key
equations. A detailed solution algorithm is
reviewed in Section A.3 of the Appendix.
If power is applied to the train on the interval

and if we assume that the speed


txk>
xk+l)
increases throughout the interval then there is
precisely one solution xk < xk+* to (26) with
u(xk) = vk < wp < I$+, = i+&+l). If the coast
control is used and we assume that the speed
decreases throughout
the interval then once
again there is precisely One SdUtiOn xk < xk+, ,
but in this case u(xk) = uk > w, > v,,, = u(&+,).
In each of the above cases the interval (xk, xk+,)
is uniquely determined.
Therefore, when the gradient is not too large,
we can expect the speed to oscillate about the
critical value w, as the control is switched
between power and coast. As the number of
phases is increased it is intuitively reasonable to
suggest that the length of each phase is reduced
and hence the critical speed w, can be
interpreted as an approximate holding speed.
Thus, for non-steep track, the strategy of
optimal type will contain an initial power phase,
followed by an approximate speedhold phase
near the critical speed We, a semi-final coast
phase and a final brake phase.
In practice, the track gradient may be
sufficiently large to disrupt this general pattern.
Thus we may have situations where maximum
power is applied to the train but the speed
decreases. On the other hand, there may be
situations where the coast control is used and the
speed increases. When these situations occur, we
shall say that the track is steep. Note that this
definition is relative to the actual speed of the
train, and whether or not a track is classified as
steep will depend on the desired holding speed.
In such cases it is possible that there will be
more than one solution to the key equations, and
consequently
it is possible that the interval
xk,
xk+
I)
Will
not
be UniqUdy
determined. For
(
this reason, we reiterate that the key equations
give necessary conditions for a strategy of
optimal type but that these conditions may not
be sufficient. In practice, it is often relatively
easy to decide which solutions are superfluous.
Nevertheless, on steep track the strategy of
optimal type will have the same form as on
non-steep track, except that the approximate
speedhold phase may be disrupted by phases of
maximum power in the vicinity of steep inclines
and by phases of coast in the vicinity of steep
declines. These disruptions can be interpreted as
necessary adjustments to keep the speed as near
as possible to the desired holding speed. The
theoretical difficulties associated with steep track
are not discussed in this paper.
We shall illustrate our work with several
realistic examples that give some indication of
the general nature of a strategy of optimal type.
In Section A.5 of the Appendix it is shown

525

Train control
that any segment of non-negative measurable
control can be approximated by a sequence of
coast-power
pairs. Thus we can construct a
strategy of optimal type with approximately
minimum fuel consumption.
5. EXAMPLES

FOR NON-STEEP

TRACK

The following examples are based on data


obtained
from train models used by the
Scheduling and Control Group at the University
of South Australia. Length is measured in metres
and time in seconds. We consider a journey with
X = 18 000 and T = 1500. We assume that
p(u) = a + bu + cu*,

(29)

where a=1.5~10-*,
b=3X10P5
and C=~X
lo-. We take H = 1.5 and K = 1, and assume
only two allowable rates of fuel supply with
f0 = 0 and fi = 1. For each x E [0, X], we shall
take
g(x) = ax(X - 2x)(X - x),
(30)
where (Yis a constant. Outside the interval [0, X]
we assume that g(x) = 0. In all of the examples
the value of (Yis so small that the gradient would
not be readily apparent to the naked eye. In fact,
the height of the track is determined by the
formula

h(x)=-gJ
5(X .

25)(X - 5) d5

-&*(x-x)

(31)

for x E [0, X], with h(x) = 0 otherwise. Hence, if


la1 = 2 x 10-14, we have @(x)1 ~6.7 for all x.
Thus in a total journey of 18 km there is a rise or
fall of only 6.7 m. The strategies of optimal type
in these examples all have the basic power,
approximate speedhold, coast, brake form, with
the approximate speedhold segment constructed
using alternate phases of coast and power.
Nevertheless, we shall show that even these very
small gradients have a dramatic effect on the
position of the switching points, and in particular
that the extent of the crucial semi-final coast
phase is drastically changed.
There are some important general observations that should be made at this stage. There is
only a small decrease in the fuel consumption
when a strategy of optimal type using only one
coast-power pair is replaced by a strategy of
optimal type using nine coast-power pairs. Thus
a seemingly rudimentary approximation to the
idealised minimum cost strategy can be very
energy-efficient. Hence, in practice, we do not
need a large number of coast-power control
pairs to construct an approximate speedhold

strategy. Indeed, the examples confirm that in


practical terms coast-poiver control is almost as
close as we please to continuous control.
In our examples we have used very simple
gradient profiles. There are two critical factors
that determine
an energy-efficient
strategy.
Firstly, it is necessary to keep the train speed
close to the selected holding speed during the
approximate speedhold phase, and secondly it is
important to choose the correct switching point
to begin the semi-final coust phase. We have
found that manual selection of the switching
points is difficult because of extreme sensitivity
to small changes in gradient. In our simple
examples selection of the correct switching point
to begin the final phase is extremely important.
On metropolitan
railcars with apparently flat
track and relatively small distances between
stations we have found that even the most
experienced drivers could not choose this point
effectively.
This was demonstrated
in our
Metromiser trials, where audited fuel savings of
fourteen
percent and improved timekeeping
were achieved when Metromiser was used to
select the point where the semi-final coust phase
begins.
5.1. Strategies of optimal type on level track
We begin by considering a level track. Thus
we let (Y= 0. We consider a strategy with one
coast-power
pair and a strategy with nine
coast-power pairs. It is significant for practical
purposes that fuel consumption is reduced only
marginally by the more elaborate strategy.
Example
1. (One coast-power pair on level
truck.) The detailed results are shown in Table

1. We set A = 2.05969 X lop2 and p = 5.51350 X


10w2. The critical speed is given by wfl = 15.832
and the fuel consumption by J = 202.15.
Example 2. (Nine coast-power pairs on level
truck.) The detailed results are shown in Table

2. We set h = 2.06798 X 10e2 and p = 5.85768 X


lo-*. The critical speed is given by w, = 16.170
and the fuel consumption by J = 201.89. The
speed profile is sketched in Fig. 4.
Table 1. Strategy of optimal type with one COW-powerpair
on level track
k

j(k + 1)

1
0
1
0
-1

1
2
3
4
5

tk

0.000
134.060
440.430
508.525
1497.363
1500.000

xk

0.000
16%.581
6 583.762
7 684.230
17 996.470
18000.000

vk

0.000
18.588
13.374
18.588
2.677
0.000

526
Table

P. Howlett
2. Strategy

i(k + 1)

fh

1
0
1
0
I

0.000
104.513
152.528
163.348
211.364

623.214
634.034
1 497.209

i8
19
20
21

of optimal
type with
pairs on level track

0
-1

nine

J-k

15oo.ooo

0.000
1 176.226
1 952.765
2 127.808
2 904.347
9 585.425
9 740.468
17 996.048
18000.000

coasf-power

Table 3. Level track strategy


i(k + 1)

th

0
l
2
3
4

1
0
I
0
1

0.000
104.085
151.024
161.468
207.116

0.000
1 176.226
1 952.765
2 127.808
2 904.347

o.ooo
16.778
16.331
17.181
16.853

567.857
577.188
1 363.733
1 363.733

9 585.425
9 740.468
16 988.205
16 988.205

18.477
19.038
0.000
0.000

0.000
16.583
15.764
16.583
15.764

lb
19
20
21

0
-1

xh

Vk

ff = (-2) x lo-14. The switching points will be


the same points used in Example 2.

5.2. Level track strategies applied to non-level


track with small gradients

We consider essentially the same problem on


non-level track, but with very small gradients.
The gradients are so small that they would not
be apparent to the naked eye. We shall show
what happens if the driver attempts to use a
strategy designed to be optimal on a level track.
There are three alternative ways in which the
level track strategy could be implemented. The
most natural way is to select the same switching
points, but it is also possible to select the same
switching times or the same switching speeds. In
all cases the strategies can be shown to be
incorrect.
5.2.1. A small valley. In the first case we
consider a small valley. To define a track with an
initial downhill section and a final uphill section
we let (Y= 2 X 10-14. The switching points will be
the same points used in Example 2.
Example 3. (Level track strategy over a small
valley). The strategy is not feasible and the final

phase is degenerate.

Vh

15.764
16.583
2.833
0.00

over a small valley

The train stops at time

Example 4. (Level
hill.) The strategy

track strategy over a small

is not feasible and is not


energy-efficient.
The train stops at time t =
1461.125 and at distance x = 18 028.070, and the
cost of the strategy is J = 218.18. To make this
strategy feasible, it would be necessary to
decrease the time spent under power. The
detailed results are given in Table 4.
5.3. Strategies of optimal type on non-level track
with small gradients

Once again, we consider essentially the same


problem on non-level track. This time, however,
we calculate strategies of optimal type.
5.3.1. A small valley. As before, we let
cr = 2 x 10p14. We consider a strategy with nine
coast-power
pairs. Although
we have not
presented the detailed calculations, it is significant for practical purposes that fuel consumption is reduced only marginally by using a
large number of coast-power pairs.

t = 1363.733 and at distance x = 16 988.205, and

Example 5. (Strategy of optimal type over a


small valley.) The results are shown in Table 5.

the cost of the journey is J = 19058. To make


this strategy feasible, it would be necessary to
extend the time spent under power. The detailed
results are given in Table 3.
5.2.2. A small hill. We now consider a small
hill. To define a track with an initial uphill
section and a final downhill section, we let

We set A = 2.25335 X lo-* and p = 4.76702 X


lo-. The critical speed is given by We = 15.047
and the fuel consumption by J = 200.18. The
speed profile is sketched in Fig. 5.
5.3.2. A small hill. As before, we let (Y=
(-2) x 1op4. We consider a strategy with nine
Table 4. Level track strategy

k
~~___

16.58
15.76

Fig. 4. Stylised

0
I
2
3
4

speed profile

of optimal

type on level track.

18
19
20
21

over a small hill


-r!f

i(k + 1)

1,

1
0
1
0

0.000
104.948
154.124
165.361
216.156

0.000
~1 176.226
I 952.765
2 127.808
2 904.347

0.000
16.386
15.177
15.967
14.599

704.273
717.137
1453.182
1461.125

9 565.425
9 740.468
17 996.048
18 028.070

12.480
14.214
8.064
0.000

i
0
-1

vh

527

Train control
Table 6. Strategy of optimal type over a small hill

Table 5. Strategy of optimal type over a small valley


k

j(k + 1)

0
1
2
3
4

1
0
1
0
1

0.000
88.986
185.874
194.573
328.719

0.000
931.336
2 389.459
2 520.418
4 539.073

0.000
15.533
14.647
15.454
14.666

i8
19
20
21

i
0
-1

ssi.105
897.321
1497.917
1500.000

12 853.961
13 098.264
17 997.797
18000.000

14.k
15.642
2.116
0.000

tk

x/e

vk

i(k + 1)

is
19
20
21

Example 6. (Strategy of optimal type over a


small hill). The results are shown in Table 6. We

set A = 1.60209 X lo-* and p = 9.08786 X lo-.


The critical speed is given by wP = 18.839 and
the fuel consumption by J = 209.52. The speed
profile is sketched in Fig. 6.
5.3.3. A brief comparison of the strategies of
optimal type. The most significant difference in
the strategies of optimal type described in the
above examples is the length of the semi-final
coast phase. In a total journey of 18 km we have
a semi-final coast phase of 4.3 km when
travelling over the small valley and a semi-final
coast phase of 12.7 km when travelling over the
small hill. Since the rise and fall of the track in
each case is less than 6.7 m, it is clear that this
difference could not be determined precisely by
the driver alone. Because there is no fuel
consumption while coasting we can see that an
error in estimating the length of the semi-final
coast phase can have a significant effect on the
cost.
We have also shown that ad hoc strategies,
even those with the correct overall structure, can
result in non-feasible journeys or journeys where
fuel consumption is increased significantly.

vk

xk

1
0
1
0
1

0.000
148.575
161.447
168.021
180.563

0.000
1952.211
2 194.708
2 318.567
2 5!4.841

0.000
19.011
18.667
19.012
18.666

i
0
-1

313.289
320.002
1494.409
1500.000

5 055.356
5 181.827
17 984.142
18 OOO.ooo

18.k66
19.013
5.673
0.000

6. AN EXAMPLE
WITH LARGE
GRADIENTS

coast-power

pairs. Although
we have not
presented the detailed calculations, we again
note for practical purposes that fuel consumption
is reduced only marginally by using a large
number of coast-power pairs.

tk

TRACK

In practice, there are many sections of a track


on a typical long-haul journey where the
gradients are classified as steep. Even in an
apparently flat continent such as Australia, this is
the case. Under these conditions, the predominant speedhold mode is interrupted
by
segments of coast and power. With complicated
track gradient profiles, there may be a number of
extended coast or power phases. Selection of
appropriate switching points will be important in
defining the extent of each of these phases.
Inappropriate selection will mean that excessive
fuel is consumed in returning the train to the
desired holding speed.
The following example shows a strategy of
optimal type on a track with an isolated steep
incline. We consider a journey with X = 50 000
and T = 3000. As before, we assume that
p(v) = a + bv + cv*,

(32)

where a=1.5XlO-*,
b=3X10P5
and c=6x
10e6. We take H = 1.5 and K = 1, and assume
only two allowable rates of fuel supply, with
f0 = 0 and f, = 1. For x E [0, X], we take

g(x) =
1

0
-0.1

if
if

O<x<20000,
20000<x<25000,

if

25OOO<x<X,

(33)

and outside the interval [0,X] we assume that

0
Fig. 5. Stylised speed profile of optimal type over a small
valley.

XZ

Fig. 6. Stylised speed profile of optimal type over a small


hill.

528

P. Howlett

g(x) = 0. We consider a strategy with nine


coast-power pairs. We observe that the strategy
of optimal
type contains
an approximate
speedhold segment that is interrupted by a power
phase in the vicinity of the steep incline. At an
intuitive level, it seems reasonable to remark
that the control segments are chosen so that the
average speed over the steep interval is kept as
close as possible to the desired holding speed. In
the case of a steep incline this is done by
increasing the train speed before the climb is
commenced. Of course, this may seem intuitively
reasonable, but our point once again is that
correct selection of the appropriate switching
points is very difficult and that small mistakes
can result in a dramatic error such as stalling or
in the consumption of excess fuel.
Example 7. (Chengs climb). The detailed results are shown in Table 7. We set A =
2.3030148 X lo- and p = 10.03085 X 10m2. The
critical speed is given by w, = 19.495 and the
fuel consumption
by J = 924.39. The speed
profile is sketched in Fig. 7.
7. CONCLUSIONS

AND

FUTURE

DEVELOPMENTS

For a prescribed sequence of fuel supply rates,


it has been shown that a strategy of optimal type
depends on two real-number parameters. These
determine the switching points and ultimately
determine the distance travelled by the train and
the time taken for the journey. By adjusting the
values of the parameters, it has been possible to
construct a feasible strategy in several realistic
examples.
We have found a minimum-cost strategy given
that a prescribed sequence of fuel supply rates
must be used. We have shown that even a small
track gradient exerts a dramatic influence on the
switching points and on the corresponding
switching speeds. The solution given in this

Table
k

7. Strategy
;(k + 1)

II
1
2
3
4

I
0
I
0
I

4
10
II
12

0
1
0
I

18
19
20
21

i
0
-I

of optimal
fh

type for Chenga


_Lk

climb
I/A

O.WO
2 390.834
4 X75.674
5 630.6 I3
x 115.454

0.000
20.656
I X.380
20.656
18.380

x3j.2x.i
960.636
1425.214
I S37.356

I5 34Y.95 I
17 834.7) 1
26 555.060
2X 742.692

2O.hSh
18.380
20.5 15
IX.51 1

1975.736
2009.721
2YYS.709
3OOO.000

37 247.4 I2
37.Y61.353
4Y 990.656
so 000.000

IX.51 1
20.5 IS
4.356
O.OOO

0.000
16Y.396
2Yh.749
335.367
462.72 1

1I

23.75

Fig. 7. Stylised

speed

profile of optimal
climb.

type

for Chengs

paper is believed to be the first effective solution


of this important practical problem.
Simulated
performance
of our proposed
algorithm shows that the practical impact of our
methods will be in the affirmation that, even on
undulating track, a speedholding strategy is the
most energy-efficient strategy. If the locomotive
is unable to maintain a speedholding strategy
over steep terrain then correct selection of the
regions where power should be applied or where
coasting should be used will result in significant
fuel savings.
One further point is also worthy of comment.
The argument (Cheng and Howlett, 1993) that
intermediate throttle settings are less efficient is,
in essence, still valid on tracks with non-constant
gradient. Although we have assumed a linear
relationship between power and fuel supply, we
also point out that in practice the maximum fuel
supply rate is usually the most power-efficient
control
setting. Thus intermediate
settings
should be used only if speed fluctuations are
avoided and if the power efficiency of the
intermediate
setting is comparable
with the
power efficiency at the maximum setting.
It is important to emphasise that the driving
strategies recommended in this paper are not
used effectively in practice. Where automatic
control systems are used, the strategies are much
more concerned with safe driving practices. We
believe our strategies can be effectively used in
conjunction
with such systems to produce
significant fuel savings. Where automatic control
systems are not used, it is demonstrably wrong to
assume that an experienced driver can make
effective judgements about the timing of control
changes.
Although we have shown that our solution can
be efficiently computed,
there are several
practical problems that remain.
Some of these pose no serious difficulty. For
example, in this paper we have nominally
considered a point-mass train. By considering a
real train as a distributed mass, we can construct
a modified gradient proofile that allows us to
treat the train as if it were a point mass. This

Train control
matter is discussed in another paper (Howlett et
al., 1994a). It is also anticipated that the analysis
can be extended to incorporate speed limits. At
this stage it is not easy to decide whether the
difficulties posed by steep gradients will be a
serious compututational
problem. It should be
pointed out that the difficulties do not relate to
isolated steep grades. Difficulties arise when
individual steep grades are so close together that
it is not easy to decide whether the grades should
be treated separately
or as a group. The
difficulties can usually be resolved by personal
intervention,
but is not so easy to do this
automatically.
The Scheduling and Control Group have
developed a prototype computer program that
calculates the idealised strategy of optimal type
for any given journey. This program has been
used with realistic data supplied by Australian
National to calculate minimum cost strategies for
a typical long-haul journey.
would like to thank my colleague
Peter Pudney for his assistance with the numerical
calculations required for the examples. I should also like to
thank the other members of the Scheduling and Control
Group at the University of South Australia.

Acknowledgements-I

REFERENCES
Asnis, I. A., A. V. Dmitruk and N. P. Osmolovskii (1985).
Solution of the problem of the energetically optimal
control of the motion of a train by the maximum principle.
USSR Comput.

Maths Math. Phys., 25(6), 37-44.

Benjamin, B. R., A. M. Long, I. P. Milroy, R. L. Payne and


P. J. Pudney (1987). Control of railway vehicles for energy
conservation and improved timekeeping. In Proc. of Conf
on Railway
Engineering,
Institute
of Engineering
Australia, Perth, Western Australia, pp. 41-47.

of

Benjamin, B. R., I. P. Milroy and P. J. Pudney (1989).


Energy-efficient operation of long-haul trains. In Proc. 4th
International
Heavy Haul
Engineers
of Australia,

Railway

Brisbane,

Conf,

Institute

Queensland,

of

pp.

369-372.
Cheng Jiaxin and P. Howlett (1990a). Critical velocities for
the minimisation of fuel consumption in the control of
trains. University of South Australia, School of Mathematics, Report 1.
Cheng Jiaxin and P. Howlett (1990b). Optimal strategies for
the minimisation of fuel consumption in the control of
trains. University of South Australia, School of Mathematics, Report 3.
Cheng Jiaxin and P. Howlett (1992). Application of critical
velocities to the minimisation of fuel consumption in the
control of trains. Automatica, 28, 165-169.
Cheng Jiaxin and P. Howlett (1993). A note on the
calculation of optimal strategies for the minimisation of
fuel consumption in the control of trains. IEEE Trans.
Autom. Control, AC-38,1730-1734.
Howlett, P. (1984). The optimal control of a train. University
of South Australia, Study Leave Report.
Howlett, P. (1988a). Existence of an optimal strategy for the
control of a train. University of South Australia, School of
Mathematics, Report 3.
Howlett. P. (1988b). Necessary conditions on an optimal
strategy for the control of a train. University of South
Australia, School of Mathematics, Report 4.
Howlett, P. (1990). An optimal strategy for the control of a
train. J. Aust. Math. Sot., Ser. B, 31,454-471.

Howlett, P. (1993). Determination of optimal strategies for


the control of a train on a track with non-constant
gradient. University of South Australia, School of
Mathematics, Report 7.
Howlett, P., P. Pudney and B. Benjamin (1992). Determination of optimal driving strategies for the control of a train.
In B. J. Noye, B. R. Benjamin and L. H. Colgan (Eds),
Proc. Computational
Techniques and applications,
91, pp. 241-248. Computational
Mathematics

CTAC-

Group,
Australia Mathematical Society.
Howlett, P. G., I. P. Milroy and P. J. Pudney (1994a).
Energy-efficient
train
control.
Control
Engineering
Practice, 2, 193-200.

Howlett, P. G., J. Cheng and P. J. Pudney (1994b). Optimal


strategies for energy-efficient train control. In Proc. SIAM
Symp. on Control Problems in Industry, San Diego, CA,
Birkhauser, Boston.
Kraft, K. H. and E. Schnieder (1981). Optimale Trajektorien
im spurgebundenen Schnellverkehr (Optimal trajectories
for rapid transit systems). Regelungstechnik, 29.
Milroy, I. P. (1980). Aspects of automatic train control. PhD
thesis, Loughborough University, UK.

APPENDIX

A.l. The speed profiles


In this subsection we define functions that satisfy the
equations of motion. For each j = 0,1, . , m, let WIj,rl be
the unique solution to the equation Hfi - u[ p(u) - gr] = 0. If
the fuel supply rate is J and if the acceleration due to the
track gradient is g, then W(,.,) is the speed at which the net
acceleration is zero. Choose W such that W > WE,,) for all
[j, r]. A direct integration of the equation of motion can now
be used to define distance as a function of speed.
For k<n we set

u-=wdw

I
0

q*.r](u)

for u o ]O, W[j(k+l),r]),

%1(w)

(A.11

w(-l)w
~

dw

for u o (WIj(k+l).ll, WI.

F[k.r](w)

In general, the value of u, determines which of the above


definitions is used. Because the graph y = up(u) has positive
slope when u > 0, it is easy to show that xIk,,l(v) + m as
u+ Wcj(e+t).r). It now follows that if u,< W(++,),,) then
v,+, < Wtj(k+lr,rl. Alternatively if u, > WU(k+lj,rl then v,+, >
Wl,o+Ij,rl. Whichever of the above definitions is used, we
denote the inverse function by uIk,,l.
When k = n, we define
W(-l)wdw
~
%rl(W)

-qn,r,(v)=

for v E (0, W]

(A.2)

and let ul,,,l denote the inverse function. We assume that


K >g, for all r.
For all k, we claim that the actual speed on the interval
(xk, xk + i) is given by

-,k.r(k,](x

xk

+ X[k.r(k)](Vk))
for

u(x) =

yk,r](x

hr

for
v[k.,(k+l)]@

cxk,

h,(k)+,)>

@,,

hr+d

qk.rl(ur))

hr(k+,j
for

f-l

bkr

Xk+dt

(A3)

X[k.,(k+,)](,(k+l)))

X E (hr(k+,),Xk+,).

Note that

Xik.,](V)=-

uik,rl(X) = F,k;;;t$))

F[k.rl(u)'

and also that

the time

taken

to traverse

(A.4)
the interval

P. Howlett
given and that we seek a strategy of optimal type from the
subset Y({j(k + l)}k=,,,l, ,,,). Following
the work of Cheng
and Hewlett (1993). we assume that n is even and that
m

;(k+l)-

(A.5)

(A.6)

(0

if k is even,

(A.16)

if k is odd

for k <n. We stress that this assumption


is not a necessary
part of the proposed
algorithm
but note that in seeking a
strategy
that minimises
fuel consumption
it is sufficient to
restrict our attention
to control subsets of this form. This
point is further explained in Section AS.
We assume that the track is not steep. That is, we assume
that the speed increases
during
each power
phase and
decreases
during each coast or brake phase. Although
the
proposed
algorithms
can still be used on steep track, the
existence and uniqueness of solutions is a more difficult issue.
The speed profile can be calculated
in the following way.
In general.
we consider an interval [xkr x,,,]
and suppose
that V, i , and xk +, are known. In the first case we suppose
that j(k+l)=m.
For each IE[~,(~+,),x~+,],
we define
u = V,,,,,,+,,,(x)
as the solution to the equation

when k =n. Since VI = V,(.$,, &.


, &), it follows that
u, = u,(&, &.
, &,) for each r with r(k)irsr(k
-+ I),
and hence r,, , = rk+,(5,, &,
, &I.
A.2. Thr construinrs
Since the initial and final speeds
that
v, = v,,

are zero.

it is necessary

(A.17)
and in particular
we note that I,,(* I ,) = V,k,rCk+,),(hrCk+ ,)).
We now use a recursive definition for u, when r <r(k + 1).
We define v = u, as the solution to the equation
I\*,

+/ = 0.

wdw

-=h,,,

(A.7)

It is important
to realise that the variable
t,, i, can be
eliminated
from the problem using the equation
V,, +, = 0. In
fact, we deduce that

r(k+l)p2,....r.
u = V,k,rI(x) as the

for

each
.s=r(k+I)-1.
]h,. h, +I 19 we define
equation

(A.18)
For
solution

to

xt
the

(A.]))

and hence

the distance

can be rewritten

(5, , !L,.

constraint

in the form

h ,(,, / 1) + q,u(,,,
Since

Note that u, = VIk,rl(h,), u, / , = V,*,,,(h,+,) and that u, < u <


u, , , for I E (h,. h,, ,). A similar
calculation
procedure
applies when j(k + 1) = 0 or j(k + 1) = -1, but in each of
these cases v, > IJ > u,,, for x t (h,, h,, 0.
To solve the key equations
(24)-(26).
we begin by noting
the useful recursive relationships

1,,(O) .- xI,,,,(,, , , ,I(ue, +IJ = X.


depends

only

I,I(% +I,I
, [,,I, we can define

X[,,,,(,, i

on

(A.101
[ -

45) = k,,, +,, + ~~l,,.r,,,


+,,,(O)- .rl,,,r(,,t I,~(+,, +I,). (A.1 1)
in which case the distance

constraint

becomes

X(5) = x.
The time constraint

is expressed

simply

&A+&

(A.12)
as
(A.13)

{r,},(h,)

=&[JuJ
r

~ E,(u,+,)l

+{%(hri,).

The key equations


can now be solved as follows. We
assume that the parameters
A and p are known. To begin the
calculation,
we consider the final interval [x,,, x,,,,]. We note
that x,, +, = X and V,, +, = 0. We then apply the recursive
procedure
described below.
In general, we consider an interval [xk, I~+,], where x/, , ,
and V, , , are known from the previous stage but where xk
and V, are unknown. We use the recursive calculation
(54) to
find r = r(k) such that
(A.22)

(r;,},(h,)>E,(V,.,)>(~~},(h,+,).
Since r, , , = r, +,([, , &.
only on 5. we can define

, & , , ) and since {,, , , depends

t(t)=
and the time constraint

i r,*,.
!. -0

If we assume in the first instance


that j(k + 1) = m then
u, <u,+,.
Now we find u = V, t [u,~,,.,, urCkJ such that

(A.14)
(A.23)

becomes
r(t) = T.

(A.211

(A.15)

A.3. An algorithm for solving the kyv equurions


In this section we shall review an algorithm for solution of
the key equations.
This algorithm has been fully described in
a separate paper (Howlett et al.. 1992), but it is convenient
to
present the basic ideas here.
We suppose that the control sequence {j(k + I)}, (,,,, ,, is

Note that V,<w,<V,.,.


second
instance
that
u = V, E [u,(~,+,. urckI).
Ih ,(I )( h,,k,+t) from

Alternatively,
if we assume in the
j(k + I ) = 0 then
IJ, > II,+,
and
we
calculate
x =I~ E
Finally.
w dw

F b(k)(W)
Note that the general

procedure

must be modified

(A.24)
on the first

531

Train control
two intervals [x,, x,+,1 and [xn_, , x,,], where the form of the
key equations is atypical.
This calculation determines a strategy of optimal type. It is
necessary to adjust the parameters A and p to obtain a
feasible strategy. It is anticipated that these adjustment
procedures can be adapted from previous work (Cheng and
Howlett, 1992). In general, we note that the critical holding
speed w& can be estimated approximately from the average
speed for the journey. In fact, w& will be greater than the
average speed. Thus we could use an initial estimate for p
given by
(A.25)
On level track the value of A must be greater than E,(w,),
and so we can begin with

By comparing the two equations, it follows that


u*+,

](9k - u,4) + &+I -

Iu*

rk)l-

f dt

[%.iP(%,i) -

and, by choosing a sufficiently fine subdivision, we can ensure


that uO., is as close as we please to u, with
(A.34)

(4k - uk) + &+I - rk) It follows that

(A.26)

A,,, = EJ+S,(W*eS,)
+ 6.

The Scheduling and Control Group has developed a


prototype computer program that is designed to perform the
above calculations on board a long-haul freight train in real
time.
A.4. The properties of { %P}k(~)
In this section we simply emphasise that the function
{9}&) is continuous and has a continuous derivative in the
region x E [xk, Q+,]. When x E (h,, h,+,), it is obvious that
these properties are true and that
(A.27)
where u = u(x). By using this formula and considering the
appropriate left- and right-hand limits, it is easy to show that
{S&,h(x) and {5$,};(x) are continuous at x = h,.
A.5 Approximate coast-power strategies
It has been shown elsewhere (Howlett et al., 19946) that
any time interval [u,, ub] of non-negative measurable control
f(t) can be approximated as accurately as we please by a
sequence of coast-power
controls. It is convenient to review
the argument. In essence, we replace the actual speed profile
on each small time subinterval [uk, u*+J by the speed profile
resulting from a power-coast-power
strategy, where the
control changes are timed so that the distance travelled in the
given interval is the same as the distance travelled with the
original control. We need to show that the fuel consumption
for the alternative strategy is essentially the same as the fuel
consumption for the original strategy.
On each time interval [uk, uk+,] we have
(A.28)
AJO., = (qk - u,J + (uk+] - rk),

(A.29)

where [ukr ~~1, [qk, rk] and [rkr uk+i] are the time
subintervals for the new power-coast-power
controls. By
integrating the appropriate equations of motion, we obtain
equivalent energy balance equations. If we write
G(x)=-lg&=-Igudt

(A.30)

and hence the fuel consumption for the alternative strategy


can be made arbitrarily close to the fuel consumption for the
original strategy.
A.6. Necessary conditions for a strategy of optimal type
We define a Lagrangian function of the form
S5

A, IL) = HJ(5) + A]X -r(5)]

$=O
k
and the complementary
A[X -x(t)]

(A.36)

for all k

slackness conditions
= 0,

p[t([)

- T] = 0.

(A.38)

If we weaken the equality constraints to read x(t) 2 X and


t(c)5 T then we can also guarantee that A and p are
non-negative. It is intuitively obvious that the weakened
problem has the same solution. We shall assume that 5 lies in
an open set such that V,(t) # WI,~k+,~,r~k~l and such that
u,(.$) # W,,Ck+,j,rl for each r with r(k) + 1 r r 5 (k + 1) and
each k = 1,2,
. , n.
In order to apply the Kuhn-Tucker
equations, it is
necessary to calculate a number of partial derivatives.
Because the relevant formulas are rather complicated, it has
been convenient to introduce a number of gradient-weighted
averages that enable us to write these expressions in a
reasonably
coherent
way. Direct application
of the
Kuhn-Tucker conditions gives a set of equations that can be
simplified by an essentially inductive argument. An outline of
the argument is given below. Details of the derivation are
quite tedious, and are contained in an expanded version of
this paper, which is available as a University of South
Australia Research Report (Howlett, 1993).
For convenience, we shall use the notation

The condition
[Hf -u,&)]

dt,

(A.31)

atO
z-

and for the alternative strategy we have

(A4

can be simplified to give

]:u& * G(x,,)ll::+l = H](qk - UL)


u/l+,
+ (ak+l -

- T],

(A.39)

k+I
I t

+ ME)

where J(t) is the cost of the journey, x(E) is the distance


travelled, t(t) is the time taken and A and p are Lagrange
multipliers. To find the minimum cost strategy for the given
sequence of throttle settings, we apply the Kuhn-Tucker
conditions

then for the original strategy we have


[:u; + G(x~)]~::+ =

UjP(Uf)I
dt, (A.33)

~O.d~O.ddf. (A.32)
r.dl - I
Uk

(-l)A+p+=O.

(A.41)

532

P. Howlett
If we assume that

When the condition


(A.42)

(1 -

Qd%[ (p +Hf;v+,d(~-+-j
+&r(h+l)lWh+l)I =

is combined with (A.41) we can deduce that

H[f;(h+t~-hd
v,

(A 46)

for each h with k < h <n - 2 then the additional equation


(A.43)
In similar fashion. we can combine the condition

a$_,
as,-

(A.47)

allows us to deduce that

with (A.41) and (A.43) to deduce that

(A.45)

Therefore the hypothesis is established for h = k. We can


now deduce that (A&) is valid for all h = 1,2,.
, n - 2. It
is a straightforward matter to show that (A.41), (A.43) and
(A.48) can be rewritten in the equivalent forms (19), (20)
and (21) respectively.

Anda mungkin juga menyukai