Professional Documents
Culture Documents
University of Exeter
U.K.
Important features of cost are the electricity A description of standard genetic algorithms
tariff structure, the relative efficiencies of the can be found in [2,3]. The paper now goes on
available pump sets, the head through which to describe the application of genetic algo-
they pump and marginal treatment costs. Im- rithms to pump scheduling and to discuss the
portant constraints include consumer demand, possibilities of improving their performance.
First, some facts about modelling the supply Initially a standard GA was adopted contain-
system must be mentioned. It is normi~llysuf- ing:
ficient to consider scheduling over a one day
period, taking the historic demand paltern for 0 a fitness function which calculates the
an average day as the predicted requirement. fitnesses of the population members di-
Pumps can supply water from one or more rectly from their costs
sources to one or more reservoirs, from which e the genetic operators
the demands are drawn. For continuity, the * one-, two-, or multiple point cross-
reservoirs must finish their daily cyclie at the over
same level from which they started. ‘The day 4 mutation
has to be divided up into time intervals, during * an ellitist strategy
which each pump can either be switched on or
not. The pump schedule derived will specify This GA works but performs relatively ineffi-
the time intervals for which each pump has to ciently, as the calculation times even for small
be switched on. Only fixed speed pumps are systems are quite long with a lot of genera-
considered in this study. For all possible pump tions needed to find optimal solutions. Some
combinations, the pump efficiencies and improvement can lbe achieved by tailoring the
head/flow characteristics have to be defined. optimization paraimeters such as population
Furthermore the usable reservoir capacities size, probability of mutation and crossover to
have to be determined, and maximum and the specific grobllem. The most significant
minimum levels set. improvement, however, comes from the intro-
duction of a ranking function for the members
The next step is to define the cost fiinction, for calculating their relative fitnesses, as de-
which has to include all the essential expendi- scribed litter in the paper.
ture caused by the pumping process. The main
component will be the cost of electricity but For pump scheduling, the use of a standard
other costs such as special charges for peak binary coding is the obvious choice, since each
monthly electrical demand or a maintenance binary value can represent one pump that is
charge dependent on the number of pump either on or off during a particular time inter-
switches made, can be included. val. The use of other forms of coding were
considered to have no advantages for this type
To ensure that the system constraints are not of problem.
violated in the solution, it is necessary to in-
troduce penalty functions for unfulfililed de- Ranking of the Population
mand, violation of reservoir capacity, etc. To In the standard GA the fitness of each member
define suitable values for the penalties is an is calculaited as an inverse function of the cost.
important task in order to make the search in The relative fitness and prolbability of being a
the solution space as efficient as possible. Too parent are then derived. The relative fitness
small a penalty will make it likely that a large can, however, cause difficulties. If the fit-
part of the search is performed outside the real nesses axe quite similar there will not be a
solution space and too high a penalty can proper selection between good and less good
make it difficult for the genetic algorithm to solutions, but if, on the other hand, the rela-
find the optimal solution if it is near or even tive fitnesses of a very few members are too
on a boundary introduced by the system con- high ciornpared to the others, this will soon
straints. lead to a lack of diversity in the population.
The cost function, now including all h e dif- For this application, the problem was solved
ferent cost components and constraint viola- by introducing a ranking procedure [4]. The
tion penalties is generally non-linear aind non- population members are ranked in order ac-
differentiable. It depends on several ldiscrete cording to their costs. Each then receives a
valued variables and so the task of optimka- fitness dependent on its position within the
tion is computationally difficult. An attempt to ordered list. For simplicity, a fitness equal to
apply genetic algorithms to this optinnisation the order number is used, tlhe most expensive
problem is detailed in the next section. solution getting a fitness of 1, the next 2 etc.
402
The introduction of the ranking function leads length of the complete binary string for the
to a remarkable increase of the efficiency of solution space is 24 x 4 = 96. With this binary
the genetic algorithm. The extra computing string all the 296 (8x1QZ8)theoretically possi-
time needed each generation to sort the popu- ble pump schedules are represented. A1 the
lation, (using a heap sort algorithm [SI), is important data about the pump c o ~ b i ~ a ~ ~ o n s
small compared to the time saved by finding their efficiencies and the influence of the
the optimal or near optimal solu!ion in a pump head are precalculated and summarised
greatly reduced number of generations. Fur- in Table 1. The binary coding is also defined
thermore the ranking function makes the ge- there, 1 meaning switched on and zero
netic algorithm more robust against factors switched off.
such as inappropriate sizing of penalty func-
tions. Cost function
The main cost is for the energy to pump the
4 Computational Considerations for an water into the reservoirs. The amount of elec-
Example System tricity consumed by each pQmp combination,
which is shown in table 1 has to be multiplied
The example used for demonstration of the by the unit costs for electricity. It is assumed
method is a simple system with 4 different that the night-time tariff is 7 money-units per
pumps delivering water to a single reservoir. electricity-unit and that the day-time tariff is
Consequently there are 16 different possible double that price. The higher daytime tariff
pump combinations. The day was divided up has to be paid from 7 a.m. until 8 p.m. It is
into 24 time intervals, each of one hour, with easy to include other types of cost like previ-
the demand drawn from the reservoir chang- ously mentioned maintenance costs which
ing for each hour. The difference between the may depend on the number of times per day
rate at which water is pumped to the reservoir that pumps are switched, or any other costs
and the demand is accommodated by changes e.g. different costs for taking the water from
in the volume of water stored in the reservoir, different sources. However, to include a
subject to maximum and minimum storage monthly peak demand electricity tariff, as re-
volumes. To represent the 16 different pump quired for some systems, makes it necessary to
combinations, a binary string of length 4 is extend the optimisation period from one day to
required, and as there are 24 time intervals the one month. In this case the demand prediction
0 0 0 0 0 0
0 0 0 1 10 12
0 0 1 0 30 30
0 0 1 1 40 42
0 1 0 0 50 44
0 1 0 1 60 56
0 1 1 0 80 74
0 1 1 1 90 86
1 0 0 0 100 80
1 0 0 1 110 92
1 0 1 0 130 111
1 0 1 1 140 124
1 1 0 0 150 127
1 1 0 1 160 141
1 1 1 0 180 165
1 1 1 1 190 182
becomes a very difficult task, becaust: it is no One can see that the optimal, or near optimal,
longer possible to take an average daiy as the solution is found after about 10,000 genera-
only basis for the cost calculation: in addition tions. The cost progression of the average
it is necessary to consider the wcirst day, member is in general ternas quite similar to
which is far less predictable. that of the best member. For this example the
cost of the average population member is, ex-
As well as electricity charges, it is necessary to cept at the very beginning, about 5 % higher
include penalty functions for system constraint than the current optimum. If this difference
27
26
25
24
23
’ 22
t
21
ferent from the volume pumped in the night, Simulation of changes to the system
there is a huge oscillation in the reservoir Apart from the direct advantage of savings in
level. If the reservoir were bigger, the volume energy costs, the availability of an easy to use
pumped during the day would decrease fur- optimal pump scheduling package enables
ther. However in this case, the most efficient changes in the system to be simulated and
pump b u m p 1) is used nearly the whole day examined for efficiency. Critical hardware
so that the water level does not drop too far compofients can be identified, such as the
towards the end of the expensive electricity most frequently used pumps or the reservoirs
tariff period. The less efficient pumps are jus1 that are used for their maximum capacity
used during the night. It must be mentioned ranges. The effect of modifying these compo-
that there are a number of equally good opti- nents, such as replacing a pump with a more
mal solutions, as the pumped volumes speci- efficient unit or enlarging a reservoir can tie
fied for some hours can be switched around examined in a systematic manner to maximise
wi~hou~ affecting costs or vioiating con- benefit. A GA based pump scheduling model
straints. can easily cope with changes in the system
components, as the optimization routines are
5 Conclusions and further ideas largely independent of the structure of the
modelled system.
The question of whether pump scheduling
optimisation with genetic algorithms is supe- Similarly, if the supply network is to be ex-
rior to the other optimisation methods men- tended to meet the demands of new housing or
tioned in chapter one cannot be answered de- industrial development, the GA approach is
finitively at this stage as no formal compari- flexible enough to cope with the system
son has been made. However it is clear that changes, and to help in the identification of
GAS and other evolutionary programming the optimum arrangement for the modified
800 t------
700
t
00
508
400
300
200
100
0 . . . .
0 2 4 6 8 10' 42' 44' 16' 18' 20' 22' 24
time (hours)
Figure 2: Stored volume vs. time
200
175
150
125
180
75
50
mands, it is likely that on-line control will 2. Michalewicz, Z., Genetic Algorithms +
become increasingly implemented. One impor- Data Structures = Evollution Programs.
tant aspect of the GA approach currently being Springen Verlag, Berlin, Heidelberg, 1992.
investigated is the ability of the GA to adapt to
slightly changed constraint conditioils, e.g.. 3. Goldberg, D.E., Genelic Algorithms in
on-line changes in predicted demand ilnd CUT- Searchi, Optimization and Machine Learning.
rent reservoir levels. Maintaining a family of Addison-Wesley, 1989.
diverse yet near optimal solutions as a starting
point for the evolution may allow rapid iden- 4.Baker, J.E., Adaptive selection methods for
tification of the optimum schedule suilable for genetic algorithms, in J.J. Grefenstette (ed.),
on-line control. Pmceeding of an international conference on
Genetic Algorithms, pp. 101-111, Lawrence
6 References Earlbaurn, 1985.
1. Water Research Centre, Pump Scheduling 5. Press, W.H., Teukolsky, S.A., Vetterling,
in Water Supply, Report TR232, Swindon, W.T., Flannery, B.P. Numerical Recipes in C:
UK, 1985. the art of scientific comlputing. University
Press, Cambridge, 1992.