You are on page 1of 7

Resources, Conservation & Recycling 129 (2018) 168–174

Contents lists available at ScienceDirect

Resources, Conservation & Recycling

journal homepage:

Full Length Article

Modeling energy-related CO2 emissions from office buildings using general MARK
regression neural network

Hong Yea,b, , Qun Rena,c, Xinyue Hua,b, Tao Lina,b, Longyu Shia,b, Guoqin Zhanga,b, Xinhu Lia,b
Key Lab of Urban Environment and Health, Institute of Urban Environment, Chinese Academy of Sciences, Xiamen 361021, China
Xiamen Key Laboratory of Urban Metabolism, Xiamen 361021, China
Xiamen University, Xiamen 361102, China


Keyword: Carbon dioxide (CO2) emissions from urban office buildings energy usages (BEC) constitute a substantial
Office building component of anthropogenic greenhouse gas emission, and are set to rapidly increase with further urbanization.
Energy usage carbon emission Establishing a concise, accurate, and realistic model that can predict future emissions is challenging but essential
Model construction for strategies to develop low carbon construction and sustainable development in urban areas. In this paper, the
General regression neural network
operational energy use for 294 office buildings across China was collected and analyzed. We focus on four main
Scenario analysis
variables, and analyze a further ten second-level variables, to elucidate the role that a building’s occupants, its
structural characteristics, and localized natural conditions play in determining energy consumption. Using
general regression neural network (GRNN), the factors’ direct and indirect effects on energy consumption were
tested. A building’s structural attributes had the most impact on energy-related CO2 emissions, followed by the
relevant socioeconomic conditions, the micro-climate, and finally the regional climate. A version of the model
that was constructed with interaction between the four main variables was found to be the most precise. GRNN
combined with urban development scenarios was used for the prediction of cities’ future CO2 emissions.
Economic development and improving standards in the construction industry could have significant impacts on
future CO2 emissions. This study provides a detailed method that could be used to explore the dynamics of office
energy use and competing options for the construction of low-carbon office buildings.

1. Introduction (GHG) emission targets.

Office buildings provide space for many socioeconomic activities,
Carbon dioxide (CO2) emissions are an important cause of climate and play an important role in supporting sustainable urban develop-
change, and how to decrease these, especially those arising from energy ment (Lin and Liu, 2015). Currently, per m2 floor area office buildings
use, is a great challenge. The continued use of finite fossil fuels miti- consume 22% of the total urban electricity, approximately 10–20 times
gates efforts to fight climate change and avert energy crises. In 2008, more than residential buildings of China (Construction, 2007). Previous
China became the largest CO2-emitting country as the country’s rapid studies (Gao et al., 2017; Thomas and Azevedo, 2013a, 2013b) have
development has led to significant growth in the amount of energy shown that rebound effects (e.g. government financial support may lead
used. Indeed, energy use in 2008 was 9.408 × 104 GJ (China, 2009), to more energy wastage) can lead to more energy consumption. Across
which increased to 1.249 × 105 GJ in 2014 (China, 2015) with an- a building’s lifecycle, 85% of total energy-related CO2 emissions arise
nually 5.128% increasing rate. The industrial, transport, and con- during the operational period, far larger than the 13% and 2%, re-
struction sectors contributed more than 70% to total energy-related spectively, from the construction and demolition stages (Luo et al.,
emissions with construction playing an increasingly large role. Build- 2016; Peng, 2016). Decreasing office buildings’ operational energy
ings contributed about one-third of total CO2 emissions in 2012 (Hirst, consumption is therefore critical for the development of low-carbon
2013). The situation in China is similar with buildings responsible for cities in China (Wu et al., 2012).
28% of the total emissions in 2011, a figure that is expected to reach Since the 1960s, the energy performance of buildings has been
35% by 2020 owing to a drive for improved living conditions (Chen widely studied (March, 1972). At the building level, many energy
et al., 2015). The increasing carbon footprint of the construction sector models and tools have been applied to precisely investigate the relevant
places great pressure on achieving the government’s greenhouse gas thermodynamics, especially from a construction point of view.

Corresponding author at: Key Lab of Urban Environment and Health, Institute of Urban Environment, Chinese Academy of Sciences, Xiamen 361021, China.
E-mail addresses:, (H. Ye).
Received 8 July 2017; Received in revised form 14 October 2017; Accepted 16 October 2017
0921-3449/ © 2017 Elsevier B.V. All rights reserved.
H. Ye et al. Resources, Conservation & Recycling 129 (2018) 168–174

Nomenclature EE Expenditure on education

HDD Heating degree days
Abbreviation CDD Cooling degree days
BF Building floors
FA Occupied floor area DCBD Distance to CBD (commercial business district)
ECM Energy consumption membership BUI Building’s attributes
GRP Gross regional production SEC Socioeconomic conditions
Wage Average wage of employed staff MAC Regional climate conditions
EST Expenditure on science and technology MIC Micro-climate conditions

Typically, census and energy use data are combined by models to es- ANN models (for example, back propagation (BP) calculations) are used
timate a building’s potential energy demand (Clarke, 2001). Statistical to solve complicated problems and have many advantages over linear
correlations (e.g. linear regression models) are sometimes used as an regression models (Abu Qdais et al., 2010; Pombeiro et al., 2017).
alternative method of predicting the variations in energy demand pat- However, BP neural network has the disadvantages of slow con-
terns (Liu and Sweeney, 2012). Moreover, most models require time- vergence and easily converging to a local minimum (Yu, 1992). The
consuming geometric modeling and painstaking data compilation to accuracy of predictions made using BP neural networks is difficult to
produce valid assessments (Jaffal and Inard, 2017). Even so, the un- quantify owing to the long training process that is required. Proposed in
certainties associated with numerous input parameters can result in 1991 by Specht, general regression neural networks (GRNNs) involve a
misleading outputs. The impacts on energy consumption in office one-pass learning algorithm, are able to solve non-linear problems, and
buildings from construction attributes, climatic and socioeconomic have been widely used in signal processing, energy prediction and de-
conditions have been investigated individually using a range of models cision making. GRNNs offer the advantages of needing fewer training
(Robati et al., 2017; Tian et al., 2016; Yuan et al., 2017; Zhou et al., samples, having a flexible network structure, and a high fault-tolerance.
2015). However, few tools are available to evaluate the impact on en- Moreover, the GRNN has few parameters that need to be artificially set
ergy consumption caused by changes and interactions between these in advance, so it can learn the potential relationship between variables
variables (Ratti et al., 2005; Reinhart et al., 2013). This work aims to fill (Li et al., 2017). Therefore, in this study, we focused on constructing a
this gap by integrating parameters that have impacts on a building’s GRNN model to predict a building’s energy-related CO2 emissions.
thermal and lighting demands to attain a more robust method for Strict selection of the input variables is used to make the model rea-
predicting energy demand in office buildings. listic. To avoid magnifying errors during the computation process, the
An artificial neural network (ANN) is a mathematic model that deals input variable was concisely selected also. Furthermore, when in-
with information processing in a manner that simulates the behavior of tegrated with an analysis of a future scenarios, such a model can be
neurons in the brain (Antanasijević et al., 2015; Benedetti et al., 2016; used to simulate various projections of regional CO2 emissions.
Jing et al., 2017; Khayatian et al., 2016). Work has shown that some

Fig. 1. Samples’ distribution and attribute.

H. Ye et al. Resources, Conservation & Recycling 129 (2018) 168–174

2. Data collection where CEF is the total carbon emission from fuels including coal and
natural gas. Cei denotes the consumption amount of energy i (kg or m3);
We hypothesize that it is better to consider four aspects (a building’s Gei is the capacity conservation of energy i (MJ/kg or MJ/m3); Iei is
attributes (BUI), regional climate (MAC), socioeconomic conditions carbon containment per unit according to IPCC estimate (gC/MJ). The
(SEC), and micro-climate (MIC)) not for their individual influence on CO2 emissions associated with electricity and heating consumption
energy consumption but to consider their combined, interrelated effect were calculated using the appropriate emissions factors. The origin (and
(Yuan et al., 2016). Yearly energy consumption for the people’s bank of therefore emissions factor) for electricity and heating was selected ac-
China and the other 293 branch banks in 2011 across China were ac- cording to the procedure in the Guidelines for Provincial Greenhouse
quired (Fig. 1). Gas Inventories. In particular, the electricity emissions factor was
To define each building’s attributes, we collected data reporting the chosen to reflect the average emissions factor for plants in the local
occupied floor area (FA) and the energy consumption membership region (NDAR, 2011).
(ECM) in each bank by surveying their basic information and energy
type in 2011. The per capita gross regional product (GRP), average 3.2. GRNN model construction
wage of employed staff (Wage), expenditure on science and technology
(EST), and expenditure on education (EE) were collected through China As shown in Fig. 2, a GRNN consists of four layers: input, pattern,
City Statistical Yearbook of 2012 to represent socioeconomic condi- summation; and output (Specht, 1991). The neuron numbers of the
tions. Data on local weather conditions was collected from the Global input layer and output layer is equal to the number of input and output
Surface Summary of Day (GSOD). We calculated the heating degree parameters that have been used in the GRNN model (Bendu et al.,
days (HDD) and cooling degree days (CDD) by using GIS tools to esti- 2017). The training data were aggregated in the pattern layer, which
mate the regional climatic conditions. A building’s location, vertical had a neuron number equal to that of training data. And the summation
attributes such as layer and its surface quality such as the conditions of layer had an extra neuron compared to the output variable
surrounding underlying insulation surface were representative to (Antanasijević et al., 2013). In this research, the neuron numbers of
micro-climate conditions. In this research, we chose the distance to the input layer varied from two to ten because we tried several types of
commercial business center (DCBD), which was estimated using a GRNN model by using various influencing factors. The pattern layer
nightlight remote image, and the building floor (BF), which was also had 244 neurons since we separate all the 294 samples into 244 for
collected from the survey of the bank, were used to assemble the factor training and 50 for testing, the summation layer was composed of two
relating to each building’s micro-climate. neurons and the output layer had one neurons.
There were a linear function linking between input layer and pat-
3. Methodology tern layer. While the transfer function between pattern layer and
summation layer was shown by Formula (3).
3.1. Calculating energy-related CO2 emissions
f (D) = exp(− )
2σ 2 (3)
To calculate a building’s operational emissions, we included emis-
sions produced to provide the energy used by lighting, cooling, and where D is the distance of training patterns using Euclidean method,
heating through the consumption of coal, electricity, and natural gas. and σ is a parameter of Gaussion curve, here we call it Smoothing
For each building, its total CO2 emissions were then calculated using Factors (SF) (Siljic et al., 2015).
the Eq. (1): SF was also considered as an important factor determining the ac-
curacy of GRNN. It was greater than 0 and usually range from 0.01 to 1.
The determination of SF could be made manually by testing different SF
Where CE is the carbon emissions, CEELE is the carbon emissions comes values until achieving a satisfactory agreement on predicted and ob-
from electricity usage. CECOAL CENG, CEGAS and CEHE represent the served output value (Antanasijević et al., 2013).Both small SF and large
carbon emissions from coal, heating, natural gas, gasoline and heating SF were not proper for training because a large SF may reduce the
consumption respectively. fitting performance while a small one may reduce the generation per-
Based on the method recommended by the IPCC’s for the calculation formance (Antanasijević et al., 2014).
of national GHG emissions, we used Eq. (2) to convert consumption of a Try to demonstrate the changes and the interaction among 4 main
primary energy source (coal or natural gas) to CO2 emissions (IPCC, factors for regulating building carbon emission, we input these factors
2007). and trained the model at the individual and (combined) group level.
n Matlab 8.0 software was used to design, train, and validate the emis-
CEF = ∑ Cei × Gei × Iei sions predicted using the GRNN model. The root mean squared error
i=1 (2) (RMSE), mean absolute percentage of error (MAPE), coefficient of

Fig. 2. The GRNN architecture.

H. Ye et al. Resources, Conservation & Recycling 129 (2018) 168–174

determination (R2), and Willmott’s index of agreement (IA) were se- confirming results we reported previously (Ye et al., 2015; Ye et al.,
lected as indicators of the model’s precision (Csábrági et al., 2017) 2013; Ye et al., 2011). A large variation in emissions was observed
(Formula (4)–(7)). depending on a building’s relative distance to the city center (the lowest
value of 9.04 × 10−3 was found in Yan’an city in Shangxi Province
1 while the highest was 0.994 in Chanye city in Shanxi Province city) and
∑ (Oi − Pi )
i=1 (4) the layer (lowest value of 2 in Jiujiang city, Jiangxi Province and
highest value of 36 in Baoding city in Hebei Province) which would fall
1 o −p into different quality categories to demonstrate micro-climate quality.
MAPE = ∑ i o i × 100%
n i=1 i (5)
n 4.2. GRNN modeling for energy usage mechanism
⎡ ⎤
⎢∑ (Pi − P )(Oi − O ) ⎥
⎣ i=1 ⎦ Using the GRNN model, we exploited our buildings dataset to study
R2 = n n
how various factors impacted a building’s energy consumption, con-
∑ (Pi − P )2 ∑ (Oi − O )2 centrating particularly on the interactions between climatic conditions,
i=1 i=1 (6)
membership attribute, and observable conditions such as building outer
n micro-climate quality. We also used the model to investigate links be-
⎡ ∑ (Pi − Oi)2 ⎤
⎢ ⎥ tween changes in energy consumption and overall economic activity
i = 1
IA = 1 − ⎢ n ⎥ and the prevalence of energy-saving behavior.
⎢ 2⎥
⎢ ∑ ( p i − p + oi − o ) ⎥ To understand the most effective way to limit CO2 emissions from
⎣ i=1 ⎦ (7) office buildings, the four main factors were input into the model and
simulated individually. We then gradually added vectors of other at-
where Pi and Oi represent the predicted and observed value of sample i.
tributes to the model. The resulting output demonstrated a factor’s di-
P and O represent average predicted and observed values respectively.
rect and indirect (via its impact on other factors) contributions to en-
ergy-related CO2 emissions. Of the 294 buildings in our sample, we
3.3. Scenario analysis
selected 244 (over 80%) to form the training dataset with the remaining
50 buildings forming the test dataset (see Table 1 for detailed in-
We developed a plausible scenario to investigate the impact of
formation). The buildings in the two datasets included samples from
urban development on CO2 emissions, to reflect how uncertain influ-
five building climate areas. The relative error (RE) and MAPE were used
encing factors affect future trend, and to predict the future of an office’s
to test the model’s accuracy (Bakshan et al., 2017).
energy-related emissions in 2020. The scenario assumed no changes in
The results for modelling the factors individually showed that the
the regional or micro-climates nor in the construction area. Meanwhile,
most important predictor of energy usage was a building’s attributes,
GRP, income, EST and EE were assumed to vary according to the na-
with larger buildings consuming more energy than smaller ones (Fig.
tional and regional economic development rates. The energy con-
A1, Table A1). These finding agreed with our previous studies that
sumption membership was determined by the regional employment
showed an increased energy requirement for the cooling and heating of
buildings with larger floor spaces.
Following the building’s attributes, the next most important con-
4. Results and discussion tributions came from socioeconomic, micro-climate, and regional cli-
mate factors (Fig. A2, Table A2). Office buildings in our samples were
4.1. Variation in energy consumption and CO2 emissions differentiated products whose energy utility was supported by a special
government fund. Personal duty free may increase energy requirements
The average value for a building’s CO2 emissions was 1355 T/year. through the seeking of higher quality environments (better lighting and
The CO2 emissions from electricity accounted for 84.45% of the total. more focus on aesthetics and comfort). Our previous work showed that
Emissions from the raw consumption of coal and natural gas accounted these rebound effects were impacted by the level of education or the
for 8.26% and 5.11%, respectively. Of the buildings in our sample, economic conditions of the users (Ye et al., 2017b).
those located in areas characterized as severely cold approximately The energy consumption of office buildings in more urbanized areas
167.73% more coal than the average, while those buildings in mild was slightly greater than that of those in less urbanized areas. Some
climates showed a 12.15% increase in electricity consumption. researchers have asserted that this increment in energy use is a result of
Buildings were separated into five classes depending on their con- changes in a building’s surrounding micro-climate, for example, heat
struction characteristics (Code for Design of Civil Buildings (CDCB)).
Fig. 1 clearly indicates that the regional climate factor affected energy-
related emissions, with CO2 emissions lowest in the mild area with
0.20 kg/m2/person, increasing to the highest value of 0.34 kg/m2/
person and 0.49 kg/m2/person in the cold and severely cold areas
where central heating is required. Energy-related CO2 emissions varied
from 0.581 × 103 T CO2 per year in areas with hot summers and cold
winters to 2.26 times that value for buildings in areas with hot summers
and warm winters.
The average floor area of the buildings ranged from 400 to
90,100 m2. Most of the buildings had a medium floor area (between
5000 and 20,000 m2). Otherwise, 8 buildings, which were located in
less developed cities, had a small floor area (< 5000 m2) while 57
buildings in developed cities had a large floor area (> 20,000 m2). The
number of employees varied from 12 in Jinhua City to 2400 in Beijing.
In general, office buildings located in provincial capitals had a larger
floor area and more employees. From Fig. 3 we see that both floor area
Fig. 3. The positive relationship between building Floor area and CO2 emissions.
and number of employees are positively correlated with CO2 emissions,

H. Ye et al. Resources, Conservation & Recycling 129 (2018) 168–174

Table 1
Training dataset of the GRNN model.

Input indicators Unit Training dataset Test dataset

Mean Min Max Mean Min Max

Floor area (FA) hectare 1.71 400 8.70 1.91 0.19 9.01
Energy consumption membership (ECM) person 271.36 12 2400 291.32 31 882
Gross regional production (GRP) thousand RMB/person 50.81 10.25 251.49 55.38 9.34 166.65
Average wage (Wage) thousand RMB/person 38.07 17.21 320.63 37.84 25.52 758.34
Expenditure on science and technology (EST) million RMB 967.59 11.96 21849.75 812.99 18.64 1830.75
Expenditure on education (EE) million RMB 5546.77 200.48 54923.93 496.89 79.85 52007.78
Heating degree days (HDD) °Cd 2398.52 216.63 7326.18 2436.39 33.13 5717.02
Cooling degree days (CDD) °Cd 1021.37 43.10 2228.97 1068.29 86.12 2403.95
Building floor (BF) floor 11.52 2 33 11.78 4 36
Distance to CBD (DCBD) – 0.084 0.0090 0.99 0.076 0.0098 0.52

Output indicator Unit Training dataset Test dataset

Mean Min Max Mean Min Max

CO2 ton 1338.20 33.64 9593.54 1439.65 121.73 11095.02

island impacts become more important in more densely developed city space per membership size. The large-scale auxiliary equipment are
centers (Li et al., 2013; Sun and Chen, 2017; Zhou et al., 2011). The required for heating in these areas led to large energy consume dis-
coefficient on city development rate indicated that there was economics tance. This was confirmed by the increased consumption of buildings
of scale in heating and cooling buildings (Fig. A3, Table A3). located in areas with average winter temperatures lower than 20 °C.
National benchmarks for energy consumption in buildings highlight We also studied the relationships between the main factors
that heating, cooling, and lighting account for the majority of an office (building attributes, regional and micro-climate, and socioeconomic
building’s energy consumption and that this level of consumption is conditions) (Fig. 4a-1). The MAPE value for the models suggested that a
heavily dependent on local climatic conditions. As noted in Section 4.1, model constructed using these four main factors as inputs could accu-
the regional climate factor in the model shows that buildings in cold rately simulate reality (Table A4).
and severely cold environments consume large energy per unit of floor

Fig. 4. Model Quality Comparison between the GRNN and the MLR.

H. Ye et al. Resources, Conservation & Recycling 129 (2018) 168–174

4.3. GRNN modeling for future predictions carbon emission targets. Since a building’s attributes show greatest
impact on CO2 emissions, it is important to control the heating or
Before using the GRNN model to make predictions for the future, we cooling area of office buildings to the least scale. For example, areas in
used the original dataset to compare the GRNN model outputs with lobby, corridor and elevator, installing less air-conditioners or con-
those using multiple linear regression (MLR). The R2 value obtained trolling the space conditioner temperature would help CO2 decrement
using the MLR model was 0.7001, lower than that for the GRNN model (Kim, 2017). In addition, the increment of water body and green space
(0.7673) (Fig. 4). The GRNN model also performed well in predicting are recommended to distribute in scientific quality and quantity be-
CO2 emissions with values of MAPE of 2.53%, RE of 5.40%, RMSE of cause they are essential for controlling micro-climate which would
0.40, and IA of 0.9017. subsequently improving building comfort according to our previous
Using the last decade’s growth as a basis, the values of EST, EE, studies (Ye et al., 2017a). More comprehensive office building con-
Wages and GRP in 2020 were calculated for the 294 cities and then struction rules for different climate such as concentrative heating fa-
combined with the other variables to carry out the simulation. Fig. 5 cilities, heating and warming insulation material and surrounding land
shows the GRNN model’s predictions for office buildings’ energy con- surface construction should be carefully formulated.
sumption. CO2 emissions from the buildings in 2020 reached
7.660 × 104 tons, 1.19 times more than that in 2011. The increase was 5. Conclusions
mainly derived from provincial or sub-provincial level cities such as
Chengdu, Urumqi, Shenzhen, Wuhan and Zhengzhou. The model pre- A simple, flexible, and intelligent model of a building’s operational
dicts that each of these five cities will contribute around 2.0 × 103 tons energy consumption was developed to predict future energy con-
of carbon in 2020, owing to rapid development of each of their sumption and CO2 emissions. Taking into account the building’s attri-
economies. The model suggested that most of the selected cities will see butes, the regional and micro-climate impacts, and relevant socio-
their CO2 emissions increase. However, 66 cities could see emissions economic conditions, a GRNN model was constructed to estimate the
decrease if improvements in industry frame construction and other low- energy requirements for heating, cooling, and lighting in 294 office
carbon strategies, especially in cities who largely depend on coal pro- buildings in China. Indicators of second-level effects were selected to
duction and use such as Taiyuan in Shanxi Province, are employed. provide simple but realistic representation of the model’s performance
It caused GRNN only several seconds to train the data by integrating and output. We estimate the direct and indirect impact of each of the
all the four main factors including building’s attributes, socioeconomic, variables’ contribution to CO2 emissions. A building’s attributes were
climate and micro-climate conditions, which would improve the found to have the most impact on its energy consumption, followed by
model’s efficiency. Meanwhile, more concise than other model such as the socioeconomic conditions, and then the micro- and regional cli-
MLR showed GRNN’s advantage for simulating and predicting the en- mate. A version of the model that included these four main factors’
ergy usage and related CO2 emissions. The robust and concise pro- fixed effects was tested to capture variations in energy consumption and
duction originated from GRNN would give more help for CO2 emission CO2 emissions and to evaluate whether it could be used to predict fu-
reduction, and would help for low carbon city construction and citizen ture energy-related CO2 emissions. The GRNN model was shown to be
wellbeing improvement. more representative of reality than a MLR model. The scenario analysis
However, the CO2 emissions in 2020 also exerted great pressure on predicted future CO2 emissions from a range of office buildings,

Fig. 5. Predicted CO2 emission by the GRNN model.

H. Ye et al. Resources, Conservation & Recycling 129 (2018) 168–174

providing useful information for the development of low-carbon urban manufacturing building. Resour. Conserv. Recycl. 123, 249–254.
Li, X., Zhou, W., Ouyang, Z., 2013. Relationship between land surface temperature and
construction strategies. spatial pattern of greenspace: what are the effects of spatial resolution? Landsc.
Future work in this area should extend the studies carried out Urban Plann. 114, 1–8.
herein. First, more concise and precise indicators are needed to better Li, T., Shen, H., Zeng, C., Yuan, Q., Zhang, L., 2017. Point-surface fusion of station
measurements and satellite observations for mapping PM2.5 distribution in China:
reveal characteristics related to a building’s micro-climate and the re- methods and assessment. Atmos. Environ. 152, 477–489.
levant socioeconomic conditions. Second, the GRNN model could be Lin, B., Liu, H., 2015. CO2 emissions of China’s commercial and residential buildings:
used to predict CO2 emissions from other building types. evidence and reduction policy. Build. Environ. 92, 418–431.
Liu, X., Sweeney, J., 2012. Modelling the impact of urban form on household energy
demand and related CO2 emissions in the Greater Dublin Region. Energy Policy 46,
Acknowledgements 359–369.
Luo, Z., Yang, L., Liu, J., 2016. Embodied carbon emissions of office building: a case study
of China’s 78 office buildings. Build. Environ. 95, 365–371.
Thanks for the fund from the National Key Research and
March, L., 1972. Elementary models of built forms. Urban Space Struct. 55–96.
Development Program of China (2016YFC0502700), Projects NDAR, D., 2011. Guidelines for Provincial Greenhouse Gas Inventories.
41771570, 41671444, 71403258, 41371540 supported by National Peng, C., 2016. Calculation of a building’s life cycle carbon emissions based on Ecotect
Natural Science Foundation of China and Project2015J05086 supported and building information modeling. J. Clean. Prod. 112, 453–465.
Pombeiro, H., Santos, R., Carreira, P., Silva, C., Sousa, J.M.C., 2017. Comparative as-
by Natural Science Foundation of Fujian Province. sessment of low-complexity models to predict electricity consumption in an institu-
tional building: linear regression vs. fuzzy modeling vs. neural networks. Energy
Appendix A. Supplementary data Build. 146, 141–151.
Ratti, C., Baker, N., Steemers, K., 2005. Energy consumption and urban texture. Energy
Build. 37 (7), 762–776.
Supplementary data associated with this article can be found, in the Reinhart, C., Dogan, T., Jakubiec, J.A., Rakha, T., Sang, A., 2013. Umi-an Urban
online version, at Simulation Environment for Building Energy Use, Daylighting and Walkability. In:
13th Conference of International Building Performance Simulation Association.
Chambery France. pp. 476–483.
References Robati, M., Kokogiannakis, G., McCarthy, T.J., 2017. Impact of structural design solutions
on the energy and thermal performance of an Australian office building. Build.
Environ. 124, 258–282.
Abu Qdais, H., Bani Hani, K., Shatnawi, N., 2010. Modeling and optimization of biogas
Siljic, A., Antanasijevic, D., Peric-Grujic, A., Ristic, M., Pocajt, V., 2015. Artificial neural
production from a waste digester using artificial neural network and genetic algo-
network modelling of biological oxygen demand in rivers at the national level with
rithm. Resour. Conserv. Recycl. 54 (6), 359–363.
input selection based on Monte Carlo simulations. Environ. Sci. Pollut. Res. 22 (6),
Antanasijević, D.Z., Ristić, M.Đ., Perić-Grujić, A.A., Pocajt, V.V., 2013. Forecasting
human exposure to PM10 at the national level using an artificial neural network
Specht, D.F., 1991. A general regression neural network. IEEE Trans. Neural Netw. 2 (6),
approach. J. Chemom. 27 (6), 170–177.
Antanasijević, D.Z., Ristić, M.Đ., Perić-Grujić, A.A., Pocajt, V.V., 2014. Forecasting GHG
Sun, R., Chen, L., 2017. Effects of green space dynamics on urban heat islands: mitigation
emissions using an optimized artificial neural network model based on correlation
and diversification. Ecosyst. Serv. 23, 38–46.
and principal component analysis. Int. J. Greenh. Gas Control 20, 244–253.
Thomas, B.A., Azevedo, I.L., 2013a. Estimating direct and indirect rebound effects for US
Antanasijević, D., Pocajt, V., Ristić, M., Perić-Grujić, A., 2015. Modeling of energy con-
households with input–output analysis part 1: theoretical framework. Ecol. Econ. 86,
sumption and related GHG (greenhouse gas) intensity and emissions in Europe using
general regression neural networks. Energy 84, 816–824.
Thomas, B.A., Azevedo, I.L., 2013b. Estimating direct and indirect rebound effects for US
Bakshan, A., Srour, I., Chehab, G., El-Fadel, M., Karaziwan, J., 2017. Behavioral de-
households with input–output analysis part 2: simulation. Ecol. Econ. 86, 188–198.
terminants towards enhancing construction waste management: a bayesian network
Tian, W., Liu, Y., Heo, Y., Yan, D., Li, Z., An, J., Yang, S., 2016. Relative importance of
analysis. Resour. Conserv. Recycl. 117, 274–284.
factors influencing building energy in urban environment. Energy 111, 237–250.
Bendu, H., Deepak, B.B.V.L., Murugan, S., 2017. Multi-objective optimization of ethanol
Wu, G., Liu, L.-C., Han, Z.-Y., Wei, Y.-M., 2012. Climate protection and China’s energy
fuelled HCCI engine performance using hybrid GRNN–PSO. Appl. Energy 187,
security: win–win or tradeoff. Appl. Energy 97, 157–163.
Ye, H., Wang, K., Zhao, X., Chen, F., Li, X., Pan, L., 2011. Relationship between con-
Benedetti, M., Cesarotti, V., Introna, V., Serranti, J., 2016. Energy consumption control
struction characteristics and carbon emissions from urban household operational
automation using Artificial Neural Networks and adaptive algorithms: proposal of a
energy usage. Energy Build. 43 (1), 147–152.
new methodology and case study. Appl. Energy 165, 60–71.
Ye, H., Qiu, Q., Zhang, G., Lin, T., Li, X., 2013. Effects of natural environment on urban
Chen, H., Lee, W.L., Wang, X., 2015. Energy assessment of office buildings in China using
household energy usage carbon emissions. Energy Build. 65, 113–118.
China building energy codes and LEED 2.2. Energy Build. 86, 514–524.
Ye, H., He, X., Song, Y., Li, X., Zhang, G., Lin, T., Xiao, L., 2015. A sustainable urban form:
China, N.B.o.S.o.t.P.s.R.o., 2009. China Statistical Yearbook. China Statistical Publishing
the challenges of compactness from the viewpoint of energy consumption and carbon
emission. Energy Build. 93, 90–98.
China, N.B.o.S.o.t.P.s.R.o., 2015. China Statistical Yearbook. China Statistical Publishing
Ye, H., Hu, X., Ren, Q., Lin, T., Li, X., Zhang, G., Shi, L., 2017a. Effect of urban micro-
climatic regulation ability on public building energy usage carbon emission. Energy
Clarke, J.A., 2001. Energy Simulation in Building Design. Routledge.
Construction, M.o., 2007. Opinions on Strengthening the Implementation of Energy-
Ye, H., Ren, Q., Hu, X., Lin, T., Xu, L., Li, X., Zhang, G., Shi, L., Pan, B., 2017b. Low-carbon
saving Management of State-owned Office Buildings and Large-scale Public
behavior approaches for reducing direct carbon emissions: household energy use in a
coastal city. J. Clean. Prod. 141, 128–136.
Csábrági, A., Molnár, S., Tanos, P., Kovács, J., 2017. Application of artificial neural
Yu, X.-H., 1992. Can backpropagation error surface not have local minima. IEEE Trans.
networks to the forecasting of dissolved oxygen content in the Hungarian section of
Neural Netw. 3 (6), 1019–1021.
the river Danube. Ecol. Eng. 100, 63–72.
Yuan, L., Ruan, Y., Yang, G., Feng, F., Li, Z., 2016. Analysis of factors influencing the
Gao, L., Wang, S., Li, J., Li, H., 2017. Application of the extended theory of planned
energy consumption of government office buildings in qingdao. Energy Proc. 104,
behavior to understand individual’s energy saving behavior in workplaces. Resour.
Conserv. Recycl. 127, 107–113.
Yuan, X.-C., Sun, X., Zhao, W., Mi, Z., Wang, B., Wei, Y.-M., 2017. Forecasting China’s
Hirst, N., 2013. Build. Clim. Change 23–30.
regional energy demand by 2030: a bayesian approach. resources. Conserv. Recycl.
IPCC, 2007. Intergovernmental Panel on Climate Change.
127, 85–95.
Jaffal, I., Inard, C., 2017. A metamodel for building energy performance. Energy Build.
Zhou, W., Huang, G., Cadenasso, M.L., 2011. Does spatial configuration matter?
151, 501–510.
Understanding the effects of land cover pattern on land surface temperature in urban
Jing, R., Wang, M., Zhang, R., Li, N., Zhao, Y., 2017. A study on energy performance of 30
landscapes. Landsc. Urban Plann. 102 (1), 54–63.
commercial office buildings in Hong Kong. Energy Build. 144, 117–128.
Zhou, X., Yan, D., Hong, T., Ren, X., 2015. Data analysis and stochastic modeling of
Khayatian, F., Sarto, L., Dall’O’, G., 2016. Application of neural networks for evaluating
lighting energy use in large office buildings in China. Energy Build. 86, 275–287.
energy performance certificates of residential buildings. Energy Build. 125, 45–54.
Kim, J.J., 2017. Economic analysis on energy saving technologies for complex