Review on Energy Resilience

Energy & Buildings 186 (2019) 80–97

Contents lists available at ScienceDirect

Energy & Buildings

journal homepage: www.elsevier.com/locate/enbuild

Incorporating machine learning with building network analysis to

predict multi-building energy use

Xiaodong Xu a , ∗, Wei Wang a , b , Tianzhen Hong b , ∗, Jiayu Chen c

a School of Architecture, Southeast University, 2 Sipailou, Nanjing, Jiangsu Province, China b Building Technology and Urban Systems Division, Lawrence Berkeley National Laboratory, 1 Cyclotron Road, Berkeley, CA 94720, USA c Department of Architecture and Civil Engineering, City University of Hong Kong, Y6621, AC1, Tat Chee Ave, Kowloon, Hong Kong

a r t i c l e i n f o

Article history:

Received 30 June 2018

Revised 26 November 2018

Accepted 1 January 2019

Available online 22 January 2019

Keywords:

Multi-building

Energy use prediction

Social network analysis

Artificial neural networks

Machine learning

Building network

a b s t r a c t

Predicting multi-building energy use at campus or city district scale has recently gained more attention;

and more researchers have started to define reference buildings and study inter-impact between build-

ing groups. However, how to integrate the relationship to define reference buildings and predict multi-

building energy use, using significantly less amount of building data and reducing complexity of predic-

tion models, remains an open research question. To resolve this, this study proposed a novel method to

predict multi-building energy use by integrating a social network analysis (SNA) with an Artificial Neural

Network (ANN) technique. The SNA method was used to establish a building network (BN) by identifying

reference buildings and determine correlations between reference buildings and non-reference buildings.

The ANN technique was applied to learn correlations and historical building energy use, and then used

to predict multi-building energy use. To validate the SNA-ANN method, 17 buildings in the Southeast

University campus, located in Nanjing, China, were studied. These buildings have three years of actual

monthly electricity use data and were grouped into four types: office, educational, laboratory, and res-

idential. The results showed the integrated SNA-ANN method achieved average prediction accuracies of

90.67% for the office group, 90.79% for the educational group, 92.34% for the laboratory group, and 83.32%

for the residential group. The results demonstrated the proposed SNA-ANN method achieved an accuracy

of 90.28% for the predicted energy use for all building groups. Finally, this study provides insights into

advancing the interdisciplinary research on multi-building energy use prediction.

© 2019 Elsevier B.V. All rights reserved.

e

s

s

a

n

c

a

i

c

u

m

s

g

d

p

a

1. Introduction

Buildings are the main energy consumer, demanding more than

40% of primary energy usage [1] ; while in cities, buildings can

consume up to 75% of total primary energy usage [2] . In particular,

electricity use is a main driver. The latest Electric Power Monthly

data reported in January 2018 by United States Department of

Energy (DOE) indicated that electricity consumption from both

commercial and residential buildings represented 77.5% of all

the electricity produced in the U.S. [3] . The International Energy

Agency (IEA)’s Energy in Buildings and Communities (EBC) Pro-

gramme annexes discussed methods to analyze total energy use

in buildings to reduce energy use and associated emissions [4,5] .

The use of building energy modeling has significantly improved

building energy efficiency and reduced environmental impact [6,7] .

A considerable number of studies have been conducted to develop

∗ Corresponding authors. E-mail addresses: [email protected] (X. Xu), [email protected] (T. Hong).

r

f

p

https://doi.org/10.1016/j.enbuild.2019.01.002

0378-7788/© 2019 Elsevier B.V. All rights reserved.

fficient energy models for single buildings [6,8,9] . In recent years,

ome researchers have recognized the importance of energy use

tudies in large-scale areas with distributed building groups to

nalyze distributed building energy use patterns and optimize

et-zero building or distribution energy systems [10,11] , also, for

ity-scale buildings through benchmarking building energy use

nd reducing city building emissions [12,13] . Focus on analyz-

ng and modeling urban building energy use at the large scale

an potentially provide insights into large-scale building energy

se patterns and opportunities to save energy [14,15] . Also, in

odeling large-scale building energy use, more researchers have

tarted to study the impact and interrelationship between building

roups. The concept of the Inter-Building Effect (IBE) was intro-

uced to understand the complex mutual impacts within spatially

roximal buildings [16–18] . Han et al. explored mutual shading

nd reflection for IBE on building energy performance with two

ealistic urban contexts in Perugia, Italy [19] . Han and Taylor

urther simulated the IBE on energy consumption by embedding

hase change materials into the building envelope [20] .

X. Xu, W. Wang and T. Hong et al. / Energy & Buildings 186 (2019) 80–97 81

U

d

i

d

f

o

s

i

o

b

c

r

u

s

p

p

p

a

2

2

t

u

r

e

T

a

a

d

b

s

g

[

(

e

c

f

s

t

t

[

b

b

f

d

b

s

s

b

e

C

a

m

o

a

C

m

a

(

o

C

o

s

a

t

s

p

s

l

v

b

i

l

c

a

i

M

i

w

a

d

t

s

r

S

u

Nomenclature

BN building network

ANN artificial neural network

X input layer vector

x i value of input neuron i

H hidden layer vector

h j value of hidden neuron j

Y output layer vector

y k value of output neuron k

c _ i _ j correlation index of energy use between building i, j

EU energy use vector

e i energy use of building i

�EU change of energy use vector

�e i change of energy use of building i

NS number of story vector

ns i number of story value of building i

YB year-built vector

yb i year-built value of building i

CT construction type vector

ct i construction type value of building i

C correlation vector

V weight vector between the input layer and hidden

layer

v j weight value of hidden neuron j

W weight vector between the hidden layer and output

layer

w k weight value of output neuron k

f ( x ) activation function of ANN algorithm

d ground truth vector

d k ground truth value of output neuron k

E error function

�v jk deviation of weight from the input layer to hidden

layer

�w ij deviation of weight from the hidden layer to the

output layer

η learning rate of algorithm MAPE mean absolute percentage error

RMSE root mean squared error

Li et al. analyzed 51 high-performance office buildings in the

.S., Europe, and Asia using portfolio analysis and individual

etailed case studies based on actual energy use data of build-

ngs [21] . Pang et al. brought together real-time data sharing, a

atabase for assessing past and present weather data, a network

or communicating energy-saving strategies between building

wners, and a set of modeling tools for real-time building energy

imulation, all in an effort to promote large-scale energy efficiency

n neighboring buildings [22] . Fonseca and Schlueter proposed

ne integrated model for the characterization of spatiotemporal

uilding energy consumption patterns in neighborhoods and

ity districts. The model calculated the power and temperature

equirements for residential, commercial, and industrial sectors

sing spatial (building location using geographic information

ystem, GIS) and temporal (hourly) dimensions of analysis [23] . To

redict energy use of a large group of buildings, Panao and Brito

resented a bottom-up building stock energy model [20] . They

redicted hourly electricity consumption of residential buildings

nd validated the model by using smart meter data of roughly

50 dwellings [24] . Kalogirou et al. utilized the electricity data of

25 buildings and applied back propagation of neural networks

o predict the required heating load of buildings [25] . Constantine

sed data-driven prediction models, including linear regression,

andom forest, and support vector regression, to predict city-scale

lectricity and natural gas usage in New York City buildings [22] .

he project encompassed 23,0 0 0 buildings, with model validation

t the building and ZIP code levels [12] .

Similarly, Hsu studied multi-family buildings in New York City

nd used clusterwise regression and cluster validation methods to

etermine building energy use [13] . Jain et al. applied a sensor-

ased forecasting approach coupled with support vector regres-

ion modeling and examined the impact of temporal and spatial

ranularity on to energy consumption of multi-family buildings

26] . Hawkins et al. applied statistical and artificial neural network

ANN) method to predict energy use determinants in UK higher

ducation buildings [27] , resulting in 34% of mean absolute per-

entage error for electricity use prediction and 25% for heating

uel use prediction. Kavgic applied Monte Carlo method to predict

pace heating energy use of Belgrade’s housing stock [28] and fur-

her analyzed uncertainty for a city-scale domestic energy model

o address the impact of sensitivity on the modeled energy use

29] .

Those machine-learning based models, usually called “black

ox”, provide users high accuracy by measuring the data of the

uilding systems input and output and fitting a mathematical

unction to the data, even although such models ignore the un-

erstanding of the system physics with poor generalization capa-

ilities. On the other hand, “white box” models implementing the

ystem physics can use the building parameters for modeling the

ystem dynamics [30] . For example, innovative software or web-

ased applications have been developed to analyze and predict the

nergy use of multiple buildings in distributed or urban areas. The

ity Building Energy Saver (CityBES), an Energyplus-based web

pplication, provides a visualization platform, focusing on energy

odeling and analysis of a city’s building stock to support district

r city-scale building energy efficiency programs [31–33] , as well

s to predict energy use for informing building retrofits. Based on

ityBES, Chen et al. analyzed the impacts of building geometry

odeling on urban building energy models to understand how

group of buildings perform together [33] . City Energy Analyst

CEA) provides a computational framework for the analysis and

ptimization of energy systems in neighborhoods and city districts.

EA has a unique interface to facilitate the spatiotemporal analysis

f energy patterns for energy savings [34] . Usually, with these

oftware or web-based applications, every building is explicitly

nd detailed modeled in EnergyPlus. While it can be accurate, it is

ime consuming and requires absorbent amounts of data.

To reduce the complexity of urban building energy models,

ome studies advocate reduce-order building models or building

rototype models. Felsmann used reduced order building energy

ystem modeling, e.g., district heating or cooling systems, to create

arge-scale urban energy simulations [35] . Heidarinejad et al. de-

eloped a framework to rapidly create urban scale reduced-order

uilding energy models relying on the contributions of different

nfluential variables to the internal, external, and system thermal

oads [36] . Then the framework was validated by applying typi-

al building geometries for simulations [36] . Zhao et al. developed

reduced order building energy model to estimate single build-

ng energy performance; then applied regression and Markov chain

onte Carlo techniques to integrate physics-based energy model-

ng to replicate the single building model [37] . The resulting model

as an efficient energy model development at the city scale [37] .

One method to reduce data demands includes the development

nd replication of prototype building models. The U.S. DOE has

eveloped a suite of prototype building models covering 80% of

he commercial building stock in the U.S. to support the analy-

is of urban energy use. This database includes 16 commercial

eference building types across different climate zones [38,39] .

imilarly, Mastrucci et al. analyzed six types of dwellings by

sing a GIS-based statistical downscaling approach and adopted a

82 X. Xu, W. Wang and T. Hong et al. / Energy & Buildings 186 (2019) 80–97

2

w

u

t

(

n

2

o

m

g

t

t

t

y

W

u

d

X

W

c

c

d

t

p

�

a

t

2

b

s

f

t

u

i

b

o

a

b

F

e

a

h

p

(

l

g

s

multiple linear regression model for estimating energy savings at

the city scale [40] . Caputo et al. used four archetypes to charac-

terize the energy performance of the built environment in a city

or neighborhood, and to evaluate the effects of different energy

strategies [41] . Such prototype buildings or archetypes extend

the knowledge beyond individual buildings for efficient energy

models of neighborhoods or cities. Furthermore, city-scale building

energy benchmarking policy provides a holistic dataset foundation

and enables comparison of energy performance between similar

buildings [13,42,43] . Holistic building energy consumption data

can be used for defining reference buildings by investigating the

closeness of building groups, for which, cluster analysis is one

of the most efficient methods. Deb and Lee [44] studied the

determining key variables influencing energy consumption in 56

office buildings through cluster analysis. The clustering approach

focused on a small number of representative, reference buildings

from a large building dataset [45,46] . Gaitani et al. [47] applied

several variables, including the heated floor area, building age,

insulation of the building envelope, number of classrooms and stu-

dents, operation hours, and age of heating system, using principal

component and cluster analysis methods to establish the reference

buildings. Tardioli et al. developed a novel framework utilizing a

combination of building classification, clustering, and predictive

modeling to identify a total of 67 representative buildings out of a

dataset of 13,614 mixed-use buildings in the city of Geneva [48] .

However, two challenges arise: (1) how to capture the impact

and interrelationship between multi-buildings to define reference

buildings from an existing building stock, and (2) how to use ref-

erence building energy datasets with machine learning techniques

to learn and predict multi-building energy use. To address these

prediction gaps, this study presents a novel data-driven method,

integrating social network analysis based building network and

artificial neural network (SNA-ANN) techniques, to predict multi-

building energy use. Energy use patterns between buildings are

leveraged to identify reference buildings and create building

networks with the theory of social network analysis. The building

networks are created based on the correlation coefficients of en-

ergy use between any two buildings, which consist of correlation

coefficients between the energy use of the reference buildings and

the total energy use of all buildings, and correlation coefficients

between the energy use of the reference buildings and that of the

non-reference buildings. Built on the network, the SNA-ANN model

aims to apply energy use and building features from a small refer-

ence group (e.g., n buildings) to accurately and efficiently predict

energy use of a larger group (e.g., n + m buildings). To validate the technique, the proposed approach was evaluated using campus

buildings at Southeast University, China. Three-years of monthly

energy use data from 2015 to 2017, was used. Seventeen buildings

were selected, covering four use types namely: office build-

ings, educational buildings, laboratory buildings, and residential

buildings.

The main contribution of this work is in the unique inter-

disciplinary method of combining social network analysis to

create building network and reference buildings, and artificial

neural network to learn multi-building energy use patterns. This

technique efficiently learns the building feature and network for

multi-building energy use and provides a framework for analyzing

building energy use patterns in large-scaled areas. Moreover, the

proposed algorithm is validated using building groups with actual

data to demonstrate significant accuracy in the results. While

energy prediction is critical, the data-driven energy modeling

also opens many other applications, such as performance mon-

itoring, control and optimization of building groups, distributed

energy systems and micro-grids implementation, which need

co-operations between buildings.

f

. Methods

To establish the BN-ANN relationship, three main components

ere first conducted: (1) the feature selection for building energy

se prediction, (2) the extraction of the reference buildings and

he networks between buildings with social network analysis, and

3) the integration of the building network-based artificial neural

etwork algorithm.

.1. Feature selection

Before implementing the SNA-ANN method, pre-processing

f raw data is necessary to eliminate erroneous or missing

easurements in the energy use data. In this study, the La-

range polynomials for the interpolation filter is applied due

o its computational efficiency and causality which are impor-

ant in time-series applications. If we have time-series data as

( x 1 , y 1 ) , ( x 2 , y 2 ) , . . . , ( x n , y n ) , we can formulate the interpola-

ion for the default measurement as shown in Eq. (1 ).

= n ∑

i =0 y i

n ∏ j =0 , j � = i

x − x j x i − x j

(1)

here, y is the interpolation value and n is the size of the data

sed for interpolation.

After the dataset is filtered, the time series data of energy use

ata for one building can be described in Eq. (2 ).

= ( x 1 , x 2 , . . . , x t , . . . , x n ) T (2) here, 1, 2, …, n is the discrete time step.

For feature-based prediction, the feature is a variable which

ontains the information relevant for object recognition. In fore-

asting energy use, it should include the use, trend, and the

etermined factors of building energy. Therefore, we considered

wo kinds of features, the value and the change of energy use

arameter. The change of a parameter was equated using Eq. (3 ).

x = x t − x t−1 (3) To better predict building energy use, some determined factors

re also considered in this study, including the year-built, construc-

ion type, number of stories, building area, and roof type.

.2. Network extraction

Predicting the total energy use or the demand in a distributed

uilding group is difficult and complicated, especially at the city

cale. Moreover, the task of collecting historical energy use datasets

or large-scale building groups is a big issue. One way to overcome

hese complexities is to simulate or estimate the total energy use

sing typical reference building geometry. However, such methods

gnore actual energy use patterns which are influenced not only by

uilding geometries, but also occupancy, operation mode, and so

n. Therefore, it is necessary to define the reference buildings with

ctual building energy use pattern and consider the correlation

etween the reference buildings and the non-reference buildings.

urther consideration should include building physical information,

.g., building floor area, year-built, construction type, and so on.

In this study, the theory of social network analysis method was

pplied to extract and enhance building networks through their

istorical energy use patterns. Social network analysis (SNA) is the

rocess of investigating connections between networked nodes

individual actors, people, or things) and ties, edges, or links (re-

ationships or interactions) to connect nodes using networks and

raph theory [49] . The SNA is the process of investigating social

tructures of a group to build relation of participants in the group

or network analysis and has contributed to various academic

X. Xu, W. Wang and T. Hong et al. / Energy & Buildings 186 (2019) 80–97 83

d

n

p

a

t

s

e

u

w

b

t

t

r

b

b

t

S

d

t

c

w

c

t

e

p

b

c

W

E

2

m

s

N

r

c

o

p

a

t

a

t

d

i

w

d

v

r

b

b

s

i

X

W

E

�

N

Y

C

C

d

o

f

o

H

Y

V

W

W

p

n

a

l

l

d

T

e

i

h

y

t

C

s

d

E

s

m

isciplines as well as practical applications such as social media

etworks [50] , information system [51] , and prefabricated building

roject [52] , and so on. This method allows to interact the network

nd illustrate the participants’ interaction with other members of

he group. The network analysis has also been applied in energy

aving projects, e.g., using social networks for promoting domestic

nergy technologies adoption [53] , using network analysis to

nderstand wave energy policy [54] , and advantage of social net-

orks to diffuse energy-efficiency innovations for households [55] .

In the SNA method, two main approaches are mainly used to

uild the networks: (1) the distance method (e.g., Euclidean dis-

ance), which is usually used to calculate the difference between

wo participants, and (2) the correlation method (e.g., Pearson cor-

elation coefficient), which is usually used to find the similarity

etween the two participants. Therefore, to reduce the number of

uildings used, this study establishes connection and relationship,

hat’s the network, between buildings in a building group with the

NA method using the building energy use dataset. Considering the

istance method infers the closeness of energy use rather than the

endency of building energy use pattern, therefore, to build the

onnections of individual buildings using BN analysis in this study,

e use the Pearson correlation coefficient method to calculate the

onnections between buildings shown in Eq. (4 ). Two steps are

aken to extract the networks. The first step is to identify the refer-

nce buildings by using Eq. (4 ). The reference buildings are used to

redict building energy use. The second step is to build networks

etween the reference building and the non-reference buildings.

_ i _ j = E ( E U i E U j

) − E ( E U i ) E

( E U j

) √

E ( E U i

2 )

− ( E ( E U i ) ) 2 − √

E ( E U j

2 )

− ( E ( E U j

))2 (4) here, c _ i _ j is the correlation coefficient between building i and j.

U i and EU j are the energy use dataset of building i and j .

.3. Building network based artificial neural network model (BN-ANN

odel)

To predict the energy use of multi-buildings using a building

ubset, this study proposed the Building Network based Artificial

eural Network (BN-ANN) model shown in Fig. 1 . The ANN algo-

ithm is used to solve problems similar to the human brain. ANN

onsists of a network of simple neuron elements connecting the

utput to the input with the directed and weighted graph. The ca-

abilities of the ANN algorithm fall within the realm of regression

nalysis including time series prediction and modeling, classifica-

ion, including pattern recognition and sequential decision making,

nd so on. Meanwhile, since this study considers multi-features in

he prediction model, the ANN model is also well fit to tackle the

ifferent scale of feature datasets.

The BN- ANN algorithm has four layers: the feature layer, the

nput layer, hidden layer, and the output layer. In the feature layer,

e selected the building information property dataset and the BN

ataset to represent the compiled dataset composing the input

ector for the input layer of BN-ANN model.

Suppose that EU represents the energy use vector, �EU rep- esents the energy use difference vector, NS represents the num-

er of building story vector, YB represents the year-built vector for

uildings, CT represents the construction type vector, and C repre-

ents the correlation network vector between buildings, then, the

nput vector for the input layer can be illustrated in Eq. (5 ).

input = ( x 1 , x 2 , . . . , x i , . . . , x N ) T = ( E U, �E U, NS, Y B, CT , C ) T (5)

here, N is the size of input layers.

Suppose n is the number of buildings, then,

U = ( e 1 , . . . e 2 , . . . , e i , . . . , e n ) (6)

EU = ( �e 1 , �e 2 , . . . , �e i , . . . , �e n ) (7)

S = ( n s 1 , n s 2 , . . . , n s i , . . . , n s n ) (8)

B = ( y b 1 , y b 2 , . . . , y b i , . . . , y b n ) (9)

T = ( c t 1 , c t 2 , . . . , c t i , . . . , c t n ) (10)

= ( c _ 1 _ 2 , c _ 1 _ 3 , . . . , c _ 1 _ n . . . , c _ i _ j, . . . , c _ ( n − 1 ) _ n ) (11) To eliminate the impact caused by different scales of the feature

ataset, we need to normalize the different f eature vectors. The

utput of the hidden layer, the output of the output layer, weights

rom the hidden layer, and weights from the hidden layer to the

utput layer are defined in Eqs. (12 )- (15) , respectively.

= ( h 1 , h 2 , . . . , h j , . . . , h m

)T (12)

= ( y 1 , y 2 , . . . , y k , . . . , y l ) T (13)

= ( v 1 , v 2 , . . . , v i, j , . . . , v m,n ) (14)

= ( w 1 , w 2 , . . . , w j,k , . . . , w m,l ) (15) here, m and l are the length of the hidden layer and the out-

ut layer, respectively. The v i, j donates the weight vector from x th

eural cell of the input layer to the j th neural cell of hidden layer

nd w j, k donates weight vector from j th neural cell of the hidden

ayer to the k th neural cell of output the layer. The length of input

ayer the is determined by the number of elements of the input

ata while the length of the hidden layer ( m ) is randomly selected.

he mathematic information transfer between each layer can be

xpressed in Eqs. (16 ) and (17 ). The activation function is shown

n Eq. (18 ).

j = f (

n ∑ i =0

v i j x i

) , j = 1 , 2 , . . . , m (16)

k = f (

m ∑ j=0

w jk h j

) , k = 1 , 2 , 3 (17)

f ( x ) = 1 1 + e −x (18)

In this study, the measured building energy use dataset can be

he ground truth of energy output, which is presented in Eq. (19 ).

ompared with prediction results from the output neuron, Eq. (20 )

hows the squared error function.

= ( d 1 , d 2 , . . . , d k , . . . , d l ) T (19)

= 1 2

l ∑ k =1

( d k − y k ) 2 (20)

The gradient descent approach computes the derivative of the

quared error function and iterates through different weights to

inimize the error shown in Eqs. (20 )–( 22 ) show the adjustment

84 X. Xu, W. Wang and T. Hong et al. / Energy & Buildings 186 (2019) 80–97

Fig. 1. The construction of the BN-ANN model.

a

f

w

u

r

(

t

3

d

o

f

a

i

a

c

a

a

n

t

d

s

d

t

c

t

w

process of weight v jk and w ij and η is learning rate the of gradient descent.

�v jk = η (

l ∑ k =1

( d k − y k ) y k ( 1 − y k ) w jk

) h i ( 1 − h i ) x i (21)

�w i j = η( d k − y k ) y k ( 1 − y k ) h i (22)

3. Case study

3.1. Description of the case study

A distributed building group from Southeast University (SEU)

was selected to validate the proposed multi-building energy con-

sumption prediction algorithm. The SEU is located in the center

of Nanjing City, Jiangsu Province, China. The SEU has a total area

of 3.9 km 2 and consists of 53 buildings on the main campus,

including office buildings, laboratories, educational buildings,

multiple-use buildings, residential buildings, and other building

types. A group of other buildings includes some auxiliary service

buildings, such as a kindergarten, an elementary school, retail

stores, canteen, and so on. The multi-use buildings usually consist

of classrooms, research rooms, office rooms, lab areas, etc. With

only two multi-use buildings on the campus and the pattern of

their energy use hard to identify, the multi-use buildings and

a few other buildings were not considered. The four types of

buildings analyzed include office, educational, laboratory, and res-

idential. The dataset of the four building groups, provided by the

General Affairs Department of SEU, includes energy use, year built,

construction type, wall material, total building floor area, use type,

nd the number of stories. Energy use data were collected monthly

rom 2015 to 2017. The buildings without complete three-year data

ere excluded, resulting in: six office buildings (O1-O6), four ed-

cational buildings (E1-E4), four laboratories (L1-L4), and three

esidential buildings (R1-R3) being used for validation purposes

Fig. 2 ). Details of each office, educational, laboratory and residen-

ial building groups are presented in Tables 1–4 , respectively.

.2. Model configuration and assessment

To eliminate the impact of different scales embedded in the

ifferent building features, it was necessary to configure the inputs

f the BN-ANN model before model training. For the year-built

eature, the age varied from 1922 to 1994 with a clear separation

round 1960 to 1970. The model sets a binary value of 0 for build-

ngs built before the year 1965, otherwise 1. Construction type

lso required two values. The model sets the value of “reinforced

oncrete structure” as 0 and the value of “brick-concrete structure”

s 1. Similarly, the value of “flat roof” and “sloping roof” was set

s 0 and 1, respectively. The value of the number of stories was

ormalized as in [0, 1], while the building floor area was used

o calculate the building energy use intensity. Table 5 shows the

etails of different scales of feature data for model input. Fig. 3

hows the learning process of the BN-ANN model. With the EUI

ataset, Eq. (4) was used to identify reference buildings.

To dynamically update the network between buildings, one

ime window with length �T was applied in the model to cal- ulate the correlations. During model training, the initial parame-

ers for the neural networks were defined randomly, including the

eights and bias of each neural between each layer. The gradient

X. Xu, W. Wang and T. Hong et al. / Energy & Buildings 186 (2019) 80–97 85

Fig. 2. Diagram of the footprint and location of the building groups.

Table 1

Office building group.

Building Type Year Built Construction Type No. of story Building Area Roof Type

Office Building 1 (O1) 1980 Reinforced concrete structure 16 16,910 Flat roof

Office Building 2 (O2) 1927 Reinforced concrete structure 3 5072 Sloping roof

Office Building 3 (O3) 1957 Brick-concrete structure 4 3938 Sloping roof

Office Building 4 (O4) 1922 Brick-concrete structure 2 4500 Sloping roof

Office Building 5 (O5 1990 Brick-concrete structure 8 11,748 Flat roof

Office Building 6 (O6) 1991 Reinforced concrete structure 4 7106 Flat roof

Table 2

Educational building group.

Building Type Year Built Construction Type No. of story Building Area Roof Type

Education building 1 (E1) 1980 Brick-concrete structure 3 3630 Sloping roof

Education building 2 (E2) 1987 Brick-concrete structure 6 5595 Flat roof

Education building 3 (E3) 1982 Brick-concrete structure 6 7482 Flat roof

Education building 4 (E4) 1982 Brick-concrete structure 3 2859 Flat roof

Table 3

Laboratory building group.

Building Type Year Built Construction Type No. of story Building Area Roof Type

Laboratory 1 (L1) 1994 Brick-concrete structure 4 2993 Flat roof

Laboratory 2 (L2) 1955 Reinforced concrete structure 6 10,902 Flat roof

Laboratory 3 (L3) 1957 Reinforced concrete structure 1 949 Sloping roof

Laboratory 4 (L4) 1957 Reinforced concrete structure 1 1421 Sloping roof

86 X. Xu, W. Wang and T. Hong et al. / Energy & Buildings 186 (2019) 80–97

Table 4

Residential building group.

Building Type Year Built Construction Type No. of story Building Area Roof Type

Residence building 1 (R1) 1980 Reinforced concrete structure 4 1313 Flat roof

Residence building 2 (R2) 1990 Reinforced concrete structure 16 12,906 Flat roof

Residence building 3 (R3) 1980 Reinforced concrete structure 15 9980 Flat roof

Table 5

Details of the BN-ANN model input.

Feature Source/Scope Description

Year Built Discrete, Binary

Construction Type Reinforced concrete structure, Brick-concrete structure Binary

No. of story [1,16] Discrete, Normalization

Roof Type Flat roof, Sloping roof Binary

Building Area [1313, 16,910] To calculate Energy Use Intensity

Energy Use Intensity N.A. Normalization

Correlation Eq. (4) Normalization

Fig. 3. Overview of the learning process of the BN-ANN model.

p

a

m

W

E

W

4

4

e

F

e

y

b

descent rule was applied to learn and update the weights and bias

shown in Eqs. (23) and (24) until the errors between predicted val-

ues and actual values were minimized. In validation, this study se-

lected the cross-fold validation method, which splits the dataset

into 70% for training and 30% for test. In comparison, this study

compared the predicted energy use proposed by the BN-ANN and

ANN models with the measured energy use data. The former inte-

grates the building network into the ANN algorithm as the pre-

diction model, which considers physical information, energy use

intensity of the reference buildings, and, the physical information

of non-reference buildings and correlations between the reference

and the non-reference buildings and the total EUI, respectively.

While for the ANN algorithm, only physical information and energy

use intensity of the reference buildings are applied in the predic-

tion model without the building networks. Fig. 3 shows the learn-

ing process of the proposed BN-ANN model. To assess the model

erformance, three indices were used to compare the results for

ccuracy, the mean absolute percentage error ( Eq. (25) ), the root

ean squared error ( Eq. (26) ) and the Q-Q plot curve.

a. Mean Absolute Error shows the mean error between the

predicted EUI and the actual EUI of the building.

MAE ( EU I p ) = 1 N

N ∑ i =1

∣∣E U I a i

− E U I p i

∣∣ (23) b. Mean Absolute Percentage Error (MAPE) shows the mean

percentage error between the predicted EUI and the actual

EUI of the building.

MAP E ( EU I p ) = 1 N

N ∑ i =1

∣∣(E U I a i

− E U I p i

) /EU I a

i

∣∣ (24) c. Root Mean Squared Error (RMSE) shows the magnitude of

the estimation error.

RMSE ( EU I p ) := √

N ∑ i =1

( E U I a

i − E U I p

i

)2 /N (25)

here, the EUI a and EUI p are the actual and predicted building

UI, respectively. N is the sample size.

d. Standard deviation of the absolute percentage error shows

the error variation between the results of the predicted EUI

and the actual EUI of buildings.

St d APE = √

1

n − 1 n ∑

i =1 ( AP E i − MAP E ) 2 (26)

here AP E = ( E U I a i

− E U I p i ) /EU I a

i

e. Q-Q plot curve is a graphical plot used to compare the

true positive rate and the false positive rate as the criterion

changes.

. Results and assessment

.1. Building energy use intensity results

The energy use intensity measurement results of the office,

ducational, laboratory, and residential buildings are shown in

igs. 4 (a–d), Figs. 5 (a–d), Figs. 6 (a–d), Figs. 7 (a–d), respectively. For

ach building group, the energy use intensity distribution in the

ear 2015, 2016, 2017, and the total building energy use intensity

ox plot are presented.

X. Xu, W. Wang and T. Hong et al. / Energy & Buildings 186 (2019) 80–97 87

Fig. 4. (a–d). EUIs of the office building group (a, b, c) and total building energy use (d).

Table 6

EUIs of the office building group (kWh/m 2 ).

Min. Mean Std. Max.

O1 2.01 6.14 4.35 21.64

O2 1.05 2.13 0.83 3.97

O3 6.89 10.76 2.25 15.51

O4 0.46 2.00 1.21 4.66

O5 0.54 6.81 5.35 20.89

O6 4.51 8.01 2.88 14.50

O_total 2.86 6.15 2.29 12.46

c

O

o

b

b

t

f

d

l

r

2

a

s

a

1

fi

a

b

T

c

t

i

E

t

Table 7

EUIs of the educational building group (kWh/m 2 ).

Min. Mean Std. Max.

E1 1.06 1.83 0.49 3.11

E2 3.63 7.84 2.54 14.34

E3 0.97 2.51 0.90 4.86

E4 0.79 1.58 0.45 3.22

E_total 1.89 3.77 1.13 6.85

Table 8

EUIs of the laboratory building group (kWh/m 2 ).

Min. Mean Std. Max.

L1 0.72 3.58 3.11 10.87

L2 6.36 8.48 1.44 11.40

L3 0.16 4.75 1.53 7.27

L4 0.04 0.89 1.05 6.01

L_total 4.67 6.70 1.29 10.01

t

1

i

l

n

1

r

L

a

f

f

3

i

s

e

o

t

The results show in Fig. 4 (a–c) that the three biggest energy

onsumers are O1, O3, and O6 in 2015 and 2016, and O3, O5, and

6 during 2017. Fig. 4 (d) shows the combined total of the EUI

f the office building group. The EUI trends show that the office

uildings generally consumed more energy in the winter (Decem-

er and January), and summer (July and August) months, due to

ypical seasonal patterns. The results show the EUI of O1 varies

rom 2 to 22 kWh/m2, with an average of 6 kWh/m2 and a stan-

ard deviation of 4 kWh/m2. The O2 and O4 buildings consumed

ess energy and had minimum EUIs of about 1 and 0.5 kWh/m2,

espectively. The average EUIs for O2 and O4 buildings is about

kWh/m2 (for both), with maximums of 4 and 4.7 kWh/m2 and

standard deviation of 0.8 and 1.2 kWh/m2, respectively. Ob-

erved from Table 6 , O2 and O4 have smaller floor areas and are

lso older, than O1 and O3. O3 has the biggest average EUI of

1 kWh/m2 and a minimum EUI of 7 kWh/m2. For the entire of-

ce building group, the EUI varies from 3 to 12.5 kWh/m2 with an

verage of 6 kWh/m2.

Fig. 5 (a–d) and Table 7 show the EUIs and the combined total

uilding energy use intensity box plot of the educational buildings.

he educational buildings are relatively new and smaller ( Table 2 )

ompared with the office buildings and the EUI results are reflec-

ive of this fact. Additionally, the educational buildings empty dur-

ng summer (July and August) and winter (February) breaks. For

1, E2 and E4, the EUI varies (minimum to maximum) from 1

o 3 kWh/m2, from 1 to 5 kWh/m2, from 1 to 3 kWh/m2, respec-

ively. While E2, the biggest energy consumer, varied from 4 to

4 kWh/m2.

Fig. 6 (a–d) and Table 8 show the EUIs of the laboratory build-

ngs with no unclear trend. The average EUI for L4 is substantially

ess than the EUIs of the other laboratory buildings. L4 varied from

early zero (0.04 kWh/m 2 ) to 6 kWh/m 2 and with an average of

kWh/m 2 . Although L3 has a similar EUI maximum to minimum

ange as L4 (varying from 0.2 to 7.3 kWh/m 2 ) the average EUI of

3 was 4.8, far higher than that of L4. L2 consumes the largest

mount of energy and has the largest building area. Its EUI varies

rom 6.4 to 11.4 kWh/m 2 , with an average of 8.5 kWh/m 2 . While

or L1, its EUI ranged from 0.7 to 11 kWh/m 2 , with an average of

.6 kWh/m 2 .

Fig 7 (a–d) and Table 9 shows the EUIs of the residential build-

ngs. The EUI trend for the student residential buildings is very

imilar to the educational buildings, as they are utilizing the same

ducational schedule. During the summer and winter breaks, the

ccupancy and operation of the residential buildings decreases,

hus the building energy use decreases accordingly. This study se-

88 X. Xu, W. Wang and T. Hong et al. / Energy & Buildings 186 (2019) 80–97

Fig. 5. (a–d). The EUIs of the educational building group (a, b, c) and total building energy use (d).

Fig. 6. (a–d). The EUIs of the laboratory building group (a, b, c) and total building energy use (d).

4

i

l

d

b

v

lected three adjacent residential buildings which were built in the

same year, with the same construction wall and roof type. The

R2 and R3 are high-rise buildings with 16 and 15 stories, respec-

tively. R2 has a floor area of 12,906 m2, larger than R3. The EUIs

of R2 vary from 0.2 to 17 kWh/m2 with an average of 4 kWh/m2;

while EUIs of R3 vary from 0.2 to 6 kWh/m2 with an average of

2 kWh/m2. While for R1, its EUI is from 0.7 to 10 kWh/m2 with an

average of 3.5 kWh/m2.

2

.2. Network analysis and prediction accuracy

This section discusses the applications of the SNA based build-

ng network modeling and the building network-based machine

earning technique, as well as presents the multi-building pre-

iction accuracy. Fig. 8 (a–d) shows the networks between the

uildings in each group, by calculating the correlations of indi-

idual building EUI with the total building group EUI, in the year

015 to 2017.

X. Xu, W. Wang and T. Hong et al. / Energy & Buildings 186 (2019) 80–97 89

Fig. 7. (a–d). The EUIs of the residential building group (a, b, c) and total building energy use (d).

Fig. 8. (a–d). The building network analysis results of (a) office, (b) educational, (c) laboratory and (d) residential buildings.

90 X. Xu, W. Wang and T. Hong et al. / Energy & Buildings 186 (2019) 80–97

Table 9

EUIs of the residential building group

(kWh/m2).

Min. Mean Std. Max.

R1 0.77 3.54 2.25 9.81

R2 0.16 3.82 4.03 17.24

R3 0.21 1.89 1.59 5.85

R_total 0.27 3.35 3.12 13.08

r

a

t

p

b

t

n

s

w

g

c

3

d

A

w

s

b

a

t

3

t

A

o

E

c

h

p

f

h

9

1

B

c

r

d

p

c

1

1

b

m

c

A

f

w

t

b

b

b

t

n

E

t

F

A

m

w

e

e

t

r

f

t

t

In the office building group, buildings O1, O3, and O6 were

identified as the reference buildings ( Fig. 8 a). The correlations be-

tween O1, O3, O6 and the total building’s EUI are 0.7, 0.6, and 0.7,

respectively. Building O1 shows the most relevant trend to the to-

tal building EUI. Observed from the networks between the non-

reference buildings and the reference buildings, it is found that

most non-reference buildings do not have much relevancy to the

reference buildings and their EUI trend correlations are generally

less than 0.6 (except for O2 and O3 with a correlation of 0.7). This

is especially true for O1 and O5 as they shared a negative network

correlation.

For the educational building group building E4 is the only non-

reference building ( Fig. 8 b). The building with the most relevant

trend to the total building’s EUI trend is building E2 with a high

correlation of 0.98. The building E3 is also highly correlated to the

total building’s EUI trend with a correlation of 0.91.

While considering the networks in the laboratory buildings,

buildings L1 and L2 are the reference buildings with correlations

of 0.72 and 0.92, respectively ( Fig. 8 c). For the non-reference build-

ings, building L4 is negatively relevant to both buildings L1 and L2,

showing opposite EUI trends between L4 and L1, L2, respectively.

Meanwhile, building L3 shows a much lower relation of EUI trend

to the reference buildings L1 and L2.

In the residential buildings group, buildings R2 and R3 are iden-

tified as the reference buildings with high correlations of 0.82 and

0.96 ( Fig. 8 d). Building R1 shows a positive relationship between

the two reference buildings, but with low correlations.

After calculating the networks between buildings, the building

dataset consisted of the networks, building physical information,

and the building EUI. To validate the algorithms, 70% of dataset

was used to train the models while 30% of the dataset was used

to test the models. Fig. 9 –12 show the training and test results of

the BN-ANN and the ANN for four types of the building dataset. To

compare the proposed BN-ANN model, two baseline results were

applied ( Fig. 13 ). The first one compares the actual building EUI to

the predicated EUI using the BN-ANN model for a six-month pe-

riod from July to December. The second compares the actual EUI

with the predicted building EUI generated by applying the refer-

ence buildings in the ANN model while ignoring networks between

the buildings. In most cases the actual and predicted values are

reasonable. To further understand the prediction performance Q-Q

plots were generated.

The Q-Q plot represents the quantiles of the actual data set

compared against the quantiles of the predicted data set. Fig. 14

presents an assessment of the results for the office and educational

buildings. The results from the office buildings Q-Q plot validate

the predicted BN-ANN model results with R 2 of 1. Moreover, the

BN-ANN model performed better than using just the ANN model

(R 2 of 0.6309). Additionally, the BN-ANN predicted results gravi-

tate closer to the line Y = X, which indicates the BN-ANN model can predict more accurately the actual EUI. However, for the ed-

ucational buildings, although the two models achieved good pre-

diction performance (R 2 of 0.9578 for the BN-ANN model and R 2

of 0.988 for the ANN model), the ANN model predicted results far

better when assessing the line Y = X.

Fig. 15 presents the Q-Q plot results for the laboratory and

esidential buildings. Observed from the results, the ANN model

chieved better predicted results in both building groups than

he BN-ANN model as indicated by the R 2 values. However, com-

ared with the ANN model, the BN-ANN model has more predicted

uilding energy results falling with the line Y = X. This means he BN-ANN model achieved more accurate prediction results. A

umerical comparison of the ANN and BN-ANN results are pre-

ented in Table 10 with MAPE and RMSE calculations. Compared

ith the ground truth, the BN-ANN model for the office building

roup showed greater than a 23% improvement in prediction ac-

uracy over the ANN model for both the MAPE (9.33% BN-ANN,

2.6% ANN) and the RMSE (9.12% BN-ANN, 36.1% ANN). The stan-

ard derivation of absolute percentage error ( Std APE ) for the BN-

NN model is 1.62% while ( Std APE ) for the ANN model is 21.51%,

hich proves that proposed BN-ANN model can improve accuracy

ignificantly and reduce the variation of error. For the educational

uildings, The MAPE results for the BN-ANN and the ANN are 9.21%

nd 3.6%, respectively; and the RMSE results for the BN-ANN and

he ANN are 9.56% and 3.66%, respectively. For Std APE , 10.35% and

.3% are for the BN-ANN and the ANN. The results show that al-

hough the BN-ANN model achieved an acceptable accuracy, the

NN model can lead to a better total building EUI prediction with-

ut considering the networks between the non-reference building

4 and the reference buildings E1, E2, and E3. This might be be-

ause there are already two reference buildings (E1, E2) that are

ighly related to the total building energy EUI. It is adequate to

redict the total building EUI with the reference buildings. While

or the laboratory buildings, we can find the BN-ANN model can

ave a better performance with the MAPE and RMSE of 7.66% and

.04%, respectively, while the two assessment indices are 12.1% and

2.2% for the ANN model. For Std APE , 5.01% and 8.80% are for the

N-ANN and the ANN. The BN-ANN model has a slight advantage

ompared with the ANN model for the laboratory buildings. In the

esidence buildings, the RMSE results show the BN-ANN model

oesn’t improve a lot on the robustness of the total building EUI

rediction compared with the ANN model. However, the BN-ANN

an improve the accuracy of EUI prediction results from 27.46% to

6.68% compared with the ANN model and Std APE from 16.9% to

0.94. Finally, the overall assessment shows the BN-ANN model

ased prediction accuracy with MAPE is 10.72% while the ANN

odel-based accuracy with MAPE is 18.94%. Also, the robustness

an be improved from 26.06% to 14.52% when comparing the BN-

NN with the ANN model. The BN-ANN model can reduce Std APE rom 17.98% to 8.25%. It can be concluded that the building net-

ork analysis can greatly improve the prediction performance of

he ANN model by integrating the networks between the reference

uildings with the total building energy use and the non-reference

uildings.

In the proposed BN-ANN model, three networks are vital for the

uilding group EUI prediction. They are: (1) the network between

he reference buildings’ EUI and the total buildings’ EUI, (2) the

etwork between each other of the individual reference buildings’

UI, and (3) the network between the reference buildings’ EUI and

he non-reference buildings’ EUI. For office buildings, it shows in

ig. 13 that the predicted building EUI results based on the BN-

NN and ANN models are less than the actual building EUI. This

ay be attributed to two important reasons, that the three net-

orks are weak in the office buildings group and the sum of ref-

rence buildings’ EUI is far less than the total building’ EUI. How-

ver, the accuracy using the BN-ANN model is improved substan-

ially when compared with the ANN model, which only used the

eference buildings’ EUI to predict the total buildings’ EUI. While

or the educational buildings, the findings are on the contrary;

he networks between each reference building’s EUI and the to-

al buildings’ EUI and the networks between reference buildings,

X. Xu, W. Wang and T. Hong et al. / Energy & Buildings 186 (2019) 80–97 91

Fig. 9. Training and test results with the office building dataset for the BN-ANN and the ANN models.

Fig. 10. Training and test results with the education building dataset for the BN-ANN and the ANN models.

92 X. Xu, W. Wang and T. Hong et al. / Energy & Buildings 186 (2019) 80–97

Fig. 11. Training and test results with the laboratory building dataset for the BN-ANN and the ANN models.

Fig. 12. Training and test results with the residence building dataset for the BN-ANN and the ANN models.

X. Xu, W. Wang and T. Hong et al. / Energy & Buildings 186 (2019) 80–97 93

Fig. 13. The building EUI prediction results for the office, educational, laboratory and residential buildings.

Fig. 14. The Q-Q plot for the office and educational buildings.

94 X. Xu, W. Wang and T. Hong et al. / Energy & Buildings 186 (2019) 80–97

Fig. 15. The Q-Q plot for laboratory and residential buildings.

Table 10

Comparison of the predicted results from the BN-ANN and ANN models.

BN-ANN model ANN model

MAE MAPE RMSE Std APE R 2 MAE MAPE RMSE Std APE R

2

Office buildings 0.75 9.33% 9.12% 1.62% 1.0 2.57 32.6% 36.1% 21.51% 0.6309

Education buildings 0.30 9.21% 9.56% 10.35% 0.9578 0.12 3.6% 3.66% 3.3% 0.988

Laboratory buildings 0.60 7.66% 9.04% 5.01% 0.7853 0.83 12.1% 12.2% 8.80% 0.9627

Residence buildings 1.26 16.68% 23.55% 10.94% 0.8142 1.60 27.46% 24.19% 16.90% 0.9592

Total 0.73 10.72% 14.52% 8.25% 0.8591 1.28 18.94% 26.06% 17.98% 0.56

m

w

t

g

t

i

d

5

b

a

d

b

b

w

e

are both strong so that the accuracies based on the BN-ANN and

the ANN models are both significant with the latter being a lit-

tle better. This indicates that the ANN model, using only the ref-

erence buildings’ EUI, is good enough to predict the total building

energy use.

For laboratory buildings, the accuracy using the BN-ANN model

is also improved. Although the network between the reference

buildings’ EUI and the total buildings’ EUI is strong, it shows that

most of the networks between the reference buildings’ EUI and

the non-reference buildings’ EUI are negative. This phenomenon

causes the predicted results based on the BN-ANN model to be less

than those based on the ANN model. In the residential building

group, the accuracy is improved and networks between each other

of reference buildings’ EUI, and the reference buildings’ EUI and

the total buildings’ EUI are both strong. Considering the standard

deviation of each building group’s EUI pattern, we found that the

standard deviation of the office building group and the residential

building group are relatively high. Applying the BN-ANN model in

those two groups can improve accuracy when predicting the total

building EUI.

s

In conclusion, the findings indicate that the BN-ANN model is

ore suitable and accurate for those buildings, in which the net-

orks (or correlations) between reference buildings’ EUI and the

otal buildings’ EUI are weak and standard deviation of building

roups’ EUI is relatively high. While for other building groups, e.g.,

he educational buildings, in which the three networks are strong,

t might infer that the ANN model will be accurate enough to pre-

ict the total building EUI.

. Discussion

This study presented interdisciplinary research integrating SNA-

ased building network analysis and an artificial neural network

lgorithm. First, we applied the theory of SNA method to con-

uct the building network analysis and identify: (1) the reference

uildings, the building’s energy use that closely matched the total

uildings energy use (correlation coefficient > = 60%), (2) the net- orks between the reference buildings, (3) the total building en-

rgy use trend, and (4) the non-reference buildings. In the second

tep, we integrated machine learning techniques with the building

X. Xu, W. Wang and T. Hong et al. / Energy & Buildings 186 (2019) 80–97 95

n

d

v

r

e

d

b

e

e

n

e

t

t

i

e

e

b

t

E

o

h

T

t

b

m

t

e

a

s

t

n

a

w

s

n

fi

s

o

a

o

i

n

a

n

f

t

o

w

f

p

a

i

o

n

w

p

t

c

h

6

m

w

a

a

(

I

n

w

m

w

l

p

u

l

c

r

t

l

r

r

p

b

i

a

A

p

a

t

m

b

a

o

b

m

E

t

b

c

p

t

l

m

f

S

w

e

C

A

d

T

(

N

(

t

c

t

s

etworks to predict the total building energy use. The final vali-

ation step used buildings from Southeast University that were di-

ided into four building types, including office, educational, labo-

atory and residential groups. Notably, each group had unique op-

rational hours resulting in different peak energy use. The results

emonstrated the proposed SNA-ANN model predicted the multi-

uilding EUI with satisfactory accuracy.

For city planners and energy policymakers, understanding

nergy use dynamics is critical to (1) knowing where and how

nergy is being consumed across the morphologic and socioeco-

omic contours of the city, (2) providing situational awareness of

nergy use to better allocate resources and target policy interven-

ions, and (3) identifying cost-efficient savings opportunities across

he city. Campuses consisting of a big group of buildings are an

mportant component of a city, and optimization of distribution

nergy systems in a campus can help reduce energy use and GHG

missions in the city. In this view, although the SNA-ANN is a

lack-box model, it provides insights into which buildings are

he most significant users and to predict the dynamic building

UI. The SNA-ANN model provides policymakers the prediction

f multi-building energy use with much less required data which

elps allocate energy resources and prioritize energy retrofit.

his study provides interdisciplinary framework outlining how

o define reference buildings and apply them to predict multi-

uilding energy. Finally, the multi-building energy prediction

odel can estimate the energy use, which is one key feature of

he grid-interactive efficient buildings [56] .

This study has some limitations. Firstly, it only selected South-

ast University as a case study with a limited number of buildings

nd did not test the method in other types of building groups,

uch as the multi-use buildings. This study also does not validate

he performance of the proposed SNA-ANN model to predict

on-campus building energy use at the city scale. Secondly, when

pplying the building network analysis method, three networks

ere created and inputted into the SNA-ANN model. However, this

tudy did not investigate if all the networks are needed or which

etwork is better for multi-building energy prediction. Also, the

ndings in this study indicated that the SNA-ANN model is more

uitable for those building groups with a higher standard deviation

f EUI patterns. Therefore, more effort and richer building datasets

re needed to support further research. Thirdly, the attributes

f buildings used to create the building networks are limited

n the scope of building physical information; and the building

etworks are created by calculating the correlations between those

ttributes of two buildings; however, two limitations should be

oted. One is that more attributes of buildings (e.g., occupancy in-

ormation, schedule) are recommended to be included in forming

he networks of buildings, which can improve the representation

f the reference buildings for the non-reference buildings as

ell as the accuracy of the prediction results. Another potential

or future study is the contribution of each attribute to the EUI

rediction and how to determine and select the most effective

ttributes to reduce the number of model inputs. Those limitations

nspire future work of attribute selection during the modeling

f multi-building energy use regardless of statistical and engi-

eering methods. Finally, further research is needed to determine

hether the proposed SNA-ANN model can be adopted for energy

rediction of larger building groups, e.g., city-scale buildings. For

his, the modeling to identify all the reference buildings at the

ity scale can be intensive and thus requiring cloud computing or

igh-performance computing to handle large-scale problems.

. Conclusions

This study proposed a multi-building energy use prediction

odel by integrating social network analysis based building net-

ork modeling and machine learning techniques. Social network

nalysis method was used to identify the reference buildings

nd establish correlations between the reference buildings and

1) the total building energy use and (2) non-reference buildings.

mportant building property information, like the building height,

umber of stories, year built, roof type, and construction type

ere considered in the model. To validate the proposed SNA-ANN

ethod, this study selected Southeast University as a case study,

ith four building groups tested including office, educational,

aboratory, and residential groups. To test the performance of the

roposed SNA-ANN model, we selected the ground truth energy

se data and the ANN-based predicted energy use as two base-

ines. The results show the proposed SNA-ANN and ANN models

an achieve the prediction MAPE accuracies of 9.3% and 32.6%,

espectively for office buildings; 9.2% and 3.6%, respectively for

he educational buildings; 7.7% and 12.1%, respectively for the

aboratory buildings; and 16.7% and 27.5%, respectively for the

esidential buildings. Considering the robustness of the RMSE

esults, the proposed SNA-ANN and ANN models can achieve the

rediction accuracies of 9.1% and 36.1%, respectively for the office

uildings; 9.6% and 3.6%, respectively for the educational build-

ngs; 9.0% and 12.2%, respectively for the laboratory buildings;

nd 23.5% and 24.2%, respectively for the residential buildings.

lso, observed in the overall results, the SNA-ANN model can

redict the multi-building energy use with an accuracy of MAPE

nd RMSE of about 10.72% and 14.52%, respectively, demonstrating

hat the proposed model can efficiently and accurately predict the

ulti-building energy use.

Findings indicate that the SNA-ANN model is more suitable for

uildings, where the networks between reference buildings’ EUI

nd the total buildings’ EUI are weak and the standard deviation

f building groups’ EUI is relatively high. While for the other

uilding groups, where all the three networks are strong, the ANN

odel proved to be accurate enough to predict the total building

UI. The proposed interdisciplinary SNA-ANN model, presented in

his study, provides a new, attractive empirical approach to urban

uilding energy use prediction. Further practical applications

an help identify reference buildings, which have a more similar

attern to the total district building energy use intensity. With

he wider adoption of smart meters and micro-grid at the campus

evel, such method proposed in this study might help facility

anagers quickly estimate total energy use with less data and

ormulate optimal energy-saving strategies. In future work, the

NA-ANN model could be applied to predict the energy use of the

hole campus or a larger group of buildings in city districts or an

ntire city depending on more energy use data.

onflict of interest

There is no conflict of interest.

cknowledgments

This work was sponsored by National Natural Science Foun-

ation of China (NSCF # 51678127 ), National Scientific and

echnological Support during the 12th Five-Year Plan Period

#2013BAJ10B13), China Scholarship Council (CSC # 201706095035 ),

ational Natural Science Foundation of China for Young Scholars

NSFC #51408122 ), and Beijing Advanced Innovation Center for Fu-

ure Urban Design (UDC #016010100). Any opinions, findings, con-

lusions, or recommendations expressed in this paper are those of

he authors and do not necessarily reflect the views of the spon-

oring committees.

96 X. Xu, W. Wang and T. Hong et al. / Energy & Buildings 186 (2019) 80–97

[

References

[1] L. Pérez-Lombard, J. Ortiz, C. Pout, A review on buildings energy consumption

information, Energy Build. 40 (2008) 394–398, doi: 10.1016/j.enbuild.2007.03.

007 . [2] City Energy Project A Joint Project of NRDC + IMT, http://www.

cityenergyproject.org/ (accessed May 3, 2018). [3] U.S. Energy Information Administration (EIA), Electric Power Monthly with

data for January 2018, 2018 (accessed April 4, 2018). https://www.eia.gov/ electricity/monthly/ .

[4] T. Hong, IEA EBC annexes advance technologies and strategies to reduce en-

ergy use and GHG emissions in buildings and communities, Energy Build. 158 (2018) 147–149, doi: 10.1016/J.ENBUILD.2017.10.028 .

[5] H. Yoshino, T. Hong, N. Nord, IEA EBC annex 53: total energy use in buildings— Analysis and evaluation methods, Energy Build. 152 (2017) 124–136, doi: 10.

1016/J.ENBUILD.2017.07.038 . [6] H.X. Zhao, F. Magoulès, A review on the prediction of building energy con-

sumption, Renew. Sustain. Energy Rev. 16 (2012) 3586–3592, doi: 10.1016/j.rser. 2012.02.049 .

[7] Y. Wei, X. Zhang, Y. Shi, L. Xia, S. Pan, J. Wu, M. Han, X. Zhao, A review

of data-driven approaches for prediction and classification of building energy consumption, Renew. Sustain. Energy Rev. 82 (2018) 1027–1047, doi: 10.1016/j.

rser.2017.09.108 . [8] Z. Zheng, Z. Zhuang, Z. Lian, Y. Yu, Study on building energy load prediction

based on monitoring data, Proc. Eng. 205 (2017) 716–723, doi: 10.1016/j.proeng. 2017.09.894 .

[9] Kadir Amasyali, Nora M. El-Gohary, A review of data-driven building energy

consumption prediction studies, Renew. Sustain. Energy Rev. 81 (2018) 1192– 1205, doi: 10.1016/J.RSER.2017.04.095 .

[10] J. Kneifel, D. Webb, Predicting energy performance of a net-zero energy build- ing: a statistical approach, Appl. Energy 178 (2016) 46 8–4 83, doi: 10.1016/J.

APENERGY.2016.06.013 . [11] Y. Sun, G. Huang, X. Xu, A.C.-K. Lai, Building-group-level performance evalua-

tions of net zero energy buildings with non-collaborative controls, Appl. En-

ergy 212 (2018) 565–576, doi: 10.1016/J.APENERGY.2017.11.076 . [12] C.E. Kontokosta, C. Tull, A data-driven predictive model of city-scale energy use

in buildings, Appl. Energy 197 (2017) 303–317, doi: 10.1016/j.apenergy.2017.04. 005 .

[13] D. Hsu, Comparison of integrated clustering methods for accurate and stable prediction of building energy consumption data, Appl. Energy 160 (2015) 153–

163, doi: 10.1016/j.apenergy.2015.08.126 .

[14] W. Li, Y. Zhou, K. Cetin, J. Eom, Y. Wang, G. Chen, X. Zhang, Modeling urban building energy use: a review of modeling approaches and procedures, Energy

141 (2017) 2445–2457, doi: 10.1016/j.energy.2017.11.071 . [15] C.F. Reinhart, C. Cerezo Davila, Urban building energy modeling - A review of

a nascent field, Build. Environ. 97 (2016) 196–202, doi: 10.1016/j.buildenv.2015. 12.001 .

[16] Y. Han, J.E. Taylor, A.L. Pisello, Toward mitigating urban heat island effects: In-

vestigating the thermal-energy impact of bio-inspired retro-reflective build- ing envelopes in dense urban settings, Energy Build. 102 (2015) 380–389,

doi: 10.1016/J.ENBUILD.2015.05.040 . [17] A.L. Pisello, V.L. Castaldo, J.E. Taylor, F. Cotana, Expanding inter-building effect

modeling to examine primary energy for lighting, Energy Build. 76 (2014) 513– 523, doi: 10.1016/J.ENBUILD.2014.02.081 .

[18] A.L. Pisello, J.E. Taylor, X. Xu, F. Cotana, Inter-building effect: simulating the

impact of a network of buildings on the accuracy of building energy perfor- mance predictions, Build. Environ. 58 (2012) 37–45, doi: 10.1016/J.BUILDENV.

2012.06.017 . [19] Y. Han, J.E. Taylor, A.L. Pisello, Exploring mutual shading and mutual reflection

inter-building effects on building energy performance, Appl. Energy 185 (2017) 1556–1564, doi: 10.1016/J.APENERGY.2015.10.170 .

[20] Y. Han, J.E. Taylor, Simulating the Inter-Building Effect on energy consumption from embedding phase change materials in building envelopes, Sustain. Cities

Soc. 27 (2016) 287–295, doi: 10.1016/J.SCS.2016.03.001 .

[21] C. Li, T. Hong, D. Yan, An insight into actual energy use and its drivers in high-performance buildings, Appl. Energy 131 (2014) 394–410, doi: 10.1016/j.

apenergy.2014.06.032 . [22] P. Xiufeng, H. Tianzhen, Mary ANN Piette, Improving building performance at

urban scale with a framework for real-time data sharing, in: Proceedings of the Symposium on Simulation for Architecture & Urban Design, 2013, p. 221.

https://dl.acm.org/citation.cfm?id=250 0 0 07 .

[23] J.A . Fonseca, A . Schlueter, Integrated model for characterization of spatiotem- poral building energy consumption patterns in neighborhoods and city dis-

tricts, Appl. Energy 142 (2015) 247–265, doi: 10.1016/j.apenergy.2014.12.068 . [24] M.J.N. Oliveira, P. Ao, M.C. Brito, M.J.N. Oliveira Panão, Modelling aggregate

hourly electricity consumption based on bottom-up building stock, Energy Build. 170 (2018) 170–182, doi: 10.1016/j.enbuild.2018.04.010 .

[25] S. Kalogirou , C. Neocleous , C. Schizas , Building heating load estimation us-

ing artificial neural networks, in: Proceedings of the Seventeenth International Conference on Parallel Architectures and Compilation Techniques, 1997, pp.

1–8 . [26] R.K. Jain, K.M. Smith, P.J. Culligan, J.E. Taylor, Forecasting energy consumption

of multi-family residential buildings using support vector regression: Inves- tigating the impact of temporal and spatial monitoring granularity on perfor-

mance accuracy, Appl. Energy 123 (2014) 168–178, doi: 10.1016/j.apenergy.2014.

02.057 .

[27] D. Hawkins, S.M. Hong, R. Raslan, D. Mumovic, S. Hanna, Determinants of en- ergy use in UK higher education buildings using statistical and artificial neural

network methods, Int. J. Sustain. Built Environ. 1 (2012) 50–63, doi: 10.1016/j. ijsbe.2012.05.002 .

[28] M. Kavgic, A. Summerfield, D. Mumovic, Z. Stevanovic, Application of a Monte Carlo model to predict space heating energy use of Belgrade’s housing stock, J.

Build. Perform. Simul. 8 (2015) 375–390, doi: 10.1080/19401493.2014.961031 . [29] M. Kavgic, D. Mumovic, A. Summerfield, Z. Stevanovic, O. Ecim-Djuric, Uncer-

tainty and modeling energy consumption: sensitivity analysis for a city-scale

domestic energy model, Energy Build. 60 (2013) 1–11, doi: 10.1016/J.ENBUILD. 2013.01.005 .

[30] A. Afram, F. Janabi-Sharifi, Black-box modeling of residential HVAC system and comparison of gray-box and black-box modeling methods, Energy Build. 94

(2015) 121–149, doi: 10.1016/j.enbuild.2015.02.045 . [31] Building technology and urban systems division at lawrence berkeley national

laboratory, City Building Energy Saver. https://citybes.lbl.gov/ .

[32] Y. Chen, T. Hong, X. Luo, B. Hooper, Development of city buildings dataset for urban building energy modeling, Energy Build. 183 (2019) 252–265, doi: 10.

1016/j.enbuild.2018.11.008 . [33] Y. Chen, T. Hong, M.A. Piette, Automatic generation and simulation of ur-

ban building energy models based on city datasets for city-scale building retrofit analysis, Appl. Energy 205 (2017) 323–335. http://www.sciencedirect.

com/science/article/pii/S0306261917310024 . (accessed October 25, 2017) .

[34] J.A . Fonseca, T.A . Nguyen, A . Schlueter, F. Marechal, City energy analyst (CEA): Integrated framework for analysis and optimization of building energy systems

in neighborhoods and city districts, Energy Build. 113 (2016) 202–226, doi: 10. 1016/j.enbuild.2015.11.055 .

[35] C. Felsmann , S. Robbi , E. Eckstädt , Reduced order building energy system mod- eling in large-scale, in: Proceedings of the Thirteenth Conference of Interna-

tional Building Performance Simulation Association, 2013, pp. 1216–1223 .

[36] M. Heidarinejad, N. Mattise, M. Dahlhausen, K. Sharma, K. Benne, D. Macum- ber, L. Brackney, J. Srebric, Demonstration of reduced-order urban scale build-

ing energy models, Energy Build. 156 (2017) 17–28, doi: 10.1016/j.enbuild.2017. 08.086 .

[37] F. Zhao, S.H. Lee, G. Augenbroe, Reconstructing building stock to replicate energy consumption data, Energy Build. 117 (2016) 301–312, doi: 10.1016/J.

ENBUILD.2015.10.001 .

[38] U.S. Department of Energy, Commercial prototype building models, Build. En- ergy Codes Program. (2016) 4–7. https://www.energycodes.gov/development/

commercial/prototype _ models#90.1 . [39] M. Deru, K. Field, D. Studer, K. Benne, B. Griffith, P. Torcellini, B. Liu, M. Halver-

son, D. Winiarski, M. Rosenberg, M. Yazdanian, J. Huang, D. Crawley, U.S. de- partment of energy commercial reference building models of the national

building stock, Golden, CO, United States, 2011. doi: 10.2172/1009264 .

[40] A. Mastrucci, O. Baume, F. Stazi, U. Leopold, Estimating energy savings for the residential building stock of an entire city: a GIS-based statistical downscaling

approach applied to Rotterdam, Energy Build. 75 (2014) 358–367, doi: 10.1016/ j.enbuild.2014.02.032 .

[41] P. Caputo, G. Costa, S. Ferrari, A supporting method for defining energy strate- gies in the building sector at urban scale, Energy Policy 55 (2013) 261–270,

doi: 10.1016/j.enpol.2012.12.006 . [42] Z. Yang, J. Roth, R.K. Jain, DUE-B: data-driven urban energy benchmarking of

buildings using recursive partitioning and stochastic frontier analysis, Energy

Build. 163 (2018) 58–69, doi: 10.1016/J.ENBUILD.2017.12.040 . [43] N. Gaitani, C. Lehmann, M. Santamouris, G. Mihalakakou, P. Patargias, Us-

ing principal component and cluster analysis in the heating evaluation of the school building sector, Appl. Energy 87 (2010) 2079–2086, doi: 10.1016/J.

APENERGY.20 09.12.0 07 . 44] C. Deb, S.E. Lee, Determining key variables influencing energy consumption

in office buildings through cluster analysis of pre- and post-retrofit build-

ing data, Energy Build. 159 (2018) 228–245. https://www.sciencedirect.com/ science/article/pii/S0378778817330244#! . (accessed April 5, 2018) .

[45] European Commission, Commission delegated regulation (EU) No 244/2012 of 16 January 2012 supplementing Directive 2010/31/EU of the European Par-

liament and of the Council on the energy performance of buildings by es- tablishing a comparative methodology framework for calculating, 2012 (ac-

cessed April 6, 2018). https://publications.europa.eu/en/publication- detail/- /

publication/40347d51- cd2d- 4935- 9ae1- 293171ba12d2/language- en . [46] R. Arambula Lara, G. Pernigotto, F. Cappelletti, A. Gasparella, Energy audit of

schools by means of cluster analysis, Energy Build. 95 (2015) 160–171, doi: 10. 1016/J.ENBUILD.2015.03.036 .

[47] N. Gaitani, C. Lehmann, M. Santamouris, G. Mihalakakou, P. Patargias, Us- ing principal component and cluster analysis in the heating evaluation of

the school building sector, Appl. Energy 87 (2010) 2079–2086, doi: 10.1016/J.

APENERGY.20 09.12.0 07 . [48] G. Tardioli, R. Kerrigan, M. Oates, J. O’Donnell, D.P. Finn, Identification of rep-

resentative buildings and building groups in urban datasets using a novel pre- processing, classification, clustering and predictive modelling approach, Build.

Environ. 140 (2018) 90–106, doi: 10.1016/J.BUILDENV.2018.05.035 . [49] E. Otte, R. Rousseau, Social network analysis: a powerful strategy, also

for the information sciences, J. Inf. Sci. 28 (2002) 441–453, doi: 10.1177/

016555150202800601 . [50] M. Grandjean, A social network analysis of Twitter: Mapping the digital hu-

manities community, Cogent Arts Humanit. (2016) 3, doi: 10.1080/23311983. 2016.1171458 .

X. Xu, W. Wang and T. Hong et al. / Energy & Buildings 186 (2019) 80–97 97

[

[

[

[

[

[51] M. Aram, G. Neumann, Multilayered analysis of co-development of busi- ness information systems, J. Internet Serv. Appl. (2015) 6, doi: 10.1186/

s13174- 015- 0030- 8 . 52] L. Luo, Q. Shen, G. Xu, Y. Liu, Y. Wang, Stakeholder-associated supply chain

risks and their interactions in a prefabricated building project : a case study in Hong Kong, J. Manag. Eng. (2018), doi: 10.1061/(ASCE)ME.1943- 5479.0 0 0 0675 .

53] C.S.E. Bale, N.J. McCullen, T.J. Foxon, A.M. Rucklidge, W.F. Gale, Harnessing so- cial networks for promoting adoption of energy technologies in the domestic

sector, Energy Policy 63 (2013) 833–844, doi: 10.1016/j.enpol.2013.09.033 .

54] A. Vantoch-Wood, P.M. Connor, Using network analysis to understand public policy for wave energy, Energy Policy 62 (2013) 676–685, doi: 10.1016/j.enpol.

2013.07.066 . 55] M. McMichael, D. Shipworth, The value of social networks in the diffusion of

energy-efficiency innovations in UK households, Energy Policy 53 (2013) 159– 168, doi: 10.1016/j.enpol.2012.10.039 .

56] ENERGY.GOV, Buildings and the grid 101: why does it matter for energy efficiency, 2017. https://energy.gov/eere/buildings/articles/

buildings- and- grid- 101- why- does- it- matter- energy- efficiency .