Professional Documents
Culture Documents
4, DECEMBER 2016
Abstract This paper proposes a novel approach to power GHI Global horizontal irradiance.
wireless communication network base stations (BSs) in order to
achieve higher resilience. In the proposed dc microgrid, resilience
CTF Cycles to failure.
is improved with the coordinated operation of renewable energy DOD Depth of discharge.
sources, energy storage devices, and BS traffic using an integrated QoS Quality of service.
controller that adjusts BS traffic and distributes stored energy PSNR Peak signal-to-noise ratio.
among cell sites based on weather forecast and other operational LTE Long-term evolution.
conditions. The essential role played by power electronics active C1 Configuration 1.
power distribution nodes in this resilient microgrid is also C2 Configuration 2.
explained. Resilience improvements are evaluated quantitatively
with metrics analogous to that of availability, which considers
CTF Cumulative distribution function.
energy storage levels, power generation, and load. The analysis MDT Mean downtime.
shows that the proposed system allows reducing service restora-
tion times by a factor of 3 or improve resilience by double-
I. I NTRODUCTION
digit percentage points, while the battery life is extended by
about 10%. Additionally, battery bank size needs and PV array
footprints can be reduced without compromising resilience. The
proposed microgrid is implemented for wireless communica-
T HIS paper explores an integrated energy management
controller for a dc microgrid that improves power sup-
ply resilience in wireless communication networks. In this
tion networks because both energy and communications are
identified in the U.S. Presidential Policy Directive 21 as the paper, the concept of resilience is considered based on the
two infrastructures that are especially critical for community definition provided by the U.S. PPD-21 [1], which indicates
resilience. Nevertheless, this same technology can also be used that resilience is the ability to prepare for and adapt to
in other applications, such as residential neighborhoods or changing conditions and withstand and recover rapidly from
industrial campuses. disruptions. That is, resilience is characterized not only by
Index Terms Active power distribution nodes (APDNs), how quickly an infrastructure system can recover from a
dc microgrids, energy sharing, power electronics, resiliency, disruptive event but also by how well it can withstand such
traffic shaping, wireless communications.
event. Effective planning contributes to improved resilience by
affecting preparedness and adaptation. Recognizing the impor-
N OMENCLATURE
tance of communications and energy infrastructures during
PPD-21 Presidential Policy Directive 21. natural disasters, we will focus this paper on the power supply
BS Base station. of wireless communication networks. This importance is also
RWCM Resilient wireless communication microgrid. recognized in [1] by identifying energy and communications
QoE Quality of experience. systems as uniquely critical due to the enabling functions they
DSM Demand side management. provide across all critical infrastructure sectors. Recently,
DR Demand response. there has been increased interest in studying wireless com-
SOC State of charge. munication network resilience during natural disasters, in
APDNs Active power distribution nodes. part, motivated by several intense extreme events that in the
NB Nave Bayes. past decade affected all corners of the world. Some notable
Manuscript received December 30, 2015; revised March 13, 2016 and examples include hurricanes Katrina, Ike, and Sandy in the
June 10, 2016; accepted September 11, 2016. Date of publication USA, and earthquakes and tsunamis that affected Chile, Japan,
September 28, 2016; date of current version October 28, 2016. This work and New Zealand. In all these events, important service dis-
was supported by NSF under Award 1331788. Recommended for publica-
tion by Associate Editor Sayed Ali Khajehoddin. (Corresponding author: ruptions were documented in both power and communication
Alexis Kwasinski.) infrastructures [2][7]. In particular, the two main causes of
Y. Kwon is with The University of Texas at Austin, Austin, TX 78701 USA. communication network disruptions were power supply issues
A. Kwasinski is with the University of Pittsburgh, Pittsburgh,
PA 15261 USA (e-mail: akwasins@pitt.edu). from extensive electric grid outages and network congestion
A. Kwasinski is with the Rochester Institute of Technology, Rochester, due to increased traffic.
NY 14623 USA. In recent past natural disasters, wireless communication
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. network BSs were especially vulnerable to electric grid power
Digital Object Identifier 10.1109/JESTPE.2016.2614425 outages [2], [5], [6]. That is, as Fig. 1 exemplifies, although
2168-6777 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
KWON et al.: COORDINATED ENERGY MANAGEMENT IN RESILIENT MICROGRIDS FOR WIRELESS COMMUNICATION NETWORKS 1159
B. Control Architecture
Fig. 4. Example of a basic topology of an APDN. Fig. 5 shows the general architecture of the RWCM energy
management and system control platform. As Fig. 5 shows,
QoE changes that are near imperceptible for the end user or the proposed controller has a hierarchical structure in which
modify QoE so that it still is within the range that corresponds the top level is used to coordinate the operation of all power
to acceptable performance. sources, energy storage devices, and loads in order to optimize
RWCM operation. At a lower level, local decentralized con-
A. Power Distribution Architecture trollers located at each node (usually at a cell site, although it
As it was indicated, Fig. 3 shows an example of a possible is possible to have nodes without a load) are in charge of acting
power distribution architecture for the RWCM that intercon- on its corresponding BS traffic and/or local power generation
nects a few nodes, most of them BSs, through a dc microgrid. or energy storage devices based on the regulation set points
In it, power electronic dcdc converters are used to interface received from the high-level controller. As indicated in [15],
each node with the rest of the microgrid. Although these it is assumed that these lower level local controllers operate
converters could be realized with conventional single-input based on droop laws so that in the case of a failure in the
single-output topologies, a more cost effective and opera- communication link with the high-level controller or in the
tionally flexible approach is to use multiple-input multiple- high-level controller itself, the local controllers are still able
output (MIMO) bidirectional dcdc converters called APDNs, to manage local resources and load at their corresponding
exemplified in Fig. 4 and studied in [48][51]. As it was BSs but at a suboptimal level. Since decentralized control
discussed in these papers, APDNs allow power flow control approaches for microgrids in general and for APDNs in
in all of their ports, which is a very desirable feature for the particular have been presented in the past [15], [40], [49],
RWCM because controlled power flow among nodes enables a the focus of this paper is on the higher level controller that
more effective energy management and operation. Moreover, optimizes the operation of the RWCM for enhanced resilience.
APDNs can incorporate embedded energy storage, which,
as described in [48], creates the possibility of improving C. Planning and Forecasting
availability. Hence, the use of APDNs enables operating each Development of a resilient microgrid involving the use of
node with a differentiated availability target, so operation renewable energy sources requires considering weather fore-
of more critical nodes, e.g., a BS serving a hospital, can casting both for long-term planning and for shorter term oper-
be prioritized over other nodes in the RWCM. This is an ation. Anticipating weather profiles when planning an RWCM
important feature within the context of realizing a resilient configuration serves to ensure that the microgrid will have
microgrid, because as will be discussed later, the metrics used sufficient resources, particularly energy storage, in order to
to measure resilience are analogous to those used to evaluate meet availability and resiliency requirements without oversiz-
availability [17]. Furthermore, APDNs are essential system ing the system. Short-term weather forecasting serves to ensure
components in order to implement the energy sharing algo- that existing resources are used effectively. For example,
rithm explained later in this paper. In order to provide an equal two-day-ahead weather forecasting influencing controller
basis for comparison of all control approaches discussed in actions serve in an RWCM to keep batteries with adequate
this paper, APDNs are assumed to have ideal power efficiency. charge levels while anticipating potential future weather con-
Nevertheless, at the proposed power levels, APDNs, such as ditions that would require more intense use of batteries, while
those in Fig. 4, can be designed with high efficiency. Moreover, at the same time reducing chances of early battery failure due
the system cabling layout is expected to be designed following to inadequate usage profiles.
the communication network practices that avoid fully cascaded In the proposed RWCM, the controller achieves higher
power distribution architectures. This design in addition to the resiliency by optimizing the use of batteries based on weather
use of a dc power distribution architecture leads to a reduced forecasting in combination with the integrated management
1162 IEEE JOURNAL OF EMERGING AND SELECTED TOPICS IN POWER ELECTRONICS, VOL. 4, NO. 4, DECEMBER 2016
of BSs power consumption and of power output from renew- meteorological agencies do not produce a forecast for the GHI,
able sources. Power output from each renewable source (solar wind speed forecast can be typically directly obtained from the
and wind) is predicted based on an NB classifier model anticipated weather information. Then the proposed NB model
presented in [18] and implemented based on the National can be used to establish the relationship between forecasted
Oceanic and Atmospheric Administration (NOAA) weather wind speed and actual wind power.
information [52]. This weather information is used to antici-
pate GHI and wind speed conditions and the main variables D. Resiliency Metrics
influencing PV and wind power generation, respectively. In the past, it has been proposed that microgrids can
The used NB model [18] is an effective probabilistic classi- improve resilience [34][38]. However, such statements were
fication algorithm. Based on Bayes theorem, the NB method not supported with quantitatively metrics that allowed assess-
performs the classification with the assumption that the fea-
ing by how much resilience was improved or that provided
tures (the weather variables) are independent of each other in a a way for evaluating different technological alternatives in an
given class (e.g., the value of the GHI). The fundamental idea objective way. In this paper, based on the aforementioned def-
of this application of Bayes theorem is that the probability of
inition of resiliency from PPD-21 [1], power supply baseline
a class C {1, . . . , l} can be estimated based on observations resilience B of a BSi.e., the resilience of the proposed
features. The value of l is an integer indicating the class levels, microgrid as seen by each of the BSs in the RWCM without
i.e., the set of values that GHI can take; in this paper, l is accounting for the contribution of energy storage devicesis
assumed being equal to 100. In order to apply the proposed measured based on a metric analogous to that of availability
algorithm, four weather variables are considered: temperature,
as [17]
relative humidity, dew point, and sky coverage. In the case
TU
of GHI prediction, the proposed NB model has three steps. B = (3)
In steps 1 and 2, the training data sets (weather variables) TU + TD
are partitioned into several hourly subsets and the subsets are where TU is the time when the BS is receiving the necessary
filtered based on the sky coverage variable, respectively, in power from the microgrid and TD is the downtime when the
order to avoid an overlapping classification and improve the BS is not receiving power from the microgrid either due to
classification accuracy. In step 3, the NB model applies Bayes failure, damage, maintenance, or repairs. As detailed in [15],
theorem based on the assumption of independent variables as this metric of resiliency analogous to that of availability has
follows: been independently proposed in both [53] and [54]. In these
P(W1 |C) P(W2 |C) P(W3 |C) P(C) works, it is explained how TU and TD relate to the attributes
P(C|W1 , W2 , W3 ) = of resiliency indicated in PPD-21, such as withstanding capa-
P(W1 , W2 , W3 )
(1) bility and recovery speed, respectively. Both TU and TD are
evaluated over the entire period of consideration, which is
where the weather variables, W1 , W2 , and W3 , represent the represented by the sum of TU and TD in the denominator
temperature, relative humidity, and dew point, respectively. of (3). In order to consider the effect of renewable energy
Additionally, the prior probability of C is indicated by P(C) sources on availability, TU is associated with the time when the
and P(C|W1 , W2 , W3 ). Each conditional probability term of power generated by the local sources is equal or higher than
P(Wi |C) is then estimated with a Gaussian kernel [18], and the load and TD corresponds to the time when the opposite
finally, the posterior probability of P (C|W1 , W2 , W3 ) in (1) condition is observed. Although (3) seems to be indicating an
is updated based on [18] availability measurement, the baseline resilience metric in (3)
is analogous only to the mathematical definition of availability.
GHINB = arg max P(C) P(Wi |C) (2) The main difference between availability and resilience is that
C{1,...,l} i availability is computed based on mean uptime and mean down
where GHINB stands for the target class value, chosen to be time calculated over a very large number of failure and repair
the one maximizing the forecast probability as indicated in (2). cycles, whereas resilience in (3) may be calculated over one
On a clear day, the values of GHI over time are deterministic cycle of potential failure and operation during a disruptive
(a bell-shaped curve), increasing until noon and decreasing event or may also be calculated over a very large number of
thereafter until sunset. However, on a cloudy day, GHI shows cycles. Hence, in this paper, the term availability is still used
only a partly deterministic characteristic that is difficult to when considering periods of time sufficiently long, such as
be explained with a simple parametric family distribution, in planning assessments, because in such conditions, metrics
such as a normal distribution N(, 2 ). Thus, in order to for availability and resilience are indistinct. In the same way,
predict GHI more accurately, nonparametric kernel function resilience is defined in an analogous way to availability;
is used, which can deal with large variations and, therefore, in this paper, fragility F is considered to be analogous to
achieve a more accurate GHI prediction. In addition, since the unavailability. That is, F = 1 B . This definition for
GHI classification is a nonlinear problem, the kernel density fragility originates in the lexical definitions of both resilience
function may be preferable to the estimation of GHI variations. and fragility. However, it is relevant to indicate that in some
Wind speed classification could be performed with a similar past works [55], the term fragility has been associated with a
methodology. However, contrary to the case of the GHI failure probability depending on a hazard intensity. Likewise,
prediction that requires using four weather variables, because it is important to note that the term resilience is distinct
KWON et al.: COORDINATED ENERGY MANAGEMENT IN RESILIENT MICROGRIDS FOR WIRELESS COMMUNICATION NETWORKS 1163
bank SOC. Specifically, a real-time cellular traffic shaping levels than when not doing traffic shaping. Limiting the traffic
algorithm is investigated in order not only to increase the use resulted in increased data delay and real-time video with lower
of renewable sources but also to decrease potential excessive quality (because it needs to be compressed more), which
battery bank discharges. For simplicity of calculations, this results in end users experiencing a controlled reduction in
paper assumes an RWCM with two BS and the focus is on their QoE. As mentioned earlier, the metric used to measure
the high-level controller. QoE depends on the type of traffic. For a real-time video, the
QoE shall measure the playback quality of the video, and for
A. Net Power Generation Model data traffic, QoE is measured in terms of delay. Reducing the
traffic through the BS with a setting of [n] < 1 results in a
The net power generated in the RWCM at any given time reduction of the QoE to active calls. For a real-time video, a
is defined as [n] < 1 reduces the playback quality (measured here as the
Pnet [n] = PPV [n] + PWind [n] PBS [n] (8) PSNR) because video is compressed more. For data traffic,
[n] < 1 increases the delay because it yields longer transmit
where PWind [n] is the power generated by wind turbines, times and queuing waiting times. Nevertheless, as is the case
PPV [n] is the power generated in PV arrays, and PBS [n] is in this paper, by judiciously setting [n] , the QoE can be
the power consumed by the BS, all of them considered at time reduced in a way that is nearly imperceptible or acceptable
step n. If the effects of temperature on PV cell performance for end users.
are neglected in order to simplify the calculations, the power Following (10), the power consumed by the BS can be
generated at a PV module of the RWCM is given by controlled by setting [n]. The proposed traffic management
PPV = (GHI)(A)(PV ) (9) improves resilience in two ways. On the one hand, reduced
load makes Tbank in (4) to increase. On the other hand, reduc-
where A is a total PV module area (m2 ) and P V is the ing BS load by limiting traffic increases Pnet as given in (8).
conversion efficiency. For simplicity in the calculations, it is As a result, the uptime tends to increase when traffic is limited
assumed that all arrays in the RWCM receive the same GHI and the downtime tends to decrease. Limiting traffic may
and uses the same PV module, a USA MX60-240 PV module, be beneficial even when applying fixed preset values during
which has a size of 1.7 m2 , with an efficiency of 14% and a power outages events. Indeed, as discussed in [62], traffic
maximum power output of 240 W. In order to characterize shaping can achieve important reductions in battery bank size
PWind , a theoretical turbine wind speedpower output curve is without reducing autonomy or in PV array footprint without
used [60]. Like the PV modules, it is assumed that all the wind compromising performance. Although these improvements are
turbines used in the RWCM are the same: a Bergey BWE10 beneficial, it is desirable to realize more flexible controllers
wind turbine is considered, which composes of a three-blade that provide similar advantages without having to continuously
horizontal axis (7-m rotor diameter) with a 9-km/h cut-in wind limit traffic. This is the motivation for the controller discussed
speed and a maximum power of 10 kW at 41.8 km/h. next.
For the BS, it is assumed that the RWCM is operating in an
urban area in which each cell cite has a coverage area given by B. Traffic Shaping Strategy
a circle with a radius of 250 m. As a macro-BS, transmitting The previous paragraphs describe the value of integrated
three sectors with 2 2 MIMO (two transmitters and two traffic and energy management in order to improve resilience
receivers) antenna configuration and the transmitting power while still providing acceptable QoE to network users. How-
of 40 dBm are assumed [61]. In addition, the frequency reuse ever, network operators are also interested in optimizing oper-
factor of 3 and a 10-MHz system bandwidth are assumed as ation to prevent inadequate energy storage sizing or reduced
in [61]. In this paper, a BS power consumption is calculated battery life. Such an optimized operation is achieved through
using the model from [14], which can shape the traffic at the RWCM higher level controller. The general controller
the load based on the predicted renewable energy and battery strategy based on (8) and (10) is summarized in Table I
bank SOC. The power consumption is expressed by and represented in Fig. 7. In this control strategy, a Markov
PBS [n] =NRT (PT [n] [n]+PB ) (10) chain is used in order to estimate the likelihood of batteries
being at each of the possible SOC levels. That is, in the
where NT R is the number of active radio transmitters (in this Markov chain, each state represents a discrete SOC level,
case is 6; the result of multiplying the number of antennas per from state #1, representing the discharged state, to state #N,
sector, 2, with the number of sectors per BS, 3), [n] is the representing the fully charged state. States #2 to # N 1
normalized traffic profile at the nth time period, PT represents represent intermediate SOC levels, usually spaced based on
a factor (in this paper considered equal to 97) that accounts an equal energy difference of W that can be considered to
for the power consumed at the BS depending on the traffic equal a power difference P times a time step period T , which
profile, and PB = 65 W is a constant power term accounting equals the duration of one control action step or period. Hence,
for pilot signals, cooling, processing, base band interface, and the battery capacity equals [19]
so on.
Cbank = (N 1) P T. (11)
The traffic shaping factor [n] controls the volume of
cellular traffic handled by each BS. Settings with [n] < 1 Then, the information about the likelihood of batteries being
indicate that the traffic through a BS is limited to smaller at each of the possible SOC levels is used to adjust the load
KWON et al.: COORDINATED ENERGY MANAGEMENT IN RESILIENT MICROGRIDS FOR WIRELESS COMMUNICATION NETWORKS 1165
TABLE I where v[n] is the battery bank SOC at time step n. Thus, the
T RAFFIC S HAPING S CENARIO vector v[n + M] represents
v[n + M] = [v 1 [n + M], v 2 [n + M], . . . , v N1
[n + M], v N [n + M]]1N (13)
where each v i [n + M] represents the probability of each of
the SOC states. Once v [n + M] in (13) is calculated, the
maximum probability value in v[n + M] is selected as the
maximum likelihood battery bank SOC at time step n + M.
This value is then used to determine the most suitable value
for the traffic shaping factor in the next control decision step,
[n + 1], in order to prevent excessive battery discharges.
In order to evaluate the value of [n + 1], an algorithm
summarized in Table I is used. For simplicity in the explana-
tion, consider that a BS station is powered only by a PV array
coupled with batteries. Then, the battery energy charge level
at time step n, S[n], is given by
n
S[n] = S[0]+T (PP V [k] NTR (97 [k] [k] + 65)). (14)
k=1
Then, at each decision time n, the high-level controller con-
siders a target change in the energy stored in the batteries,
s S[n] S[n 1], which is given by
s( [n]) T (PPV [n] NTR (97 [n] [n] + 65)). (15)
That is, s( [n]) = Pnet [n]T . Values of s( [n]) > 0
represent cases when the batteries are charged and
s( [n]) < 0 are cases when the batteries are discharged.
Hence, a large positive setting for s( [n]) corresponds to a
situation when the battery is rapidly charged, likely implying
Fig. 7. Real-time iteration algorithm for a maximizing utilization of energy that traffic is shaped and is significantly reduced or that traffic
storage. is already low due to low network usage at the time. Since
the former of these two cases is the one of interests in terms
of controller design, this is the case considered at this point.
using the traffic shaping factor [n + 1] in order to optimize When the former case happens, the power consumption at the
battery use while keeping users QoE within acceptable levels. BS is reduced at the expense of reducing video quality and
In this paper, for simplicity, T is considered to be equal to 1 h increasing data delay. Hence, it may be desirable to charge
although other time step durations can easily be considered the batteries at a slower rate so that end user experience is not
instead. affected as much as when the batteries are rapidly charging.
Fig. 7 represents the real-time iterative process using a On the contrary, a small value for s( [n]) likely imply
Markov chain model used to choose the control traffic shaping minimal traffic shaping, or no traffic shaping at all. A value
factor [n + 1]. In order to consider the potential effect much smaller than zero implies that the batteries may be
of future weather conditions, this process begins by calcu- discharging excessively fast when no or little traffic shaping
lating the net power profiles up to M hours in the future is likely being applied. Thus, a maximum value for s( [n]),
(Pnet [n+1], . . . , Pnet [n+M]) based on the proposed NB model smax , is defined as in [14], representing the maximum value
summarized in Section II-C, which also includes estimation of Pnet [n + 1] that is set with the maximum traffic shaping
for BS power consumption based on (8) and (10). The (corresponding to the smallest setting of [n + 1] = 0.3).
integer Mis determined corresponding to the battery bank This is because the battery SOC is low and the batteries need
capacity [14], [19] and these profiles are used to calculate to be charged fast or because the batteries are being discharged
the net power transition probabilities pi . Then, the M-step excessively fast and so the load would have to be reduced by
net power transition matrix P M , which includes M-hour- applying maximum traffic shaping (in which case smax 0).
ahead anticipated battery bank SOC transition information is Hence, as indicated in Table I, the higher level controller
constructed in order to determine the state probability vector algorithm first evaluates whether smax is greater than 0 as
for a Markov chain in which each state represents a discretized an Event 1 trigger condition. Once Event 1 is evaluated,
SOC level each control decision of determining [n + 1] can be chosen
considering three cases: when the SOC[n] < SOClim , when
v[n + M] =v[n] P M (12) SOC[n] SOCc , and when SOClim SOC[n]<SOCc , in
1166 IEEE JOURNAL OF EMERGING AND SELECTED TOPICS IN POWER ELECTRONICS, VOL. 4, NO. 4, DECEMBER 2016
which for simplicity the SOC at time step n of a node is aggressive approach when reducing QoE in order to reduce
defined based on a linear assumption as environmental impact of powering wireless networks.
However, when smax 0 and batteries are discharged, the
S[n]
SOC[n] = (16) algorithm follows a 1-time-step look-ahead strategy based
Cbank on equally spaced intervals in which traffic is limited more
In these cases, SOClim represents a lower limit for battery bank as SOC decreases. This strategy is chosen over the logistic
SOC that is selected based on various aspects, such as battery curve strategy because although both strategies have the same
technology, load criticality, and battery replacement cost. That 1-time-step look-ahead horizon, at higher SOCs, the intervals-
is, although SOClim does not represent a fully discharged based strategy limits traffic more aggressively than the logistic
battery bank, it indicates a condition in which its charge level curve-based strategy, whereas at a lower SOC, both strategies
is considered at a critically low level. On the other end, SOCc perform similarly. The Markov-chain-based strategy could also
and all higher SOCs represent SOC levels sufficiently high that have been used during discharges, but the intervals-based
allows operating the microgrid without implementing traffic strategy was chosen for two reasons. Since the Markov-
shaping. chain-based approach is an M-time-step look-ahead strategy,
As shown in Table I, the control strategy when the current it tends to limit traffic more aggressively than the 1-step look-
SOC, SOC[n], is below SOClim (case 1) or above SOCc ahead intervals-based strategy. That is, during discharges, the
(case 3) is straightforward. In case 1, the command sent by the intervals-based strategy tends to affect QoE in an intermediate
higher level controller is to limit traffic as much as possible, way compared with the effects of the logistic curve strategy
so [n + 1] takes the minimum set value. The minimum value or the Markov chain strategy.
for [n + 1] is chosen by considering the subjective nature Still, there is one more important reason for the proposed
of QoE and that communication networks may operate in approach. In reality, the previous discussion assumes perfect
conditions where the QoE for end users is reduced to levels weather forecasting, whereas in reality, weather forecast is
that, while not as good as in normal conditions, they are still subject to uncertainties that may lead to wrong forecasting.
acceptable to end users if the reduction in QoE is transient. Such uncertainties and forecasting errors increase as the time
In this paper, the minimum value for [n + 1] was chosen to horizon for the forecasting increases. That is, in the Markov
be 0.3. This is following the results in [14] that show that for chain model, as M increases, the chances for the controller of
this minimum setting, video traffic distortion and data traffic making unnecessarily aggressive or not sufficiently aggressive
delay would be at the limit of QoE acceptability. In case 3, traffic shaping decisions also increases. For example, if the
when the battery SOC is sufficiently highin this paper, it is M-time-step look-ahead forecast suggests that the SOC may
considered that the battery has at least 90% of its maximum fall below SOClim and the observed weather matches the
chargethen traffic is not limited and [n + 1] is set equal forecast, then the Markov chain strategy would be desirable.
to 1. Note that a traffic shaping scheme that avoids setting However, if the weather ends up being better than the forecast,
[n + 1] to its minimum all the time is needed because end then QoE would have been unnecessarily reduced. Likewise,
users would not accept a service with the minimum acceptable if the forecast anticipate favorable weather conditions so that
QoE all the time when an operation at better-than-acceptable SOC[n + M] > SOClim , but the actual weather is worse than
QoE is possible. In this way, the control strategy serves the the forecast, then the batteries may end up being discharged
purpose of reducing the impact of traffic shaping to end more than if a 1-step look-ahead strategy was used. Hence,
users. the choice for a method to estimate the weather forecast,
For case 2, the proposed algorithm includes both short- such as the NB approach discussed above, is an important
term (1-time-step) and longer term (M-time-step) strategies. aspect affecting system operation. Although other weather
Consider first when smax > 0 so that batteries are being prediction methods can be chosen instead of the NB approach
charged. In this situation, if the longer term evaluation of the considered here, the NB method has shown to be considerably
SOC at time step n + M indicates that the SOC will be above more accurate than other methods presented in the past [18].
SOClim , then a 1-step look-ahead traffic shaping strategy based Moreover, weather data accuracy may present some issues.
on a logistic function and introduced in [14] is implemented. In our case, use of weather data from the USA is considered
However, if the evaluation of the SOC at time n+M yields that to present a more challenging design environment than the
the SOC[n + M] < SOClim , then the algorithm recognizes that one found in other regions, such as Europe, which is broadly
although within the next time step, batteries are being charged, considered to be more accurate. Still, discussion of the forecast
in M time steps not only batteries would have discharged estimation errors is outside of the scope of this paper, and for
but they would have discharged below SOClim . As a result, a the NB method, it has been presented in [18].
better strategy than using the logistic function is to limit traffic Evidently, the key issue involving strategy selection is that
more aggressively using the M-steps look-ahead Markov chain the controller makes a decision about traffic shaping before
strategy, which makes the battery to be charged faster in an the forecast is confirmed or that different weather conditions
attempt to prevent the SOC to fall below SOClim in M time are observed. Although applying a more aggressive M-time-
steps. The use of this more aggressive traffic shaping approach step look-ahead strategy like the one based on a Markov chain
than just using the logistic function introduced in [14] is may be more rewarding than a 1-step look-ahead strategy like
consistent with the focus of this paper on resilient operation at the one based on intervals, if the observed weather follows
the expense of a reduced user QoE because [14] sought a less the forecast. Although these rewards may be greater with
KWON et al.: COORDINATED ENERGY MANAGEMENT IN RESILIENT MICROGRIDS FOR WIRELESS COMMUNICATION NETWORKS 1167
Fig. 9. Estimated energy harvested for each configuration during the Fig. 11. CDF for traffic shaping in Configuration 1.
operational evaluation period.
TABLE III
C OMPARISON B ETWEEN N O -T RAFFIC S HAPING AND
T RAFFIC S HAPING IN T WO C ONFIGURATIONS
V. C ONCLUSION
This paper has discussed the realization of a dc microgrid
in order to provide a highly resilient power supply to wireless
communication network BSs. In this way, this microgrid
addresses the main sources of communication network dis-
ruptions during natural disasters. The focus of this paper is
on wireless communication networks applications because the
proposed system can be readily used by this industry. The
reason for this direct application is that its use is not dependent
on modifying regulatory or safety codes, because in this
application, the proposed microgrid is a private industrial-type
power distribution grid accessible only by trained personnel.
The RWCM presented here is intended to facilitate the use
of renewable sources in wireless communication networks
and, thus, reduce vulnerabilities affecting resilience by avoid-
ing lifeline dependencies. The microgrid power distribution
architecture uses APDNs, which are power electronic cir-
cuits placed at the nodes of the power distribution grid.
These APDNs are essential components enabling the control
approach proposed in this paper.
In addition to proposing a novel architecture to power
Fig. 14. Summary of resilience results for configurations C1 and C2. wireless communication cell sites, another fundamental con-
tribution of this paper is addressing issues in the application of
renewable energy sources that limit their application to power
sharing algorithm allows the improving resiliency for C2 from BSs. These limitations of renewable energy sources are large
0.79 to 0.91, while discharges are shallower, with the average footprint and variable output. These issues are addressed by
DOD being reduced from 96.32% to 80.42%. Moreover, a high-level controller presented in this paper that operates
the downtime over an 8-month period improves from 50 to power generation, energy storage, and load power consump-
21 days. In terms of resilience and considering the same tion in an integrated way. This controller uses a combination
aforementioned equivalent natural disaster scenario in which of short term, 1-step-aheadlook-ahead strategies, and longer
an electric power grid outage lasts at least a week, the use of term, M-time-step look ahead strategy, in order to control
the energy sharing algorithm and traffic shaping algorithm in the load so that batteries have their autonomy extended, and
C2 allows improving the uptime from 132.72 to 154 h and thus thus, resiliency is increased. The load is controlled by limiting
improving resiliency metrics. This control strategy applied to traffic on the BSs while still maintaining acceptable user QoE.
C2 also achieves important improvements in terms of number An M-time-step look-ahead strategy based on Markov chains
of cycles, CTF, throughput, and life of batteries. As indicated, using NB weather estimates is also discussed in detail.
improvements when the energy sharing algorithm is applied The analysis supported by quantitative resilience metrics
to C1 are not as significant because average discharges when and actual BS and weather data indicates that the proposed
implementing traffic shaping to C1 are already relatively control approach improves power supply resilience by extend-
shallow. Hence, improvements with energy sharing with C1 are ing battery autonomy and increasing uptimes. Moreover, since
relatively minor compared with those with C2 in which dis- the analysis shows that installed power generation capacity
charges without energy sharing are relatively deep, even when limits how much resilience can be improved by adding bat-
traffic shaping is implemented. It is important to highlight the teries, the proposed control approach allows for improved
role of APDNs in achieving these significant improvements resilience without adding more batteries. In the same way,
to C2, because, as explained, APDNs are the key RWCM the analysis explains how the proposed approach allows
components necessary for implementing an effective energy achieving the same levels of resiliency than conventional
sharing algorithm. approaches while reducing the footprint of PV arrays. Addi-
As indicated, both the traffic shaping and energy sharing tionally, as Fig. 14 summarizing the analytic results show, the
algorithms are implemented by the high-level controller, which inclusion of an energy sharing algorithm in the controller also
sends the necessary commands to each node to regulate their contributes to improve resiliency and to extend battery life.
1172 IEEE JOURNAL OF EMERGING AND SELECTED TOPICS IN POWER ELECTRONICS, VOL. 4, NO. 4, DECEMBER 2016
Future work includes experimental validation using the [20] A. Kwasinski and A. Kwasinski, Increasing sustainability and resiliency
Vienna MATLAB LTE-A System Level Simulator, which of cellular network infrastructure by harvesting renewable energy, IEEE
Commun. Mag., vol. 53, no. 4, pp. 110116, Apr. 2015.
allows simulating and measuring performance for a multicell [21] E. Altman, M. K. Hanawal, R. El-Azouzi, and S. Shamai (Shitz),
LTE network. The simulation is detailed up to the point Tradeoffs in green cellular networks, ACM SIGMETRICS Perform.
to be possible to measure the resource allocation for the Eval., vol. 39, no. 3, pp. 6771, Dec. 2011.
[22] Z. Niu, Y. Wu, J. Gong, and Z. Yang, Cell zooming for cost-
transmission of each frame in each BS (eNB). efficient green cellular networks, IEEE Commun. Mag., vol. 48, no. 11,
pp. 7479, Nov. 2010.
R EFERENCES [23] K. Son, H. Kim, Y. Yi, and B. Krishnamachari, Base station oper-
ation and user association mechanisms for energy-delay tradeoffs in
[1] Presidential Policy DirectiveCritical Infrastructure Security green cellular networks, IEEE J. Sel. Areas Commun., vol. 29, no. 8,
and Resilience, accessed on Dec. 25, 2015. [Online]. Available: pp. 15251536, Sep. 2011.
https://www.whitehouse.gov/the-press-office/2013/02/12/presidential- [24] D. Brubaker, Optimizing performance and efficiency of PAs in wireless
policy-directive-critical-infrastructure-security-and-resil base stations: Digital pre-distortion reduces signal distortion at high
[2] A. Kwasinski, Telecommunications outside plant power infrastructure: power levels, Texas Instrum. Incorporated, Dallas, TX, USA, White
Past performance and technological alternatives for improved resilience Paper SLWY002A, Feb. 2008.
to hurricanes, in Proc. INTELEC, 2009, pp. 121. [25] M. Varela, L. Skorin-Kapov, and T. Ebrahimi, Quality of service
[3] A. Kwasinski, W. W. Weaver, P. L. Chapman, and P. T. Krein, Telecom- versus quality of experience, in Quality of Experience, S. Moller and
munications power plant damage assessment for Hurricane KatrinaSite A. Raake, Eds. Switzerland, Springer, 2014, pp. 8596.
survey and follow-up results, IEEE Syst. J., vol. 3, no. 3, pp. 277287, [26] M. H. Albadi and E. F. El-Saadany, A summary of demand response
Sep. 2009. in electricity markets, Electr. Power Syst. Res., vol. 78, no. 11,
[4] A. Tang, A. Kwasinski, J. Eidinger, C. Foster, and P. Anderson, pp. 19891996, Nov. 2011.
Telecommunication systems performance: Christchurch earthquakes, [27] L. Zhao and B. Zeng, Robust unit commitment problem with demand
Earthquake Spectra, vol. 30, no. 1, pp. 231252, Feb. 2014. response and wind energy, in Proc. IEEE PES General Meeting,
[5] A. Kwasinski, Effects of notable natural disasters from 2005 to 2011 Jul. 2012, pp. 18.
on telecommunications infrastructure: Lessons from on-site damage [28] S. Shao, M. Pipattanasomporn, and S. Rahman, Grid integration of
assessments, in Proc. INTELEC, Oct. 2011, pp. 19. electric vehicles and demand response with customer choice, IEEE
[6] A. Kwasinski, Effects of hurricanes Isaac and sandy on data and com- Trans. Smart Grid, vol. 3, no. 1, pp. 543550, Mar. 2012.
munications power infrastructure, in Proc. IEEE INTELEC, Oct. 2013, [29] Z. Tan, P. Yang, and A. Nehorai, An optimal and distributed demand
pp. 16. response strategy with electric vehicles in the smart grid, IEEE Trans.
[7] L. Dueas-Osorio and A. Kwasinski, Quantification of lifeline system Smart Grid, vol. 5, no. 2, pp. 861869, Mar. 2014.
interdependencies after the 27 February 2010 Mw 8.8 offshore Maule, [30] J. M. Lujano-Rojas, C. Monteiro, R. Dufo-Lpez, and
Chile, earthquake, Earthquake Spectra, vol. 28, no. S1, pp. S581S603, J. L. Bernal-Agustn, Optimum residential load management strategy
Jun. 2012. for real time pricing (RTP) demand response programs, Energy Policy,
[8] A. Kwasinski, Lessons from Field Damage Assessments about Commu- vol. 45, pp. 671679, Jun. 2012.
nication Networks Power Supply and Infrastructure Performance During [31] H. A. Aalami, M. P. Moghaddam, and G. R. Yousefi, Demand response
Natural Disasters With a Focus on Hurricane Sandy, document 11-60, modeling considering interruptible/curtailable loads and capacity market
FCC, Feb. 2013. programs, Appl. Energy, vol. 87, no. 1, pp. 243250, Jan. 2010.
[9] D. Robinson et al., Impact of distributed energy resources on the [32] V. L. Erickson and E. E. Cerpa, Occupancy based demand response
reliability of a critical telecommunications facility, Sandia Nat. Lab. HVAC control strategy, in Proc. 2nd ACM Workshop Embedded Sens.
Nat. Renew. Energy Lab., Tech. Rep. SAND2006-1277, Mar. 2006. Syst. Energy-Efficiency Building (BuildSys), 2010, pp. 712.
[10] G. M. Grant et al., Emergency diesel engine generator power system [33] M. Houwing, R. R. Negenborn, and B. De Schutter, Demand response
reliability 19871993, Idaho Nat. Eng. Lab., Idaho Falls, ID, USA, with micro-CHP systems, Proc. IEEE, vol. 99, no. 1, pp. 200213,
Tech. Rep. INEL-9510035, Feb. 1996. Jan. 2011.
[11] A. Kwasinski and P. T. Krein, Telecom power planning for natural [34] P. Li, P. Degobert, B. Robyns, and B. Franois, Participation in the
and man-made disasters, in Proc. Int. Telecommun. Energy Conf. frequency regulation control of a resilient microgrid for a distribu-
(INTELEC), Sep./Oct. 2007, pp. 216222. tion network, Int. J. Integr. Energy Syst., vol. 1, no. 1, pp. 15,
[12] A. Kwasinski, V. Krishnamurthy, J. Song, and R. Sharma, Availability Jan./Jun. 2009.
evaluation of micro-grids for resistant power supply during natural [35] C. M. Colson, M. H. Nehrir, and R. W. Gunderson, Distributed multi-
disasters, IEEE Trans. Smart Grid, vol. 3, no. 4, pp. 20072018, agent microgrids: A decentralized approach to resilient power system
Dec. 2012. self-healing, in Proc. 4th Int. Symp. Resilient Control Syst. (ISRCS),
[13] A. Kwasinski, Quantitative evaluation of DC microgrids availabil- Aug. 2011, pp. 8388.
ity: Effects of system architecture and converter topology design [36] C. M. Colson, M. H. Nehrir, and R. W. Gunderson, Multi-agent
choices, IEEE Trans. Power Electron., vol. 26, no. 3, pp. 835851, microgrid power management, in Proc. 18th IFAC World Congr., Milan,
Mar. 2011. Italy, Sep. 2011, pp. 36783683.
[14] A. Kwasinski and A. Kwasinski, Role of energy storage in a microgrid [37] L. Che and M. Shahidehpour, DC microgrids: Economic operation and
for increased use of photovoltaic systems in wireless communication enhancement of resilience by hierarchical control, IEEE Trans. Smart
networks, in Proc. IEEE INTELEC, Sep./Oct. 2014, pp. 18. Grid, vol. 5, no. 5, pp. 25172526, Sep. 2014.
[15] A. Kwasinski and A. Kwasinski, Operational aspects and power archi- [38] V. Meyer, C. Myres, and N. Bakshi, The vulnerabilities of the power-
tecture design for a microgrid to increase the use of renewable energy in grid system: Renewable microgrids as an alternative source of energy,
wireless communication networks, in Proc. IEEE Int. Power Electron. J. Bus. Continuity Emergency Planning, vol. 4, no. 2, pp. 142153,
Conf. (IPEC ECCE-Asia), May 2014, pp. 26492655. Mar. 2010.
[16] A. Kwasinski, Examination of power supply options for communication [39] H. Kanchev, D. Lu, F. Colas, V. Lazarov, and B. Francois, Energy
sites operating in grid-islanded environments, in Proc. IEEE INTELEC, management and operational planning of a microgrid with a PV-based
Sep./Oct. 2012, pp. 18. active generator for smart grid applications, IEEE Trans. Ind. Electron.,
[17] A. Kwasinski, Numerical evaluation of communication networks vol. 58, no. 10, pp. 45834592, Oct. 2011.
resilience with a focus on power supply performance during natural [40] J. M. Guerrero, J. C. Vasquez, J. Matas, L. G. de Vicua, and M. Castilla,
disasters, in Proc. INTELEC, Oct. 2015, pp. 16. Hierarchical control of droop-controlled AC and DC microgrids
[18] Y. Kwon, A. Kwasinski, and A. Kwasinski, Microgrids for base A general approach toward standardization, IEEE Trans. Ind. Electron.,
stations: Renewable energy prediction and battery bank management vol. 58, no. 1, pp. 158172, Jan. 2011.
for effective state of charge control, in Proc. INTELEC, Oct. 2015, [41] V. Nasirian, S. Moayedi, A. Davoudi, and F. L. Lewis, Distributed
pp. 16. cooperative control of dc microgrids, IEEE Trans. Power Electron.,
[19] J. Song, V. Krishnamurthy, A. Kwasinski, and R. Sharma, Develop- vol. 30, no. 4, pp. 22882303, Apr. 2015.
ment of a Markov-chain-based energy storage model for power supply [42] B. Patterson, DC, come home: DC microgrids and the birth of the
availability assessment of photovoltaic generation plants, IEEE Trans. Enernet, IEEE Power Energy Mag., vol. 10, no. 6, pp. 6069,
Sustain. Energy, vol. 4, no. 2, pp. 491500, Apr. 2013. Nov./Dec. 2012.
KWON et al.: COORDINATED ENERGY MANAGEMENT IN RESILIENT MICROGRIDS FOR WIRELESS COMMUNICATION NETWORKS 1173
[43] H. Kakigano, Y. Miura, and T. Ise, Configuration and control of a Youngsung Kwon (S11) received the B.S. degree
DC microgrid for residential houses, in Proc. Transmiss. Distrib. Conf. (Hons.) in electrical engineering from Soongsil Uni-
Expo., Asia Pacific, Oct. 2009, pp. 14. versity, Seoul, South Korea, in 2009, and the M.S.E.
[44] I. Aharon and A. Kuperman, Topological overview of powertrains for and Ph.D. degrees in electrical and computer engi-
battery-powered vehicles with range extenders, IEEE Trans. Power neering from the University of Texas at Austin,
Electron., vol. 26, no. 3, pp. 868876, Mar. 2011. Austin, TX, USA, in 2012 and 2015, respectively.
[45] L. Solero, A. Lidozzi, and J. A. Pomilio, Design of multiple-input He is currently with LG Electronics, Seoul, where
power converter for hybrid vehicles, IEEE Trans. Power Electron., he is involved in designing and developing electric
vol. 20, no. 5, pp. 10071016, Sep. 2005. vehicle charging system.
[46] P. Rodatz, G. Paganelli, A. Sciarretta, and L. Guzzella, Optimal
power management of an experimental fuel cell/supercapacitor-powered
hybrid vehicle, Control Eng. Pract., vol. 13, no. 1, pp. 4153,
Alexis Kwasinski (S03M07SM14) received
Jan. 2005.
the B.S. degree in electrical engineering from
[47] R. E. Hebner et al., Technical cross-fertilization between terrestrial
the Buenos Aires Institute of Technology (ITBA),
microgrids and ship power systems, J. Modern Power Syst. Clean
Buenos Aires, Argentina, in 1993, the Graduate
Energy, vol. 4, no. 2, pp. 161179, Apr. 2016.
Specialization degree in telecommunications from
[48] M. Kim, A. Kwasinski, and V. Krishnamurthy, A storage integrated
the University of Buenos Aires, Buenos Aires, in
modular power electronic interface for higher power distribution avail-
1997, and the M.S. and Ph.D. degrees in electri-
ability, IEEE Trans. Power Electron., vol. 30, no. 5, pp. 26452659,
cal engineering from the University of Illinois at
May 2015.
UrbanaChampaign, Champaign, IL, USA, in 2005
[49] M. Kim and A. Kwasinski, Decentralized hierarchical control of active
and 2007, respectively.
power distribution nodes, IEEE Trans. Energy Convers., vol. 29, no. 4,
From 1993 to 1997, he was with the Telefonica
pp. 934943, Dec. 2014.
de Argentina, Buenos Aires, designing and planning telephony outside plant
[50] R. Zhao and A. Kwasinski, Controller analysis for active distribution
networks. Then he was with Lucent Technologies Power Systems (later Tyco
nodes in advanced dc power systems, in Proc. IEEE COMPEL,
Electronics Power Systems) for five years as a Technical Support Engineer and
Jun. 2012, pp. 17.
a Sales Technical Consultant in Latin America in Mesquite, TX. He was also
[51] M. Kim and A. Kwasinski, Distribution interface for microgrid oper-
a Part-Time Instructor and was in charge of the ITBA Telecommunications
ation and expansion with local energy management, in Proc. IEEE
Laboratory for three years Buenos Aires, Argentia. From 2007 to 2014,
PEDG, Jun. 2014, pp. 18.
he was a Faculty Member with the University of Texas at Austin, Austin,
[52] National Oceanic and Atmospheric Administration (NOAA),
TX, USA, where he reached the rank of Tenured Associate Professor. He
accessed on Mar. 1, 2014. [Online]. Available: http://www.
is currently an Associate Professor and a R. K. Mellon Faculty Fellow
forecast.weather.gov/MapClick.php?lat=30.19452690400044&lon=-97.
with the University of Pittsburgh, Pittsburgh, PA, USA. His current research
6698695449997#.Vn_wNxV96Cr
interests include power electronic systems, distributed generation (microgrids),
[53] Measurement Frameworks and Metrics for Resilient Networks and
renewable and alternative energy, smart grids, and the analysis of the impact of
Services: Challenges and Recommendations. Heraklion, Greece, Eur.
natural disasters on critical power infrastructure, which included site damage
Netw. Inf. Secur. Agency (ENISA), 2010.
assessments after several natural disasters in Japan, New Zealand, the U.S.,
[54] A. Kwasinski, Field technical surveys: An essential tool for improving
and Chile.
critical infrastructure and lifeline systems resiliency to disasters, in
Dr. Kwasinski was a member of the Executive Committee of the Argentine
Proc. IEEE GHTC, Oct. 2014, pp. 7885.
Electrotechnical Association in 1994 and 1995. He was a recipient of
[55] M. Panteli and P. Mancarella, The grid: Stronger, bigger, smarter?:
the Joseph J. Suozzi INTELEC Fellowship in 2005, the Best Technical
Presenting a conceptual framework of power system resilience, IEEE
Paper Award at International Communications Energy Conference in 2007,
Power Energy Mag., vol. 13, no. 3, pp. 5866, May/Jun. 2015.
the National Science Foundation CAREER Award in 2009, and an IBM
[56] A. Kwasinski. Quantitative Model and Metrics of Electrical
Faculty Innovation Award in 2011. He is an Associate Editor of the IEEE
Grids Resilience Evaluated at a Power Distribution Level, Ener-
T RANSACTIONS ON E NERGY C ONVERSION.
gies, vol. 9, no. 2, pp. 127, Feb. 2016. [Online]. Available:
http://dx.doi.org/10.3390/en9020093
[57] H. Bindner, T. Cronin, P. Lundsager, J. F. Manwell, Andres Kwasinski (S98M04SM13) received
U. Abdulwahid, and I. Baring-Gould, Lifetime modelling of lead the Diploma degree in electrical engineering from
acid batteries, Ris Nat. Lab., Roskilde, Denmark, Tech. Rep., the Buenos Aires Institute of Technology, Buenos
Apr. 2005. Aires, Argentina, in 1992, and the M.S. and Ph.D.
[58] T. Lambert, P. Gilman, and P. Lilienthal, Micropower System degrees in electrical and computer engineering from
Modeling With Homer. New York, NY, USA: Wiley, 2006, the University of Maryland at College Park, College
pp. 379418. Park, MD, USA, in 2000 and 2004, respectively.
[59] U.S. Battery, accessed on Mar. 1, 2015. [Online]. Available: He is currently an Associate Professor with the
http://www.usbattery.com/wp-content/uploads/2015/02/usb_250_group_ Department of Computer Engineering, Rochester
web_2015b.pdf Institute of Technology, Rochester, NY, USA. He has
[60] Bergey Wind Power, accessed on Mar. 1, 2015. [Online]. Available: co-authored over 60 publications in peer-reviewed
http://bergey.com/documents/2012/05/excel-10-swcc-summary-report. journals and international conferences, and the books Cooperative Com-
pdf munications and Networking (Cambridge University Press, 2009) and 3-D
[61] G. Auer et al., Energy Efficiency Analysis of the Reference Sys- Visual Communications (Wiley, 2013). His current research interests include
tems, Areas of Improvements and Target Breakdown, document smart infrastructures, with an emphasis on the use of renewable energy,
INFSO-ICT-247733 EARTH, Energy Aware Radio and network cognitive radios and wireless networks, cross-layer techniques in wireless
tecHnologies (EARTH) Project deliverable D2.3, 2010. communications, multimedia communications, and networking and operation
[62] A. Kwasinski and A. Kwasinski, Tradeoff between quality-of- of the Internet-of-Things.
service and resiliency: A mathematical framework applied to LTE Dr. Kwasinski is an Area Editor of the IEEE Signal Processing Magazine,
networks, in Proc. IEEE Int. Conf. Commun. (ICC), May 2016, and an Editor of the IEEE T RANSACTIONS ON W IRELESS C OMMUNICA -
pp. 348353. TIONS and the IEEE W IRELESS C OMMUNICATIONS L ETTERS .