Intelligent Control for Sustainable Energy Management
in Underground Stations
Hongliang Guo and Alfons Salden
Almende, B. V., Rotterdam, The Netherlands
Keywords: Energy Efficiency, Constrained Optimization, Model Predictive Control, Newton’s Method, Barrier Interior
Method, Backtracking Line Search.
Abstract: We present the scientific approach of the EU-FP7 SEAM4US project to the problem of sustainable energy
management of underground systems, in particular the optimal and scalable control of single metro stations
and their surroundings that will yield at least a 5% saving in non-traction electricity consumption,
equivalent to that consumed by more than 175 households. To this end we first formulate the sustainable
energy management problem as a constrained optimization problem and then present a modified Newton’s
method as solution. Preliminary simulation results of the model-in-the-loop are delivered and promising.
1 INTRODUCTION
Underground transportation systems are big energy
consumers (e.g. 631 million kWh / year), and have
significant impacts on energy consumptions at a
regional scale (Anderson et al., 2009).
Approximately one third of the metro networks’
energy is required for operating the subsystems of
metro stations and surroundings, such as ventilators,
lifts, escalators, and lighting (Oscar, 2011).
Although a relatively small percentage of energy
can be saved with the optimal management of these
subsystems, a large energy saving in absolute terms
can be obtained on a regional scale. The EU-FP7
project, SEAM4US (Sustainable Energy
mAnageMent 4(for) Underground Systems) is to
develop advanced technologies for optimal and
scalable energy consumption control of metro
stations that will yield a 5% saving in non-traction
electricity consumption, equivalent to that consumed
by more than 175 households.
The objective of the SEAM4US project is to
develop an intelligent control system for metro
stations, which is adaptive on the basis of
environmental factor forecasts and occupancy flow
patterns. Most of the works are ongoing; related
hardware and software deployment in the pilot
station are supposed to be implemented before
October 2014.
A metro station is a very complex system. It
involves, among others, multi-storey underground
spaces with multi-faceted thermal behaviours, e.g.,
intricate air exchange dynamics with the outside,
heat conduction with the surrounding soil and high
variable internal gains due to travelling passengers
and trains. Processes that occur in metro stations,
such as the arrival and departure of trains, passenger
transit, commercial activities, surface traffic and
weather take place on different spatio-temporal and
dynamic scales (Ansuini et al., 2012). Furthermore,
a typical metro station is a very large environment.
The modelling of the environmental conditions
requires analysis at the urban block scale, which
means a size up to thousands of meters. It is well
known that at, this dimensional scale, fluid dynamics
finite element models (FEM) are pushed to their
limits (Franke et al., 2004).
Thus the overall modelling task of SEAM4US
project is very complex; it involves user behaviour
modelling, environmental factor modelling, and
optimal controller design. We will present the
scientific position of the project in particular on the
controller design.
The structure of the paper is organized as
follows. We first introduce related work in Section
2, followed by a mathematical problem formulation
for sustainable energy management of a metro
station in Section 3. Section 4 proposes our model-
in-the-loop framework as solution. Simulation
results are presented in Section 5, followed by
discussions and future works in Section 6.
566
Guo H. and Salden A..
Intelligent Control for Sustainable Energy Management in Underground Stations.
DOI: 10.5220/0004590705660571
In Proceedings of the 10th International Conference on Informatics in Control, Automation and Robotics (ICINCO-2013), pages 566-571
ISBN: 978-989-8565-71-6
Copyright
c
2013 SCITEPRESS (Science and Technology Publications, Lda.)
2 RELATED WORK IN ENERGY
CONTROL
Energy efficiency has been gaining increasing
research interest in the past two decades. As
economic crisis continues, people are keen to design
energy efficient systems and apply them to various
application areas.
Energy efficiency is a traversal problem across
numerous application domains such as sensor
networks (Cui et al., 2004), building management
(Lamoudi et al., 2011), (Samuel, et al., 2011), and
data centre management (Lakshmi, 2012) requiring
sophisticated approaches. Cui (Cui et al., 2004)
showed, for instance, that cooperative multi-input-
multi-output (MIMO) transmission and reception
simultaneously achieve both energy savings and
delay reduction in radio application of sensor
networks.
Buildings account for 40% of worldwide energy
use (US department of energy, 2008). Many EU
projects focus on the energy performance of
buildings, like Adapt4EE (SEC-288150). Model
predictive control (MPC) methods have been applied
to minimize the energy consumption in buildings
(Lamoudi et al., 2011).
3 SCIENTIFIC PROBLEM
FORMULATION
The SEAM4US project is about (1) acquiring
optimized energy consumption minimizing
strategies (2) given a certain context determined by
outside temperature, airflow status, passenger
density, train schedule, etc., (3) while satisfying
various constraints, such as comfort-levels and
operational constraints.
Consequently, SEAM4US defines the control
task as a constrained optimization problem, i.e., find
a distributed, but coordinated, control strategy
i
w
,
which minimize the total energy consumption across
the target metro station.
()
i
i
t
ew dt
(1)
Subject to comfort level constraints:
Temp_L Temp (x, t) Temp_H
_
L (x, t) _HAirflow Airflow Airflow
Temp_L Temp (x, t) Temp_H
(2)
Hum_L H (x, t) H _Hum um
Co2_L Co2(x, t) Co2_H
Lum_L L (x, t) L _Hum um
and operational constraints:
Ctwtw
ii
||)()1(||
(3)
Where is the frequency of fan or any other
subsystem i, is the energy consumption rate of
fan , lighting or any other subsystems given input
frequency or lighting luminance level, and where
Lxx _
and
Hxx _
refer to the lower bound and
upper bound of the referred context variable,
respectively. For instance,
LTemp _
refers to
minimal requirement of temperature.
Note that passenger density (user modelling) will
influence the Temp, Airflow, humidity, CO2, etc.
Furthermore, all context variables (temperature,
humidity, level of pollutants, airflow rate) are
functions of passenger density (spatial-temporal)
distribution, train effects, and other context variables
such as outside wind, outside temperature, etc.
Therefore, the modelling task is trying to establish
and quantify the relationships between the fan
frequency, lighting luminance level and the
environmental and thermal factors and the passenger
behavioural patterns as part of the contexts such as
temperature, humidity, CO2 concentration, etc.
For constrained optimization the interior point
method (Alizadeh, 1991) is usually used to unify the
inequality constraints into the objective function.
There are two types of interior functions that we can
use, barrier interior function (Gill et al., 1986) and
primal-dual interior function (Alizadeh et al., 1998).
When the constraints are box-like constraints,
meaning that we want to bound the constraints
within a range, barrier interior functions are
typically used. When the constraints are single sided
constraints or change as time goes on, the primal-
dual interior method is often used.
After unifying the constraints into the objective
function, we reach an unconstrained optimization
problem.
If the objective function is twice differentiable,
then Newton’s method is a good candidate to learn
the optimal point. When the objective function is
differentiable, but not twice differentiable, we can
use gradient-based methods (Boyd and
Vandenberghe , 2004), such as steepest gradient
method. When the objective function is not
differentiable, we can use the sub-gradient method
(Boyd and Vandenberghe, 2004) for optimization.
)(
i
we
i
w
IntelligentControlforSustainableEnergyManagementinUndergroundStations
567
In the following section, we will elaborate our
modified Newton’s method to solve the sustainable
energy management problem.
4 MODEL-IN-THE-LOOP
SOLUTION
The modified Newton’s learning method is used to
reach the optimized solution (minimized fan energy
consumption) of the problem. The modified method
is divided into three steps, namely, (1) unifying
objectives with the constraints, (2) determining the
Newton step for next-time-step fan frequency, (3)
backtrack line search to determine the actual update
step.
4.1 Unifying Objectives
with Constraints
Since all of the constraints are box constraints,
which means that we are interested in keeping the
target variable
x
within a range . We use
logarithmic-barrier function to transform the
constraints into objectives. Logarithmic-barrier
function (Den Hertog et al., 1990) is defined as
follows:
10
log( ) log( )
()
x
xxx
x
01
x
xx
else
(4)
The unified objective function for the preliminary
problem becomes
1
(1) (( (,1))
((,1)))
obj e t Temp x t
Airflow x t
(5)
Where
is a meta-parameter that gauges the
preference weight between the main objective
(energy minimization) and the associated objectives
(keeping context variables within a target range), as
increases to
, the transformed problem
becomes the same as the constrained problem.
4.2 Newton’s Method
After unifying the constraints into the minimization
objectives, we successfully transformed the
constrained optimization problem into an
unconstrained optimization problem. We use
Newton’s method (Boyd and Vandenberghe , 2004)
to learn the direction and step size. Before
introducing Newton’s method, we would like to
display the general objective of unconstrained
optimization. In order to optimize a function
)(xf
,
we are in fact searching for a
*
x
which makes the
first order derivative
0)(
*
'
xf
. In case that there
are multiple
*
x
that makes
0)(
*
'
xf
, we select
the minimal
)(
*
xf
as the solution.
Newton's method attempts to construct a
sequence from an initial guess that converges
towards
*
x
such that
0)(
*
'
xf
. This
*
x
is called a
stationary point of
(.)f
. The second order Taylor
expansion
)(xf
of function
(.)f
around (where
n
xxx
) is:
'"2
1
( ) () () ()
2
nnnn
f
xxfxfxxfxx
(6)
attains its extremism when its derivative with respect
to
x
is equal to zero, i.e. when
x
solves the linear
equation:
0)()(
"'
xxfxf
nn
(7)
(Considering the right-hand side of the above
equation as a quadratic in
x
, with constant
coefficients.)
Thus, provided that
)(xf
is a twice-
differentiable function well approximated by its
second order Taylor expansion and the initial guess
0
x
is chosen close enough to
*
x
, the sequence
)(
n
x
defined by:
'
"
()
()
n
n
n
f
x
xxx
f
x
'
1
"
()
()
n
nn
n
f
x
xx
f
x
0,1,...n
(8)
will converge towards a root of , i.e.
*
x
for which
0)(
*
'
xf
.
Back to our optimization problem, the definition of f
is the objective function that we are trying to
minimize:
1
( ( 1)) ( 1) ( ( ( , 1))
((,1)))
fwt et Tempxt
Airflow x t
(9)
4.3 Backtrack Line Search
Although Newton learning guarantees that we can
find a
*
x
that makes
0)(
*
'
xf
, the Newton step
'
f
n
x
n
x
0
x
ICINCO2013-10thInternationalConferenceonInformaticsinControl,AutomationandRobotics
568
might be really large in reality. However, in real
world application, the fan frequency has a limited
range; it cannot go to infinity. Therefore, we use
backtrack line search (Boyd and Vandenberghe,
2004) algorithm to find the optimal step for learning.
Suppose that we have an initial guess of
displacement obtained from Newton’s method
x
.
We evaluate
)( xxf
to see if it satisfies all the
operational constraints and if it did minimize the
energy consumption, if so, we let
x
as it is. If not,
we update
x
according the following rule
xx
:
, where
is a pre-specified learning
rate. We evaluate the new
)( xxf
. We stop
updating
x
until we have a feasible and improved
)( xxf
or
x
is smaller than a certain threshold
.
5 SIMULATION RESULTS
The final goal of the SEAM4US project is to
develop an advanced control system and run it on an
actual metro station. However, in the design phase,
we should better only simulate the behaviour of
metro stations and test the algorithm on the virtual
station.
We have developed a simulator for the metro
station in dymola, and the modified Newton’s
method is tested in the simulated virtual station
environment. In the dymola model, we currently
only simulate one controllable entity, the fan. Other
facilities such as lighting and escalators will be
added in the later stage. Environmental models and
user models are used to predict the future context
variable changes such as future temperature, airflow
rate, and occupancy density level. However, we do
not concentrate on how those models are developed
in this position paper. We focus on the effectiveness
of the controller.
Fig. 4.1 shows the Model-In-The-Loop
framework, and how we do energy minimization
while considering comfort level and operational
constraints.
We first start with a fan frequency
0
w
, and through
unifying objectives and constraints, we are able to
represent the constrained optimization problem as an
unconstrained optimization problem. From
Newton’s method, we are able calculate out a
displacement (
x
), going into the backtrack line
search box, we are able to search out the ‘best’ fan
frequency. The best fan frequency will go into the
virtual station; the executed results together with
new predictions from Bayesian Networks, and user
models will trigger another round of Newton
learning and backtracking line search. Thus, the
modified Newton’s method, when including the
virtual station in the loop, is an online learning
scheme, which is able to adapt its policies in real
time.
Figure 4.1: Model-In-The-Loop Scheme.
In the simulation, we specify
in Eq. (1) as 10
(theoretically, we should specify
as large as
possible, however, that would make the qualified fan
frequency range very short, and it would therefore
be hard for the fan control agent to reach an optimal
fan frequency). The learning rate
is set to be 0.9,
threshold
is set to be 0.2 (
corresponds to the
granularity of the fan frequency control,
=0.2
means that we can increase/decrease the fan
frequency by a minimum of 0.2). The upper and
lower bound of the target temperature range is 25
and 35 respectively, and the upper and lower bound
of the target airflow rate is 40 and 80 respectively.
The starting fan frequency is 35, and the fan
frequency feasible range is from 0 to 50. We used
the model-in-the-loop framework for simulation, and
we simulated the behaviour of the controller in a
single day from 5am to 11pm.
Fig. 5.1 shows the fan frequency update over the
day. From Fig. 5.1 we can see that when adopting
our modified Newton learning strategy, we can
always have a lower fan frequency output than the
normal fan policy. Fig. 5.2, Fig. 5.3 and Fig. 5.4
show the expected energy consumption, expected
airflow rate and expected temperature at different
hours of the day. In the figure, we can see that, after
several steps of Newton Learning, we can decrease
the energy consumption of the subsystems, while
maintaining the environment factors such as
temperature, airflow rate within the pre-specified
range.
IntelligentControlforSustainableEnergyManagementinUndergroundStations
569
Figure 5.1: Fan frequency at different hours.
Figure 5.2: Expected energy consumption rate at different
hours.
Figure 5.3: Airflow rate at different hours.
Figure 5.4: Platform temperature at different hours.
6 CONCLUSIONS
AND FUTURE WORKS
In this position paper, we have presented a
mathematic problem formulation and tentatively
solved the scientific problem through modified
Newton’s method. Preliminary and promising results
within the model-in-the-loop framework are
presented but need further experimental verification.
Therefore, we are currently deploying sensor
networks in the pilot station to gather metro system,
passenger density and environmental data. After the
data collection, we will first validate and improve
the virtual station control model in dymola.
An alternative approach to the control problem
based on fuzzy control is currently investigated for
subsystem control. Furthermore, we are up to
develop distributed but coordinated control solutions
at multiple scales to tackle robustness and
computational issues. By the end of 2014, we will
have implemented algorithm and deployed the
SEAM4US system to the pilot station in Barcelona.
ACKNOWLEDGEMENTS
The authors would like to thank UniVPM for the
virtual station simulation development in dymola
and European Union for the financial support under
grant number 285408.
REFERENCES
Anderson, R. J., Maxwell, R., Harris, N. 2009.
Maximizing the potential for metros to reduce energy
consumption and deliver low-carbon transportation in
6 8 10 12 14 16 18 20 22
0
10
20
30
40
50
60
70
fan frequency
hours
adjusted fan frequency
original fan frequence
6 8 10 12 14 16 18 20 22
0
500
1000
1500
energy consumption rate
hours
adjusted energy consumption
original energy consumption
6 8 10 12 14 16 18 20 22
30
40
50
60
70
80
90
100
Airflow rate
hours
adjust airflow rate
original airflow rate
6 8 10 12 14 16 18 20 22
20
22
24
26
28
30
32
34
36
38
40
temperature
hours
adjusted temperature
original temperature
ICINCO2013-10thInternationalConferenceonInformaticsinControl,AutomationandRobotics
570
cities. Working paper for the CoMET and Nova Metro
Benchmarking Groups.
Franke, J., Hirsch, C. Jensen, A. G., Krüs, H. W.,
Schatzmann, M., Westbury, P. S., Miles, S. D., Wisse,
J. A., Wright, N.G., 2004. Recommendations on the
use of CFD in wind engineering. In: van Beeck,
J.P.A.J. (Ed.), Proc.of the International Conference on
Urban Wind Engineering and Building Aerodynamics.
COST Action C14. Von Karman Institute, Sint-
Genesius-Rode, Belgium.
Ansuini R., Larghetti R., Vaccarini M., Carbonari A.,
Giretti A., Ruffini S., Guo H., Lau S. L. (2012).
Hybrid Modeling for Energy Saving in Subway
Stations. In: Jonathan Wright and Malcolm Cook,
(Editors); 2012, “Proceedings of the 2012 Building
Simulation and Optimization Conference”; 10-11
September 2012, Loughborough, Leicestershire, UK;
ISBN 978-1-897911-42-6. Published by
Loughborough University.
Cui S.; Goldsmith, A. J.; Bahai, A., "Energy-efficiency of
MIMO and cooperative MIMO techniques in sensor
networks," Selected Areas in Communications, IEEE
Journal on , vol.22, no.6, pp.1089,1098, Aug. 2004
doi: 10.1109/JSAC.2004.830916
Den Hertog D., Roos C. and Terlaky T., (1990) On the
Classical Logarithmic Barrier Function Method for a
Class of Smooth Convex Programming Problems, In
Journal of Optimization Theory and Applications,
volume 73, Issue 1, pp 1-25.
Boyd S.and Vandenberghe L., Convex Optimization
(2004), Cambridge University Press.
Alizadeh F., Haeberly J.-P. A., and Overton M. L., Primal-
dual interiorpoint methods for semidefinite
programming: Convergence rates, stability and
numerical results. SIAM Journal on Optimization,
8(3):746–768, 1998.
Alizadeh F.. Combinatorial Optimization with Interior-
Point Methods and Semi-Definite Matrices. PhD
thesis, University of Minnesota, 1991.
Gill P. G., Murray W., Saunders M. A., Tomlin J. A., and
M. H. Wright. On projected newton barrier methods
for linear programming and an equivalence to
Karmarkar’s projective method. Mathematical
Programming, 36:183–209,1986.
US Department of Energy, Energy Efficiency Trends in
Residential and Commercial Buildings (2008), in
Energy Efficiency and Renewable Energy.
Lamoudi M. Y, Alamir M. and Béguery P., "Distributed
constrained Model Predictive Control based on bundle
method for building energy management,". 2011 50th
IEEE Conference on Decision and Control and
European Control Conference (CDC-ECC) Orlando,
FL, USA, pp. 8118-8124, Dec 2011.
Samuel P., Jan S., Lukas F., Jiri C., Model predictive
control of a building heating system, The first
experience, in Journal of Energy and Buildings, 2011,
volume 43, issue 2-3, pp 564-572.
Lakshmi G., Data Center Energy Management, PhD
dissertation, Cornell University, 2012.
Oscar P., private communication, TMB, Barcelona,
October, 2011.
SEAM4US consortium. (2013, 04 01). SEAM4US website.
Available: http://www.seam4us.eu.
IntelligentControlforSustainableEnergyManagementinUndergroundStations
571
