www.elsevier.com/locate/solener

Solar Energy 81 (2007) 102–110

The London Heat Island and building cooling design

Maria Kolokotroni *, Yuepeng Zhang, Richard Watkins

Brunel University, Mechanical Engineering, School of Engineering and Design, Uxbridge, Middlesex, UB8 3PH, UK

Received 3 October 2005; received in revised form 1 June 2006; accepted 7 June 2006Available online 24 August 2006

Communicated by: Associate Editor Matheos Santamouris

Abstract

London’s urban heat island increases the mean air temperature which affects the demand for heating and cooling buildings. Measuredair temperature data have been used as input to a building energy simulation computer program to assess the heating and cooling load ofa typical air-conditioned office building positioned at 24 different locations within the London Heat Island. It is found that the urbancooling load is up to 25% higher than the rural load over the year, and the annual heating load is reduced by 22%. The effect of raisedtemperature and urban context are assessed separately, and the sensitivity of the net impact to the internal gains in a building is deter-mined. For the estimation of peak cooling demand, we propose hourly temperature corrections based on radial distance from London’scentre to be applied to standard published temperatures for the region. For more detailed investigations over the cooling season a rangeof models is available. These are reviewed in this paper and we describe preliminary results of an Artificial Neural Network (ANN) modelthat predicts location specific hourly temperatures for London, taking into account radial distance from central London, hourly air tem-perature measured at the meteorological station and associated synoptic weather data.� 2006 Elsevier Ltd. All rights reserved.

Keywords: Heat island; London; Energy demand; Building design

1. Introduction

Heat islands are well-established consequences of theurban environment. In general, urban centres are warmerthan the surrounding area and this can be beneficial, ornot, in terms of the energy used in providing comfortableconditions in buildings. In hot or cold climates the annualbalance of the impact of a heat island may be clear, but intemperate regions the reduced heating season loads maysignificantly offset the higher cooling loads in the summer.

This paper addresses this impact in the case of London.First and in order to demonstrate the effect, the heatingand cooling loads for a typical office building have beenassessed using measured hourly air temperature data. Themethodology of data collection and quantification of the

0038-092X/$ - see front matter � 2006 Elsevier Ltd. All rights reserved.

doi:10.1016/j.solener.2006.06.005

* Corresponding author. Tel.: +44 1895 266 688; fax: +44 1895 256 392.E-mail address: maria.kolokotroni@brunel.ac.uk (M. Kolokotroni).

London Urban Heat Island has been recently reported pre-viously in Watkins et al. (2002a) and Kolokotroni et al.(2006). Using these measured air temperature data as oneof the input parameters together with additional weatherdata and building construction and operational character-istics, the energy performance of a typical air-conditionedoffice building has been modelled using an energy simula-tion programme in a variety of urban contexts to determinethe effect of the heat island over a year. It will be shownthat location specific air temperature has a marked effecton energy consumption and therefore, designers shouldtake this into account.

The paper provides information of how location specificurban air temperatures in London can be considered forthe calculation of peak cooling demand (Section 4) andproposes a model for hourly air temperature calculationbased on available meteorological station data and loca-tion of site within the Urban Heat Island (Section 5).

M. Kolokotroni et al. / Solar Energy 81 (2007) 102–110 103

2. Modelling the energy demand of a typical office building

Much research has been carried out investigating theeffect of increased air temperature in urban areas on energydemand by buildings. In London, Chandler, using datafrom 1951–60, found a reduction in annual heating degreedays (base 15.6 �C) of about 10% between central Londonand a rural area (Chandler, 1965). More recently, climatechange has been linked to Urban Heat Island and itsimpact on building environmental design. Graves et al.(2001) have presented temperature data for designers andguidelines for reducing the effect of the heat island in Lon-don. GLA (2002) discusses the effect of climate changeincluding Urban Heat Island on London and CIBSE(2005) examines in detail the effect of climate change onthe indoor built environment in the UK.

Elsewhere, in Munich, Brundl reported that annualheating degree days (base 15.0 �C) were 14% lower in a cen-tral urban area compared to a suburban one (Brundl andHoppe, 1984). Energy use for cooling was not consideredbecause summer cooling was not prevalent. Energyrequirements for air-conditioning are considered higherthan for heating (Landsberg, 1981), and both Landsberg(1981) and Taha (1997) concluded that the elevation ofurban temperatures imposes a net energy penalty in severalAmerican cities because of increased cooling requirements.Shading effects of surrounding buildings are oftenneglected in calculating the loads in a building. This canlead to over-sizing air-conditioning equipment, and in gen-eral, neglects a basic aspect of urbanization. In central Ath-ens, the annual cooling load for an apartment block withwindows shaded 50% of the time was found to be 15–50% higher than when modelled using weather data froman open site on a hill 2 km away (Hassid et al., 2000). Alsoin Athens, where the mean heat island intensity exceeds10 �C, it was found that the cooling load of urban buildingsmay be doubled and the peak electricity load for coolingpurposes may be tripled especially for higher set point tem-peratures. During the winter, the heating load of centralurban buildings is found to be reduced up to 30% (Santa-mouris et al., 2001). Akasaka et al. (2002) report that dueto the heat island the cooling load of Tokyo has increasedabout 20% since 1900 with a corresponding decrease ofabout 40% for heating load. Strategies to reduce coolingenergy use in buildings due to the heat island are proposedby Akbari and Konopacki (2005) classified to direct (reduc-ing heat gain through the building shell) and indirect(reducing the ambient air temperature). They have devel-oped summary tables for the US sorted by heating andcooling degree-days based on simulations to quantify theeffect of these strategies.

In this research project, a commercial building energysimulation computer program (TAS) has been used to pre-dict the impact of the heat island and the urban environ-ment on energy demand by buildings. Some preliminaryresults were reported in Watkins et al. (2002b) and Koloko-troni et al. (2004). In this paper, additional analysis is

included (Section 3) and new data on proposed tempera-ture corrections suitable for summer cooling design assess-ment are presented (Section 4).

2.1. External conditions

One year’s measured hourly air temperature data from24 locations within London were combined with regionalweather data to form 24 weather files for the buildingenergy simulation model. Wind speed and direction datawere obtained from Heathrow airport, 23 km WSW of cen-tral London, and humidity, cloud cover, and solar radia-tion data from the London Weather Centre, centralLondon. These regional data were assumed to be applica-ble to all 24 sites – an approximation. Wind speed canaffect infiltration and convective heat transfer. In this studyof an air-conditioned building, both the infiltration andventilation rates are scheduled and are thus independentof wind-speed. The convective heat loss coefficient is variedby the model according to the hourly regional wind speed,but this is not modified here for urban settings.

2.2. Typical Office building

The dominant type of air-conditioned building in Lon-don is the office, and to represent this a typical designhas been selected from a set of designs widely used in theUK in comparative energy studies. It is a standard, as dis-tinct from prestige, air-conditioned office termed theECON 19/3 building; taken from the Energy ConsumptionGuide 19 (BRECSU, 1999). It is a three storey open planbuilding, 9 m high, 30 m long and 15 m wide orientatedwith the longer sides facing north:south, with 60% glazingon these facades. There is clear double glazing with noshading. The end walls are unglazed. Walls and roof areconcrete with insulation. Intermediate floors are of con-crete with false ceilings, and the ground floor is uninsu-lated. The walls have a solar absorptance of 40% and theroof 65%. The surrounding land was set to have 20%ground reflectance to solar radiation. An air-conditioningsystem (vapour compression) and heating system (gas-fired) operates from 06:00 to 18:00 (to include pre-condi-tioning) maintaining the internal air temperature between20� and 24 �C. Fresh air is supplied during occupied hourswith a total air change rate of 1.1/hour (including infiltra-tion of 0.5 ach/hour at all times). Internal gain from lights,occupants and plant is 43 W/m2.

2.3. Urban contexts

In urban areas, buildings usually experience a degree ofover-shadowing which reduces solar gain. This effect hasbeen modelled by surrounding the test building with neigh-bouring blocks to the same height (9 m) at a varying dis-tance depending on the appropriate site categorization(Table 1). The spaces formed between the buildings, thestreet gorges, were given a height to width ratio that varied

Table 1Criteria used for categorizing sites

Ct H/W ratio of street, x Description

1 x = 0 Rural fields, or large park, or trees2 x = 0 Housing near park or field3 x = 0 Urban derelict or unbuilt area4 0 6 x < 0.3 Low density residential area5 0.3 6 x < 0.5 Medium density urban area6 0.5 6 x < 1 High density urban area7 1 6 x < 2 Very high density urban area8 P2 Exceptionally high density urban area

104 M. Kolokotroni et al. / Solar Energy 81 (2007) 102–110

from zero (no over-shadowing) to 2.0 (heavy over-shadow-ing for category 8 – Table 1).

2.4. Internal gains

Internal gains affect the ratio of heating to cooling load,and the relative significance of a change in solar gain withover-shadowing. To determine the effect of this the build-ing was modelled with four alternative levels of internalgain. Base 43 W/m2, +25% 54 W/m2, +50% 65 W/m2 and�25% 32 W/m2. These higher and lower gains are not arti-ficial and are quite likely to be found in practice (BRECSU,1999).

y = -1.40x + 55600r2 = 0.56n = 24

30000

32000

34000

36000

38000

40000

42000

12000 13000 14000 15000 16000Heating load, kWh

Co

olin

g lo

ad, k

Wh

Fig. 2. The relationship between simulated annual heating and cooling

3. Energy demand simulation results

3.1. Main results – base gain

3.1.1. Annual cooling load and mean temperature

Fig. 1 shows the relationship between the measuredmean 24 h temperature of each location over the yearand the annual cooling load.

Approximately two thirds of the variance in coolingload is associated with varying mean temperature(r2 = 0.67, Fig. 1). The cooling loads at three sites are atsome distance from the regression line. Site A is 1.6 kmfrom the centre, has over-shadowing (category 6, Table1) and is 250 m from a large park. Site B is in the city of

y = 3275x - 1873

r2 = 0.67

30000

32000

34000

36000

38000

40000

42000

10.5 11.0 11.51 2.0 12.5 13.0

Mean annual temperature, ˚C

An

nu

al c

oo

ling

load

, kW

h

B

A

C

D

Fig. 1. The relationship between measured annual mean temperature andsimulated cooling load.

London and warmer, but has heavy over-shadowing (cate-gory 7, Table 1). Both these sites have particularly lowcooling loads for their mean temperatures. Site C is3.2 km from London in the centre of the heat island, warm,but in a more open site (category 5, Table 1), and has ahigh cooling load for its temperature. Site D is at the cen-tre, warm and has over-shadowing (category 6, Table 1).Thus at the three warmest sites (B, C and D) the order ofdecreasing cooling load follows the order of increasing

urbanization, i.e., categories 5, 6 and 7 (Table 1).

3.1.2. Annual heating and cooling loads

Fig. 2 shows how locations with higher cooling loadstend to have lower heating loads. A reason for scatter inthe relationship is that some sites with higher mean temper-atures (increasing cooling load), experience over-shadow-ing which reduces cooling load.

3.1.3. Annual total load

Fig. 3 shows the total (heating + cooling) load at eachlocation and how this varies with site category. The lineshows the mean of the values for each category. The trendis for total load to increase with increasing urbanization, atleast to site category 5 (where it is about 8.5% higher thanthe total rural load). There appears to be a reduction

load at each location.

47000

48000

49000

50000

51000

52000

53000

54000

0 2 4 7Site category

To

tal a

nn

ual

load

, kW

h

1 3 5 6 8

Fig. 3. Simulated annual total load at each location, separated by sitecategory.

47000

48000

49000

50000

51000

52000

53000

54000

0 10 12 14 16 18 20 22 24 26 28 30

Radial distance Xr, km

To

tal a

nn

ual

load

, kW

h

2 4 6 8

Fig. 4. Simulated annual total load at each location, separated by distancefrom the city centre.

-8.0

-6.0

-4.0

-2.0

0.0

2.0

4.0

6.0

8.0

1 6

Site category

% c

han

ge

wrt

cat

-1

BASE-25% gain+25% gain+50% gain

2 3 4 5 7

Fig. 5. Simulated percentage change in total annual load with sitecategory alone for different internal gains. Rural weather.

M. Kolokotroni et al. / Solar Energy 81 (2007) 102–110 105

(down to 2% of total rural load) after this point, withincreasing over-shadowing, but the small sample makesthis speculative at this stage.

Fig. 4 shows the same total loads, but how these varywith the distance of each site from the city centre. The lineshows the mean of the values at each distance. Total loadtends to increase closer to the city, but the range of scatteralso increases. This may be associated with the increasingcontrast in environments as one approaches a city centre:hard surfaces (open, or heavily over-shadowed) on theone hand, contrasted with large parks. At a radial distanceof 3.2 km (Xr = 3.2), the highest total demand is from ahard surfaced site that has the highest annual cooling loadof all sites. The lowest total demands at Xr = 3.2 are at asite next to a park and a site in a hard area of the Cityof London but heavily over-shadowed. At a radial distanceof 1.6 km the site is near Regent’s Park and has a low totaldemand. This point is responsible for the sudden fall in themean demand line on the graph and may be atypical ofload at this distance.

3.2. Separation of the temperature and urban context effects

– variable gain

There are two main factors affecting load: higher urbantemperatures increasing cooling load, and higher over-shadowing decreasing it, with the opposite effects for heat-ing. The size of these effects when operating independentlyare evaluated here. The sensitivity of the results to vari-able internal gain is also given here; the higher the internalgain the more the annual loads shift from heating tocooling, affecting the balance of the effect of the heatisland.

3.2.1. The effect of context alone

The building energy simulation program has been runusing rural weather acting on the typical air-conditionedoffice (ECON 19/3) building in each of the seven contexts(site categories). This introduces the effects of over-shad-owing, while controlling for external temperature.

Fig. 5 shows the change in total annual load (heat-ing + cooling) with site category. The change is plotted asa percentage of the load in a rural setting (category 1, Table1).

The following observations can be made:

• The effect of site category depends on the level of inter-nal gain.

• For a low internal gain building, increasing over-shad-owing leads to an increase in total annual load, evenat the densest urban category 7.

• For the BASE gain building, greater over-shadowingleads to an increase in total load, but starts to fall forthe most urban site type.

• Higher gain buildings have a reduced total load with anylevel of over-shadowing.

• The BASE gain building’s total load increases by about2% at the urban site (category 6), before falling toalmost the same as in a rural context, at the mostover-shadowed site (category 7).

As noted above, for the BASE and low gain buildings,increasing context density leads initially to an increase intotal annual load, and then a reduction. This is probablybecause of the angle of the sun being higher during thecooling season compared to the heating season. Applyinga wide context affects gain during the heating season muchmore than in the cooling season.

The trend of reducing total annual load with increasingsite category is apparent as early as category 4 for the highgain building. This suggests that this effect, first observed inFig. 3 where temperature was uncontrolled, and for justone site and therefore speculative, is in fact a real effectand not an anomaly.

3.2.2. The effect of temperature aloneThe building energy simulation program has also been

run using the measured hourly temperatures at each loca-tion, but keeping the typical air-conditioned office (ECON19/3) building in the open, as if in a rural context. Thisintroduces the effects of temperature, while controllingfor over-shadowing.

Fig. 6 shows the change in total annual load (heat-ing + cooling) with the annual mean temperature foreach location. The change is plotted as a percentage ofthe load at the rural site temperature (10.7 �C). Polynomial

-8.0

-6.0

-4.0

-2.0

0.0

2.0

4.0

6.0

8.0

10.0

12.0

14.0

10.0 10.5 11.0 11.5 12.0 12.5 13.0

Annual mean temperature, ˚C

% c

han

ge

wrt

Tru

ral

BASE-25% gain+25% gain+50% gain

Fig. 6. Simulated percentage change in total annual load with meantemperature alone for different internal gains. Rural site category.

Table 2Proposed air temperature corrections based on radial distance from citycentre

Hour Distance from city centre Rural reference

0–3 km 3–10 km 10–23 km

0 1.9 1.0 �0.9 �2.31 1.9 1.1 �0.9 �2.52 1.9 1.1 �0.7 �2.03 1.7 0.9 �0.7 �2.34 1.5 0.8 �0.7 �2.35 1.4 0.7 �0.6 �2.36 1.7 1.3 �0.4 �2.5

106 M. Kolokotroni et al. / Solar Energy 81 (2007) 102–110

regression lines have been fitted to help visually in follow-ing the trend of the different sets of points.

The following observations can be made:

• The effect of site temperature depends on the level ofinternal gain.

• In all but the lowest gain building, higher mean temper-atures lead to increased total load.

• In the low gain building, higher temperatures have onlya small effect on total load and at the highest tempera-tures total load decreases.

• The BASE gain building’s total load increases by about7% at the warmest site in the heat island compared tothe load at rural temperatures.

With both increasing site category and mean tempera-ture (both general concomitants of increasing urbaniza-tion), it can be concluded that the size and sign of theeffect on total annual load depend on the level of internalgains. Note that the level of solar gain can also vary, givena fixed degree of over-shadowing, if the facade design is dif-ferent: higher or lower percentage glazing, shading fea-tures, etc. The percentage glazing used here is as given inthe specification for the ECON 19/3 office. The effect ofvarying solar gain is likely to be similar to that of varyinginternal gain.

7 1.5 1.9 0.8 �1.38 1.7 2.0 1.3 �0.79 1.8 2.2 1.5 �0.410 1.6 2.0 1.4 0.711 1.4 1.8 1.3 0.012 1.7 2.1 1.3 �0.213 1.6 2.2 1.4 �0.214 0.6 1.1 0.4 0.915 0.9 1.4 0.5 0.016 0.9 1.4 0.4 �0.117 0.7 1.0 0.2 �0.218 0.4 0.8 0.0 0.819 0.4 0.4 �0.4 �0.520 0.6 0.3 �0.9 �1.921 1.0 0.3 �1.4 �2.722 1.5 0.6 �1.3 �2.723 1.5 0.6 �1.3 �2.8Average 1.3 1.2 0.0 �1.1

4. Summer design temperature

The results presented above suggest that air temperaturecontributes the highest percentage of change in energydemand and indicate that temperature variations withinthe London Heat Island should be considered in the designof buildings. In particular there are specific implications forpassive cooling techniques due to increased temperatures,(Geros et al., 2005).

In order to estimate peak cooling loads, it is desirable tomodify published temperature values (CIBSE, 2006) toaccount for the position of a building within the LondonHeat Island. Such a method was first proposed in Graveset al. (2001). CIBSE (2006) includes hourly design data

for typical days for each month. These days are derivedfrom the 97.5 percentile daily global irradiation exceedencedata. From available data (July and August 1999 and 2000)Graves et al. (2001) selected 6 days (5% of data) with thehighest solar radiation and calculated the temperature dif-ference (site air temperature minus Bracknell air tempera-ture). From these, tables were proposed containing airtemperature correction data for locations within the GreatLondon Area.

In order to simplify the procedure, the authors haverepeated this study with one adjustment. Heathrow airtemperatures have been used instead of Bracknell’s as Hea-throw is the met station used for climatic data for London(CIBSE, 2001).

The London region has been divided into three concen-tric annular zones, and the mean heat island intensitieswith respect to Heathrow computed for each zone andhour. Different radii were examined for significant changesin the zone means before selecting: 0–3 km (core tempera-ture), 3–10 km (semi-urban), and 10–23 km (suburban).Beyond this lie rural areas. We propose these hourly tem-perature corrections to Heathrow data for summer peakcooling estimation based on radial distance (CIBSE,2006). Table 2 presents the proposed adjustments to beadded to Heathrow design data. Heathrow is towards theedge of the London Heat Island but significantly warmerthan a rural site. The heat island air temperature correc-tions in Table 2 are significantly lower than if they weregiven relative to a true rural site. The real Urban Heat

M. Kolokotroni et al. / Solar Energy 81 (2007) 102–110 107

Island Intensity can be obtained by subtracting the ruralreference column from an annular column in Table 2.For example, for a site located in the core temperature area(0–3 km) at 03.00 h, the real Urban Heat Island Intensity is4 K (1.9 � (�2.3)).

5. Model for predicting location specific hourly temperatures

However, as discussed in Section 3 site category (micro-climate) has been shown to have a marked effect on theheat island intensity and additional work is required toenable us to include the effect of physical site characteristicsinto urban heat island prediction algorithms. We firstreview available techniques for developing models andtheir suitability for use by building designers. Based on thisreview, an Artificial Neural Network (ANN) model is pro-posed as the most suitable and some preliminary results arepresented.

5.1. Review of models

The Urban Heat Island Intensity (UHII) has been pri-marily studied by urban climatologists. Their researchfocuses on microclimatology and boundary-layer climatol-ogy and a comprehensive review on the advances over thelast twenty years has been recently published (Arnfield,2003). It includes ten recommendations for furtherresearch, one of which is that ‘simple models are needed

to estimate UHII within urban areas, as a function of time,

weather conditions and structural attributes, for practicalapplications such as road climatology, energy conservation

and weather forecasting’.The following categories of models have been developed

to predict the UHII:

1. Climatology models: These are detailed but are com-plex and require specific expertise for accurate predic-tions. For example the Colorado State Universitymesoscale model (CSUMM) has been developed topredict UHII (Taha, 1999; Leuzzi and Monti, 2004;Taha et al., 2000). The model requires the knowledgeof soil parameters such as density, specific heat, ther-mal diffusivity, moisture availability, roughness lengthand albedo in addition to knowledge of the large scalegeostrophic wind and the initial profiles of temperatureand specific humidity. Such models might be too com-plicated for the purpose of estimating air temperaturesto assess specific building cooling requirements. Meth-ods of coupling urban models and building thermalperformance models have been proposed (Flor andDominguez, 2004; Tanimoto et al., 2004) which couldbe useful in estimating building cooling loads accu-rately but still the detail of required inputs is quitecomplex. Climatologic models are necessary in casesof complex terrain, where large temperature differencesare formed due to changes in elevation or in land useor both.

2. The UHII has also been studied by building physicistswho have developed empirical models. For example,the Cluster Thermal Time Constant (CTTC) (Swaidand Hoffman, 1990) model has been developed for pre-dicting the UHII in hot-arid climates and has also beenextended to semi-arid climates (Elnahas and William-son, 1997). Runnals and Oke (2000) describe an empir-ical model to disentangle the multiple controls onVancouver UHII, including time of the day, wind speed,cloud cover and rural thermal admittance. Such modelsdepend on measured local physical parameters and arevery sensitive to empirically measured coefficients.Developed empirical models use experimentally mea-sured coefficients which are site-specific and thereforeonly applicable to certain urban environments and cli-matic conditions. Interpolation and application of thesemodels to different urban environments would not bepossible without large errors. Sensitivity of results tosmall variations of coefficients would also prohibitinterpolation.

3. Computational Fluid Dynamics (CFD) models havealso been used, mainly to predict air flow within theurban environment but also temperature fields (forexample Tabahashi et al., 2004; Assimakopoulos et al.,2006). Such models, although reasonably accurate inmany cases, require a large amount of input data withthe corresponding uncertainties. Models are usuallybased on the wind and temperature and can not takemany other important factors as inputs, althoughimprovements to basic CFD models have been proposed(Tabahashi et al., 2004).

4. The UHII has also been studied using analysis of vari-ance and regression, primarily to examine the effect ofweather conditions on the UHII (as reported by Arnfield(2003)), including recent studies (for example Morrisand Simmonds, 2000). Many of these studies confirmedthe first empirical generalisation offered by Runnals andOke (2000). UHII decreases with increasing wind speedand cloud cover, UHII is greatest during anticyclonicconditions and is best developed in the summer or warmhalf of the year. The authors have used statistical tech-niques in the case of London to study the effect ofweather conditions and distance from the thermal centrewith success (Watkins, 2002). However, it was alsofound that the UHII is dependent on site category andits physical characteristics. Probability methods suchas the Monte Carlo method has been also used to pre-dict UHII (for example Montavez et al., 2000; Sailorand Fan, 2002).

5. The UHII as affected by synoptic climatic parametershas also been studied using Artificial Neural Networks(ANNs) (Mihalakakou et al., 2002; Kim and Baik,2002). ANN models are suited to the proposed researchbecause they are particularly good at representing anycomplex non linear functions whose analytical formsare difficult or impossible to obtain. The UHII problemis complicated and some relationships between different

108 M. Kolokotroni et al. / Solar Energy 81 (2007) 102–110

factors are still not fully understood. Kim and Baik(2002) have compared the ANN model results to a mul-tiple linear regression model and concluded that theANN model gave improved predictions by up to 6.5%.

5.2. Proposed model

A preliminary study was carried out using similar tech-niques (ANNs) to investigate the effect of synoptic climaticconditions experienced over London. In addition, distanceof location from the centre of the heat island was includedas one of the parameters. To start with, standard neuralnetwork architecture procedures were followed, whichinvolve the following:

1. prepare the network database;2. design the neural network;3. train the network;4. test and diagnostic check;5. adjust the parameters of the neural network;6. repeat steps 3–5, until the performance of network is

satisfactory.

5.2.1. Input data

As mentioned before the aim is to predict hourly ambi-ent air temperature at specific locations. Therefore, hourlyambient air temperature of the meteorological (reference)station and hourly synoptic data of London are two setsof the input data. As indicated before, radial distance fromcentral London is one of the parameters to influence loca-tion specific air temperature. Therefore, radial distance isthe third input for the network.

Ten stations were selected from the available databasefor which full historical hourly data for one year (01/01/2000 to 30/09/2000) is available. Initially, one day was con-sidered (from 00:00 to 23:00) as a whole cycle. As datafrom the period 01/01/2000 to 30/09/2000 were used, 274sets of data for each station were produced which includethe following:

Synoptic weather data: These include cloud cover,humidity, wind speed, radiation diffuse and radiation glo-bal; London Weather Centre hourly data were used.

Ambient air temperature of meteorological station: Lon-don Heathrow is used as the reference station – this isbecause Heathrow weather data are typically used for ther-mal modelling for London locations.

Radial distance from Central London: The BritishMuseum has been used as the centre point of London inthis project – this is the centre of the monitored data.Therefore the radial distances for each of the ten investi-gated locations are as follows to cover the range of moni-tored data distances:

Great Russell street: 0.24 kmWest Smithfield: 1.6 km

North Road: 3.2 kmFitzjohn’s Avenue: 4.8 kmSt.Mark’s Road: 6.4 kmIvy Road: 9.6 kmHallywell Crescent:12.8 kmRiver Road: 16.0 kmKent Avenue: 19.2 kmSevenoaks Road: 22.4 km

5.2.2. Design of the network

The choice of the type of network and learning rule isvery important for successful results and its choice isdependent on the type of available data for input and theform of the output required. In this case, a definite functionbetween the inputs and outputs could not be identified andtherefore the feed-forward network seems the most suitabletype of network. There are a number of learning rulesavailable, of which back-propagation has been used in sim-ilar work before (Mihalakakou et al., 2002). Back-propaga-tion was created by generalising the Widrow–Hoff learningrule to multiple-layer networks and nonlinear differentiabletransfer functions. Input vectors (data) and the corre-sponding output vectors (data) are used to train a networkuntil it can approximate a function, associate input vectorswith specific output vectors, or classify input vectors in anappropriate method as defined by the trainer. Trials withthe available data indicated that this is a suitable networkand learning rule. Three layers were designed; input layer,hidden layer and output layer. The next step was to choosea suitable training algorithm; nine promising algorithmswere tested and the scaled conjugate gradient back-propa-gation (trainscg) was selected for further development; thisproved to have best performance in function approxima-tion for this particular network design. All parameters weredetermined by following the repeating procedure of ‘train –test – adjust’ parameters, until the performance was satisfy-ing; the final learning rate was 0.2, error goal is 0.5, nervesin input layer 7, nerves in hidden layer 19, nerves in outputlayer 1.

As the inputs are hourly data, building the model for24 h results to 145 elements in input vectors, and 24 ele-ments in output vectors. This treatment produced a rela-tively large network compared with the number ofelements in input vectors; therefore the total number ofsamples (data) becomes too small to train such a network.The solution was to divide it into 24 small networks, onefor each hour of the day. In the following sections, we pres-ent the results of one example (the network for 15:00 h) toexplain the process.

6. Results

Two important factors, which could be used to evaluatethe trained network, are training error and predictionerror. 100 sets of trained data were selected as the inputfor the network and simulative results were produced.

Fig. 7. ANN training phase: correlation between simulative and measuredresults of the ANN model. Trained sets of data are used (measured results)as input vectors and output vectors (simulative results) are predicted. Thetraining phase aims for the measured results to be the same as thesimulative results or within a small training error (less than 0.3 in this case;mean square errors varied in the range of 0.1–0.3).

M. Kolokotroni et al. / Solar Energy 81 (2007) 102–110 109

These results were compared with the corresponding mea-sured data to find the training error. Fig. 7 shows that thereis good agreement between simulative and measuredresults. Most relative error values are less than 0.3. Meansquare errors varied in the range of 0.1–0.3.

From the 274 sets of data available from each station,200 sets were then used for training and 74 for testing thepredictions. The 200 sets of untrained data were randomlyselected (20 from each station) for prediction testing. Fig. 8shows that the network also performed well in prediction.85% of the simulative result has a good agreement withthe corresponding measured data, while relative error val-ues are less than 0.6 and the mean square errors rangebetween 0.2 and 0.5.

These preliminary results demonstrate that the proposednetwork is successful in considering synoptic weather data

Fig. 8. ANN prediction phase using 200 sets of untrained data. 85% of thesimulative results have good agreement with the corresponding measureddata; relative error values are less than 0.6 and the mean square errorsrange between 0.2 and 0.5.

and distance from the centre of the heat island to predictsite-specific air temperature in London. Work continuesto improve the accuracy of predictions as described in Sec-tion 8.

7. Discussion

The effect of the London Heat Island on energy used forheating and cooling depends on the degree of urbanizationin a particular location, radial distance from the centre(depth within the heat island) and the relative contributionof solar gain to total gains in a building. Heating dominatedbuildings will tend to benefit from a heat island whereas theopposite is the case for cooling dominated ones. Previousresults from an analysis of variance of daytime temperatureshowed a fairly consistent ordering of mean temperature (orheat island intensity) with site category, when controllingfor radial distance (Watkins et al., 2002a). It should benoted that the site categories and associated street gorgeratios (Table 1) reflect the urban densities in London, andmay not be appropriate for other cities where, for example,very much narrower streets are the norm.

For a typical air-conditioned office building (ECON 19/3) operating with internal gains of 43 W/m2, the annualheating load decreases by 22% from a rural site, with cool-ing load increasing 25%. Mean summer daytime tempera-tures for different sites are found to be associated with asite categorization based on the nature of surfaces andstreet gorge ratio. For a given radial distance, daytime tem-peratures tend to peak at category 5 sites. The annual pri-mary energy use is also associated with site category, againtending to peak at category 5 sites, and reducing at moreurbanized (over-shadowed) sites, despite them beingwarmer.

When the ECON 19/3 building has 25% lower internalgain, total annual energy demand tends to be less sensitiveto position within the heat island. This is because the heat-ing and cooling loads are more balanced and a reducedheating load, deeper within the heat island, is matched byan increased cooling load. With higher gains, +25% and+50%, the pattern followed is similar to the base gainbuilding.

Therefore, estimation of cooling load due to externalconditions will depend on the location of the buildingwithin the Urban Heat Island. Prediction of locationdependent ambient temperature will improve accuracy insuch calculations.

A simplified method for adjusting temperatures basedon Heathrow peak cooling calculations’ design data andradial distance from the centre of the London Heat Islandis proposed (Section 4).

However, for low energy design strategies, availability oflocation specific hourly air temperatures available for ayear would contribute to more accurate predictions fortheir effectiveness. It was shown that an ANN model mightbe a suitable model to predict such values. Preliminaryresults from an ANN model based on meteorological air

110 M. Kolokotroni et al. / Solar Energy 81 (2007) 102–110

temperature, synoptic weather data and distance from thecentre indicate a good agreement. Such a model can beused for more detailed prediction of location specifichourly air temperature. In this model, 85% of the simula-tive result agrees with the corresponding measured dataand this is consistent with findings by Kim and Baik(2002) who reported a 6.5% improvement in comparisonto a multiple linear regression model. In our case, it hasproduced a marked improvement in comparison to linearregression analysis. This is because radial distance has beenincluded as one of the inputs in addition to weather data.However, the match is not perfect yet indicating that theremight be other parameters affecting the result.

8. Concluding remarks

Site category (microclimate) has been shown to have aneffect location specific air temperatures within the LondonUrban Heat Island. More work is needed to understand themain parameters for including such physical site character-istics into the prediction model. Therefore, future work willinvestigate how to classify and quantify physical urbancharacteristics with the aim to incorporate these into theANN model. We believe that by incorporating locationrelated weighting factors within the ANN model, the accu-racy of predicting location specific air temperatures wouldbe improved both for design applications using typicalweather data and for forecasting applications using realweather data.

References

Akasaka, H., Nimiya, H., Soga, K., 2002. Influence of heat islandphenomenon on building heat load. Building Services EngineeringResearch and Technology 23 (4), 269–278.

Akbari, H., Konopacki, S., 2005. Calculating energy-saving potentials ofheat island reduction strategies. Energy Policy 33 (6), 721–756.

Arnfield, A.J., 2003. Two decades of urban climate research: a review ofturbulence, exchange of energy and water, and the urban heat island.International Journal of Climatology 23 (1), 1–26.

Assimakopoulos, V.D., Georgakis, C., Santamouris, M., 2006. Experi-mental validation of a computational fluid dynamics code to predictthe wind speed in street canyons for passive cooling purposes. SolarEnergy 80 (4), 423–434.

BRECSU, 1999. Energy use in offices (Energy Consumption Guide 19),BRECSU.

Brundl, W., Hoppe, P., 1984. Advantages and disadvantages of the urbanheat-island – an evaluation according to the hygro-thermic effects.Theoretical and Applied Climatology 35 (1–2), 55–66.

Chandler, T.J., 1965. The Climate of London. Hutchinson&Co (Publish-ers) Ltd.

CIBSE, 2006. Guide A, Environmental Design, CIBSE.CIBSE, 2001. Guide J, Weather, Solar and Illuminance Data, CIBSE.CIBSE TM36, 2005. Climate Change and the Indoor Environment:

Impacts and Adaptation, CIBSE.Flor, F.S., Dominguez, S.A., 2004. Modeling microclimate in urban

environments and assessing its influence on the performance ofsurrounding buildings. Energy and buildings 36 (5), 403–413.

Geros, V., Santamouris, M., Karatasou, S., Tsangrassoulis, A., Papa-nikolaou, N., 2005. On the cooling potential of night ventilationtechniques in the urban environment. Energy and Buildings 37 (3),243–257.

GLA , 2002. The impact of Climate Change on London: summary report,Greater London Authority, Available from: <http://www.lon-don.gov.uk/gla/publications/environment/londons_warming02.pdf>.

Graves, H., Watkins, R., Westbury, P., Littlefair, P., 2001. CoolingBuildings in London – Overcoming the heat island, BR431, CRC.

Elnahas, M.M., Williamson, T.J., 1997. An improvement of the CTTCmodel for predicting urban air temperatures. Energy and Buildings 25,41–49.

Hassid, S., Santamouris, M., Papanikolaou, N., Linardi, A., Klitsikas, N.,2000. The effect of the Athens heat island on air conditioning load.Energy and Buildings 32, 131–141.

Kim, Y.H., Baik, J.J., 2002. Maximum urban heat island intensity inSeoul. Journal of Applied Meteorology 41 (6), 651–659.

Kolokotroni, M., Watkins, R., 2004. In: The impact of the London heatisland on building heating and cooling loads, ‘Contemporary Problemsof Thermal Engineering’, Gliwice-Ustron, Poland, pp. 233–240.

Kolokotroni, M., Giannitsaris, I., Watkins, R., 2006. The effect of theLondon urban heat island on building summer cooling demand andnight ventilation strategies. Solar Energy 80 (4), 383–392.

Landsberg, H., 1981. Urban Climate. Academic Press.Leuzzi, G., Monti, P., 2004. Numerical modelling of flow and dispersion

in Roma area. In: 9th Int. Conf. on Harmonisation within Atmo-spheric Dispersion Modelling for Regulatory Purposes.

Mihalakakou, G., Flocas, H.A., Santamouris, M., Helmis, C.G., 2002.Application of neural networks to the simulation of the heat islandover Athens, Greece, using synoptic types as a predictor. Journal ofApplied Meteorology 41, 519–527.

Montavez, J.P., Jimenez, J.I., Sarsa, A., 2000. A Monte Carlo model ofthe nocturnal surface temperatures in urban canyons. Boundary-LayerMeteorology 96 (3), 433–453.

Morris, C.J.G., Simmonds, I., 2000. Associations between varyingmagnitudes of the urban heat island and the synoptic climatology inMelbourne, Australia. International Journal of Climatology 20 (15),1931–1954.

Runnals, K.E., Oke, T.R., 2000. Dynamics and controls of the near-surface heat island of Vancouver, British Columbia. Physical Geog-raphy 21, 283–304.

Sailor, D.J., Fan, H.L., 2002. Modelling the diurnal variability of effectivealbedo for cities. Atmospheric Environment 36 (4), 713–725.

Santamouris, M., Papanikolaou, N., Livada, I., Koronakis, I., Georgakis,C., Assimakopoulos, D.N., 2001. On the impact of urban climateto the energy consumption of buildings. Solar Energy 70 (3), 201–216.

Swaid, H., Hoffman, M.E., 1990. Prediction of Urban Air TemperatureVariations using the Analytical CTTT Model 14, 313–324.

Tabahashi, K., Yoshida, H., Tanaka, Y., Aotake, N., Wang, F., 2004.Measurements of thermal environment in Kyoto city and its predictionby CFD simulation. Energy and Buildings 36 (8), 771–779.

Taha, H., 1997. Urban climates and heat islands: albedo, evapotranspi-ration, and anthropogenic heat. Energy and Buildings 25, 99–103.

Taha, H., 1999. Modifying a mesoscale meteorological model to betterincorporate urban heat storage: a bulk-parameterization approach.Journal of Applied Meteorology 38 (4), 466–473.

Taha H., Chang, S., Akbari, H., Heat Island Group 2000. Meteorologicaland air quality impacts of heat island mitigation measures in three UScities. Environmental Energy Technologies Division, Lawrence Berke-ley National Laboratory, Berkeley, CA 94720.

Tanimoto, J., Hagishima, A., Chimklai, P., 2004. An approach forcoupled simulation of building thermal effect and urban climatology.Energy and Buildings 36 (8), 781–793.

Watkins, R., 2002. The impact of the urban environment on the energyused for cooling buildings. Ph.D. Thesis, Brunel University.

Watkins, R., Palmer, J., Kolokotroni, M., Littlefair, P., 2002a. TheLondon heat island – results from summertime monitoring. BSER&T23 (1).

Watkins, R., Palmer, J., Kolokotroni, M., Littlefair, P., 2002b. Thebalance of the annual heating and cooling demand within the Londonheat island. BSER&T 23 (4), 207–213.