Professional Documents
Culture Documents
1 s2.0 S0378778802000713 Main
1 s2.0 S0378778802000713 Main
Abstract
The design of buildings is a multi-criterion optimization problem, there always being a trade-off to be made between capital expenditure,
operating cost, and occupant thermal comfort. This paper investigates the application of a multi-objective genetic algorithm (MOGA) search
method in the identification of the optimum pay-off characteristic between the energy cost of a building and the the occupant thermal
discomfort. Results are presented for the pay-off characteristics between energy cost and zone thermal comfort, for three design days and three
building weights. Inspection of the solutions indicates that the MOGA is able to find the optimum pay-off characteristic between the daily
energy cost and zone thermal comfort. It can be concluded that multi-criterion genetic algorithm search methods offer great potential for the
identification of the pay-off between the elements of building thermal design, and as such can help inform the building design process.
# 2002 Elsevier Science B.V. All rights reserved.
Keywords: Building energy optimization; HVAC systems; Supervisory control; Multi-criterion optimization; Genetic algorithms
0378-7788/02/$ – see front matter # 2002 Elsevier Science B.V. All rights reserved.
PII: S 0 3 7 8 - 7 7 8 8 ( 0 2 ) 0 0 0 7 1 - 3
960 J.A. Wright et al. / Energy and Buildings 34 (2002) 959–972
developed, but the form of GA implemented in this work is to the multi-criterion optimization of buildings but also the
derived from the so called ’’simple genetic algorithm’’ [5]. A single criterion building optimization [9].
detailed description is beyond the scope of this paper, but the The single criterion genetic algorithm is derived from the
form of the GA implemented can be summarized as being, a same ‘‘simple genetic algorithm’’ as used for the multi-
simple binary encoded GA with ‘‘roulette wheel’’ selection criterion genetic algorithm, the difference between the two
and a non-overlapping population. relating to the fitness formulation. In particular, the fitness
formulation developed in this work addresses the need to
2.1. The multi-criterion genetic algorithm automatically find an initial feasible solution starting from a
randomly generated initial population. It is beyond the scope
GAs are unconstrained search methods that were origin- of this paper to describe the method in detail. However, the
ally designed to optimize a single criterion as represented algorithm uses a normalized sum of the constraint violations
by the fitness of each individual in the population. Many of as a measure of a solutions infeasibility (the normalization
the variations to the basic GA are concerned with the des- being described in [10]). The solutions infeasibility is used
cription of the individuals fitness; this is particularly the to rank the solutions in the population, the ranking forming
case for the multi-criterion optimization. Several app- the fitness of the solution. For instance, if all of the solutions
roaches to defining the fitness for a multi-criterion genetic are infeasible then the ranking is obtained for each indivi-
algorithm optimization exist [4]. The approach implemen- dual by counting the number of solutions with a higher
ted here is the multi-objective genetic algorithm (MOGA) (worse) infeasibility. Since the GA seeks to maximise the
[6,7]. This algorithm employs the Pareto ranking scheme fitness of the population, bias is given to the solutions with
illustrated in Fig. 1, the Pareto rank then being used to form the higher rank (lower infeasibility) as the search progresses.
the fitness of each solution (solutions of equal rank having A two part ranking (objective and infeasibility) is applied
equal fitness). However, since GAs seek to maximize fitness once a feasible solution is found.
and we seek to find the pay-off that minimizes the criteria,
the Pareto rank is inverted to form the fitness of the
individuals. An exponential weighting is also applied dur- 3. Example problem formulation
ing the inversion to give extra weight to the non-dominated
(Pareto 0) solutions. The multi-criterion optimization of building thermal sys-
Since many problems are constrained (as is the case in the tems is investigated here through the design of a single zone
thermal design of buildings), the MOGA includes an ‘‘all outside air’’ HVAC system. The system consists of a
approach to handling the constraint functions. In short, regenerative heat exchanger, cooling coil, heating coil, and
the constraints are treated as criteria and ‘‘goal restraints’’ supply fan (Fig. 3). The control system is open loop to the
applied to force the solutions into the desired feasible region zone, since the optimization seeks to determine the supply
(by penalizing the Pareto rank of the infeasible solutions). air temperature (ys ), and flow rate (Ms ), in each hour such
Treating the constraints as criteria has the advantage that the that the zone conditions remain comfortable. This control
pay-off between constraint function bounds and the true configuration has been selected since it can be applied to
design criteria can be examined. However, it has the dis- buildings having ventilated floor slabs as well as conven-
advantage that the pay-off characteristic becomes increas- tional air conditioned buildings. The controller has two
ingly difficult to interpret as the number of criteria increases setpoints, one for the supply air temperature (ysp ) and a
(for instance, the pay-off curve becomes a pay-off surface as second for the supply air flow rate (Msp ). For a given supply
the number of criteria is increased from two to three). For air setpoint (ysp ), the control supervisor determines if the
this reason, a means of aggregating the constraints into a heat recovery device should be operational, and whether
single criterion has been developed [8]. As well as reducing active heating or cooling is required in order to meet the
the number of criteria, the method also allows the aggregated
constraint ‘‘criterion’’ to be removed from the search once
feasible solutions are obtained, further reducing the dimen-
sionality of the pay-off information.
The model initialization is ensured by running the model control schedule and size of HVAC system). Three weights
five times for each day of boundary data; however, only the of building are used in the example problems, the construc-
results from the last day in the sequence are used in the tions for which are given in Table 3. The corresponding
optimization. Five consecutive runs of the model has been material properties for each construction element are given
found to result in acceptable initialization of the model, even in [17]. In all the examples, a low emissivity glass with 20%
for a ‘‘heavy-weight’’ building construction. glazed area has been specified.
The zone air temperature and radiant star temperatures
are used in the calculation of the room thermal comfort 4.1. Summer design day optimization
(PPD). All other thermal comfort parameters (metabolic
rate, air velocity, and zone vapour pressure), have been The summer design day optimization study has been
been assigned constant values. conducted for the medium-weight building only. Fig. 5,
The HVAC system performance has been simulated using gives the optimum pay-off characteristic between the daily
steady state component models. Both the heating and cool- energy cost and occupant thermal discomfort for a summer
ing coil models have been based on a transfer unit, capacity design day. As would be expected, an improvement in
ratio, effectiveness relationship for a counter flow heat thermal comfort requires an increase in energy use. A
exchanger. The overall conductance for the coils is based comparison between Figs. 6 and 7 illustrates that the higher
on typical values derived from manufacturer’s performance energy cost is driven largely by changes in plant operation
data [16]. The latent heat transfer in the cooling coil has not during the afternoon. The solid lines in both the figures
been modelled. The fan model is a simple polynomial represent the variable bounds in the optimization; the dashed
curvefit of the manufacture’s normalized performance data line for the temperature setpoints is the ambient temperature
[13]. The pressure drop due to the ductwork has been at that hour.
modelled as a simple quadratic resistance; the quadratic Fig. 6 illustrates that the energy cost is practicably unaf-
resistance of the coils is also a function of the number of coil fected by the system operation at 10:00 a.m., with the
rows. The heat recovery device has been modelled as having temperature setpoint being equal to the ambient temperature
a fixed temperature effectiveness, regardless of the air flow and the flow rate tending to the minimum allowable flow rate
rate. Although in some instances, simplified models have (0.2 m3 /s). Conversely, the changes to both setpoints at
been implemented (such as ignoring latent heat transfer), the 4:00 p.m. (Fig. 7), are clearly contributing to the increase
models retains sufficient detail and characteristics to allow in energy cost (the increase in flow rate resulting in more fan
evaluation of the optimization procedure. power and the supply air temperature demanding increasing
levels of mechanical cooling). The trend of little change in
setpoints during the morning and increased cooling in the
4. Results and analysis afternoon is driven by the load characteristic for the building.
In general, the building is comfortable during the morning,
The performance of the multi-criterion genetic algorithm and it is not until the afternoon that the thermal loads increase
has been investigated here for the solution of two optimiza- to a level where mechanical cooling is necessary to maintain
tion problems. In each case, for a given building construc- comfort. Hence, the greatest coupling between the thermal
tion, the pay-off is sort between the energy cost and occupant comfort and the energy use occurs during the late afternoon
thermal discomfort. However, in the first optimization pro- when the thermal loads are the highest.
blem, only the summer design day of operation is consid- It is informative to investigate the setpoint solutions for
ered, which reduces the scale of the optimization problem the extreme points on the curve, since these give an insight to
and consequently makes it easier to solve. In the second the extremes of system operation and to some extent, the
optimization problem, all three design days are included in degree to which the solutions are optimal. Fig. 8 illustrates
the analysis with the pay-off between energy cost and the supply air setpoints for the lowest energy cost solution
occupant discomfort being sort for all 3 days. (and the highest discomfort solution). It would be expected
The building construction is fixed for each optimization that the lowest energy cost would result from a setpoint
example, (with the problem variables relating only to the schedule that demanded no mechanical cooling and flow
Table 3
Building construction
External wall Facing brick air gap, gypsum Facing brick, airgap, insulation, Facing brick, air gap, insulation,
sheathing, insulation, wall board concrete block, plaster solid concrete block, plaster
Internal wall Wall board insulation, wall board Plaster, concrete block, plaster Plaster, concrete block, plaster
Floor and ceiling Floor tile, concrete Floor tile, concrete Floor tile, concrete
J.A. Wright et al. / Energy and Buildings 34 (2002) 959–972 965
rates set to the minimum allowable value. Fig. 8 illustrates it would be expected that energy use would increase during
that this is the case for the solution obtained from the the afternoon as the thermal load increases (hence more
optimization. The corresponding profile of thermal discom- mechanical cooling is required in order to minimize the
fort (Fig. 9), throughout the day indicates that the maximum thermal discomfort). Fig. 10 indicates that this is the case,
discomfort during the occupied period (8:00 a.m. and with the increase in thermal load first being offset by an
5:00 p.m.) occur late in the afternoon when the summer increase in flow rate. As the load increases further however,
thermal loads are the highest. it can only be offset by a lower supply air temperature. This
The solution for the highest energy cost (and lowest occurred at 1:00 p.m., the change in temperature setpoint
thermal discomfort) is somewhat harder to predict, although being so marked that the supply air flow rate was also
reduced so as to not over-cool the building. Note also that the during the occupied period (Fig. 11), indicates that the
supply air temperature during the early part of occupancy is control schedule is effective in maintaining a near minimum
above the ambient temperature; this occurs since during (5% PPD) discomfort throughout the occupied period,
summer operation, the building maybe slightly cool during which is in stark contrast to the lowest energy cost solutions
the early part of occupancy until the thermal gains increase. (Fig. 9).
Hence, the supply air temperature setpoint is higher than the Inspection of the solutions for the extreme points on the
cool morning ambient temperature so as to not over-cool the pay-off curve suggests that they are consistent with the
building. The corresponding profile of thermal discomfort perceived operation of the HVAC system and are optimal.
The extent to which the pay-off curve is optimal is examined The energy cost and the thermal discomfort are very close
further by comparing a solution for 10% PPD with that for the two optimizations. The multi-criterion optimization
obtained from the single criterion optimization. The corre- resulted in an energy cost of £0.3125 and the single criterion
sponding solution on the pay-off curve was obtained from optimization, £0.3120; clearly an insignificant, if not
the single criterion minimization of the energy cost, with the unquantifiable difference. The maximum thermal discom-
maximum thermal discomfort of 10% PPD set as a con- fort was 9.99% PPD for the multi-criterion optimization and
straint on the optimization. 9.96% for the single criterion optimization.
Fig. 12, illustrates the supply air setpoints from multi- made some use of free cooling during the off-peak period,
criterion optimization closest to 10% PPD. As expected, whereas the multi-criterion optimization has not. The solu-
the highest demand for mechanical cooling is in the late tion from the multi-criterion optimization also uses more
afternoon, the temperature setpoints following the ambient energy through a generally higher flow rate setpoint,
temperature during the early part of occupancy. Fig. 13 although this is compensated for by a slightly higher supply
illustrates that the corresponding solution obtained from the air temperature setpoint in the late afternoon.
single criterion optimization. The main differences between It can be concluded, that for this example, the multi-
the two solutions is that the single criterion optimization has criterion optimization is finding a near optimal pay-off
characteristic between the daily energy cost and zone ther- ‘‘genetic’’ operators to influence the selection of the solu-
mal discomfort. The extreme points on the curve are con- tions used in the production of each new generation.
sistent with the expected behaviour of the system, and Fig. 14 illustrates the energy cost discomfort pay-off
comparison with a solution found from a single criterion curves for each design day. Not surprisingly, the moderate
optimization confirmed that the difference in energy cost ambient conditions of the swing season day, result in least
was unquantifiable. As to which solution the ‘‘Decision pay-off between energy-cost and discomfort, with the worst
Maker’’ should choose, it would seem that the solution with discomfort only being in the order of 8% PPD. Conversely,
the lowest energy cost is a good solution, considering that the high thermal gains associated with the summer design
the corresponding thermal discomfort is only 15.0% PPD, day result in the greatest pay-off between energy cost and
(this not withstanding that the sensitivity of the solution has occupant discomfort.
not been investigated). The optimality of this set of solutions is confirmed here
by comparing the pay-off curve for the summer design
4.2. Simultaneous three design day optimization day with the curve obtained from the single season opti-
mization, (the optimality of the solution from the single
In the second optimization problem, the pay-off charac- design day optimization having been confirmed by inspec-
teristic between energy cost and discomfort is sort for the tion of the results). Fig. 15 illustrates that the two pay-off
three design days simultaneously. Since the days are apart in curves are very similar, which suggests that the solutions
time (and in season), the control strategy for one day does obtained for the three design day optimization are near
not influence the energy cost and discomfort on any other optimal, (strictly, this would only be confirmed by further
day. The pay-off between energy cost and discomfort for inspection of the results, particularly the control variable
each design day is however coupled via the size of the HVAC values).
system. Since the majority of the optimization problem Fig. 16 illustrates the pay-off curves for different building
variables are concerned with the control strategy, the size weights. In this figure, the energy cost and occupant dis-
of the HVAC system can be considered to have a secondary comfort are given as the sum of the values obtained for the
effect, particularly as the search approaches the region of the three design days. Each curve was obtained from a separate
optimum pay-off. This has led to an enhancement to the run of the multi-criterion optimization. The pay-off curves
ranking scheme employed in the multi-criterion optimiza- obtained for the medium-weight and heavy-weight buildings
tion. It is beyond the scope of this paper to discuss this in are almost indistinguishable. This might be expected, con-
detail, but in short a four part Pareto ranking scheme has sidering that the two constructions are very similar (Table 3).
been implemented, three rankings based on the energy cost The light-weight building construction is however, very
and discomfort for each design day, and a fourth ranking that different from the medium-weight and heavy-weight build-
accounts for the coupling due to the size of the HVAC ings, which has resulted in much larger discomfort for the
system. The separate rankings are used in the algorithms the low energy cost solutions.
970 J.A. Wright et al. / Energy and Buildings 34 (2002) 959–972
Fig. 14. Three design day energy cost vs. PPD pay-off.
4.3. Performance of the multi-criterion genetic algorithm multi-criterion genetic algorithm was run with a population of
200 sample solutions, and was allowed to run for a 1000
The search results indicate that the multi-criterion genetic generations. It is common to allow a GA based search to run
algorithm has the potential to find Pareto optimal solutions for for a fixed number of generations since defining stopping
building design problems. It is however, prudent to examine rules for these stochastic optimizers is difficult. The fixed
the computational effectiveness of the algorithm in finding number of generations could result in a total of 200,000 trial
these solutions. In order to determine the pay-off character- solutions. However, it is likely that the number of trial
istic between the energy cost and zone thermal discomfort, the solutions required to find the Pareto optimum set was much
Fig. 16. Energy cost vs. PPD pay-off for different Building Weights.
less, since GA based optimizers exhibit an exponential con- pay-off curve. For the medium-weight building, the multi-
vergence towards the solution. This is illustrated in Fig. 17, in criterion genetic algorithm found 120 solutions along the
which starting from an initial randomly generated population pay-off curve. The maximum number of trial solutions was
of solutions, the search converged to the region of the 200,000, whereas the single criterion genetic algorithm
optimum within 20 generations. It should also be noted that provided a single optimum solution (for 10% PPD), in
all solutions in the initial population were infeasible, but that maximum of 80,000 trial solutions (a population of 80
feasible solutions were found within four generations. individuals with a maximum of 1000 generations). Hence,
The power of the multi-criterion genetic algorithm is also 119 more solutions were obtained for 2.5 times the number
apparent from the number of solutions obtained along the of trial solutions.