Anda di halaman 1dari 14

OPTIMAL DESIGN OF TRANSMISSION TOWERS BY

DYNAMIC PROGRAMMING
DAVID J. SmPPti.and ANDREWC. PALMER
Engineering Department, Cambridge University, Cambridge, England

Ah&act-Dynamic programming has been found a useful technique for the synthesis of optimal layouts
for braced frameworks, provided that the interconnection between different parts of the structure is relatively
simple. Its application to real structures has been explored by a study of the optimal design of the steel
towers used to carry overhead lines. Many practical features complicate the problem, and make it inaccessible
to classical techniques such as the theory of Michell structures. There are strong geometric constraints on
layout, and yet many topologically distinct possible configurations, there are several alternative loadings
to be taken into account, and in addition the design of many of the bars is constrained by buckling rather
than by a iixed allowable stress level. It has proved possible to include these and similar factors in a com-
puter program which uses dynamic programming to synthesise optimal designs. The program generates
designs significantly lower in weight than existing designs.

INTRODUCTION
A NEWapproach to the optimal design of certain kinds of structure uses the mathematical
technique known as dynamic programming. It has now been applied to optimal plastic
design of frames and continuous beams [I, 21 and to optimizing the layout of plane trusses
[3]. Dynamic programming itself was first used in problems of operations research, particu-
larly in scheduling, allocation and inventory control [4]. Much of the mathematical theory
is due to Bellman [5].
In this paper we take as our starting point some relatively academic work on the
optimization of plane truss layouts [3], and go on to apply it to practical problems of design-
ing transmission towers for overhead electricity distribution lines.

DYNAMIC! PROGRAMMIN GANDOPTIMAL TRUSS LAYOUT


Consider the braced cantilever pin-jointed truss of Fig. 1. Points on the truss are
identified by coordinates x and y. A vertical load W acts at the origin, and the truss is
supported by a rigid foundation at (x., y,) and (x,, -y,). The truss is symmetrical about
the x-axis. It is composed of a number of K-braced panels, numbered from the left 1

panel 1 2 3 4
FIG. 1. Plane K-braced cantilever truss.

455
456 DAVID J. SHEPPARDand ANDREW C. PALMER

through n. The corners of the i-th panel in the upper chord are at (xi_ Ir yi_ I) and (xi, yi),
and the corners in the lower chord at (x~_~, -Y~_~) and (Xi, -y,).
Now consider the forces in the bars of the i-th panel. The two vertical bars joining
(Xi-13 yi-1) to (Xi-l, 0) and (Xi-13 0) to (Xi-13 -yi_l) are considered to belong to panel i
and not to panel i- 1. It is easy to show that the forces in the six bars which make up
panel i depend only on xi_ I, Xi, yi-1 and yi, and of course on W, but not on the location
of bars elsewhere in the truss. If the cost of that panel is the sum of the costs of individual
bars, and the design of a single bar depends only on its length and the force it transmits,
then the cost of panel i is a function of Xi_ 1, xi, Y~_~,y, and W. This will still be true
even if fabrication and connection costs are included, provided that we can neglect, for
instance, the influence on the complexity of the joint at (xi, yi) of the direction of the bar
joining (xi, yi> to (xi+ 1, .Yi+1). Suppose now that in a particular design problem the hori-
zontal extent of each panel is predetermined, and that W is given, but that the vertical
coordinates of the panel points have still to be chosen so as to minimize the cost of the
structure. The cost of panel i is then a function gi(yi_l, yi, W) of Y~_.~,yi and W. The
minimum cost of the first i panels depends on the location of the right-hand end of this
segment of the whole truss, and is therefore a function ci(yi, W) of yi and W. Therefore

= nlin (BitYj- I, Yi, W)+ y1 ,‘?i:,_ iSIt YlY W+ * * * +Si-ICYi- Yi-13 W) t2)
Yi-1 ,. I L 2

since g, is independent of yl, yZ, . . . , yi-2. The minimum with respect to yl, . . . , yi-2
on the right-hand side of (2) is the minimum cost of the first i- 1 panels if the i- 1-th
panel has its right-hand panel points at (x~-~, JJ,_.~) and (Xi-l. -yi_l), and is therefore
ci- l(yi- 1, W). Accordingly

ci(Yi, W)= min (gi(yi-1, .Yi?W)+ci-l(Yi-l~ W)> * (3)


Yi-1

Dynamic programming relies on the systematic use of equations of this kind, and in
this case equation (3) can be used to find the optimal design for the cantilever truss. The
process is explained in detail elsewhere [3, 61. It proceeds by first working along the
structure from left to right to find the minimum cost functions ci(yi, W), and then working
back to find the optimal structure. Thus, cl, the cost of the first panel, is a simple function
of Yi; setting i=2 in (3), and knowing g, from an analysis and design procedure for a
single panel, the function c2 can be determined. Then cg can be determined once c2 is
known, cq once cg is known, and so on until c, is reached. The value of y, which minimises
c, is the optimal location of the right-hand end of the structure. Giving y,, this value, the
optimal value of y,_ i is the one which minimises g,(y,_ 1, y,,, IV) + c,_ I(y,_ 1, W). This
value can be used to find Y”_~, that to find Y,,-~, and so on.
Computationally, the process requires a means of storing the functions Ciy so that a
value can be calculated for any argument of interest, and a means of carrying out the
minimisations. A function is represented either by interpolation on a table of values at
particular points [3, 71 or by a polynomial [8, 91. Any standard method of minimising a
function of a single variable can be used.
Optimal Design of Transmission Towers by Dynamic Programming 457

TRANSMISSION TOWID!
In this preliminary work we were trying out dynamic programming as a technique for
layout o~tim~tion. It was natural to look first at plane st~ct~es, with a relatively
simple and perhaps slightly artificial layout, and to make severe idealizations. A technique
is of course of little genuine importance until it can be applied to real structures, and so
we next sought out such a structure. It soon became clear that transmission towers (collo-
quially “pylons”), which carry high-voltage overhead transmission lines, were the most
attractive subject for further work. There are several reasons for this. Firstly, transmission
towers are of relatively simple construction, being made from galvanized steel angle sections
connected by black bolts. Their foundations are also simple, and the costs of erection and
inspection are small, and accordingly the cost of the structural steel forms a high proportion
of the cost of the whole structure, in contrast to steel-framed buildings, for instance. This
means that even relatively small economies in steelwork carry through as significant pro-
portional reductions in the cost of complete towers, In addition, large numbers of towers
are made to a single design: the additional design cost made necessary by a search for a
more sophisticated design can be covered by the saving on a thousand towers, whereas it
could hardly be justified if only a single tower were to be built.
Secondly, dynamic programming is best adapted to the optimization of “chain-like”
sequential systems, in which the interaction be‘tween different stages is relatively simple.
The amount of computation is p~m~ily dependent on the comple~ty of this interaction,
and only to a much smaller extent on the length of the system, A tower composed of a
number of quite simple panels is therefore well suited to such an analysis.
Thirdly, we were able to benefit from the interest of an important manufacturer of
transmission towers, British Insulated Callenders Construction Ltd., and their cooperation
much assisted our work. Our attitude to the optimization problem is relevant here. One
can ask either of two distinct questions. The first is the more ambitious: what is the best
of all possible tower designs ? To answer this demands an extensive study of possible
materials, from timber to plastics, of construction techniques, and of users’ reactions to
possible radically different designs. Instead we deliberately asked a much narrower question :
using the existing technology (steel angle sections bolted together) and the now traditional
general layout (a self-supposing braced tower tapering from a broad base to three cross-
arms), can anything be done to reduce the cost of the steel work by modifications in detail?
A typical tower design is illustrated in Fig. 2. Four legs support the tower, sloping
upward and inward from the foundation to the peak. At any horizontal cross-section the
corners of the tower lie on the corners of a rectangle. The legs are connected by lattice
bracing. Here we shall be concerned with the optimization of the structure between the
ground and the lowest cross-arm, but with almost no mo~~cation the same method can
be applied to the peak above the highest cross-arm, to the intermediate sections of the
tower, and to the cross-arms themselves. The bracing divides the tower into a number of
“cells”, a cell being the three-dimensioilal counterpart of a panel in a plane truss. Each
cell can have several configurations, and these are illustrated in Fig. 3. A tower usually
contains cells of more than one configuration, and the optimization will have to select the
optimal cell confi~ration at each level.
The notation used in describing tower geometry is shown in Fig. 4. Certain dimensions
are prescribed by functional requirements, and are not free to be altered during an optimiza-
tion process. The height of the lowest cross-8sm is determined by an electrical clearance
between the suspended cables and the ground, and other electrical clearances fix the cross-
458 and ANDREWC. PALMER
DAVID J. SHEPPARD

FIG.2. Transmission tower: general layout.

FIG. 3. Alternative bracing indurations.

arm breadths. In addition, we shall suppose that the base dimensions x0 and z. are given,
as well as the dimensions x, and z, immediately below the lowest cross-arm, and that the
legs are straight. In this way the outline profile of the tower is defined. A design method
has to determine
(i) the optimal number of cells, and the heights yi of the panel points. Cells are
counted from the ground. Cell 1 has its lower panel points at ground level y =0
and its upper panel points at y=yi, cell 2 has its lower panel points at y=yl and
its upper panel points at y=y2, and so on,
OptimalDesignof TransmissionTowersby Dynamic Programming 459

cell

cell

cell

leg 4
FIG. 4. Transmissiontower: notation.

(ii) the type of bracing in each panel, a choice of one of the six distinct configurations
illustrated in Fig. 3,
(iii) the dimensions of the individual bars within each cell.
A large number of different loading conditions have to be considered, and for each
condition the external forces applied to the tower by the cables are known. If the tower
is statically determinate, the extreme values of the axial force in each individual bar are
determined only by the layout of the framework, and do not depend on the bar cross-
sections, A method must exist for designing a single bar, given
(i) the length of the bar,
(ii) the maximum compressive force it must transmit, under the most critical loading
condition for that particular bar in compression,
(iii) the maximum tensile force it must transmit, in the most critical loading condition
for that particular bar in tension,
(iv) the type of steel to be used,
(v) in the case of a bar in compression, the number of bracing bars that restrain it
against buckling, and their orientation.
Methods for designing individual bars are extensively discussed in different codes of
practice, and can be brought into the program without difficulty. It has, for instance, been
possible to include the following constraints, of which the tist is much the most important:
(a) the allowable stress in compression is a given discontinuous function of the slender-
ness ratio,
(B) a maximum allowable slenderness ratio must not be exceeded,
(JJ>any horizontal or nearly horizontal bar must be able to bear a 200 lb load at its
centre without distress. This is to allow for the weight of a man standing on the
bar during construction, and is sometimes more critical than (8).
460 DAVIDJ. SNEPPARDand ANDREWC. PALMER

The analysis that follows requires a method for designing individual bars given the
data in (i) through (iv) above, and a method for calculating the cost of a bar once it has
been designed. It will be assumed that costs are additive, so that the cost of the steelwork
is the sum of the costs of individual bars.
A key assumption now has to be made. Without it it would not be feasible to apply
dynamic programming. The assumption and its consequences are discussed first, and the
justification for the assumption follows. Recall that the loads acting at the top of the tower
and the outline of the profile are given. If the axial forces in the bars of each cell depend
only on the levels of the top and bottom of that cell, and on the bracing configuration
within the cell, then the cost of the cell for a given bracing configuration is a function of
yi_r and yi. This cost can be evaluated for all allowable bracing configurations. The
configuration which gives the smallest cost is the optimal configuration for a cell between
levels yi_l and yi, and the cost of this optimal cell will be denoted gi(yi_ r, yJ. If now
ci(vi) is the cost of an optimal design for the lowest i cells of the structure, when the i-th
cell reaches level yi, then, since the design is optimized with respect to the design variables
Y1, YZ, . . . > Yi- 1,
cj(Yi)= min
YI,Yz,..~,Yi-l
{Yi(Yi-1, YJ+ . . . +gl(O, Yl)). (4)

An identical argument to that used earlier in the paper shows that

ci(Yi)=min(Si(Yi-l, Yi)+Ci-1(Yi-1)}* (5)


Yi-1

We must now pause to examine the justification for the assumption that the minimum
cost of cell i depends only on yi_l and y,. A structure like that illustrated in Fig. 2 is of
course highly redundant, and therefore an elastic analysis will show that the axial force
in a typical bar in cell i will depend on the orientation and properties of bars elsewhere in
the structure. However, an elastic analysis will be sensitive to the idealizations required
to reduce the real structure to an analysable form. A tower is built with simple bolted
connections (not fitted bolts or HSFG bolts), and these will slip when the tower is loaded,
the amount of slip depending on details of hole alignment and bolt tightening which are
impossibIe to predict. Slip and yield around bolt holes will contribute to the marked
redistribution of forces in the tower which occurs on loading, and the actual force distribu-
tion will be very different from that predicted by elastic analysis. Instead of the design
method being based on an elastic analysis, therefore, certain statical assumptions have
been made in the analysis of forces within the structure. One can justify these assumptions
in two ways. Firstly, there exist experimental data on forces within transmission towers
of this type, since a prototype of most tower designs is tested to destruction (in contrast
to almost any other structure of civil engineering scale). These observations have been
used to develop appropriate statical assumptions, which essentially make the structure
statically determinate. It turns out that the forces consistent with these assumptions agree
with observation at least as well as those predicted by elastic analysis. An alternative
interpretation is that any statical assumptions made in the analysis will lead to a system
of bar forces which is internally in equilibrium and in equilibrium with the external loads.
Then suppose that each bar is so proportioned that the stress induced by this system of
bar forces is not greater than the allowable stress, and suppose in addition that the force-
elongation relation for each bar is that of an elastic perfectly-plastic bar. The conditions
Optimal Design of Transmission Towers by Dynamic Programming 461

of applicability of the limit theorems of plastic design [IO] are then met, and the structure
will not collapse under the external loads, even though the assumed equilibrium system of
bar forces is not identical with the actual system developed when the structure is loaded.
Imagine a horizontal section across the tower at level ys immediately above the cell
which is to be analysed. However the forces transmitted across this section are distributed
in detail, their resultant can be found by considering the equilibrium of the tower above
the imaginary section, and is described by resultant force components F,, Fy and F,, acting
through (0, yr , 0) in the X, y and z directions, and by couples C,, C, and C, (Fig. 5a). Each
of these forces and couples is now supposed to be distributed seperately between the
corners and edges of the tower at level y,, in the manner illustrated in Fig. 5b. The resultant

FIG. 5. Distribution of resultant force and moment across a section of the tower.

forces acting in the legs of the tower and in the faces of the cells are then found by recom-
bining these separately distributed forces. At the intersection of leg 1 with the section, for
instance, the resultant vertical force is

FYI4- CJ?a -I-C,/2b

and the shear force applied between leg 1 and leg 2 is

F,/2 - CJ2b.

The total vertical load acting on each corner of a cell is the vertical component of the
leg load, so the latter can be found by equilibrium. Because the leg slopes, each leg load
462 DAVIDJ.SHEPPARD and ANDREW C.PALMER

has horizontal components in the x and z directions. The net shear load on each face is
calculated by subtracting the corresponding leg components from the gross shear force
applied to that face. It is assumed that the shear components acting from the perpendicular
faces are negligible. It is also assumed that the net shear in a cell face is carried equally
by the two primary bracing members, one in tension and the other in compression. Knowing
the net shear and the cell geometry, it is then possible to calculate the primary member
axial forces.
Alternative loading systems are dealt with at this stage of the panel optimization. The
design load for the leg is obtained from the greatest of all the vertical corner loads. The
design loads for the bracing members are obtained from the greatest of all the net shear
loads in the plane of the cell face. The worst loading condition for each member is therefore
found at an early stage of the optimization of the panel, and this saves time compared
with a method which needs to calculate the member costs for each loading system. More
time can be saved by expressing each loading system as a set of three forces and three
couples acting at the origin.
Lateral loading due to wind has to be dealt with rather differently, and is discussed
later in the paper.
If equation (5) is to be used to synthesise designs, there must be a procedure for
evaluating the cell cost function gi(yi_l, y,). A subprogram to do this can be constructed
by bringing together
(a) a subroutine which makes use of the statical assumptions outlined above to deter-
mine forces in individual bars and the panel configuration,
(p) a subroutine which designs individual bars, and takes as its input the length of
the bar, its axial force, and design constraints such as the relation between slender-
ness ratio and allowable compressive stress,
(r) a subroutine which calculates costs of individual bars once they have been
designed, and adds these costs to find the total cell cost, and
(8) a procedure which goes through all the possible bracing configurations for a cell,
makes comparisons between one cell and another, and identifies the cost of the
design in the cheapest configuration with gi.
Once this has been done equation (5) can be used to locate optimal designs in the same
way as for the plane truss. The cumulative cost functions Ci(yi) are described by their
values at 10 equally-spaced grid points, previously selected to cover a reasonable range
of possible values for each level yi. If the range is carelessly chosen, this will become
obvious when the output is examined, and the range can be modified. The minimizations
with respect to yi_l are carried out by golden section search, an efficient procedure for
locating extreme values of unimodal functions [ll], and one which is close to the optimal
Fibonacci search procedure. Details of the program have been described elsewhere 161.
The program works up the tower, progressively determining the minimum cost functions
c,. When the lowest cross-arm has been reached, the optimal design is constructed by
working back down to the ground.

RESULTS
The program was used to investigate a number of variations on one design. An existing
BICC design for a 400 kV tower was used as the basis for this study, the height, top plan
and loading systems being the same, except for the absence of wind loading on the tower
Optimal Design of Transmission Towers by Dynamic Programming 463

itself. The tied design parameters were a 13.6 ft square cross-section at the lowest cross-
arm, at a height of 84 ft, and a constraint that required the bottom cell to be of configuration
6 in Fig. 3. The tower cross-section was made square throughout. Two alternative loading
conditions are described in Table 1; they are a normal loading condition with a load
factor of 2.5 and a broken-wire condition with a load factor of 1.5, and are usually the
critical conditions.

TABLE1. Loading conditions

Loading system 1: normal condition

Coordinates of point of Load components in coordinate


Load number application directions
x z x Z

$1 &I

1 0 165 0 1913 -3225 0


2 -23 147 0 18600 -28500 0
3 23 147 0 18600 -28500 0
4 -33 113 0 20300 -28500 0
5 33 113 0 20300 -28500 0
6 -27 84 0 20500 -28500 0
7 27 84 0 20500 -28500 0

Loading system 2: broken-wire condition

Coordinates of point of Load components in coordinate


Load number application directions
X Z x Z

6, &I

1 0 165 0 1148 -1935 0


2 -23 147 0 11145 -17000 0
3 23 147 0 11145 -17000 0
4 -33 113 0 12173 -17000 0
5 33 113 0 11408 -13275 25200
6 -27 84 0 12770 -17000 0
7 27 84 0 12770 -17000 0

Optimal designs were synthesised for a number of different base-widths, and for
different values of the assumed total number of cells. The variation of tower cost with
these two design variables is shown in Fig. 6, and the minimum cost is found when there
are 7 cells and the base width is 17.6 ft. Although this design is much cheaper than the
design taken as a starting-point, a direct comparison is unfair because some further factors
which have to be taken into account in practice have been left out of the design program.
464 DAVID J. SHEPPARDand ANDREW C. PALMER

The results show that the panel levels in the optimized designs depend on the base width.
In a vertical-sided tower the optimized cells are almost equal in height, but in a tower
with inward-sloping sides the lower cells become larger and the upper ones smaller, as
one might expect.

I I I I I
I I I I

20 30 feet
base width
FIG. 6. Cost of optimal design as a function of base width and number of cells.

The preliminary program described in outline above showed that dynamic program-
ming could be used to synthesise optimal designs for transmission towers, but it was
realised that it omitted several factors which significantly influence the design of real
towers. A longer program was then written, and it includes these additional factors. An
outline of the modifications follows: complete details are given in Ref. [6].

WIND LOADING
Wind loading is fundamentally different from the other kinds of external tower loading,
since its magnitude and distribution depend on the detail design of the tower. This loading
is difficult to incorporate properly in a design process. Ryle 1171suggests that wind loading
accounts for 20 per cent of the total design moment on a 132 kV double-circuit tower. An
estimate of wind loading can be obtained from an estimate of exposed area and wind
pressure, and the resulting force and moment can be included when the forces in the
structure are calculated. If necessary the design synthesis can be iterated, the exposed
steelwork areas in each successive design being used to estimate the wind loads in the next
design.

BENDS IN THE TOWER LEGS


Until now it has been assumed that the legs of the tower are straight between the
foundation and the lowest cross-arm. It is usually better to design a large tower so that
the legs slope less steeply near the ground than they do higher up (a feature called by the
ugly but expressive term “Eiffelization”). The legs therefore consist of straight segments
connected by one or two “bends”, which are located at the levels where two cells meet.
Optimal Design of Transmission Towers by Dynamic Programming 465

The level of the bend and the width of the tower at the bend have to be determined. Though
their positions could be optimized by dynamic programming, the extra computational
effort required would be substantial, and it is better to find optimal positions by repeated
trial, in each trial fixing the bend positions and optimizing the rest of the design. Since
the leg is kinked at a bend, equilibrium requires that the change in direction of the axial
force in the leg throw additional forces into the braced panels of the cell below the bend,
and an additional statical assumption is required to include this effect.

RESULTS PROM THE EXTENDED PROGRAM


The object of the program was to test the usefulness of dynamic programming, and
the production of new tower designs as such was not intended. To test the program, an
existing design for a 400 kV tower was used as a starting-point: it is shown in Fig. 7.
Some of the features of this design can still not be reproduced by our program, notably
the secondary bracing pattern in the lowest cells and the inverted-K panel between the
bottom X-panel and the top K-panel. It was decided to modify the design to one which
was included in the family of designs capable of being synthesised by the program, and
it was hoped that this modified design would provide a fair basis for judging any improve-
ments which the optimization might make. Although it would be better to use the original
design as a basis for any cost comparisons, it was felt that the extra work required to
incorporate into the program the features peculiar to this design would not be justified.
The program already contains some very similar features which were incorporated without
any difficulty.

I%. 7. Originalde&n used as a basis for this study.

The modsed design is shown in Fig. 8. It contains one bend, 40 ft above the ground,
and this was to be retained in any new design. The lowest cell was again to be of con-
figuration 6, and the loading conditions considered were those of Table 1.
The original BICC design contained three leg splices below the cross-arms. Although
splices are usually put in to make possible a step reduction in the leg area, these splices
466 DAVID J. SHEPPARDand ANDREWC. PALMER

FIG. 8. Modified design.

connect leg sections which have the same area and are made from the same steel, and are
put in so that the leg sections are not long and unwieldy. As far as the optimization process
is concerned, the splices in the BICC design are non-existent. Accordingly it was decided
to obtain two sets of optimal designs, one with a constant cross-section all the way up the
leg, and the other with three splices. These splices were to be positioned in the same panels
as those in the original design, and were to allow a reduction in leg area wherever possible.
The optimization program was first used to find the best levels for cells of the same
configurations as in the modified design. Figure 9 shows the solutions which were obtained.

a
FIG. 9. Optimal design subject to the condition that the bracing configuration is identical to
that in the modified design
(a) with leg splices, cost 22.7 units.
(b) without leg splices, cost 24.1 units.
Optimal Design of Transmission Towers by Dynamic Programming 467

By comparison with the corresponding basic designs, the savings in steel weight are 2.9
per cent for the three-splice case and 0.3 per cent for the no-splice case. The difference
between these figures is surprising. Since the legs account for about half of the steel, it
might be expected that excluding the legs from the optimization would reduce the weight
saving by approximately half. However, it seems that the original no-splice design happened
to be quite close to an optimal design. The program was next used to find optimal cell
configuration as well as optimal cell levels. Many results were obtained, using different num-
bers of panels above and below the bend level. Figure 10 shows the solutions which were
the cheapest overall. The three-splice design saves 4.9 per cent over the basic design, and
the no-splice design saves 3.8 per cent. Since this problem has more design freedom than
the first one, the gains are larger. The levels of the new solutions have diverged further
from the basic solution than did those in Fig. 9; this again is a consequence of the
greater amount of freedom.

a b

FIG. 10. Fully optimized design


(a) with leg splices, cost 22.2 units
(b) without leg splices, cost 23.2 units.

CONCLUSIONS

The results encourage further work on this application of dynamic programming.


On the Cambridge University Titan computer, three minutes computing time were required
to obtain a design in the most complex and general case, and less than one minute in simple
cases. At current commercial rates, each run of the optimization program would cost
roughly f 5. A large tower costs about f2500, and optimization of this type can be expected
to save between 0.5 and 4 per cent of this, a saving of the order of f 10-100. Hundreds or
even thousands of towers will be built to the same design, and so substantial cost savings
should be possible.
468 DAVID J. SHEPPARD and ANDREWC. PALMER

The program is also a good means of achieving a preliminary design for a tower: it
is probably faster and cheaper than a manual process, and produces a more efficient
solution. A complete optimization and design process should select the tower profile as
well as the bracing pattern, and cost savings should then approach 5 per cent. This is a
comparatively large saving in a type of structure whose basic design was decided on more
than 40 years ago and has since then been repeatedly improved by iterative design methods
based on trial and error.

Acknowledgements-The work of one of us (D.J.S.) was supported by the Science Research Council. We
wish to thank Mr. D. C. J. Williams of British Insulated Callenders Construction Ltd. for his help and
advice.

REFERENCES
[I] A. C. PALMER,Optimum structural design by dynamic programming. J. struct. Div. Am. Sot. civ. Engs.
ST8,1887-1906 (1968).
[2] A. B. C. CAMMAERT, Optima1 design of multi-storey frames by dynamic programming. Ph.D. disserta-
tion, Cambridge University (1971).
[3] A. C. PALMERand D. J. SHEPPARD,Optimizing the shape of pin-jointed structures. Proc. Inst. civ.
Engrs 47,363-376 (1970).
[4] M. J. BECKMANN, Dynamic Programming of Economic Decisions. Springer, Berlin (1968).
[5] R. E. BELLMAN,Dynamic Programming. Princeton University Press (1957).
[6] D. J. SHEPPARD,Structural design optimization by dynamic programming. Ph.D. dissertation, Cam-
bridge University (1970).
[7] R. E. BELLMANand S. E. DREYFUS,Applied Dynamic Programming. Princeton University Press (1962).
[8] R. E. BELLMANand S. E. DREYFUS,Functional approximation and dynamic programming. Math.
Tables and other Aids to Computation 13, 247-251 (1959).
[9] A. C. PALMER,Limit analysis of cylindrical shells by dynamic programming. ht. J. Solids Strut. 5,
289-302 (1969).
[lo] J. F. BAKERand J. HEYMAN,Plastic Analysis and Design, Vol. I. Cambridge University Press (1970).
[ll] D. J. WILDE, Optimum Seeking Methods. Prentice-Hall, Englewood Cliffs, N.J. (1964).
[12] P. J. RYLE, Steel tower economics. J. Inst. Elec. Engrs 93, 263-284 (1945).

(Received 31 August 1971)

Anda mungkin juga menyukai