Journal of KIAEBS Vol.15, No.2, April pp RESEARCH ARTICLE 머신러닝기반저층주거건물에너지소요량예측모델개발 - 단독주택, 다세대, 연립주택을대상으로 유동철 1 ㆍ김경수 2 ㆍ최창호 3 ㆍ조성은 4 ㆍ장향인 5*

Size: px

Start display at page:

Download "Journal of KIAEBS Vol.15, No.2, April pp RESEARCH ARTICLE 머신러닝기반저층주거건물에너지소요량예측모델개발 - 단독주택, 다세대, 연립주택을대상으로 유동철 1 ㆍ김경수 2 ㆍ최창호 3 ㆍ조성은 4 ㆍ장향인 5*"

해진 황보
3 years ago
Views:

1 Journal of KIAEBS Vol.15, No.2, April pp RESEARCH ARTICLE 머신러닝기반저층주거건물에너지소요량예측모델개발 - 단독주택, 다세대, 연립주택을대상으로 유동철 1 ㆍ김경수 2 ㆍ최창호 3 ㆍ조성은 4 ㆍ장향인 5* 1 미래환경플랜건축사사무소건축친환경신기술연구소책임연구원, 2 미래환경플랜건축사사무소건축친환경신기술연구소선임연구원, 3 광운대학교건축공학과교수, 4 일본동양대학교정보연계학부조교수, 5 미래환경플랜건축사사무소건축친환경신기술연구소연구소장 Development of a Machine Learning-based Low-rise Residential Building Energy Consumption Prediction Model Yoo, Dong-Chul 1 ㆍ Kim, Kyung-Soo 2 ㆍ Choi, Chang-Ho 3 ㆍ Cho, Sung-Eun 4 ㆍ Jang, Hyang-In 5* 1 Senior Researcher, Institute of Green Building and New Technology, Mirae Environment Plan, Seoul, Korea 2 Researcher, Institute of Green Building and New Technology, Mirae Environment Plan, Seoul, Korea 3 Professor, Department of Architectural Engineering, Kwangwoon University, Seoul, Korea 4 Assistant Professor, Faculty of Information Networking for Innovation and Design, Toyo University, Tokyo, Japan 5 Director, Institute of Green Building and New Technology, Mirae Environment Plan, Seoul, Korea *Corresponding author: Jang, Hyang-In, Tel: , hijang@mrplan.co.kr OPEN ACCESS Journal of KIAEBS 2021 April, 15(2): pissn : eissn : Received: December 30, 2020 Revised: February 15, 2021 Accepted: February 17, 2021 C 2021 Korean Institute of Architectural Sustainable Environment and Building Systems. This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License ( which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited. ABSTRACT The purpose of this study is to develop a prediction model that can evaluate energy consumption before and after remodeling through a reference model for low-rise residential buildings for which energy simulation evaluation is difficult due to its aging. Specifically, a prediction model to evaluate various building elements before remodeling and a model to predict savings due to the application of energy-saving technology were developed. For the objective, the energy simulation analysis of a building was performed using DesignBuilder per reference area of a Detached house, Multi-family house, and Row House. In addition, the significance of machine learning was compared and analyzed by using R 2, MSE, RMSE, CVRMSE indicators and Python s linear regression, random forest, and neural network. As a result of this analysis, both the model to evaluate the status before remodeling and the model to evaluate the reduction rate according to the energy-saving technology after remodeling showed a high determination coefficient of 0.9 or more for the neuron network. The CVRMSE was analyzed as low as 15% or less. As this is less than the index used in the M&V evaluation presented in the ASHRAE Guideline 14, it was verified that there is a statistical significance. Therefore, this aims to contribute to the basic data and green remodeling promotion project in the energy performance improvement project for the old, private low-rise residential buildings and also in the energy-saving evaluation for buildings. 주요어 : 머신러닝, 저층주거건물, 에너지소요량, 예측 Keywords: Machine learning, Low-rise Residential Building, Energy Consumption, Prediction 152 Journal of KIAEBS Vol. 15, No. 2, 2021

2 서론 연구의배경및목적정부는온실가스배출량 5억 3,600만톤을낮추기위하여 BAU 대비 37% 감축목표중국내감축량을 25.7% 에서 32.5% 로상향조정하였고, 건설분야의온실가스감축량과주요수단을발표하였다 (MOLIT, 2018). 그중신축건축물분야는단계적으로공공, 민간부분의제로에너지빌딩의무화를통해 5.5백만톤을감축목표로선정하였으며, 기존건축물의감축방향은에너지다소비공공건축물녹색건축물전환의무화추진, 그린리모델링활성화를위한중장기방안수립, 민간노후건축물에너지성능개선사업기획지정및재정지원확대등의계획으로 9.6백만톤을감축하는로드맵을발표하였다. 이러한배경으로 2020년 6월한국판그린뉴딜정책이추가적으로발표되면서공공건축물리모델링, 민간노후건축물에너지성능개선사업, 그린리모델링온실가스감축방안등을통해적극적으로추진되고있다. 그중공공건축물리모델링부문은그린리모델링지원사업및 15년이상경과된국공립어린이집, 보건소, 의료시설등 1천여동을대상으로에너지성능향상및효율개선사업을통해에너지성능평가수행이진행되고있으나, 민간노후건축물에너지성능개선사업의경우공공건축물과비교하면평가개소가월등히많으며, 해당건물을세대별로건축물에너지성능평가시개인이신청하여진행하여야하며, 전문지식이부족할경우상당시간이소요되는문제가발생한다. 또한, 해당문제와다른법적인문제로는건축물의에너지절약설계기준및친환경주택건설기준으로인해에너지시뮬레이션평가시입력자료및개선수준파악이용이한아파트및 30세대이상의공동주택과다르게저층주거건물은법적으로도냉난방면적 500 m2이하일경우에너지절약설계기준대상에서제외되며, 대부분 30세대이하임에따라, 개선수준에대한권장사항및법적인규제에대한명확한자료가없다. 그에따라법적인단열규제를적용받지않은저층주거건물은대부분건축도서정보와설비변경이력이유실되어그린리모델링평가시활용할수있는입력인자자료가부족함에따라건축주가스스로자가건축물에대한에너지성능평가시현황파악및에너지절감기술적용에대한절감량평가도어려운실정이다. 이에다양한이유로리모델링전 후에너지성능평가가어려운기존건물을평가하기위해표준건물을개발하고, 표준건물대비절감되는에너지소요량을산출또는예측하는선행연구가이루어지고있으나, 주로아파트및오피스, 상업용건물을대상으로수행되고있다. 기존연구방법은상세난방도일법 (No, 2015) 회귀분석및다중회귀분석 (Lee, 2015) 등을통해수행되었으며, 최근에는인공신경망을활용하여에너지사용량을예측하는연구 (Kim et al., 2018; Park, 2018) 가수행되고있는것으로조사된다. 그러나저층주거건물을대상으로표준모델구축및현황평가, 소요량예측에대한연구는아직까지미비한상황이다. 이에본논문은선행연구인 Kim et al. (2020) 의 저층주거건물의에너지소요량예측을위한에너지인자정의및참조모델개발 의후속연구로서선행논문에서제시한저층주거건물의참조모델을토대로평가의확장및참조모델의활용성증진의목적으로수행하였다. 이를위해저층주거건물참조모델을대상으로다양한인자특성을반영한모델확장 ( 기준면적 ) 에따른소요량예측과에너지절감기술적용에따른절감량예측을수행하였다. 이를 Journal of KIAEBS Vol. 15, No. 2,

통해개발된예측모델을활용시저층주거건물의리모델링전 후를평가할수있으며, 나아가노후건축물에너지성능개선사업등건물에너지절감평가시기초자료및간략평가자료로활용이가능하다. 이에, 해당자료를토대로기초자료가부족한저층주거건물의리모델링전 후에너지성능평가및그린리모델링활성화에기여하고자한다.

3 통해개발된예측모델을활용시저층주거건물의리모델링전 후를평가할수있으며, 나아가노후건축물에너지성능개선사업등건물에너지절감평가시기초자료및간략평가자료로활용이가능하다. 이에, 해당자료를토대로기초자료가부족한저층주거건물의리모델링전 후에너지성능평가및그린리모델링활성화에기여하고자한다. 연구의방법및절차본연구는앞서저층주거건물참조모델개발의후속연구로리모델링전 후에대한평가를위하여현황평가시기준면적평가확장및에너지절감기술에따른효과예측연구를수행하였다. 선행연구는참조모델설정을위한공공기관통계데이터의평균값또는최빈값을기준으로표준모델을정의하고, 에너지입력인자정의를통해시뮬레이션으로참조모델을구성하여해당모델의평가를통해국가통계의사용량과소요량을비교분석하였다. 본논문은앞서개발된참조모델을활용한후속연구로, 기준면적으로만평가가가능했던모델을대상으로다양한주거면적에대응하기위하여기준면적과기준면적의사이값을예측하는모델을기계학습을통해도출하였으며, 참조모델의기준면적을학습자료 (Train Data) 로사용하고, 기준면적과기준면적사이에값을검증자료 (Test Data) 로사용하였다. 또한리모델링후에너지소요량예측을위하여에너지절감기술을선정하고, 해당절감기술을복합적용시절감되는에너지소요량및절감율에대해예측하는연구를수행하였다. 학습및검증자료로활용된모든자료는 Designbuilder 에너지시뮬레이션의결과값이며, 평가시해당값의시뮬레이션결과값을참값으로가정하고오차율을평가하였다. 이에전체연구에대한프로세스및후속연구로해당되는부분을나타낸그림은 Figure 1과같다. Figure 1. Flowchart of study 해당연구를수행하기위하여활용한건물에너지시뮬레이션프로그램은 Energyplus의 Third-Party 프로그램인 Designbuilder를활용하였으며, 기준면적및절감기술변경시유저 154 Journal of KIAEBS Vol. 15, No. 2, 2021

4 의실수를방지하고자자동으로인자를변경해주는 Python의 eppy 알고리즘을활용하여에너지시뮬레이션분석을수행하였다. 수행방법으로는선형회귀분석 (Linear Regression), 랜덤포레스트 (Random Forest), 신경망 (Nueral Network) 분석방법을적용하여평가를수행하였고, 평가의검증방법은시뮬레이션결과와학습을통해예측된결과를비교하여결정계수 (R 2 ), 평균제곱근편차 (RMSE), 평균제곱근오차의변동계수 (CVRMSE) 의지표를활용하였으며, R 2 이높고, RMSE, CVRMSE 가낮게분석되는결과값이예측율이높은것으로검증을수행하였다. 이를통해도출된결과로참조모델을활용하여리모델링전 후평가에대한유효성을검증하였다. 선행연구고찰 참조모델의선정 Kim et al의 저층주거건물의에너지소요량예측을위한정의및참조모델 에서제시하는참조모델을선정하여분석을수행하였으며, 해당표준모델의기준면적및기기설정값은통계적평균값또는최빈값에의해설정된값이며, 해당내용은 Table 1과같다. Table 1. Selection result of energy factor range Variables Unit Number Building type and standard total floor area Type Building - A B C A ( m2 ) Standard total floor area B ( m2 ) C ( m2 ) Variable of Architects Plan form - 1:2 2:1 1.12:1 Number of room - 1~2 2~3 Number of bathroom Balcony form - None Expand Non-Expand Position of floors - A (D) B (E) C (F) Window to wall ratio % Azimuth Climate zone Central Southern Jeju Variable of Passive Construction year Infiltration ac/h Variable of Usage Average people n Schedule % Variable of Active Heating Eff % 80 Cooling COP % 2.88 Lighting W/ m2 10 Equipment load W/ m Type Building : A : Detached house, B : Multifamily house, C : Row house Position of floors : A : Top Floor, B : Middle (Reference) Floor, C : Bottom Floor Journal of KIAEBS Vol. 15, No. 2,

5 기계학습의선행연구고찰기계학습은세계적으로 4차산업이후다양한분야 ( 언어학습, 딥러닝, 시뮬레이션모델링, 기계번역, 소셜네트워크, 로복틱스, IoT, 이미지및영상인식등 ) 에서활용되고있다. 최근건축분야에서도, 인공지능및기계학습을통한연구는점차증가하고있으며 2017년발표된 Kang et al의 건축분야의인공지능기계학습연구동향 에서 2000년국내 2편, 국외 8 편에불과하던논문이 2016년국내 13편, 국외 37편으로증가폭이커짐에따라연구가점점확대된다고설명하였다. 이처럼, 건축환경 설비분야에서기계학습을활용한논문들이다수게재됨에따라건축분야에서기계학습분야는점차증가추세에있는것을확인하였다. 이에본연구분야와직접적관련이있는건축물에너지소요량및사용량을예측하는연구에대하여선행연구고찰을수행하였다. 수행결과, 대부분설비상태예측및건물부하예측에대한분석이주를이뤘으며, 본연구와유사한선행연구는아래의 Table 2와같다. Table 2. Literature Review on machine learning for energy consumption prediction Author Jung (2018) Bae (2019) Nam et al. (2019) Development Purpose Office Building Load Prediction Model Analysis of Energy Consumption According to Economizer Control Method Prediction of Building Energy Consumption Using Deep Learning Prediction of Energy Lee consumption in a building (2020) using life patterns of Single-Person Households Hong (2020) Prediction for Building and Housing Energy Saving Using Weather Information and Machine Learning Research Method ANN, CHAID, Random Forest Assessment Method CVRMSE CVRMSE MBE Neural CVRMSE network Error (Tensorflow) Neural network MLP, CNN, LSTM R, MSE MAE Input Output Differentiation Temperature humidity Temperature (in, out) humidity (in, out) Duct Control Outside Temperature, Solar Radiation Gender, Age, Income, Educ ation level, Occupation Building, Date, Weather data Load Prediction Result Energy consumption reduced during economizer control Cooling, Lighting, Energy Consumption Building Energy Consumption Energy Usage Cost Not Low- Rise Residential Building Characteristics of occupants, not factors of buildings Not Low- Rise Residential Building 이처럼선행연구는주로오피스및상업시설을대상으로기후, 기온, 습도에따라건물의부하및사용량을예측하는논문이대부분이다. 하지만본논문은저층주거건물을대상으로건물인자에따라서리모델링전에너지소요량을예측하고리모델링기술적용후에너지소요량을예측하는모델을개발한것이차별성으로판단된다. 구체적인인자의차별성을설명하면, 건물용도 ( 단독, 다세대, 연립 ), 장단변비, 방의개수, 화장실개수, 발코니의형태, 건물위치, 방위, 지역, 창면적비, 준공년도등의인자를모두활용하여기계학습을수행하였고, 기준연면적만평가할수있는부분을확장하여기준면적사이에있는건물도에너지소요량예측이가능하며, 리모델링후단일기술적용에의한절감량산정이아닌복합기술적용시에도절감되는에너지소요량을예측하는것이본논문의차별점이다. 156 Journal of KIAEBS Vol. 15, No. 2, 2021

6 또한연구방법으로가장많이활용된것은예측식의기본모델인회귀분석 (Linear Regression) 을주로활용하였으며, 선형회귀분석평가이후분석의정확도를높이기위하여다른기계학습알고리즘을적용하여평가를수행하거나, 기계학습에대한필요성을기재하였다. 선형회귀이후사용빈도수로보면인공신경망 (Neural Network) 의빈도가가장높았으며, 인공신경망외에는 Random Forest 방법이수행되는것을확인하였다. 이에본논문은사용률이높은선형회귀, 랜덤포레스트, 인공신경망방법을활용하여평가를수행하였으며, 비교지표로는 R 2 과 CVRMSE를활용하여비교및분석하였다. 리모델링전 후소요량예측을위한예측모델분석 예측모델구성선형회귀는기계학습이아닌회귀식에의하여 X축의예측소요량과 Y축의실제소요량을두고산점도를통해결정계수를도출하는것으로별도의학습없이단순회귀식을통해산출된다. 랜덤포레스트모델의경우, n_estimators는랜덤포레스트안의결정트리개수를정의하는것으로범위는 100~ 500으로지정하였으며, 샘플이몇개이상이어야분할할수있는지에대한규정값으로는 2, 10, 50, 100로설정하였다. 이에 Table 3과같이랜덤포레스트에대한최적매개변수를탐색하기위해그리드탐색을적용하였으며, 그리드탐색의조건은다음과같다. Table 3. Selection of random forest parameter and value Parameter Value Number of estimators 100, 200, 300, 400, 500 Minimum samples to split 2, 10, 50, 100 Minimum Gini impurity to split 0.1, 0.2, 0.3, 0.4, 0.5 인공신경망은회귀모델로일반적으로이용되는완전연결신경망 (fully-connected network) 으로구성하였으며, 독립변수의개수 (39개) 를고려하여노드수가각 16, 16, 8개인세개의레이어로구성하였다. 과적합을방지하기위해각레이어사이에배치정규화 (batch normalization) 레이어를추가하였으며, 200 epoch 이상모델의검증데이터손실 (validation loss) 이개선되지않을때학습을종료하게구성하였다. 위와같은방식을적용하여선형회귀, 랜덤포레스트, 인공신경망으로각각예측모델을구성하여연구를수행하였다. 리모델링전현황평가를위한예측모델 ( 기준면적확장예측 ) 분석선행연구의참조모델로선정된기준면적은단독주택 39, 49, 66, 99, 109 m2, 다가구 ( 세대 ) 주택 39, 54, 82 m2, 연립주택 39, 59, 66, 84 m2이다. 따라서기준면적사이에대하여예측가능여부를판단하기위하여중간면적을제외하고학습자료 (Train data) 를토대로학습을시켰으며, 총개수는 13,843개데이터로구성되어있다. Test 모델은각용도에서중간면적에해당되는면적으로총데이터는 10,459개이며, 해당결과와예측모델을통해도출된결과를토대로검증하였다. Journal of KIAEBS Vol. 15, No. 2,

7 이에 Table 4와같이지역, 준공년도, 건물위치용도, 기준면적을토대로하위변수인장단변비, 방개수, 화장실개수, 발코니형태, 기기발열별로변화되는에너지소요량을예측하여리모델링전다양한인자변경및기준면적변화에대응가능한예측모형을개발하였다. Table 4. Selection of factors and variables for predictive model Contents Value Factor Train Test Type Building Detached house 36, 66, , 99 Multifamily house 39, Row house 39, 89 59, 66 Climate zone Central, Southern, Jeju Construction year U-Value 1980, 1987, 2001 Infiltration 1.25, 2.0, 2.0 Position of floors A (D), B (E), C (F) Plan form 1:2, 2:1, 1.12:1 Number of room A (1~2), B (2~3) Number of Bathroom A (1), B (2) Balcony form None, Expand, Non-Expand Heating Eff 80% Cooling COP 2.88 Lighting 10 W/ m2 Average People (Equip load) 2 (1.97), 3 (2.17), 4 (2.37) 기준면적및설계적특성인자변경에따른시뮬레이션결과값과예측모형의에너지소요량값예측결과, 아래의 Table 5와같이분석되었다. 선형회귀의 R 2 =0.914로높게나타났으며, 랜덤포레스트의경우 R 2 =0.505로낮게분석되었다. 뉴런네트워크의경우 0.987로높은결정계수로분석되었다. Table 5. Energy consumption prediction result according to the change in the reference area Contents Linear Regression Random Forest Neuron Network R MSE 410,739 1,382,324 88,183 RMSE , CVRMSE 19.25% 35.32% 8.92% Units : kwh Graph 랜덤포레스트의경우결정계수가낮게분석된원인은단계적으로수행하는모델특성상학습되지않은기준면적을예측하는데있어조건이없음에따라, 학습의데이터가부족한것으로사료되며, 이에결과가크게차이가난것으로판단된다, 선형회귀가높게나온원인으 158 Journal of KIAEBS Vol. 15, No. 2, 2021

8 로는기준면적에따라에너지소요량이증감하는데있어기준면적인자가지배적으로작용하여에너지소요량이선형으로예측되는것으로사료된다. MSE와 RMSE, CVRMSE는선형회귀 19.25%, 랜덤포레스트 35.32%, 뉴런네트워크 8.92% 로뉴런네트워크가가장낮게분석됨에따라, 예측모델로서결정계수는높으며, CVRMSE는낮음에따라통계적으로유의한것으로판단된다. 리모델링후에너지절감량산출을위한예측모델에너지절감량산출을위한시뮬레이션분석모델은앞서도출한참조모델이기본구성이며, 리모델링전 ( 현황평가 ) 에서활용된각면적에입력된인자별기준값이 Base Model이다. 이에각기준면적별로에너지절감기술을적용하여절감율을분석하였으며, 해당기술은그린리모델링사례를통해활용되는기술로선정하였다. 이에선정된기술및값은아래의 Table 6와같다. Table 6. Selection of variables for energy saving technology Green Remodeling Tech nology Before Remodeling After Remodeling Variables Condition Condition Category Central Southern Jeju Wall Building Construction Roof Year Floor (1980, 1987, 2001) Window Reference Building Energy Conservation Design Standards [ ] Regional division is organized before remodeling LED 10 W/ m2 7 W/ m2 - High Eff Boiler 80% (Heating Eff) 91% (Heating Eff) Energy Efficiency 1 Class High Eff air conditioner 2.88 (Cooling Eff ) 3.2 (Cooling Eff ) Energy Efficiency 1 Class Photovoltaic - 3 kwh Green home Mini Photovoltaic Wh Solar map Insulation Film W/ m2k N Company Product Insulaion Paint W/ m2k S Company Product Dimming - 4 step - Inside Blind Outside Blind - - Apply Apply Winter Season 18:00~07:00 Summer Season 7:00~18:00 Inter Season 07:00~18:00 위와같이선정된기술을참조모델에적용하여분석을수행하며, 단일기술의적용뿐만아니라복합기술적용시절감되는에너지를분석하기위하여직교배열표의방법에따라각기술을그룹화하였으며세개수준으로정의하였다. 세개수준으로정의시, 기술의동시적용및오류를방지하고자유사기술을그룹으로구성하여분석을수행하였다 (ex: 창호 : 실외, 실내블라인드 ). 이에차양은실외블라인드, 실내블라인드, 조명은 LED, LED+ 디밍, 창호는교체와창호필름, 단열은단열재, 단열페인트로그룹을구성하였으며, 해당용도별기준면적당직교배열표방식을활용하여 81개 Case를선정하였으며, 내용은아래의 Table 7과같다. Journal of KIAEBS Vol. 15, No. 2,

9 Table 7. Selection of sampling cases for energy technology effect analysis Passive Location Usage Energy Saving Technology NO Construction Position Climate Azimuth Average Sche year of floors Zone People dule Shading Lighting Window Insulation a central % a central % INS a central % PAI a central % RUV LED+DIM FIL a central % RUV LED+DIM FIL INS a central % RUV LED+DIM FIL PAI a central % BLI LED WIN a central % BLI LED WIN INS a central % BLI LED WIN PAI a jeju % 0 LED FIL a jeju % 0 LED FIL INS a jeju % 0 LED FIL PAI a jeju % RUV 0 WIN a jeju % RUV 0 WIN INS a jeju % RUV 0 WIN PAI a jeju % BLI LED+DIM a jeju % BLI LED+DIM 0 INS a jeju % BLI LED+DIM 0 PAI a southern % 0 LED+DIM WIN a southern % 0 LED+DIM WIN INS a southern % 0 LED+DIM WIN PAI a southern % RUV LED a southern % RUV LED 0 INS a southern % RUV LED 0 PAI a southern % BLI 0 FIL a southern % BLI 0 FIL INS a southern % BLI 0 FIL PAI b jeju % RUV LED b jeju % RUV LED 0 INS b jeju % RUV LED 0 PAI b jeju % BLI 0 FIL b jeju % BLI 0 FIL INS b jeju % BLI 0 FIL PAI b jeju % 0 LED+DIM WIN b jeju % 0 LED+DIM WIN INS b jeju % 0 LED+DIM WIN PAI b southern % RUV LED+DIM FIL b southern % RUV LED+DIM FIL INS b southern % RUV LED+DIM FIL PAI b southern % BLI LED WIN b southern % BLI LED WIN INS b southern % BLI LED WIN PAI b southern % b southern % INS b southern % PAI b central % RUV 0 WIN b central % RUV 0 WIN INS b central % RUV 0 WIN PAI b central % BLI LED+DIM b central % BLI LED+DIM 0 INS b central % BLI LED+DIM 0 PAI b central % 0 LED FIL b central % 0 LED FIL INS b central % 0 LED FIL PAI c southern % BLI LED+DIM c southern % BLI LED+DIM 0 INS c southern % BLI LED+DIM 0 PAI c southern % 0 LED FIL Journal of KIAEBS Vol. 15, No. 2, 2021

10 Table 7. Selection of sampling cases for energy technology effect analysis (Continue) Passive Location Usage Energy Saving Technology NO Construction Position Climate Azimuth Average Sche year of floors Zone People dule Shading Lighting Window Insulation c southern % 0 LED FIL INS c southern % 0 LED FIL PAI c southern % RUV 0 WIN c southern % RUV 0 WIN INS c southern % RUV 0 WIN PAI c central % BLI 0 FIL c central % BLI 0 FIL INS c central % BLI 0 FIL PAI c central % 0 LED+DIM WIN c central % 0 LED+DIM WIN INS c central % 0 LED+DIM WIN PAI c central % RUV LED c central % RUV LED 0 INS c central % RUV LED 0 PAI c jeju % BLI LED WIN c jeju % BLI LED WIN INS c jeju % BLI LED WIN PAI c jeju % c jeju % INS c jeju % PAI c jeju % RUV LED+DIM FIL c jeju % RUV LED+DIM FIL INS c jeju % RUV LED+DIM FIL PAI BLI : Inside Blind RUV : Outside Blind, LED : LED, DIM : Dimming, WIN : Window, FIL : Film INS :Insulation, PAI : Paint 본예측모델개발을위하여학습에활용한 Data의총개수는 24,300개이며, Train Data의개수는 13,842개, Test Data 10,458개로수행하였다. 해당모델의검증은현황예측모델과동일하게중간기준면적은학습시키지않았으며, Test Model도학습시키지않은면적에서절감된에너지소요량의결과로수행하였다. 에너지절감기술적용에따른에너지소요량예측결과, 선형회귀 R 2 =0.893, 랜덤포레스트 R 2 =0.518, 뉴런네트워크 R 2 =0.984로분석되었으며, 앞서현황분석과마찬가지로뉴런네트워크알고리즘이결정계수가높은것으로분석되었고, CVRMSE는 10.58% 로낮게도출되어예측에적합한것으로판단된다 (Table 8 참조 ). Table 8. Energy consumption prediction result according to energy saving technology Contents Linear Regression Random Forest Neuron Network R MSE 416,317 1,240, ,005 RMSE , CVRMSE 21.28% 36.7% 10.58% Units : kwh Graph Journal of KIAEBS Vol. 15, No. 2,

11 이에리모델링전 후평가수행시활용가능한두가지예측모델을도출하였으며, 리모델링전 후에대한예측모델의결정계수가 (R 2 ) 0.9 이상으로분석되었고, CVRMSE 15% 이하로도출됨에따라예측모델로서검증되었으며통계적으로도유의하다고판단된다. 추가로 ASHRAE Guideline 14 (ASHRAE, 2002) 에서제시하는 M&V 평가시활용하는지표도만족하고있는것으로분석되었으나, M&V의경우, 사용량과소요량으로비교를수행함에따라다소차이가있을수있으나, 국내저층주거의통계적참조모델을통해산출된소요량을 Energyplus 또는 Designbuilder와같은시뮬레이션을수행하지않고 15% 이내로예측하는것은의미있는결과로판단된다. 이에해당예측모델을활용하여저층주거건물의도서가미비하고, 입력사항에대하여부족한자료가있어도다양한인자및기준면적에대응한통계적참조모델을활용하여본인의건물과유사한면적의에너지소요량에대한파악이가능하며, 추가적으로에너지절감기술적용에따라리모델링후절감량산출도결과로도출됨에따라, 저층주거건물의평가및활용에있어활용도가높을것으로판단된다. 예측모델검증예측모델검증으로는뉴런네트워크모델을대상으로 K-fold 검증을수행하였다. K-fold 검증은전체자료를 K개의집합으로구성하여 K-1개의집합을학습자료 (Train Data) 로사용하고나머지 1개의집합을 Test Data로활용하는방법으로과정을반복하며수행한다. 이러한방법은과적합을방지하고, 모든데이터를학습과검증의자료로활용함에따라, 최적의매개변수를구하기위한모델튜닝에주로활용된다. Table 9. K-fold validation result Contents K-fold R 2 MSE RMSE CVRMSE K fold , % K fold , % Train K fold , % K fold , % K fold , % Before K fold , % Remodeling K fold , % Test K fold , % K fold , % K fold , % Average , % K fold , % K fold , % Train K fold , % K fold , % After Remodeling K fold , % K fold , % K fold , % Test K fold , % K fold , % K fold , % Average , % 162 Journal of KIAEBS Vol. 15, No. 2, 2021

12 본논문은약 40,000개의데이터를 5개의 K-fold로구분하여수행하였으며, Train과 Test 를나누어각각시뮬레이션결과와예측값을비교하였다. 분석결과, 위의 Table 9와같으며, 리모델링전 후모든 K-fold의결정계수는 0.9 이상으로도출되었으며, 리모델링전평균결정계수 (R 2 ) 는 0.990로예측율이매우높았고, CVRMSE는 8.15% 로분석되었다. 리모델링후평균결정계수 (R 2 ) 는 0.986으로도출되었으며, CVRMSE는 10.74% 로분석되었다. 이로써 K-fold 교차검증을통해예측모델의정확성을검증하였으며결정계수와 CVRMES로통계적타당성을증명한것으로판단된다. 이를통해, 자료가부족한저층주거건물그린리모델링및노후건물에너지소요량현황평가및개선시절감되는에너지소요량평가모델의기초자료로제시한다. 결론 본논문은저층주거건물에너지성능평가와민간건축물그린리모델링활성화를위하여선행연구에서개발된참조모델을기준으로리모델링시다양한인자가반영된기준면적확장에의한소요량예측 ( 현황파악 ) 및에너지절감기술적용에따른효과소요량예측 ( 개선후 ) 모델을개발및검증하였다. 이에, 단독주택 36, 66, 109 m2, 다가구 ( 세대 ) 주택 39, 82 m2, 연립주택 39, 84 m2을대상으로학습을수행하였으며, 단독주택의경우 m2를예측하도록하였고, 다가구주택의경우 54 m2, 연립주택의경우 59, 66 m2를대상으로인자변경에따라소요량값에대한평가및검증을수행하였다. 시뮬레이션과예측모델검증결과는결정계수 (R 2 ) 및 CVRMSE 의지표로분석하였다. 이에에너지절감기술을적용하지않은리모델링현황평가예측결과는선형회귀 R 2 =0.914, CVRMSE 19.25% 랜덤포레스트 R 2 =0.505, CVRMSE 35.32%, 뉴런네트워크 R 2 =0.987, CVRMSE 8.92% 으로도출되었으며, 예측률이가장높은모델은뉴런네트워크로선정되었다. 또한에너지절감기술을적용한예측을수행한결과, 선형회귀 R 2 =0.893, CVRMSE 21.28% 랜덤포레스트 R 2 =0.518, CVRMSE 36.7%, 뉴런네트워크 R 2 =0.984, CVRMSE 10.58% 로분석되었으며, 앞서현황분석과마찬가지로뉴런네트워크알고리즘이결정계수가높은것으로분석되었고, CVRMSE 는낮게도출되어예측에적합한것으로도출되었다. 결과를종합하면, 저층주거건물의에너지소요량평가시두개의예측모델로자료가부족한저층주거건물의리모델링전 후를평가할수있으며, 두모델다뉴런네트워크의알고리즘이 R 이상, CVRMSE 15% 이하로분석됨에따라소요량예측시유의한것으로분석되었다. 이에, 통계적유의성을검증한뉴런네트워크모델을토대로 K-fold 검증결과, 리모델링전평균결정계수 (R 2 ) 는 0.990, CVRMSE는 8.15% 로분석되었다. 리모델링후평균결정계수 (R 2 ) 는 0.986으로도출되었으며, CVRMSE는 10.74% 로분석됨에따라, 예측모델의정확성및타당성을검증한것으로판단된다. 이로써 40,000개의학습데이터를통해별도의시뮬레이션수행없이, 간단한입력인자입력인자만입력하여평가할수있는정확도가높은예측모델을개발하였다. 또한저층주거건 Journal of KIAEBS Vol. 15, No. 2,

13 물의리모델링시적용기술에따른에너지절감량평가에대하여 Designbuilder 및 Energyplus 를사용하지않고, 단열에대한법적기준 (2020년기준 ) 및고효율냉난방기기적용에대한평가가가능할것으로판단된다. 이를바탕으로국내민간건축물그린리모델링, 도시개선사업, 도시재생사업시저층주거건물을대상으로, 에너지절감량을산정하고, 개략적인평가자료로활용할수있을것으로기대할수있다. 이에, 후속연구로는해당알고리즘을통합하여일원화하는연구를수행하며, 해당알고리즘을통하여저층주거건물의리모델링전 후실제사용량과소요량검증을통해예측모델의유의성을입증하는연구가향후진행되어야할것으로사료된다. 후기 본연구는산업통상자원부 (MOTIE) 와한국에너지기술평가원 (KETEP) 의지원을받아수행한연구과제입니다 (No ). References 1. Kim, J.H., Seong, N.C., Choi, W.C., Choi, K.B. (2018). An Analysis of the Prediction Accuracy of HVAC Fan Energy Consumption According to Artificial Neural Network Variables. Journal of the Architectural Institute of Korea Structure & Construction, 34(11), Kim, K.S., Yoo, D.C., Choi, C.H., Jang, H.I. (2020). Development of Energy Variable Definition and Reference Model for Predicting Energy Consumption of Low-rise Residential Buildings. Journal of the Architectural Institute of Korea, 2020, 36, Nam, Y.G., Hong, S.K., Cho, S.H., Choi, C.Y. (2019). A Study on the Prediction of Building Energy Consumption Using Deep Learning Technique. Journal of the Korean Society of Mechanical Technology, 21(6), Bae, J.G. (2019). Development of Energy Consumption Prediction Model using Economizer Method in the Office Building with Machine Learning. Master s thesis. Department of Architectural Engineering, Yonsei University, Korea. 5. Hong, S.W. (2020). Prediction for Building and Housing Energy Saving Using Weather Information and Machine Learning. Doctoral thesis, Department of Intelligent Robot Engineering, Hanyang University, Korea. 6. Jung, J.H. (2018). Electricity Load Forecasting with Weather Data for Commercial Buildings based on Machine Learning. Master s thesis. Department of Architectural Engineering, Cheongju University, Korea. 7. Lee, S.H. (2020). An Analysis and Prediction of Energy consumption in a building using life patterns of Single-Person Households. Master s thesis. Department of Architecture, Sejong University, Korea. 8. Lee, T. G. (2015). Study on Electric Power Consumption in The University Building using Multiple Regression Analysis and Energy Simulation. Master s thesis. Department of Mechanical Engineering, University of Seoul, Korea. 164 Journal of KIAEBS Vol. 15, No. 2, 2021

14 9. No, B.I. (2015). Prediction of Heating Energy Consumption of Apartment Buildings Using Performance Evaluation of Detailed Heating Degree-days. Master s thesis. Department of Architectural Engineering, Chungbuk National University, Korea. 10. Park, G.H. (2018). Machine learning for building energy management Based on short/long-term prediction algorithm Performance analysis. Master s thesis. Department of Computer and Information Science, Korea University, Korea. 11. Ministry of Land, Infrastructure and Transport (MOLIT). (2018) Greenhouse Gas Reduction Roadmap Amendment and emission permit allocation plan confirmed : Ministry of Land, Infrastructure and Transport. 12. ASHARAE (2002). ASHRAE Guideline 14 Measurement of Energy, Demand and Water Savings. American Society of Heating, Refrigerating, and Air-Conditioning Engineers, New York, USA. Journal of KIAEBS Vol. 15, No. 2,

DBPIA-NURIMEDIA

DBPIA-NURIMEDIA The e-business Studies Volume 17, Number 6, December, 30, 2016:275~289 Received: 2016/12/02, Accepted: 2016/12/22 Revised: 2016/12/20, Published: 2016/12/30 [ABSTRACT] SNS is used in various fields. Although