Applied Energy xxx (2012) xxx–xxx

Contents lists available at SciVerse ScienceDirect

Applied Energy
journal homepage: www.elsevier.com/locate/apenergy

Estimation of electrical power consumption in subway station design

by intelligent approach
Philip C.M. Leung a,b, Eric W.M. Lee a,⇑
a
Department of Civil and Architectural Engineering, City University of Hong Kong, Tat Chee Avenue, Kowloon Tong, Hong Kong, China
b
MTR Corporation, MTR Headquarters Building, Telford Plaza, 33 Wai Yip Street, Kowloon Bay, Hong Kong, China

h i g h l i g h t s

" We estimate the power consumption of subway stations by multilayer perceptron.

" Data collected from Observatory and 19 subway stations are used for model training.
" Number of hidden neurons of the model is investigated by sensitivity test.
" Correlation coefficient is used to describe the prediction performance.
" The coefficient of correlation, with 95% confidence level, can reach 0.96 or higher.

a r t i c l e i n f o a b s t r a c t

Article history: According to the records of Hong Kong rail operator, MTR Corporation, the weekly electrical consumption
Received 23 April 2012 of each railway station ranges from 18 MWh to 230 MWh. Since the electrical consumption of stations is
Received in revised form 27 June 2012 a major factor in the planning of infrastructure, a good prediction of the electrical consumption will
Accepted 2 July 2012
greatly assist in the design of the station infrastructure. This study develops an intelligent approach to
Available online xxxx
predict the energy consumption of railway stations. Multi-layered Perceptron (MLP) is adopted to mimic
the non-linear correlation between energy consumption, the spatial design of the station, meteorological
Keywords:
factors and also the usage of the 19 stations selected. Coefficient of correlation is obtained between the
Artificial neural network
Energy consumption
MLP predicted results and the actual collected data to evaluate the performance of the prediction. We
Subway station apply statistical approach to assess the performance of the developed MLP model. It shows that minimum
coefficient of correlation is 0.96 with a 95% confidence level which is considered sufficient for engineering
application. This approach is also adopted to predict the profile of the weekly electrical consumption of a
selected station. The predicted profile reasonably agrees with that of the actual consumption. This study
develops a useful tool to estimate the electrical power consumption of new MTR stations.
Ó 2012 Elsevier Ltd. All rights reserved.

1. Introduction
Subway system provides an economic and convenient transpor-
tation to the public for travelling within their cities. Especially for
metropolitan cities, majority peoples are relying on the subway
system in their daily lives. The extensive use of subway system ini-
tiates the engineers and scientists’ research on the energy con-
sumptions of the subway stations. Ampofo et al. [1] studied the
braking system of the trains and provided evidence that the ambi-
ent condition of the subway station is significantly adverse by the
heat created from the braking mechanisms of the trains. For the
facilities of a subway station, a prediction model was developed
by Ma et al. [2] to estimate the energy consumption of the escala-
⇑ Corresponding author. Tel.: +852 2194 2307; fax: +852 2788 7612.
E-mail address: ericlee@cityu.edu.hk (E.W.M. Lee).
tors. Energy audit was conducted by Fu and Deng [3] on the
Guangzhou railway station. Based on the results, they successfully
identified the potential energy conservation measures on the air-
conditioning system, power equipment system and lighting system
of the station. These pioneer works show that the prediction of en-
ergy consumption of a subway station is non-linear in nature. They
applied different mathematical models to simulate the energy con-
sumption. This paper presents the information of the energy con-
sumptions of the subway stations in Hong Kong. It also proposes
to adopt the artificial neural network model to mimic the non-lin-
ear behavior of the energy system of the subway stations in Hong
Kong by using the historical data of the stations.
Carrying more than 4.4 million passengers every weekday, the
MTR is regarded as one of the world’s leading railways for safety,
reliability, customer service and cost efficiency. The MTR Corporation
was established in 1975 as the Mass Transit Railway Corporation

0306-2619/$ - see front matter Ó 2012 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.apenergy.2012.07.017

Please cite this article in press as: Leung PCM, Lee EWM. Estimation of electrical power consumption in subway station design by intelligent approach.
Appl Energy (2012), http://dx.doi.org/10.1016/j.apenergy.2012.07.017
2 P.C.M. Leung, E.W.M. Lee / Applied Energy xxx (2012) xxx–xxx
with a mission to construct and operate, under prudent commer-
cial principles, an urban metro system to help meet Hong Kong’s
public transport requirements.
Apart from railway operations in Hong Kong, the MTR Corpora-
tion is also involved in a wide range of businesses, including prop-
erty development and management, international rail operations
joint ventures and consultancy services, advertising, and telecom-
munication services. The new urban hub with comprehensive res-
idential and commercial developments above Kowloon Station is
developed by the MTR Corporation. Some of the MTR’s property
developments have become landmarks of Hong Kong, such as the
shopping mall, Elements, and the commercial complexes, Interna-
tional Commerce Center and the International Financial Center,
Hong Kong’s two tallest buildings. With MTR’s proficiency in rail-
way operation, the Corporation now operates outside Hong Kong
including Beijing Metro Line 4, Shenzhen Metro Line 4, and Stock-
holm Metro in Sweden. MTR Corporation has also formed joint
venture companies in the United Kingdom and Australia to operate
the London Overground and the Melbourne Metro rail networks
respectively.
MTR currently operates 10 heavy rail commuter lines and 11
routes on a light rail network. The MTR network spreads over the
urban areas, and links the heart of Hong Kong with Hong Kong
International Airport, Hong Kong Disneyland Theme Park, and
the boundary with the Mainland of China.
Furthermore, MTR Corporation is extending its network with
five new railway lines in Hong Kong. These include the Guangz-
hou-Shenzhen-Hong Kong Express Rail Link, the West Island Line,
South Island Line (East), Shatin to Central Link and Kwun Tong Line
Extension.
Existing MTR stations have on-going extension and improve-
ment plans. New pedestrian subways and entrances are introduced
to connect nearby developments, and to provide more convenient
access for customers. Station assets are refurbished as needed and
sustainable technologies are applied to improve customer services
in a socially responsible manner.
Up to August 2010, MTR has 152 heavy rail stations and light
rail stops. And thus, a tool to estimate energy consumption is vital
to the development of a sustainable management system, as it can
be used to minimize energy waste, and that will have significant
impact on the performance of station buildings.
2. Pioneer works on energy consumption analysis and
prediction
Following the oil embargo of 1973, both the political and scien-
tific communities began to pay more attention to opportunities to
improve energy efficiency. Recently, Basbous et al. [4] developed a
novel combustion engine to improve fuel consumption for electric-
ity production. Kaldellis et al. [5] developed a numerical algorithm
to calculate the maximum wind power contributing to an electrical
generation system. For understanding the contributions of domes-
tic components to the increase in domestic energy consumption in
UK, Kilpatrick et al. [6] developed a methodology to filter the
domestic energy consumption profile into standby profile, cold
profile, heating element profile and residual profile. Possible en-
ergy saving opportunities can then be identified from the profiles.
Work on energy consumption forecast has become more and more
popular, as energy demand has grown dramatically when com-
pared with past decades. According to Kadoshin et al. [7], this
effect is related to two main factors, namely increase in population
and economic development. Besides, more stakeholders have
aligned their objectives with green values, and become more
conscious on sustainable energy use. Forecasting energy consump-
tion would allow a better understanding of the actual demand in
Please cite this article in press as: Leung PCM, Lee EWM. Estimation of electrical power consumption in subway station design by intelligent approach.
Appl Energy (2012), http://dx.doi.org/10.1016/j.apenergy.2012.07.017
the future. Based on the pioneer research works on energy con-
sumption forecast [8], the majority of the building energy predic-
tions are done for town planning as well as renewable energy
development purposes, such that appropriate measures can be
implemented to utilize energy use, and to ensure the ability to
backup with an adequate energy supply reservoir.
Development of forecasting model for energy consumption has
a long history. In 1978, Uri [9] forecasted energy consumption by
combining the econometric model and time-series forecasting
model to predict the peak load of a specific utility based on eco-
nomic and weather-related parameters. In 1986, Deeble and Prob-
ert [10] had developed a linear correlation to forecast annual
energy consumption. In the following year, Bodger and Tay [11]
had used growth factor based on past energy consumption to pre-
dict New Zealand’s electricity consumption. In recent years, differ-
ent types of model had been adopted to predict energy
consumption, namely autoregressive integrated moving average
(ARIMA) model [12], econometric model [13], artificial neural net-
work (ANN) model [14] etc. Computer modeling and simulation
software like TRNSYS [15] was also used to assist in the energy
consumption analysis in Cyprus. Tang et al. [16] developed an
ensemble empirical model decomposition based least square sup-
port vector machine learning paradigm for nuclear energy con-
sumption forecasting. Gutierrez et al. [17] developed a stochastic
diffusion process to model the electricity consumption in Morocco.
The factors of gross domestic product, final domestic consumption
and gross fixed capital formation were taken as inputs of the
model.
3. Intelligent approach
According to Zhang et al. [18], ANN can be used to predict rela-
tionships that are unknown or complex, as it is capable of captur-
ing fine details of the function relationship via a ‘learning’ process,
also known as ‘training’ of the network function in ANN, from his-
torical data. More importantly, ANN is a popular approach to pre-
dict non-linear functions in the past decade. Among different
models of ANN, Multi-layered Perceptron (MLP) [19] is one of the
most widely used ANN model for forecasting due to its simple
and flexible nature [8,20]. It is a supervised learning model with
feed forward connectivity. Also, MLP has been mathematically pro-
ven to be a universal function approximator [21,22]. The MLP is
proposed to be adopted to predict the energy consumption of sub-
way stations in this study.
The MLP has been successfully applied to different areas includ-
ing weather, market trends, electrical and thermal load predictions
in (something missing here?) [23]. It was also applied to predict
bankruptcy, stock prices, international airline passenger traffic,
rainfall, transportation, water demand and wind pressure profile
[18]. A lot of studies have been conducted to compare the forecast-
ing performance of traditional statistical model against MLP under
different conditions, and many of them have proven that MLP out-
performed traditional statistical models, especially forecasts for
monthly [24], quarterly [25] or even daily [18] time series data.
Countries like Saudi Arabia and Singapore have already used MLP
to predict their energy consumption as well [8].
Yalcintas and Akkurt [26] employed MLP to mimic the total
chiller plant power of a 42-storey commercial building in down-
town Honolulu, Hawaii. Their independent input variables mainly
consisted of climate data, and the model output was chiller plant
power consumption. The input parameters used in the MLP model
included dry bulb temperature, wet bulb temperature, dew point
temperature, relative humidity percentage, wind speed, and wind
direction. The hour of the day was recorded to account for
variations in occupancy throughout the day. However, data on
 P.C.M. Leung, E.W.M. Lee / Applied Energy xxx (2012) xxx–xxx 3

hourly power consumption in the chiller plant were not available
for every hour of the day. The total number of matching data items
was only 121 out of 312 for 13 of the days in the study period.
Kreider and Rabl [27] used an ANN model to predict the pre-ret-
rofit energy consumption of a building and compare it with the
measured energy consumption of the retrofitted building. The in-
put layer consisted of eight different types of input data including
weather factors, ambient dry-bulb temperature, humidity ratio,
horizontal insulation, and wind speed; and occupancy-related fac-
tors, including hour number (00:00–23:00), a weekday/weekend
binary flag (i.e. 0 and 1), chilled water consumption in the past
hour, and chilled water consumption in the previous hour.
Yezioro et al. [28] introduced an ANN model that used the
Levenberg–Marquardt training algorithm to predict heating/cool-
ing load consumption. The input parameters for the ANN model in-
cluded hourly weather data such as the outdoor temperature,
relative humidity, and set-point temperatures. The occupancy
schedule, a key parameter contributing to energy consumption,
was used as one of the model’s input parameters.
In another study, Yalcintas [29] developed an ANN model to
predict the energy savings delivered by building equipment retro-
fits. The Levenberg–Marquardt back-propagation algorithm was
used and the input layer included the weather variables and the
hour of the day. The output was the hourly electricity measure-
ments from the retrofitted equipment.
Kankal et al. [30] applied ANN to predict the energy consump-
tion of Turkey. The input parameters of the ANN model include
basic energy and demographic and socio-economic indicators
(e.g. population, GDP, import amount, export amount and
employment).
The above recent researches on application of ANN to the pre-
diction of cooling load have a common deficiency that they did
not account for the actual number of occupants inside the building.
In fact, it is extremely difficult to collect the dynamic change of the
population inside a building since most building managers of com-
mercial buildings do not intend to publicize the actual occupancy
of their buildings since it is a key indicator of their business. How-
ever, MTR is a public utility. The total number of people using the
MTR is not as sensitive a data. MTR also has a well established
computer system to record the number of operations of the turn-
stiles of the stations. It can provide the dynamic pattern of the pop-
ulation of a station for this study.
4. Data collection
The current MTR system consists of 10 lines within the bound-
aries of Hong Kong. Among the 10 lines, Kwun Tong Line, Tsuen
Wan Line and Island Line are the 3 oldest lines of the system which
were opened in 1979, 1982 and 1986 respectively. The data re-
cords of the stations in the 3 lines are extensive. Therefore, we pro-
pose to use the stations in these 3 lines for this study. Fig. 1 shows
the present MTR system.
Two kinds of stations of the 3 lines are excluded from this study.
The first kind of stations is the interchange stations which serves
two lines. Since passengers transferring line are not required to
pass through any turnstile, the total number of passengers using
the platform cannot be reflected by the turnstiles of the station.
The actual cooling load contribution from them cannot be
estimated by the record of the turnstiles. The second kind of the
stations to be excluded is those with platforms situated above-
ground and open to atmosphere. Air-conditioning is not provided
to these stations and the energy consumption is comparatively
much lower than the other stations. The summary of the 19
stations to be studied are shown in Table 1. Data is collected from
these stations for our ANN model training and performance

Fig. 1. Map of MTR system. The blue line in Hong Kong Island is the Island Line. The red and green lines in Kowloon are respectively the Tsuen Wan Line and the Kwun Tong
Line. This figure is extracted from http://www.mtr.com.hk. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this
article.)

Please cite this article in press as: Leung PCM, Lee EWM. Estimation of electrical power consumption in subway station design by intelligent approach.
Appl Energy (2012), http://dx.doi.org/10.1016/j.apenergy.2012.07.017
4 P.C.M. Leung, E.W.M. Lee / Applied Energy xxx (2012) xxx–xxx
Table 1
Data is collected from the 19 subway stations from the 3 oldest lines of the MTR
system. They are the stations not designed for line-transfer and equipped with air-
conditioning.
Line Subway stations
Island (8 Sheung Wan, Wan Chai, Causeway Bay, Tin Hau, Fortress
stations) Hill, Tai Koo, Sai Wan Ho, Shau Kei Wan
Tsuen Wan (5 Tsim Sha Tsui, Sham Shui Po, Chueng Sha Wan, Lai Chi
stations) Kok, Tai Wo Hau
Kwun Tong (6 Sek Kip Mei, Lok Fu, Wong Tai Sing, Diamond Hill, Choi
stations) Hung, Lam Tin
evaluation. Exact parameters of the stations to be adopted for this
study refers to the input and output parameters of the MLP model
described in Section 5.
5. Development of the MLP model
The MLP architectural model consists of a few layers in general
which are named input layer, hidden layer(s), and output layer.
The neurons of the layers are interconnected with the neurons at
its adjacent layers. There may be more than one hidden layer in
a model. However, it has been proven that a MLP with single hid-
den layer is a universal function approximator [21] subject to the
provision of sufficient number of hidden neurons. So, a 3-layered
(i.e. input layer, hidden layer and output layer) MLP model was
adopted in this study.
5.1. Input layer
The number of neurons in the input layer corresponds to the
number of input parameters of the model. In this study, the follow-
ing input parameters are selected to predict the energy consump-
tion of the subway station.
5.1.1. Meteorological parameters
Similar to other buildings, weather is one of the major contrib-
utive factors to the cooling load of the HVAC system of the subway
stations. The widely adopted meteorological factors in intelligent
approaches [14,26,31] are (i) outdoor dry bulb temperature and
(ii) relative humidity. The fresh air taken from the exterior of the
station by the air-conditioning system and infiltrated to the station
by the ingress of commuters are the factors directly increasing the
cooling load of the subway station since the air-conditioning sys-
tem is required to remove the excessive moisture and reduce the
air temperature to the design condition. Therefore, the tempera-
ture and the relative humidity of the outdoor air have significant
impact to the overall energy consumption of the subway station.
The daily dry bulb temperatures and daily relative humidity from
1-January-2008 to 31-December-2008 were obtained from records
of The Hong Kong Observatory. The weekly dry bulb temperature is
taken to be the average of the mean daily dry bulb temperature
over a week. Similarly, the mean weekly relative humidity is taken
to be the average of the mean daily relative humidity over a week.
5.1.2. Concourse area
Passengers enter and exit MTR stations through the concourse
area. The concourse levels of the stations are usually separated
from the platform levels, and from the concourse, passengers can
travel to different platform levels by escalators. Customer service
centers, ticketing machines and ticketing gates are also located at
the concourse. It is designed according to the population of the dis-
trict it is situated in. Therefore, the concourse area represents the
capacity of the subway stations and also the cooling load to the
centralized air-conditioning system. The concourse area of differ-
Please cite this article in press as: Leung PCM, Lee EWM. Estimation of electrical power consumption in subway station design by intelligent approach.
Appl Energy (2012), http://dx.doi.org/10.1016/j.apenergy.2012.07.017
ent subway stations ranges from 2663 m2 to 12,632 m2. They are
measured from the architectural drawings of the stations. Since
the air-conditioning and lighting are provided for all concourse
areas of this study, it is a contributive factor to the energy con-
sumption of the subway station.
5.1.3. Platform area
The platform is the area in which commuters are waiting, get-
ting into and out of the train. This platform area refers the total
platform areas of the subway station. The total platform area of dif-
ferent subway stations ranges from 1175 m2 to 2257 m2. They are
measured from the architectural drawings of the stations. The
main energy consumption of the platform area is the artificial
lighting and the air conditioning since they are provided for the en-
tire platform areas during the operation of the station. The rail
tracks are separated from the platform area by a row of continuous
full height platform screen doors. Therefore, the heat transfer be-
tween the rail tracks and the platform area is restricted and ig-
nored in this study.
5.1.4. Kiosks area
To serve passengers’ daily needs, the MTR offers the conve-
nience of a wide variety of retail outlets in all stations. The many
choices available include food and beverage as well as health and
beauty shops including bakeries, convenience stores and shops
selling fashions, cosmetics, accessories, gifts, herbal products and
health drinks, confectionery and lifestyle goods. Some concourses
are also home to dry cleaners, shoe repairers, bookshops, clinics,
travel agents, banks and financial service providers. The air-condi-
tioning systems for shops in the concourse area are separated from
the centralized air-conditioning system of the station and the de-
sign of air conditioning for the shops is also different from that
for the concourse. Different equipment are usually installed in
the shops (e.g. refrigerators, hot water urns, spot lights, cashier
machines, etc.) which are not provided in the concourse. Therefore,
the shop area represents a different electrical loading to the sta-
tion. According to MTR record, the total kiosk area of a station is
proportional to the total floor area of the station. That is, large
kiosk areas generally consume more electrical power. Therefore,
the total kiosk area of a station also serves as an input parameter
for predicting the overall power consumption. The areas are mea-
sured from the architectural drawings of the stations.
5.1.5. Plant room area
Plant rooms house mechanical and electrical systems which are
required to support the operation of the station. Large plant rooms
generally contain more equipment than small plant rooms. The
electrical power consumption of the large plant room is more than
that of the small plant room. Also, mechanical ventilation is pro-
vided to all plant rooms of a station to remove the heat generated
by the equipment and to provide fresh air to the technicians work-
ing inside the plant room. Since the ventilation rate is designed to
be proportional to the volume of the plant room, large plant rooms
require more electrical energy for its mechanical ventilation. The
total plant room areas of the stations range from 967 m2 to
11,338 m2, which are also selected as an input parameters of the
MLP model. The areas measured from the architectural drawings
of the stations.
5.1.6. Staff accommodation area
The staff accommodation area including office, pantry, control
room, washroom, etc. is the area earmarked for the activities of
staff working in the station. The requirements on lighting lumi-
nousness and air conditioning are different from other areas. The
staff accommodation area of a station is proportional to the num-
ber of staff working in the station. It represents the cooling load
 P.C.M. Leung, E.W.M. Lee / Applied Energy xxx (2012) xxx–xxx
from the staff and also the electrical energy consumption of their
ancillary activities. Therefore, this factor is also taken to be the in-
put parameters of the MLP model. The areas are measured from the
architectural drawings of the stations.
5.1.7. Weekly number of passengers using the station
The number of passengers using the stations is a major factor to
energy consumption since they contribute to the station the inter-
nal heat gain (i.e. sensible load and latent load) and other ancillary
facilities (e.g. elevators, escalators, ticketing machine, etc.). Record-
ing the movements of passengers entering and leaving on video
can accurately record the number of passengers using the station
but it may violate the Personal Data (Privacy) Ordinance in Hong
Kong. Alternatively, the operation of the turnstiles inside the sta-
tion is adopted. Each turnstile can accurately record the number
of passengers passing through it. It is assumed each passenger
passes through the turnstile each time he/she enters or leaves a
station. The weekly total number passengers using the stations
from 1 January 2008 to 31 December 2010 ranged from 178,250
to 1,950,450, with the numbers obtained through summing up
the number of turnstile operations over the week. The data is col-
lected from the recorded turnstile motions stored in the central-
ized data collection system.
5.2. Hidden layer
The number of neurons of the input and output layers respec-
tively equal to the number of input parameters and output param-
eters of the samples. It is defined by the system itself. The required
number of hidden neurons is crucial to the performance of the
model. Currently, there is no analytical approach to determine
the number of hidden neurons but different rules-of-thumb are
available to provide heuristic hints to the user. The rule-of-thumb
adopted in this study is the rule developed by Ward [32] as de-
scribed in Eq. (1) where Nh, Nin and Nout are respectively the num-
ber of neurons of hidden layer, input layer and output layer. The
value of Ns is the number of training samples.
Nin þ Nout pffiffiffiffiffiffi
Nh ¼ þ Ns ð1Þ
2 minimum validation error is taken to be the trained model. Fig. 3
Prior to the adoption of the rule-of-thumb, the sensitivity of the
number of hidden neurons to the performance of the model was
studied. That is, the change in the number of hidden neurons does
not significantly affect the model performance which is chosen to
be the coefficient of correlation between the actual outputs and
the outputs predicted by the model obtained from the model train-
ing described in the next section. The number of hidden neurons to
be examined is chosen to be Nh ± 5. For each hidden neuron, 100
trials are performed to determine the model performance and
the 95% confidence limits are obtained. If the 95% confidence inter-
vals of the models with different numbers of hidden neurons over-
lap with each other, from statistical point of view, it concludes that
the performances of the models with different numbers of hidden
neurons are comparable to each other. That is, the number of hid-
den neurons is insensitive to the model performance and the rule-
of-thumb can be adopted.
5.3. Output layer
The output of the MLP model is the weekly total electrical con-
sumption of the stations. The data is collected from the central sta-
tion management system of the subway stations. The summary of
the input and output of the MLP model is list in Table 2. The archi-
tecture of the MLP model is illustrated in Fig. 2. There are a total of
1007 samples collected from the 18 stations for model training.
Please cite this article in press as: Leung PCM, Lee EWM. Estimation of electrical power consumption in subway station design by intelligent approach.
Appl Energy (2012), http://dx.doi.org/10.1016/j.apenergy.2012.07.017
5
Table 2
Summary of input and output parameters of the MLP model.
Parameter Parameter Unit
Input Weekly average of outdoor temperature (dry-bulb) °C
Weekly average of relative humidity %
Area of concourse m2
Area of platform m2
Shops area m2
Plant room area m2
Staff accommodation area m2
Weekly number of passengers –
Output Weekly total electrical consumption MWh
The data of Cheung Sha Wan station is reserved for evaluating
the model performance in Section 9.
6. Model training
Back propagation [33] (BP) is a traditional training algorithm
used for the MLP model. It feeds back the prediction errors from
the output layer to the input layer and the weights of the links be-
tween the neurons are adjusted according to the BP algorithm.
Upon completion of the weight adjustments, a new prediction is
carried out to evaluate a new prediction error for the next epoch
of weight adjustments. These procedures are repeated numerous
times until a satisfactory prediction result is achieved. In this
study, the early-stop validation approach is adopted to monitor
and stop the BP training. Fig. 3 illustrates the concept of the
early-stop training approach.
In order to prevent ‘‘overfitted training’’, the intermediate-state
of the trained model in every training epoch is applied to the val-
idation data set to evaluate the prediction error (i.e. the validation
error). The network training process is stopped when the valida-
tion error reaches the minimum value. Since we have no prior
knowledge in the trend of the validation error, ‘‘early-stop’’ train-
ing is adopted. It records the status of the model continuously in
the course of training. When there is no reduction in the validation
error over a pre-defined number of epochs (in the paper, the No. of
epochs has been selected to be 2000), the model state with the
illustrates the process of this early-stop training.
7. Performance evaluation of trained MLP
Upon completion of the network training, the trained MLP is ap-
plied to the testing set to evaluate the performance indices by
comparing the target values of the testing set and the values pre-
dicted by the trained model. The performance indices used in this
study is the coefficient-of-correlation (r), as defined respectively in
Eq. (2) where N is the total number of samples, and ft i ; pi gNi¼1 are,
respectively, the target values and the predicted values. p  and t
are the mean value of the predicted values and the mean value
of the target values, respectively.
PN
 
i¼1 ðt i  tÞðpi  pÞ
r ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
PN PN ð2Þ
2
i¼1 ðt i  tÞ

i¼1 ðpi  pÞ
 2
It should be noted that a random process is normally involved
in the network training of an MLP model, especially when the
available samples are divided into training and validation sets. It
is thus not impossible for the random process used to result in ‘for-
tuitous’ samples that show the evaluated performance indices are
good. Instead of reporting only the best simulation result, a less-
prejudice statistical approach is adopted to minimize the effect
of randomization. We carried out the network training and perfor-
6 P.C.M. Leung, E.W.M. Lee / Applied Energy xxx (2012) xxx–xxx

Weekly Average of
Outdoor Temperature
Weekly Average of
Relative Humidity
Area of Concourse

Area of Platform
Weekly Total
Shops Area Electrical
Consumption
Plant Room Area

Staff Accommodation
Area
Weekly No. of
Passengers

Input Layer Hidden Layer Output Layer

Fig. 2. Architecture of the MLP model to be adopted in this study.

Prediction
Error

Network training stops

Training here, since there is no
error further improvement in
the validation error

Validation
error

Minimum
validation
error

Model state with minimum Training

Pre-defined number of Epoch
validation error is selected as the
epochs
trained model
Fig. 3. The early-stop validation approach stops the back propagation training when there is no further improvement in the validation error over a pre-defined number of
epochs after it has reached its minimum level. The intermediate state of the model with the minimum validation error is selected as the trained model.

mance evaluation process 1000 times. The 1000 results are statis-
tically analyzed by evaluating the mean of the results and the lim-
its of the 95% confidence intervals. This approach statistically
reveals the performance of the ANN model when this less-preju-
diced approach is taken, as illustrated in Fig. 4.
In the model training approach described above, the training,
validation and testing sample sets are obtained by random sam-
pling. The performance of the trained model might not be promis-
ing at all times, as it very much depends on the quality of the
training samples. It is possible that the sample batch favours the
network training, and thus produces a better model. In order to
verify the performance of the trained model, it is necessary to con-
duct an evaluation on the performance of the trained model by sta-
tistical approach.
Since the sample extraction process is a random process, we
carried out 1000 trials of model training and testing. In each trial,
the training, validation and testing samples are randomly grouped.
The model is trained and tested by the samples and the coefficient
of correlation of that trial is obtained. After the 1000 trials, a set of
1000 coefficients of correlation are obtained. Given that the coeffi-
cients of correlation are well bounded between 0 and 1, a beta dis-
tribution is used to represent the probability distribution of the
100 coefficients of correlation.
The limit of the one-sided 95% confidence level counting from
the right-most (i.e. x95) represents the minimum coefficient of cor-
relation with 95% confidence level. This statistical approach allevi-
ates the effect of the randomness in the presentation of the model
performance.
8. Results and discussion
The environmental control system in the public areas of MTR
stations covers 3 main areas, namely concourses, platforms and
kiosks to which artificial lighting and air-conditioning are pro-

Please cite this article in press as: Leung PCM, Lee EWM. Estimation of electrical power consumption in subway station design by intelligent approach.
Appl Energy (2012), http://dx.doi.org/10.1016/j.apenergy.2012.07.017
 P.C.M. Leung, E.W.M. Lee / Applied Energy xxx (2012) xxx–xxx 7

Network design

Random data N Y
Repeated 1000 times?
extraction

Network training
by
backpropagation
Apply trained Evaluate the mean
model to the and 95%
Save trained testing samples to confidence
network evaluate the r interval based on
the 10000 values
of r

N No reduction on Y
validation error over
2,000 epochs?

Fig. 4. This statistical training scheme alleviates the effect due to the random data extraction process.

vided. Artificial lighting is provided during the operating hours of
the station regardless of the number of commuters. Therefore en-
ergy consumption on the artificial lighting system is comparatively
constant over the year. Also the size of the artificial lighting system
is proportional to the floor areas of the concourse, platform, kiosks
and plant rooms of the stations. The sizes of the other ancillary sys-
tems (e.g. ticketing machines, escalators, elevators, etc.) of a sta-
tion are also proportional to the areas of the stations. Therefore,
their electrical consumptions of the ancillary systems are also pro-
portional to the size of the station (i.e. the floor area of the station).
The air-conditioning system is also a major consumer of elec-
tricity which adaptively varies according to the outdoor condition
(i.e. weather). In general, the electrical consumption of the air-con-
ditioning system is high in summer and low in winter. The temper-
ature control in typical underground MTR stations are regulated by
thermostats located at the occupancy zone. For typical station
arrangement, a centralized air-conditioning system is provided.
Air-handling-units (AHU) are situated at the specific ends of the
station. The thermostats would sense the average temperature of
the air returning to the corresponding AHU at the designated area.
MTR targets to maintain the concourse temperature below 29 °C.
adult is 36.8 °C ± 0.4 °C [34] which is higher than the conditioned
air inside the station (i.e. 27 °C  29 °C), people contribute signifi-
cant sensible heat gain to station spaces. The commuters also con-
tribute latent heat gain to the spaces by the moist air released from
their breathing which increases the moisture of station spaces.
Since the air conditioning in station spaces have humidity control
at around 70% relative-humidity (RH), the extra moisture coming
from the breathing of passengers is required to be removed by
applying extra air cooling for the condensation of the moisture
and extra reheating to raise the temperature of the cooled air back
to the design temperature.
The energy consumption profile of a typical station over a year
is shown in Fig. 5. It shows a large fluctuation in the energy
consumption of the station throughout the year. The energy con-
sumption is summer time (i.e. July/August) is more than 3 times
higher than that in winter (i.e. December/January). We expect that
the meteorological factors significantly contribute to the electrical
energy consumption of the station.
120
Total Energy Consumption (MW.hr)

Hence, the thermostats are preset accordingly. If the average re-

turn air temperature is detected to be 29 °C or below in a specific 100

area, the local control valve of the corresponding AHU for chilled
water supply is automatically reduced. The cooling-load of the 80

chiller operation is also decreased accordingly. As the temperature

in the designated zone gradually rises back to 29 °C by the heat 60

gains to the station, the chilled water control valve opens accord-
ingly and in turn cools down the zone. The temperature control 40

system of the platform area operates in a similar fashion to the

concourse but the target temperature of the platform is set to be 20

27 °C. Either station or designated kiosk chiller are used for supply-
ing chilled water to most of the kiosks in underground stations. 0
02/Feb 15/Mar 26/Apr 07/Jun 19/Jul 30/Aug 11/Oct 22/Nov
Individual fan coil units with their own standalone temperature
Date in Year 2008
controllers are normally used.
The number of commuters in stations is one of the major factors Fig. 5. Typical electrical energy consumption profile of an underground MTR
to the heat load burden. Since the average oral temperature of an station in 2008.

Please cite this article in press as: Leung PCM, Lee EWM. Estimation of electrical power consumption in subway station design by intelligent approach.
Appl Energy (2012), http://dx.doi.org/10.1016/j.apenergy.2012.07.017
8 P.C.M. Leung, E.W.M. Lee / Applied Energy xxx (2012) xxx–xxx

The weekly average temperature (dry-bulb) and humidity of the 250

outdoor area in 2008 were collected from Hong Kong Observatory.

PREDICTED WEEKLY CONSUMPTION (MWH)

The other parameters as listed in Table 2 were obtained from the
weekly MTR records for 2008 retrieved from the central station 200
management system. A total of 1007 samples were obtained from
the records of the 19 stations as listed in Table 1. For MLP model
training, the samples were divided up with 70% being training
samples, 20% being validation samples and 10% being testing 150
r=0.98
samples.
The number of hidden neurons is determined by the rule-of-
thumb in Eq. (1) in which the number of training samples (Ns), is 100
taken to be 70% of the available samples (i.e. 0.7  1007 = 705).
Referring to Table 2, the number of input (Nin) and output param-
eters (Nout), are 6 and 1 respectively. By applying Eq. (1), the num-
50
ber of hidden neurons is proposed to be 31.
A sensitivity study on the number of hidden neurons to the per-
formance of the model was carried out. The number of hidden neu-
rons to be investigated ranged from 26 to 36 (i.e. 31 ± 5). For each 0
0 50 100 150 200 250
hidden neuron, 100 trials of model training as described in Sec-
tion 6 and performance evaluation as described in Section 7 were ACTUAL WEEKLY CONSUMPTION (MWH)
carried out to obtain the 95% confidence intervals of 100 trials
Fig. 7. Typical performance of the MLP models.
for each hidden neuron. The results are shown in Fig. 6.
It shows that the 95% confidence intervals overlap with each
other. It indicates that the performances of the models with differ-
ent numbers of hidden neurons are comparable to each other. That
is, the number of hidden neuron within the range is insensitive to 60
the model performance. Therefore, the number of hidden neurons 150
determined by the rule-of-thumb (i.e. 31) is justified to be adopted. 50

Probability Density
The 1007 available samples are randomly grouped into training
40
Frequency

dataset (705 samples), validation dataset (201 samples) and test-

100
ing dataset (101 samples) with sample size ratio to be 70%, 20%
30
and 10% respectively. Back propagation is applied to train the
MLP model by the training samples. Early-stop validation approach 20
50
is adopted to monitor and stop the model training. Upon the com-
pletion of model training, the model is applied to the testing data- 10
set to predict the weekly electrical consumption of the 101 testing
samples. The 101 predicted consumption and the actual consump- 0 0
tion are presented in Fig. 7 of which the coefficient of correlation is 0.92 0.94 0.96 0.98 1
0.98.
Coefficient of Correlation
As explained in Section 7, the same process is repeated for 1000
times. A total 1000 numbers of coefficients of correlation are ob- Fig. 8. The discrete distribution of the 1000 numbers of coefficients of correlation is
tained. They are presented in the histogram as shown in Fig. 8. approximated by beta distribution from which it is determine that the coefficient of
Since the coefficient of correlation is well bounded between 0 correlation is at least 0.9617 with 95% confidence level.

and 1, a beta distribution as shown in Eq. (3) is proposed to be used

to describe the distribution of the 1,000 coefficients of correlation.
By applying beta distribution to describe the distribution of the
xa1 ð1  xÞb1 1,000 samples, the parameters of the distribution are estimated to
f ðxja; bÞ ¼ R 1 ð3Þ
0
ua1 ð1  uÞb1 du be a = 569.5472 and b = 15.5152 and the distribution is also plot-
ted in Fig. 8. It shows that the profile of the beta distribution rea-
sonably matches with the profile of the histogram. It is obtained
1.00
from the figure that the one-sided 95% confidence limit of the beta
CORRELATION COEFFICIENT

0.98
Table 3
Summary of input and output parameters of the MLP model.
0.96
Input parameter excluded Minimum coefficient of
from the MLP model training correlation with 95%
confidence level
0.94
Area of concourse 0.9570
Area of platform 0.9562
0.92 Shops area 0.9518
Staff accommodation area 0.9567
Plant room area 0.9574
0.90 Weekly number of passengers 0.9571
26 27 28 29 30 31 32 33 34 35 36
Weekly average of outdoor temperature 0.6032
NO. OF HIDDEN NEURONS (dry-bulb)
Weekly average of relative humidity 0.9607
Fig. 6. Performances of the models under different numbers of hidden neurons.

Please cite this article in press as: Leung PCM, Lee EWM. Estimation of electrical power consumption in subway station design by intelligent approach.
Appl Energy (2012), http://dx.doi.org/10.1016/j.apenergy.2012.07.017

Weekly Electrical Consumption (MWH)
160
140
Actual Weekly Electrical Consumption
Predicted Weekly Electrical Consumption
120
100
80
60
40
20
0
02/Feb
Fig. 9. The plot shows the predicted and actual weekly electrical consumption of the selected station (i.e. Cheung Sha Wan Station) in year 2008. The RMSE is 7.67 MWh.
distribution is 0.9617 which is the minimum coefficient of correla-
tion obtained by the trained MLP model with 95% confidence level.
Based on this result, we may conclude that the performance of the
MLP model is reasonably good.
We further investigate the significance of the eight input
parameters of the model to assess their influences to the prediction
on the weekly electrical consumption. For assessing the signifi-
cance of one input parameter, the original dataset is modified by
removing that input parameter under investigation. The MLP mod-
el is then trained and tested by the modified dataset. The model
training and testing procedures are the same as Section 6. The ob-
tained minimum coefficient of correlation with 95% confidence le-
vel represents is be used as the performance index for assessment
the reduction in performance with the absence of the input param-
eter under investigation.
Table 3 summarizes the results and it shows that the minimum
coefficient of correlation is dramatically dropped from 0.9617 to
0.6032 with the absence of the weekly average outdoor tempera-
ture (dry-bulb). It reveals the highest significance of the outdoor
temperature among the other input parameters. Therefore, the
output of the model (i.e. the weekly electrical consumption of
the station) is sensitive to the weekly average of outdoor
temperature.
Regarding the applicability of the intelligent model, by given a
weather profile, it is able to forecast the annual electrical energy
consumption profile on weekly basis. This information is useful
to the designers as a reference to adjust the design parameters of
a station to optimize the spatial design to achieve minimum elec-
trical consumption. Also, the intelligent model can be used as a
system fault detection system by which fault signal is triggered if
the actual electrical consumption of the station is significantly
deviated from the consumption predicted by the model.
9. Prediction on electrical consumption profiles of a station
Section 8 demonstrates the superior performance of the MLP
model in the prediction of the weekly electrical consumption of
the stations. It was further investigated to predict the electrical
consumption of a station by using the samples of the other sta-
tions. In this demonstration, Cheung Sha Wan station was selected
as the target station to be predicted. The samples of the other 18
stations are used to train the MLP model. They are randomly
divided into 2 groups namely training samples and validation
samples in the proportion of 70–30% respectively. The training pro-
cess is identical to the approach as described in Section 6. Upon the
completion of training, the trained MLP model is applied to the
samples of the selected station hidden from the training phase.
The predicted weekly electrical consumptions of the selected
station in year 2008 are obtained.
Please cite this article in press as: Leung PCM, Lee EWM. Estimation of electrical power consumption in subway station design by intelligent approach.
Appl Energy (2012), http://dx.doi.org/10.1016/j.apenergy.2012.07.017
P.C.M. Leung, E.W.M. Lee / Applied Energy xxx (2012) xxx–xxx 9
29/Mar 24/May 19/Jul 13/Sep 08/Nov
Date in Year 2008
Since dividing the samples into training and validation samples
in the model training is a random process, in order to minimize the
randomization effect, it is proposed to carry out the model training
for 1000 times. Therefore, a total of 1000 trained MLP models are
produced after the 1000 model training. They are applied to the
samples of the selected station. The 1000 trained MLP model cre-
ated 1000 prediction results. Bagging approach [35] is proposed
to ensemble the 1000 prediction results by simply taking their
average as the final prediction result. The room-mean-square-error
(RMSE) as shown below is used to measure the fitness of the pre-
diction to the actual data where pi and ti are respectively the pre-
dicted and target values of the weekly electrical consumption of
the station.
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P53 2
i¼1 ðpi  t i Þ
RMSE ¼
53
Fig. 9 shows that the predicted profiles of the electrical energy
consumptions of the station with RMSE = 7.67 MWh which is con-
sidered acceptable from the engineering point of view.
This approach can be used to predict the electrical consumption
profile of an air-conditioned new non-interchange MTR subway
station.
10. Conclusions
The study has developed an intelligent approach to estimate the
weekly electrical consumption of a subway station based on the
spatial design and traffic loading of the subway station. Tradition-
ally, the energy analysis is carried out by computer-based simula-
tion tools such as DOE-2, EnergyPlus and BLAST. However, the
process of establishing the simulation model is very time-consum-
ing and resource-intensive. Alternatively, this approach provides a
rapid approach for MTR engineers to estimate the electrical con-
sumption of a new station so that they can optimize the designs
of the mechanical and electrical systems for achieving maximum
efficiency.
In future, we will further explore the other factors which may
contribute to the energy consumption of a station to further im-
prove the performance of the prediction. For example, elevator
and escalator systems are believed to consume significant energy
in the stations. However, the original designs of the systems do
not equip with data collection and the management does not allow
us to alter the systems for insertion of the measuring devices. This
is the reason why this information is excluded in this study. In fact,
this information is closely correlated to the overall electrical
energy consumption of the station since the electrical consump-
tions of the elevator and escalator systems also mimic the number
of passengers inside the station which directly contributes to the
10 P.C.M. Leung, E.W.M. Lee / Applied Energy xxx (2012) xxx–xxx

heat load of the air-conditioning system of the station. Therefore, if
the electrical consumptions of the elevator and escalator systems
are included in this study, the other static factors (e.g. area factors
of the stations) may be reasonably replaced.
Acknowledgements
The authors would like to thank MTR Corporation for its valu-
able assistance in providing data for this study. The work described
in this paper was fully supported by a grant from the Research
Grants Council of the Hong Kong Special Administrative Region,
People’s Republic of China [Project No. CityU/116308].
References
[1] Ampofo F, Maidment G, Missenden J. Underground railway environment in the
UK. Part 2: Investigation of heat load. Appl Therm Eng 2004;24:633–45.
[2] Ma WW, Liu XY, Li L, Shi X, Zhou CQ. Research on the waiting time of
passengers and escalator energy consumption at the railway station. Energy
and Buildings; 41:1313–8.
[3] Fu K, Deng Z. Current situation of energy consumption in Guangzhou railway
station and analysis on the potential of energy conservation. J Sustain Dev
2009;2(1):117–20.
[4] Basbous T, Younes R, Ilinca A, Perron J. A new hybrid pneumatic combustion
engine to improve fuel consumption of wind-diesel power system for non-
interconnected areas. Appl Energy 2012;96:459–76.
[5] Kaldellis JK, Kavadias KA, Filios AE. A new computational algorithm for the
calculation of maximum wind energy penetration in autonomous electrical
generation systems. Appl Energy 2009;86:1011–23.
[6] Kilpatrick RAR, Banfill PFG, Jenkins DP. Methodology for characterising
doemstic electrical demand by usage categories. Appl Energy 2011;88:612–21.
[7] Kadoshin S, Nishiyama T, Ito T. The trend in current and near future energy
consumption from a statistical perspective. Appl Energy 2000;67:407–17.
[8] Jebaraj S, Iniyan S. A review of energy models. Renew Sustain Energy Rev 2006;
10(4):281–311.
[9] Uri ND. Forecasting peak system load using a combine time series and
econometric model. Appl Energy 1978;4(3):219–27.
[10] Deeble VC, Probert SD. Straight-line correlations for annual energy-
consumption predictions? Appl Energy 1986;25(1):23–39.
[11] Bodger PS, Tay HS. Logistic and energy substitution models for electricity
forecasting: a comparison using New Zealand consumption data. Technol
Forecast Soc Change 1987;31(1):27–48.
[12] Mustafaraj G, Chen J, Lowry G. Thermal behaviour prediction utilizing artificial
neural networks for an open office. Appl Math Model 2010;34(11):3216–30.
[13] Rehdanz K. Determinants of residential space heating expenditures in
Germany. Energy Econ 2007;29:167–82.
[14] Wong SL, Wan KKW, Lam TNT. Artificial neural network for energy analysis of
office buildings with daylighting. Appl Energy 2010;87(2):551–7.
[15] Florides GA, Kalogirou SA, Tassou SA, Wrobel LC. Modelling of the modern
houses of Cyprus and energy consumption analysis. Energy
2000;25(10):915–37.
[16] Tang L, Yu L, Wang S, Li J, Wang S. A novel hybrid ensemble learning paradigm
for nuclear energy consumption forecasting. Appl Energy 2012;93:432–43.
[17] Gutierrez R, Gutierrez-Sanchez R, Nafidi A. Electricity consumption in
Morocco: stochastic Gompertz diffusion analysis with exogenous factors.
Appl Energy 2006;83:1139–51.
[18] Zhang GQ, Patuwo B, Hu MY. Forecasting with artificial neural networks: the
state of art. Int J Forecast 1998;14:35–62.
[19] Rosenblatt F. Principles of neurodynamics. New York, USA: Spartan Books;
1962.
[20] Gardner MW, Dorling SR. Artificial neural networks (the multilayer
perceptron) – a review of applications in the atmonspheric sciences. Atmos
Environ 1998;32(14–15):2627–36.
[21] Cybenko G. Approximation by superpositions of a sigmoidal function. Math
Control Signals Systems 1989;2:303–14.
[22] Hornik K, Stinchcombe M, White H. Multilayer feedforward network are
universal approximators. Neural Networks 1989;2:359–66.
[23] Kalogirou SA. Applications of artificial neural-networks on energy systems.
Appl Energy 2000;67(1–2):17–35.
[24] Hill T, Marquez L, O’Connor M, Remus W. Artificial neural network models for
forecasting and decision making. Int J Forecast 1994;10(1):5–15.
[25] Kang SY. An investigation of the use of feedforward neural networks for
forecasting. PhD Thesis 1992. Kent State University, Kent, OH, USA; 1992.
[26] Yalcintas M, Akkurt S. Artificial neural networks applications in building
energy predictions and a case study for tropical climates. Int J Energy Res
2005;29(10):891–901.
[27] Kreider JF, Rabl A. Heating and cooling of buildings. New York, USA: McGraw-
Hill; 1994 [Chapter 8].
[28] Yezioro A, Dong B, Leite F. An applied artificial intelligence approach towards
assessing building performance simulation tools. Energy Build 2008;
40(4):612–20.
[29] Yalcintas M. Energy-savings predictions for building-equipment retrofits.
Energy Build 2008;40:2111–20.
[30] Kankal M, Akpinar A, Komurcu MI, Ozsahin TS. Modeling and forecasting of
Turkey’s energy consumption using socio-economic and demographic
variables. Appl Energy 2011;88:1927–39.
[31] Chen CS, Tzeng YM, Hwang JC. The application of artificial neural networks to
substation load forecasting. Elect Power Syst Res 1996;38:153–60.
[32] Ward. Neuroshell 2 Manual 1996; Ward Systems Group Inc. Frederick. MA;
1996.
[33] Werbos PJ. Beyond regression: new tools for prediction and analysis in the
behavioural sciences. PhD dissertation 1974. Applied mathematics. Harvard
University, Boston, MA, USA; 1974.
[34] Mackowiak PA, Wasserman SS, Levine MM. A critical appraisal of 98.6 °F, the
upper limit of the normal body temperature, and other legacies of Carl
Reinhold August Wunderlich. J Am Med Assoc 1992;268(12):1578–80.
[35] Breiman L. Bagging predictors. Mach Learn 1996;24:123–40.

Please cite this article in press as: Leung PCM, Lee EWM. Estimation of electrical power consumption in subway station design by intelligent approach.
Appl Energy (2012), http://dx.doi.org/10.1016/j.apenergy.2012.07.017