Ao

HF

a

ARRAA

KARN

1

maewn

tKsat

ea

tept

h0

Energy and Buildings 84 (2014) 160–168

Contents lists available at ScienceDirect

Energy and Buildings

j ourna l ho me pa g e: www.elsev ier .com/ locate /enbui ld

pplication of neural networks in predicting airtightnessf residential units

rvoje Krstic ∗, Zeljko Koski, Irena Istoka Otkovic, Martina Spanicaculty of Civil Engineering Osijek, University of Osijek, Drinska 16a, HR-31000 Osijek, Croatia

r t i c l e i n f o

rticle history:eceived 9 July 2014eceived in revised form 31 July 2014ccepted 7 August 2014vailable online 22 August 2014

eywords:

a b s t r a c t

The paper describes the need to investigate airtightness of existing residential units in order to achieveadequate energy efficiency. There is a brief explanation of the blower door method used in field testingof airtightness in residential units. The mentioned method was used to create a database of airtightnesstest results obtained in situ at 58 residential units in the local area. The analysis of the results of thesemeasurements indicates that there are four input parameters, which had been previously investigatedand evaluated, that had a dominant effect on the test results. The examining of the possibility of success-

irtightnessesidential unitseural networks

fully predicting airtightness was conducted by applying neural networks. The response results indicatethat there is potential for successful prediction of airtightness, which was confirmed by independentvalidation through 20 additional measurements. In order to be able to generalize the results, furtherresearch needs to be conducted, using a bigger sample of measured data, and additional configurationsof neural networks need to be examined.

© 2014 Elsevier B.V. All rights reserved.

. Introduction

Global future and development of our civilization dependainly on the provision of sufficient amount of cheap energy

nd protection of the environment. The ever growing demand fornergy requires the development of new technologies connectedith utilization of energy from renewable sources, but also a ratio-al consumption of the available energy.

Saving resources of fossil energy was the central requirement athe internationally recognised conferences in Rio de Janeiro, Berlin,yoto, The Hague and Bonn. It is expressed in the target of a 20%hare of renewable energies in overall EU energy consumption and

20% reduction of greenhouse gas emissions by 2020 adopted athe EU spring summit 2007 and re-affirmed in 2008 [1].

Buildings are complex energy systems and biggest individualnergy consumers, the European building sector is responsible forbout 40% of the total primary energy consumption [2].

This great amount of energy is partly wasted because buildingshat were constructed several decades ago do not meet the energy

fficiency requirements according to current legislation in Euro-ean Union. A significant percentage of such buildings will continueo be used for many more years, and unless they are renovated in

∗ Corresponding author. Tel.: +385 91 224 0758; fax: +385 31 274 444.E-mail addresses: hrvojek@gfos.hr, hkrstic@gmail.com (H. Krstic).

ttp://dx.doi.org/10.1016/j.enbuild.2014.08.007378-7788/© 2014 Elsevier B.V. All rights reserved.

terms of energy performance, they will continue to needlessly con-sume massive amounts of energy for heating, cooling, appliancesoperation and lighting.

When it comes to new buildings that will be built in the nearfuture, the European Directive on Energy Performance of Buildingsprescribes that all new public buildings constructed after 2018 shallbe almost zero-energy buildings, and after 2020 the same shallapply for all newly build residential buildings [3]. This especiallyhighlights the issue of existing energy-inefficient buildings whichneed to be renovated in terms of their energy performance as soonas possible. From the practical building technology point or view,it is rational to combine insulation with the air-tightening of thebuilding envelope [4], because air-tightness is an important factorrelating energy efficiency, thermal comfort and indoor air quality ofbuildings. For these reasons, understanding air-tightness is impor-tant for new design and retrofitting of existing buildings [5]. Whenmeasuring the airtightness of buildings, a blower-door is used tofind the relation between the pressure difference over the build-ing envelope, �P [Pa], and the airflow rate through the buildingenvelope, Q [m3/h] [6].

The term “building envelope” refers to several types of opaqueand transparent perimeter structures or elements which separate

the internal heated (cooled) space from the external unheated(uncooled) space. Each of these perimeter structures contributesto the overall thermal quality in proportion to their share in thetotal surface of the building envelope. With decreased energy

d Buildings 84 (2014) 160–168 161

cwiqth

abiiCttHU

ooodecttwt

2

fldimeisdibsbtba

1

••

•

o

•

•

H. Krstic et al. / Energy an

onsumption there is also less CO2 emission in the atmosphere,hich significantly contributes to environmental protection. Exist-

ng residential buildings vary significantly in terms of thermaluality of the building envelope, which is expressed by thermalransmittance coefficient U (W/m2 K) and depends on transmissioneat loss coefficient and ventilation heat loss coefficient.

The research for this paper was preceded by field testing ofirtightness of residential units, the sample of which was formedased on a classification of selected elements of residential build-

ng envelopes. The conducted research was the first of this typen the Croatia and it took place as part of the IPA Cross-borderooperation Programme between the Republic of Hungary andhe Republic of Croatia under the title “Air tightness investiga-ion of rooms from the point of view of energy and comfort”, (ID:UHR/1001/2.1.3/0009), which was co-financed by the Europeannion.

The aim of this paper is to conduct further analysis of thebtained results of measurements conducted in the area of the cityf Osijek and to explore the possibility of predicting airtightnessf residential units by using neural networks. Airtightness pre-iction provides insight into the characteristics of residential unitnvelopes without conducting field tests, which contributes to effi-iency in preparing building reconstruction projects. The results ofhe research will be used in future processes of systematic renova-ion of residential buildings in terms of their energy performance,ith the aim to achieve adequate energy efficiency and appropriate

hermal comfort and to meet the EU recommendations.

. Methods for measuring airtightness

The term “airtightness” pertains to the intensity of uncontrolledow of air through the building envelope as a result of pressureifference between interior and exterior air. Uncontrolled leak-

ng of air can occur through joints, connections between differentaterials, dilatations and other permeable points in the building

nvelope. Limiting airtightness requirements are often to be foundn building regulations [6]. The method most often used for mea-uring airtightness is the pressure difference method (the “bloweroor” test), and the application of the method is described in detail

n the HRN EN 13829:2002 standard (Thermal performance ofuildings – Determination of airtightness of buildings – Fan pres-urization method), which is prescribed in the Republic of Croatiay the Technical Regulation related to rational use of energy andhermal protection in buildings (Official Gazette No. 110/08). Eachuilding’s air tightness was measured by using a Blower Door inccordance with EN ISO 13829.

All experiments’ results are acceptable as they fulfil EN ISO3829 criteria, which require [7]:

the wind speed lower than 6 m/s,the product of maximum building height (m) and temperaturedifference between outdoor and indoor dry bulb temperature tobe lower than 500 mK, andthe building’s volume lower than 4000 m3.

In EN ISO 13829, depending on the purpose, there are two meth-ds described [8,9]:

Method A—the condition of the external building envelope needs

to be representative for the season of the year when heating orcooling systems are used, without additional sealing.Method B—all the designed openings in the building envelopeneed to be sealed.

Fig. 1. Blower Door fan with manometer DG-700.

Measurements were carried out following method A – commonbuilding use – of EN ISO 13829 while also applying a pressure dif-ference of 50 Pa in order to fulfil the requirements of EN ISO 13790.Prior to initiating the measurement of airtightness in buildings itis necessary to gather the required data, these being data on floorarea, volume and envelope surface area of the building. For the pur-poses of this research survey was conducted in order to gather thenecessary above mentioned data.

This data can be classified in four basic groups:

(1) general information about the building,(2) information about the reconstruction of the building,(3) information about the building envelope (dimensions and types

of joinery, types of window glazing, layers and layer thicknessof walls, ceilings and roofs), and

(4) meteorological data (temperature, wind speed, air pressure,relative humidity and precipitation).

The air changes per hour at 50 Pa (n50, ACH at 50 Pa) is com-monly used measure of building airtightness [10]. Measuring ofairtightness was conducted using the device “Minneapolis Blow-erDoor”, shown in Fig. 1. The device is connected to a computervia an USB port and an accompanying program called TECTITEEx-press 3.6. Once the device has been installed in accordance with themanufacturer’s recommendations, and after all the required datahas been entered into the program on the laptop computer, thedevice measures the initial pressure difference between exteriorand interior space, which may not exceed 5 Pa. If the differenceis greater than that, the measuring needs to be repeated or post-poned until the wind, which significantly influences the pressuredifference between interior and exterior space, recedes.

3. Previous research

The leakage airflow rate might be evaluated using predictivemodels determined from experimental databases. In the specificliterature, there are several databases for many countries such as:

the United States, Greece, Finland, Spain, France, Italy, Australia,Canada which are extensively used to deduce mathematical modelsfor the infiltrated air change rate for different types of buildings[11].

1 d Buil

maovwhaacawa

dbstmdmvef

m2himfiaa

easridwwicoftwsbAh1i

imtot

2a5o

62 H. Krstic et al. / Energy an

Findings based on comparing results of airtightness measure-ents without statistical analysis of data were made in Canada

nd Finland. Different types of structures and construction meth-ds were observed in Canada [12]. Based on measured airtightnessalues for 100 newly built houses, it was found that terraced housesere less airtight than detached family houses, and that familyouses with garages and several storeys have higher than aver-ge airtightness values [12]. In France information was gathered onirtightness measurements in order to determine the factors whichould be useful in improving energy performance of buildings [13],

total of 218 measurements conducted on residential buildingsere used, based on which a simulation was made which provided

n estimate of the effect of airtightness on energy consumption.In Estonia airtightness research was conducted on 32 new

etached light-framed houses, their airtightness was comparedased on the following variables: number of storeys, quality of con-truction and supervision, construction technology (being built athe site or assembled) and the type of ventilation system [14]. The

ean air leakage at the pressure difference of 50 Pa in the entireatabase was 4.2 m3/(h m2) (the minimum being 0.9 m3/(h m2) andaximum 17.9 m3/(h m2)) that is somewhat higher than standard

alue of 3 m3/(h m2). It was found that the number of storeys,xecution quality of works and supervision of construction wereactors which mostly affect the amount of airtightness.

Due to ever increasing demands for cooling buildings byechanical means, airtightness measurements were conducted on

0 houses in Greece, and they were observed in correlation with theeight and exposure of the buildings, environmental conditions and

nternal heat fluctuations [7]. A correlation between air tightnesseasurements at a 50 Pa pressure difference and the total window

rame length was noticed, mainly at the “low air tightness” build-ngs. The total frame length affects the air leakage of each housend its air tightness. However, the sample of buildings is very smallnd a large number of experiments are required.

In Finland the correlation between airtightness of the buildingnvelope, air infiltration and energy consumption was observed in

typical modern family house in cold climate conditions [15]. Thetudy was conducted as a sensitivity analysis of a simulation modelepresenting a single family house in different conditions. The typ-cal Finnish climate conditions and wind, ventilation systems andistribution of points of air leakage were observed as main factorshich affect air infiltration. Based on these factors a simple modelas adopted for a rough estimate of annual air infiltration for a typ-

cal family house. By studying the effect of air infiltration on energyonsumption, the results showed that air infiltration causes 15–30%f total energy consumption for heating and ventilation. It wasound that the average air infiltration rate and energy consump-ion for heating increase in an almost linear progression togetherith the increase of airtightness values further on according to the

imulation results, the correlation between the airtightness of theuilding envelope and the annual infiltration rate is almost linear.verage depressurization test results (n50) of the Finnish detachedouses (n = 170) with different types of construction varies from.50 ACH to 5.80 ACH and the mean value for all types of structure

s 3.70 ACH [15].Based on an analysis of similarities of residential unit character-

stics and climatic conditions of Catalonia and France, a predictiveodel was developed in Spain which uses French databases on air-

ightness. The most significant variables for correlation were typef structure, floor area, age of the buildings, number of storeys andype of thermal insulation [16].

In Great Britain airtightness measurements were conducted on

87 residential units constructed in the period after 2006 [17]. Their permeability test results of UK dwellings in this study averaged.97 m3/(h m2) at 50 Pa, which shows a considerable improvementver those built in history and investigated in previous research.

dings 84 (2014) 160–168

Comparing to the international context, the more recent new-buildUK dwellings were air-tighter than those in some other countrieslike Greece and US, but were still significantly air leakier than thosein Canada and Scandinavia, as reported in previous research [17].Airtightness was observed in correlation with various factors whichinclude method of construction, type of residential unit, design andmanagement concept, season of the year, number of significantcracks, and the values of building envelope surface area and floorarea of the residential unit. Interaction was observed between theconstruction method, management context and type of residentialunit. A predictive model was developed by using linear regression[17].

A study conducted in Ireland included airtightness measure-ments in 28 family houses built between 1944 and 2008 whichrepresented the majority of residential units which are to berenovated in the future [5]. Testing found the mean air per-meability of the pre-1975, 1980s, and 2008 dwellings to be7.5 m3/(h m2), 9.4 m3/(h m2) and 10.4 m3/(h m2), respectively. Thisstudy observed the effects of various factors affecting buildingenvelope airtightness. The focus was on type of structure, age, var-ious solutions for design details and various degrees of renovation.A comparison of the measured airtightness results and observedfactors challenged the general belief that new residential buildingshave lower airtightness values than old ones. During the courseof the study certain defects were found in design details, execu-tion and supervision of construction, and a conclusion was madethat improvement of those defects is crucial for obtaining a qualityproduct [5].

The USA has the biggest database on residential unit airtightnessmeasurements. Using this database the airtightness of residentialunit envelopes with different characteristics was analysed, andseveral models were made using regression analysis [18,19]. Theresults of measurements conducted on more than 100,000 houseswere collected in the USA in 2011, and together with previousdata they comprise the ResDB database based on which the regres-sion model for estimating the value of normalized airtightness fordetached houses in the USA was developed. Year of construction,climate zone, floor area, height of the house, type of foundations,location of ventilation systems and energy class of family houseswere the variables that were used for correlation with normalizedairtightness value.

To characterize the US housing stock, air leakage data of 134,000single-family detached homes was analysed by Wanyu R. Chanet al. [12], analysis revealed that homes that have higher NL(normalized leakage) tend to be older, occupied by low-incomehousehold, smaller in floor area, multi-storey, and with likely airleakage through its foundation and ducts. Conversely, homes thatare newer, rated for energy efficiency, larger in floor area, single-story, with minimal leakage through its foundation and ductslocated inside the conditioned space have lower NL [12]. Theseresults show that many factors can influence building envelopeairtightness, including the construction and design of the homes,and possibly how well the homes are maintained. The two fac-tors that have the largest impact on air leakage, year built andclimate zones, further point out that airtightness may be impactedby weathering of the building envelope leading to settling and leaksto develop over time. The longevity of the different components ofa home, such as insulation and air sealing, may also be a contribut-ing factor that explains the relationship between NL and year built[12].

Earlier work on airtightness estimation: The reviewed refer-ences have been grouped into three categories: (1) estimation

based on multiple regression, (2) estimation based on rough char-acteristics of the building and (3) estimation based on componentleakage and geometry of the building. In the following, emphasis isput on the component leakage methods [6].

H. Krstic et al. / Energy and Buildings 84 (2014) 160–168 163

0,00%

5,00%

10,00%

15,00%

20,00%

25,00%

30,00%

35,00%

Prior to 1919. 1919.-1945. 1946. -1960. 1961. -1970. 1971.-1980. 1981. -1990. 1991. -1995. 1996.-200 1. Other (unknown and

Share of reside n�al unit s according to year of construc�on in the total num ber of residen �al un its in the Repub lic of Cro a�a

Share of reside n�al unit s according to year of construc�on in the total num ber of residen �al un its in the Air Tightness project database

F in the

r

•

•

•

4m

ipc1

dstCittt

(avcTn

Neural networks are applicable in a great number of practi-cal problems the resolving of which requires knowledge which ishard to specify, or for systems wherein the relationship between

Table 1Results of measuring airtightness.

Period of construction Number of samples Mean value n50

Prior to 1945 5 9.691945–1965 11 8.431966–1975 13 10.87

ig. 2. Graphical presentation of comparison between the age of the housing stock

Summarized previous research regarding buildings airtightnesseveals following:

Airtightness can be affected by design and management concept,design details, type of structure, method of construction, type ofthermal insulation, the number of storeys, building envelope sur-face area, floor area, execution quality of works and supervisionof construction, season of the year, climate, how well the unitsare maintained, type of unit, joinery, frame materials, windowframe length, total frame length, age of the buildings, number ofsignificant cracks and management context [5,7,12,14,16,17].The leakage airflow rate might be evaluated using predictivemodels determined from experimental databases [11] andLiterature review in this article did not disclose any similarresearch in which neural networks are used for buildings airtight-ness prediction.

. Analysis of the database of results obtained by theeasurements

For the purposes of this study a database on residential unitsn the area of Osijek-Baranja County was made. The database com-rises 58 residential buildings, 47 of which are in the territory of theity of Osijek (81.36% of residential buildings), while the remaining1 (18.64%) residential buildings are located in suburban areas.

In order to determine whether the residential buildingsatabase is representative in terms of year of construction andhare of residential units according to year of construction in theotal number of residential units at the level of the Republic ofroatia, a comparison was made which is presented graphically

n Fig. 2. It is visible that shares of residential units according tohe year of construction as used in this study are proportionate toheir respective shares in the total number of residential units athe level of the Republic of Croatia.

Measured value of airtightness at a pressure difference of 50 Pan50) for entire database of residential units ranged between 0.76nd 19.64 (1/h). Calculation was made of the average airtightness

alues for non-renovated residential units in individual periods ofonstruction as shown in Table 1, also visible on the chart in Fig. 3.hese results show a trend of significant improvement in airtight-ess in the periods of recent construction, which could be expected

unfinis hed)

database of this study and the age of the housing stock in the Republic of Croatia.

considering the technical and technological advancements in con-struction. The given curve which reflects the changes in airtightnessof residential units throughout different periods of construction iscomparable to the curve showing the changes in average quality ofthermal protection of external walls of residential units [20].

The effect of certain observed input parameters on airtightnessresults is variable. A noticeable effect on airtightness results comesfrom the quality of transparent parts of the residential unit enve-lope, i.e. windows, as well as the quality of their installation. Anexample of this can be given in the form of measured airtight-ness values (n50) at two different locations. The same residentialunits, following a complete alteration of windows, exhibit signif-icant reduction in airtightness, from the initial 6.39 (1/h) to 1.13(1/h) and from 5.73 (1/h) to 0.94 (1/h). The analysis of results ofmeasurements provided the selection of 4 parameters that maypotentially have the biggest effect on measured airtightness. Theseare the quality of transparent and of opaque parts of residentialunit envelopes, the percentage of transparent parts in the envelopeand the percentage of residential unit envelope which borders with(i.e. is directly exposed to) exterior space. Although some previousresearch indicated that buildings age can influence buildings air-tightness, this parameter was not used due to the fact that majorityof studied buildings went to retrofitting process in order to achievebetter energy efficiency over the years.

5. Predicting airtightness by using neural networks

1976–1985 8 5.261986–1995 4 3.001996–2005 4 2.32After 2006 4 2.99

164 H. Krstic et al. / Energy and Buildings 84 (2014) 160–168

0

2

4

6

8

10

12

PRIOR TO 1945. 1945. - 1965. 1966.-1975. 1976. - 1985. 1986.-1995. 1996. - 2005. FROM 2006. TO

THI S DAY

E ME

50) ac

iiatern

5

nufoit

TP

AIR LEAKA G

Fig. 3. Mean value of airtightness (n

nput and output data has an existing, but not clearly (in theoret-cal and mathematical terms) described connection. Good resultsre obtained in analysis and prediction of results in non-linear sys-ems, which is what most real-life systems usually are. Rumelhartt al. (1986) provided first settings for a back-propagation type neu-al network [21,22], which showed great potential for predictingon-linear temporal data series.

.1. Database for neural network learning

For the purposes of testing the possibilities of using neuraletworks for predicting airtightness in residential units measuredsing the blower door method, a evaluation system was created

or evaluating input parameters which were found in the coursef the study to have an effect on airtightness results, as describedn Section 4. The input data for the database used for learning inhe neural network for predicting airtightness of residential units

able 2ossible values of corrective factors for input parameters used in predicting airtightness o

Input parameter 1

Opaque part of residential unit envelope

Wall thickness Wall thinner than 25 cm Wall thicker thanCorrective factor 0.80 1.00Ceiling/floor thickness Floor/ceiling thinner than 20 cm Floor/ceiling thickCorrective factor 0.80 1.00Wall material Full brick Masonry blocks

Corrective factor 0.90 0.90

Ceiling/floor material RC Monta/Fert

Corrective factor 1.00 0.90

Thermal insulation thickness TI from 0 to 5 cm TI 5 cm and moreCorrective factor 0.75 1.00

Input parameter 2

Transparent part of residential unit envelope

Quality of installed joinery Old joinery no maintenance Old joinery with mCorrective factor 0.40 0.60

Maintenance quality Poor Good

Corrective factor 0.40 0.60

Opening frame material Metal Wood

Corrective factor 0.80 0.90

Type of opening glazing Single Double

Corrective factor value 0.40 0.60

AN VAL UE, n5 0

cording to periods of construction.

were classified in four groups of input parameters. These are thefollowing parameters:

1. opaque part of the residential unit envelope,2. transparent part of the residential unit envelope,3. percentage of transparent parts in the residential unit envelope,

and4. percentage of residential unit envelope exposure.

The input parameters of opaque and transparent parts of res-idential unit envelopes also include different characteristics ofmaterials and structures used as the criteria for corrective factors.The corrective factor is a numerical value which is associated with

each of the mentioned input parameters 1 and 2, depending ondifferent characteristics as shown in Table 2.

Higher values of corrective factors are associated with build-ing envelope elements that theoretically have better airtightness

f residential units.

25 cm

er than 20 cm

Block brick Siporex0.80 0.80Wooden Slanted roof0.50 0.75

aintenance Newer joinery (since 2000) New joinery with maintenance0.80 1.00Very good Excellent0.80 1.00PVC1.00Double insulated glass Triple insulated glass0.80 1.00

H. Krstic et al. / Energy and Buildings 84 (2014) 160–168 165

Table 3The value range of input parameters 1 and 2 for predicting airtightness in residentialunits and output parameter, n50.

Value Inputparameter 1

Inputparameter 2

Outputparameter, n50

Minimum 0.4413 0.0576 0.7600Maximum 0.8800 0.8000 19.6400

pfleftscp

aigo

p

•

•

poe

ttmos

aviu

5

bds

TVu

INPUT LAYER

HIDDEN LAYERS

OUTPUT LAYER

Standard deviation 0.0873 0.2499 4.3715Average value 0.6885 0.5146 4.9025Median 0.0668 0.5760 3.1400

roperties i.e. thicker construction elements (walls, ceilings,oors. . .), less porous materials and high quality openings (influ-nce of glazing type, frame material type. . .). Higher correctiveactors also correspond with execution quality of works, how wellhe elements are maintained, design details and age of the con-truction elements and openings. According to abovementioned allorrective factors were obtained empirically and according to therevious research findings stated in Section 3.

The percentage of transparent parts of residential unit envelopend percentage of residential unit envelope exposure i.e. values ofnput parameters under Nos. 3 and 4 were calculated depending oneometrical and floor plan characteristics of residential units, basedn data collected using the questionnaire mentioned in Section 2.

The numerical value of the first input parameter for the opaqueart of residential unit envelope was obtained as the sum of:

Product of corrective factors pertaining to walls (wall material,wall thickness and thermal insulation) and the share of walls inthe residential unit envelope, andProduct of corrective factors pertaining to ceilings/floors(thickness and material of ceilings/floors) and the share of ceil-ings/floors in the residential unit envelope.

The numerical value of the second input parameter for the trans-arent part of residential unit envelope was obtained as the productf all corrective factors for the transparent part of residential unitnvelope.

The basic statistical indicators of input parameters 1 and 2 forhe observed sample of residential units are presented in Table 3ogether with statistical indicators for the output parameter i.e.

easured values of airtightness in residential units. Value rangef input parameters 3 and 4 with basic statistical indicators is pre-ented in Table 4.

After defining this system of valuating input parameters, anutomated pattern was made for calculating corrective factors andalues of input parameters in order to obtain a database for learn-ng of a neural network for prediction of airtightness of residentialnits.

.2. Predicting airtightness

Artificially designed neural networks have been developedased on analogy with biological nervous systems. A detailedescription of this analogy is available in the bibliography [21]. Inhort, a neuron receives an input signal as information sent from

able 4alue range of input parameters 3 and 4 for predicting airtightness of residentialnits.

Value Input parameter 3 Input parameter 4

Minimum 0.0176 0.1684Maximum 0.3637 0.9042Standard deviation 0.0473 0.2014Average value 0.0604 0.5977Median 0.0513 0.6610

Fig. 4. Configuration of the selected back-propagation neural network.

another neuron or external source, and it processes this informationand transfers it to the next neuron, or external exit point. The abilityof receiving stimuli, processing them by activating and transferring(or non-transferring) an exit signal to other neurons, comprises acomplex communication network, which is much like a biologicalnervous system. The basic analogy with the human brain, the mostcomplex biological nervous system as far as we know, is the artifi-cially created neural networks’ ability to learn and generalize basedon experience.

This paper examines the possibility of applying neural networksin predicting airtightness of residential units measured by theblower door test method. A relatively small database containing58 examples of airtightness values measured in residential unitswas used for neural network learning, so the results obtained needto be evaluated in that context. Input data for the neural networklearning database are described in Section 5.1. Output data is air-tightness measured by the blower door test method (n50), which isdescribed in more detail in Section 2.

Analysis of neural network prediction was conducted by usingthe NeuroShell2 program, and the network selection criterion isthe best generalization on the test data set. The test data set wasgenerated by random selection of 20% of data from the completedatabase, while the remaining data was used for learning in thenetwork i.e. the training data set. An independent validation ofnetwork generalization capability in Excel by using the “predict”function provided by the neural network was conducted using aset of 20 new measurements.

The best response came from the back-propagation Ward netnetwork, with three hidden layers and 5 neurons in each hiddenlayer, as shown in the network configuration in Fig. 4 and Table 5.Activation functions (Table 5) are explained in more detail in thebibliography [23].

Learning success indicators are the correlation coefficient, mean,minimum and maximum prediction error and other. Correlation

coefficient between measured data and neural network predictionis 95.37%, and the obtained mean prediction error is 0.835.

Table 5Basic characteristics of the selected neural network.

Layers Neuron number Activation function

Input layer Layer 1 4 Linear [−1.1]Hidden layers Layer 2 5 Linear

Layer 3 5 Hyperbolic tangentLayer 4 5 Gaussian compact

Output layer Layer 5 1 Logistic

166 H. Krstic et al. / Energy and Buildings 84 (2014) 160–168

0

5

10

15

20

25

30

35

40

45

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53 55 57

n50

Measured Neural network predic�on

Fig. 5. Measured airtightness n50 and airtightness obtained by prediction through Ward net network.

ve

Aaoitwrgvemsc

od

0

2

4

6

8

10

12

14

0 2 4 6 8 10 12 14 16 18 20

n50

Number of data set

Measured Neural netwo rk predi c�on

Fig. 6. Measured airtightness n50 vs. airtightness obtained by prediction.

The diagrams in Figs. 5 and 6 show the measured airtightnessalues and prediction through the Ward net neural network on thentire database.

The number of neurons in hidden layers is not strictly defined. small number of neurons reduces the ability to learn and gener-lize, while a too great number of neurons increases the danger ofver-training1 the network and also increases the network learn-ng time. Overtraining is the level of network training at whichhe generalization ability of the network declines, with the net-ork tending to remember information from the training data set

ather than to learn on them. Insufficient training results in big-er prediction error, while over-training leads to great predictionariability. Basically—during the course of learning the predictionrror for a test data set (validation data set) decreases until theoment when the network becomes too-trained, at which point it

tarts to increase, while in the training data set the prediction errorontinues to decrease.

1 Statistically, when the number of parameters in a model increases, the degreef model freedom decreases and there is possibility of over-training on the trainingata set [23].

Fig. 7. Measured airtightness n50 and airtightness obtained by prediction using avalidation data set.

Due to the small database which was available for neural net-work learning there was a danger of over-training the network.Over-training affects the generalization ability of the network, andit is most easily evaluated on an independent data set which thenetwork had no access to during the training stage. One should takeinto account that the test data set generated in the NeuroShell2 hadonly 11 data sets, independent validation was conducted using anew set of airtightness measurements subsequently conducted onresidential units, the values of which are shown in the graph inFig. 7.

An additional database containing subsequently measured air-tightness values of residential units, which was created analogouslywith the basic database for neural network learning, was used as avalidation data set for evaluating the generalization ability of theneural network which gave the best response. Validation data setused the “predict” function of the neural network which was drawnfrom Excel, ensuring that the mentioned data were not accessibleto the neural network during training.

A comparison between the measured airtightness values andairtightness values obtained by prediction using an independent

validation data set is shown in the diagram in Fig. 7.

The criterion for evaluation generalization success on thisnew, independent data set was the absolute value of mean

d Buil

prwg

6

adoccciocrc5at

wrs

ptw(itiavtoiIcat

dftbc

a(iihuns

totcba

H. Krstic et al. / Energy an

rediction error, which is 0.965. The mentioned value is compa-able to the mean prediction error obtained in NeuroShell2 (0.835),hich indicates that the selected neural network has the ability of

eneralization on an unknown data set.

. Discussion

In forming the database of selected residential units in the localrea, there was successful matching of age of residential units in theatabase with the age of residential units at the level of the Republicf Croatia in order for the database to be representative. After theonducted airtightness measurements the mean values were alsoalculated for the characteristic division according to periods ofonstruction of residential units. The obtained mean values weren line with the expectations for individual periods of constructionf the residential units, considering the technological and techni-al advancements in construction, but also the adoption of stricterules and regulations connected with energy efficiency when itomes to construction of buildings. Airtightness was measured in8 residential units, and the results of measuring the mean value ofirtightness at a pressure difference of 50 Pa (n50) range from 0.76o 19.64 (1/h).

The working group’s idea was to create a prognostic modelhich would provide basic insight into the characteristics of

esidential unit envelopes without the need to conduct field mea-urements, by using neural networks.

The database for neural network learning included four inputarameters for which it was estimated that they might poten-ially have an effect on airtightness results. These are parametershich describe the opaque (input parameter 1) and transparent

input parameter 2) parts of residential unit envelopes. The numer-cal values of input parameters were obtained by analysing theype of material, thickness of elements, characteristics of thermalnsulation and the type of transparent elements and their char-cteristics (Table 1). A part of the calculation of the numericalalues for the first two parameters includes a subjective evalua-ion of construction execution and installation, maintenance andther characteristics of residential unit envelope elements, whichs introduced by using corrective factors ranging from 0.40 and 1.0.nteractive tables were made which enable fast determination oforrective factors for any residential unit and, at the same time,utomatic determining of the value of the first two input parame-ers.

Further research should also take into account the fact thatifferent users of this method could potentially assess correctiveactors differently although those factors are limited between cer-ain boundaries as stated in Table 2. This needs to be more examinedy applying sensitivity analysis on prediction results depending onhosen values of corrective factors by different users i.e. observers.

Parameters for neural network learning under No. 3 (percent-ges of transparent parts of residential unit envelopes) and No. 4percentages of residential unit exposure) can be easily determinedf information on floor plans and geometrical characteristics of res-dential units had previously been collected. Parameters 3 and 4ave values expressed as percentages and thus they can get valuesp to 1.0. This is how each of the four input parameters for neuraletwork learning was reduced to numerical indicators of the samecale.

The database for neural network learning included 58 residen-ial units, which represents a relatively small database, so it cannly provide a basic insight into the neural network response and

est the potential of this type of prognostic model. The obtainedorrelation coefficient of 95% can in small databases be causedy over-training of the network, but then problems would haverisen in the process of independent evaluation of the network’s

dings 84 (2014) 160–168 167

generalization ability. In order to evaluate the generalization abil-ity using an unknown data set, a validation database was usedwhich was created by subsequent measurements which included20 new residential units. The database for neural network vali-dation included measurements of airtightness in 19 residentialunits in the area of the city of Osijek and one measurement in itssurroundings. The measurements for the validation data set wereconducted in accordance with the previously mentioned regula-tions and they include all the parameters like in the basic database.For numerical values of input parameters the values of correc-tive facts were determined completely analogously with the basicdatabase for neural network learning. The result of evaluation ofgeneralization ability using the validation data set shows that theneural network provides good prediction also on unknown data,which means that the prognostic model has potential which iscertainly worth further studying.

7. Conclusion

This paper examines the possibility of using neural networks increating a predictive model for measured values of airtightness ofresidential units in local conditions. This paper reports the air tight-ness test results of 58 dwellings ranging in age from 1912 to 2013.The obtained results show a good response of the neural networkfor the analysed type of problem. The selected back-propagationnetwork with three hidden layers showed the ability of generaliza-tion within a database which the neural network learned on andwas tested with (NeuroShell2), as well as an independent validationset comprising 20 new measurements which was used to test thenetwork’s generalization ability.

The results of this study need to be analysed in the context of arelatively small initial database which was used for neural networklearning, but nevertheless, the obtained results clearly show poten-tial which needs to be additionally explored. For the purposes ofthis study about two dozen different network configurations weretested, but in order to select an optimum network it is necessaryto test a bigger number of different types and configurations ofneural networks. The basic advantage of this type of prognosticmodel for airtightness in residential units is the possibility of fastassessment of the value of airtightness without the need to conductfield measurements, which are quite time consuming. The resultsobtained through the prognostic model can be applied in planningsystematic energy refurbishment of residential buildings in order toachieve adequate energy efficiency and appropriate thermal com-fort, in accordance with EU recommendations in this field.

The basic disadvantage is the relatively small number of resi-dential units which was used for neural network learning, so thecontinuation of research needs to be directed toward creating a sig-nificantly bigger database which will enable the generalization ofconclusions on the applicability of this type of prognostic model inlocal conditions.

References

[1] N. Anisimova, The capability to reduce primary energy demand in EU housing,Energy and Buildings 43 (10) (2011) 2747–2751.

[2] H. Tommerup, J. Rose, S. Svendsen, Energy-efficient houses built according tothe energy performance requirements introduced in Denmark in 2006, Energyand Buildings 39 (10) (2007) 1123–1130.

[3] European Parliament and the Council, Directive 2010/31 EU of the EuropeanParliament and the Council of 19 May 2010 on the energy performance ofbuildings, Official Journal of the European Communities (2010), L 153/21-2.

[4] Ü. Alev, et al., Renovation alternatives to improve energy performance of his-toric rural houses in the Baltic Sea region, Energy and Buildings 77 (0) (2014)

58–66.

[5] D. Sinnott, M. Dyer, Air-tightness field data for dwellings in Ireland, Buildingand Environment 51 (0) (2012) 269–275.

[6] T.-O. Relander, S. Holøs, J.V. Thue, Airtightness estimation—a state of the artreview and an en route upper limit evaluation principle to increase the chances

1 d Buil

[

[

[

[

[

[

[

[

[

[

[

[

68 H. Krstic et al. / Energy an

that wood-frame houses with a vapour- and wind-barrier comply with theairtightness requirements, Energy and Buildings 54 (0) (2012) 444–452.

[7] A. Sfakianaki, et al., Air tightness measurements of residential houses in Athens:Greece, Building and Environment 43 (4) (2008) 398–405.

[8] ISO, EN 13829:2002, Thermal Performance of Buildings—Determination of AirPermeability of Buildings—Fan Pressurization Method (ISO 9972:1996, modi-fied; EN 13829:2000), ISO, 2002.

[9] Tehnicki propis o racionalnoj uporabi energije i toplinskoj zastiti uzgradama[[nl]]Technical Regulation for Rational use of Energy and ThermalInsulation of Buildings), Narodne novine broj 110/08., 89/09., 79/13., 90/13.Available from 〈http://narodne-novine.nn.hr/clanci/sluzbeni/341775.html〉.

10] TECTITE Express Manual Ver 3.6, The Energy Conservatory. Availablefrom 〈http://materialy.wb.pb.edu.pl/irenaickiewicz/files/2013/01/TECTITE-Express-Manual-Ver-3.6.pdf〉 [cited 2013 15.10.2013].

11] V. Iordache, T. Catalina, Acoustic approach for building air permeability esti-mation, Building and Environment 57 (0) (2012) 18–27.

12] R.C. Wanyu, J. Jeffrey, H.S. Max, Analysis of air leakage measurements of UShouses, Energy and Buildings 66 (2013) 616–625.

13] A. Litvak, M. Fournier, F.R. Carrié, Envelope and ductwork air tightness data in

France: field practice, regulatory approach, energy implications and progressneeded, in: First Blower Door European Symposium, Fulda, Germany, 2006.

14] T. Kalamees, Air tightness and air leakages of new lightweight single-family detached houses in Estonia, Building and Environment 42 (6) (2007)2369–2377.

[

[

dings 84 (2014) 160–168

15] J. Jokisalo, et al., Building leakage: infiltration, and energy performance anal-yses for Finnish detached houses, Building and Environment 44 (2) (2009)377–387.

16] M.I. Montoya, et al., Air leakage in Catalan dwellings: developing an airtightnessmodel and leakage airflow predictions, Building and Environment 45 (6) (2010)1458–1469.

17] W. Pan, Relationships between air-tightness and its influencing factors of post-2006 new-build dwellings in the UK, Building and Environment 45 (11) (2010)2387–2399.

18] H.S. Max, A.M. Jennifer, Buildings X Conference—Thermal Performance of theExterior Envelopes of Whole Buildings, U.S. DOE, Oak Ridge National Labora-tory, Clearwater Beach, FL, 2007.

19] J. McWilliams, M. Jung, Development of a Mathematical Air-Leakage Modelfrom Measured Data, Measured Data, 2006.

20] Z. Koski, I. Istoka, I. Milicevic, Classification of building elements as a functionof air permeability measurements, Gra –devinar (2013) 65.

21] I.A. Basheer, M. Hajmeer, Artificial neural networks: fundamentals, comput-ing, design, and application, Journal of Microbiological Methods 43 (1) (2000)3–31.

22] P.J. Werbos, Backpropagation through time: what it does and how to do it,Proceedings of the IEEE 78 (10) (1990) 1550–1560.

23] G. Zhang, B. Eddy Patuwo, M.Y. Hu, Forecasting with artificial neural networks:the state of the art, International Journal of Forecasting 14 (1) (1998)35–62.