IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 16, NO.

9, SEPTEMBER 2017 6047

Cost-Efficient Cellular Networks Powered

by Micro-Grids
Ling Zhang, Yunlong Cai, Senior Member, IEEE, Qingjiang Shi, Guanding Yu, Senior Member, IEEE,
and Geoffrey Ye Li, Fellow, IEEE

Abstract— This paper investigates a cellular network powered I. I NTRODUCTION

by a micro-grid (MG) in the context of green communications,
which integrates the conventional generators, energy storage
devices, and renewable energy generators, so as to supply
T HE growing demand for ubiquitous network access leads
to the explosive increase of base stations (BSs), which
consume a large amount of eletricity. As a result, cellu-
electricity to base stations (BSs). Under this model, we study
the efficiency aspect of the MG-powered cellular network from lar operators are faced with drastically increasing electric-
the economical perspective. Specifically, the concept of cost ity bills. Furthermore, consuming huge amount of power
efficiency (CE) is employed to measure the sum rate delivered in the rapidly developing cellular networks deteriorates the
per dollar. Then, our goal is to maximize this CE subject to a ecological environment owing to greenhouse gas emissions.
series of constraints, including multi-variable coupling and time
coupling constraints. Particularly, we assume the zero-forcing These economic and ecological concerns drive the need for
beamforming scheme employed by the BSs. To address this “green communications” solutions [1]–[8]. In the context of
established fractional CE optimization problem, we first apply “green communications”, renewable energy resources (RES)
the Dinkelbach method, and then propose a low-complexity algo- are exploited to power communication systems [9] since
rithm based on the alternating direction method of multipliers they are cheaper and cleaner than the conventional energy
approach to jointly schedule power generation in the MG and
optimize transmit power for BSs. We introduce a number of resources, e.g., fossil fuels. Furthermore, with recent devel-
auxiliary variables to design a special variable splitting scheme opment of micro-grids (MGs), RES can be integrated into
so that the coupling inequality constraints can be separable the power grid. Therefore, developing MGs to power cellu-
among two variable sets. Consequently, the proposed algorithm lar networks has become an attractive solution. Particularly,
only incorporates simple updates in each step and thus can be MGs are power systems implemented in a small geographical
implemented in a parallel and completely distributed fashion.
Simulation results demonstrate the convergence and energy area [10]–[14] to supply electricity to local users. MGs have
scheduling performance of the proposed algorithm. many appealing features, including integration of RES, two-
way energy trading with the external market, dynamic price
Index Terms— Multi-cell systems, micro-grids, cost efficiency,
power scheduling, ADMM, completely distributed algorithm. based energy scheduling etc.
There have been a few studies on reducing energy cost in
Manuscript received November 21, 2016; revised April 24, 2017; accepted the MG-powered communication systems. Wang et al. [3] have
June 13, 2017. Date of publication June 27, 2017; date of current version developed the robust energy management for smart grid (SG)
September 8, 2017. The work of Y. Cai was supported in part by the National
Natural Science Foundation of China under Grant 61471319 and in part powered coordinated multi-point (CoMP) systems in order to
by the Fundamental Research Funds for the Central Universities. The work minimize the energy cost. Xu and Zhang [15] have aimed
of Q. Shi was supported by the National Natural Science Foundation of to minimize the total energy cost through cooperative energy
China under Grant 61671411 and Grant 61631020. The work of G. Yu was
supported by the Open Research Fund of National Mobile Communications trading in CoMP systems. Chiaraviglio et al. [17] have tackled
Research Laboratory, Southeast University, under Grant 2017D13. This paper the issue of energy-saving management for cellular access
was presented at the IEEE Globecom Conference, Washington, DC, USA, networks by reducing the number of active cells when the
December 2016. The associate editor coordinating the review of this paper
and approving it for publication was L. Musavian. (Corresponding author: network traffic is low. Jardosh et al. [18] have proposed to
Yunlong Cai.) power on or off wireless local area networks (WLAN) access
L. Zhang and Y. Cai are with the College of Information Science points (APs) dynamically according to the volume and the
and Electronic Engineering, Zhejiang University, Hangzhou 310027, China
(e-mail: [email protected]; [email protected]). location of users’ demand to reduce energy consumption.
Q. Shi is with the College of Electronic and Information Engineering, Bu et al. [19] have formulated the SG-powered cellular net-
Nanjing University of Aeronautics and Astronautics, Nanjing 210016, China work as a Stackelberg game to decide on the number of active
(e-mail: [email protected]).
G. Yu is with the College of Information Science and Electronic Engi- BSs and design the optimal electricity-purchasing scheme.
neering, Zhejiang University, Hangzhou 310027, China, and also with the Xu et al. [6] have presented novel approaches to reduce the
National Mobile Communications Research Laboratory, Southeast University, total energy cost in cellular networks by integrating renewable
Nanjing 210096, China (e-mail: [email protected]).
G. Y. Li is with the School of Electrical and Computer Engineer- energy resources and enabling two-way energy flow in the
ing, Georgia Institute of Technology, Atlanta, GA 30332 USA (e-mail: smart grid.
[email protected]). The above studies focus on saving energy or reducing
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. energy cost in the SG-powered communication systems, which
Digital Object Identifier 10.1109/TWC.2017.2718520 may not be sufficient for many applications since consuming
1536-1276 © 2017 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.
6048 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 16, NO. 9, SEPTEMBER 2017

less energy does not mean fully utilizing each unit mechanisms to improve efficiency, promote the robustness of
energy or money. Furthermore, in some circumstances, energy the system, and reduce the burden of the MGEM, we then
consumption is reduced at the cost of impairing users’ propose a low-complexity algorithm based on the alternating
utility. To address these issues, a more efficient metric is direction of multipliers (ADMM) [26]–[28] method to solve
expected both from the economical perspective and in the the transformed optimization problem in a distributed way.
interest of users. Motivated by the energy efficiency for Note that the ADMM technique has also been applied in
traditional communication systems [20]–[22], we pursue a other contexts of energy scheduling in the SG, e.g., [29].
concept of efficiency from the economical perspective for the However, the ADMM-based algorithm proposed in [29] cannot
MG-powered communication systems, which is referred to be performed in a completely distributed fashion due to the
as cost efficiency (CE). Particularly, the CE stands for the time coupling constraints. By contrast, our proposed low-
number of information bits delivered per dollar among all complexity ADMM-based algorithm allows for efficient paral-
BSs in our paper. What deserves to be mentioned is that lel and completely distributed implementation. To be specific,
Ma et al. [23] have also proposed the concept of cost with auxiliary variables introduced and a special variable
efficiency. However, our work and that in [23] address entirely splitting scheme designed, the variable coupling and time
different problem frameworks. To be specific, the work in [23] coupling constraints can be neatly decoupled. Furthermore,
has investigated the residential demand-side management in our proposed algorithm can converge to the optimal solutions
the SG while we focus on joint energy scheduling and transmit of the CE maximization problem.
power optimization in the MG-powered cellular networks. The contributions of this paper are summarized as follows:
In this paper, we formulate the CE maximization prob- i) We formulate a novel framework for the power schedul-
lem subject to a series of practiced constraints, including ing in the MG-powered multi-cell systems. Particularly,
multi-variable coupling and time coupling constraints, for the various energy resources are integrated into the MG
MG-powered cellular network. Particularly, various energy for powering BSs. Besides, the spatially and temporally
resources, i.e., RES, conventional generators (CGs) and energy varying prices are considered.
storage devices (ESDs), are integrated into the MG that ii) A new concept of efficiency metric, named CE, is pro-
powers the cellular network. Moreover, we consider spatially posed for system optimization and for the enhancement
and temporally varying electricity prices for these energy of mobile terminal (MT) users’ utility in this paper.
resources. Specifically, different kinds of energy resources To the best of our knowledge, no existing studies have
provide electricity at different prices; on the other hand, considered the CE optimization for cellular networks
different devices generating the same kind of energy may powered by MGs with RES even if the concept has been
correspond to different cost functions. In terms of time, proposed before and applied somewhere else.
the purchasing and selling market prices are also varying. iii) An efficient low-complexity ADMM-based algorithm is
Furthermore, each of the BSs, CGs, and ESDs is equipped developed to address the CE maximization problem.
with an local controller (LC) while all renewable energy With variable coupling and time coupling constraints
generators (REGs) are coordinated by one LC. These LCs decoupled via introducing auxiliary variables, the pro-
cooperate with the micro-grid energy manager (MGEM) for posed algorithm only incorporates simple updates in
information exchange and resource management in response each step and thus allows for parallel and completely
to the spatially and temporally varying electricity prices. distributed implementation. Furthermore, the proposed
We resolve this CE maximization problem via joint power algorithm is guaranteed to converge to the optimal
scheduling in the MG-powered multi-cell system. In partic- solution of the CE maximization problem.
ular, with respect to BSs, we consider the zero-forcing (ZF) This paper is organized as follows. Section II provides
beamforming scheme [24] .1 As a result, with the direction of the system model for the MG-powered multi-cell system and
the transmit beamforming vector fixed, we only optimize the formulates the CE maximization problem. In Section III,
transmit power for BSs. Moreover, owing to the randomness the Dinklebach method is first applied to the CE maximization
and intermittent characteristics of RES, the scenario of the problem and then the low-complexity ADMM-based algorithm
worst-case transaction cost is considered. For the formulated is proposed. In addition, the information exchange mecha-
CE maximization problem, it is worth pointing out that the nism for implementing the proposed algorithm is described.
fractional CE objective function as well as variable coupling Section IV presents simulation results to verify the effective-
and time coupling constraints impose new research challenges ness of the proposed algorithm. Section V summarizes the
on the energy scheduling in our considered system. To cope whole paper.
with these new challenges, we first resort to the Dinkelbach Notations: Boldface letters are used for vectors or matri-
method [25] to address this nonlinear fractional program- ces while normal font for scalar quantities. In addition,
ming by transforming it into a convex optimization prob- we use square brackets [·, ·, · · · ] to denote a row vec-
lem. Next, in order to allow more flexible electricity pricing tor or a bock matrix. The projection functions to the non-
negative and the nonpositive orthants are denoted by [a]+ =
1 For simplicity, we consider the ZF beamforming approach in this work max{a, 0} and [a]− = min{a, 0}, respectively. The superscript
since we want to emphasize the energy scheduling. However, the proposed (·)T and (·) H represents the transpose and conjugate transpose
power scheduling scheme is not limited to this ZF beamforming. Other beam-
forming techniques can also be applied by making appropriate modification (Hermitian) for a vector or a matrix, respectively. (·) denotes
to the sub-problem in (54) related to the transmit power allocation for BSs. the optimal value. We also use | · | to denote the absolute
ZHANG et al.: COST-EFFICIENT CELLULAR NETWORKS POWERED BY MICRO-GRIDS 6049

TABLE I
L IST OF N OTATIONS

the aid of the ZF beamforming scheme in [24] and only

optimize the transmit power at the BSs. We assume that the
associated normalized ZF beamforming vector for MT n is
denoted by w̃nt . Then the transmitted signal for MT n is
given by

xnt = Px,n
t w̃ t s (t),
n n (1)
where the information signal, sn (t), is assumed to be a
complex random variable with zero mean and unit variance.
Therefore, the received signal at MT n can be expressed as
ynt = ht H x t +v nt , (2)
 n n
desired signal
where the intercell interference is assumed to be completely
Fig. 1. The micro-grid powered multi-cell system. removed using the ZF beamforming scheme, and v nt is the
additive circular complex Gaussian noise at MT n with zero
mean and variance σn2 , n ∈ Nc .
value of a real scalar. Some other notations are summarized With the aid of ZF beamforming, the signal-to-interference-
in Table I. plus-noise-ratio (SINR) at MT n per time slot can be
expressed as
II. S YSTEM M ODEL AND P ROBLEM F ORMULATION t gt
Px,n n
We consider a multi-cell downlink system that is com- SINRtn = , (3)
σn2
posed of Nc cells. Each cell consists of one BS, which is
equipped with Nt antennas and serves one single-antenna MT. where gnt = |htnH w̃nt |2 , n ∈ Nc , t ∈ T . The data rate of MT n
As in Fig. 1, the multi-cell system is powered by an MG in time slot t is given by
that comprises M CGs, J ESDs, and I distributed REGs. Rnt = log(1 + SINRtn ). (4)
Furthermore, we consider a finite scheduling horizon consist-
ing of T time slots. Particularly, the slot duration is normalized B. Energy Scheduling Model
to unity in this paper, thus the terms “power” and “energy” In this part, we focus on the energy scheduling model in
can be thereby used interchangeably. For more practical appli- the MG that powers BSs.
cations, the time slot duration can be selected to be equal 1) CGs: Due to the operational limits, the change of the
to the minimum time interval between two adjacent wireless outputs of each CG in two consecutive time slots is bounded,
channel realizations. The slot-based transmission from the BSs which is typically reflected by ramping constraints:
to MTs is assumed. The details of our model and the problem t −1 up
t
PG,m − PG,m ≤ PG,m , m ∈ M , t ∈ T ,
formulation will be introduced in the following subsections. t −1 t down
PG,m − PG,m ≤ PG,m ,m ∈ M,t ∈ T . (5)
A. Multi-Cell Downlink Transmission
The quadratic CG generation cost function is given by [29]:
In this part, we introduce the multi-cell downlink transmis-
sion model. We denote the channel vector from the BS in the Cmt (PG,m
t
) = am (PG,m
t
)2 + b m , m ∈ M , t ∈ T . (6)
nth cell to its MT in time slot t as htn ∈ C Nt ×1 , n ∈ Nc , t ∈ T ,
and the associated information signal as sn (t). For simplicity, 2) REGs: Owing to the unpredictable and intermittent
we fix the direction of the transmit beamforming vector with nature of RES, we first introduce an uncertainty model.
6050 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 16, NO. 9, SEPTEMBER 2017

Define et = [E 1t , . . . , E tI ], t ∈ T , which is assumed to 

I
given E it since it is the composition of pointwise maximum
lie in an uncertainty set E t . For computational tractability, i
we consider a joint uncertainty model across all REGs during and affine functions.
t
each time slot, which is given by = Et ≤ {E it |E ti
≤ E it Ei ,
I C. Problem Formulation
t
E min ≤ i=1 E it ≤ E max
t }. That is to say, at each time slot,

the renewable energy harvested at each REG is bounded by Based on the models in Subsections II-A and II-B, we for-
t
E ti and E i while the total renewable energy harvested at all mulate the CE maximization problem in this subsection.
REGs is bounded by E min t t . All REGs are managed
and E max We aim to jointly schedule the energy generation in the MG
by one LC, which decides how much renewable energy to and optimize transmit power at the BSs so as to maximize
be delivered directly to the BSs and how much to be used the CE, which is defined as the number of information bits
for charging ESDs. The price of renewable energy can be delivered among all BSs during the considered time horizon
measured by the allowances that are paid by the MG operators for each dollar. Mathematically, this problem can be formu-
in order to encourage the use of RES. We use π E $/kWh to lated as
 t t
denote the allowances that the MG operators pay for per unit Rn (Px,n )
renewable energy. t,n
max  t t  t )+ H t (|t |) (11a)
t
3) ESDs: Denote Pch, t
j and PB, j the power drawn from the
A G (Pnet )+ Cmt (PG,m j j
t < 0) or delivered t t,m t, j
REGs only and the power drawn from (PB, j
to (PB, j > 0) other energy resources, respectively. Hence,
t subject to (5), (7) and
the stored energy obeys the dynamic equation: 0 ≤ PG,m
t
≤ PG,m
max
, m ∈ M,t ∈ T (11b)
Q tj = Q tj−1 +tj , j ∈ J, t ∈ T . (7) 0 ≤ Pch,t
j ≤ Pch, j , j ∈ J, t ∈ T
max
(11c)
J I
where tj = t
Pch, j − PB, j . It can be observed
t
from (7) that t
Pch, ≤ E it , t ∈ T (11d)
j
tj > 0 indicates an increase in the energy level and tj < 0 j =1 i=1
represents a decrease. The storage cost related to the net
− Pch, j ≤ PB, j ≤ PB,
max t
j , j ∈ J, t ∈ T
max
(11e)
charging, tj , is imposed to prevent frequent charging and
discharging behaviors, which is given by [30]  j = Pch, j − PB, j , j ∈ J, t ∈ T
t t t
(11f)
j ≤ Q j ≤ Q j , j ∈ J, t ∈ T
Q min t max
H jt (|tj |) = c j (tj )2 + dj. (8) (11g)
Px,nt gt
4) Worst-Case Transaction Cost: Note that Pnet t equals n
≥ γn , n ∈ Nc , t ∈ T (11h)
the net power delivered to the MG from the REGs and σn2
the external market so as to maintain the supply-demand 0 ≤ Px,nt
≤ Px,n max
, n ∈ Nc , t ∈ T (11i)
balance at time slot t. Therefore, the shortage energy that Pnet ≤ Pnet ≤ Pnet
min t max
, t ∈T (11j)
needs to be purchased from the market can be expressed as Nc
I t + M
t −
[Pnet i=1 E i ] , and the surplus energy that can be sold
I t − (Px,n
t
+ Pc,n ) ≤ t
PG,m
on market is given by [Pnet t −
i=1 E i ] . We assume the n=1 m=1
purchasing price to be π p and the selling price to be πs . Also J
note that we regard the renewable energy allowances as a part − tj + Pnet
t
,t ∈ T (11k)
of the transaction cost. Therefore, the worst-case transaction j =1
cost can be expressed as
 where the quantity A denotes the set of all the primal variables
I I
{Px,n
t , P t , Q t , P t , P t , t , P t }. (11b) stands for the
G t (Pnet
t
) = max π p [Pnet t
− E it ]+ +πs [Pnet
t
− E it ]− G,m j ch, j B, j j net
et ∈E t minimum/maximum limit of CG outputs, and (11c), (11e),
i i
and (11g) bound the amount of charging, discharging, and ESD
I
capacity, respectively. (11d) limits the total charging of REGs
+ πE E it , t ∈ T . (9)
to no more than the total renewable energy harvested. (11h) is
i
the set of QoS constraints for the BSs, with γn representing the
In order to provide more insights into the convexity of target SINR value at MT n per time slot. (11i) corresponds to
G t (Pnet
t ) in (9), we rewrite it as
the minimum/maximum transmit power for each BS during
 I each time slot. (11k) specifies the supply-demand balance
G t (Pnet
t
) = max (π p −πs )[Pnett
− E it ]+ constraint, that is, the total energy consumption at the BSs
et ∈E t
i plus the net charging to ESDs must be no more than the total
I I energy supply in the MG. (11j) limits the net power delivered
+ πs (Pnet
t
− E it )+π E E it , t ∈ T . (10) from REGs and the market to the range of [Pnet min , P max ].
net
i i It can be observed that the constraints (5), (7),
It can be observed that, to guarantee the convexity of G t (Pnet
t ), and (11b)-(11k) are all linear while the objective function
we should always set π p > πs . Fortunately, it is indispensable is nonlinear fractional. It is worth noting that the numerator
to impose this condition so as to avoid meaningless buy-and- of the objective function is concave and the denominator
sell activities. As a result, G t (Pnet
t ) is convex in P t for any
net is quadratic. Hence, if we transform the objective function
ZHANG et al.: COST-EFFICIENT CELLULAR NETWORKS POWERED BY MICRO-GRIDS 6051

TABLE II constraints, (11f), (11d), and (11k), and the time coupling
S UMMARY OF THE D INKELBACH M ETHOD constraints, (5) and (7). To overcome this difficulty, we first
introduce auxiliary variables and then design a special vari-
able splitting scheme so as to make the coupling constraints
separable between two optimization variable sets. Following
this, we propose the low-complexity ADMM-based algorithm.
Let us introduce some auxiliary variables to facilitate the
split of variables. From the viewpoint of information exchange,
t
Pch, j should first be decided by the REGs and then be
used to compute the net charging to ESDs. As can be seen,
t
Pch, j is shared among two constraints, i.e., (11d) and (11f),
(i.e. the fractional form) into a numerator-denominator sub- for REGs and ESDs, respectively. In order to derive separable
tractive form, then the resultant problem turns out to be a t
subproblems, a local auxiliary copy of Pch, j is introduced:
convex one. This is our basic idea to deal with the fractional t t
programming (11a), which is realized by using the Dinkelbach P̂ch, j = Pch, j (14)
method [25]. which is stored at ESDs for calculating the net charging tj
based on the dynamic equation (11f). Moreover, to tackle the
III. T HE P ROPOSED A LGORITHM FOR P ROBLEM (11) time-coupling constraints in (5) and (7), we introduce local
t t
In this section, we aim to solve the problem in (11). To this auxiliary copies P̃G,m for PG,m , and Q̃ tj for Q tj , namely
end, we first apply the Dinkelbach method to reformulate the t −1
1
P̃G,m = PG,m
0
, P̃G,m
t
= PG,m for t = 2 ∼ T , (15a)
problem in (11) as a convex optimization problem. Then we
propose a low-complexity ADMM-based algorithm to solve Q̃ 1j = Q 0j , Q̃ tj = Q tj−1 for t = 2 ∼ T . (15b)
the reformulated problem in a completely distributed fashion. 1
Note that, for the first time slot, the auxiliary variable P̃G,m
1
stores the initial CG outputs and Q̃ j stores the initial ESD
A. Dinkelbach Method t , t = 2 ∼ T stores the CG
energy level. After that, P̃G,m
In this part, we use the Dinkelbach method to transform the outputs from the last time slot and Q̃ tj , t = 2 ∼ T stores the
problem in (11) to a convex optimization problem. ESD energy level at the end of the last time slot. In addition,
Lemma 1: Define S  { A|(5), (7), (11b)−(11k)}. Then A in order to transform inequality constraint (5) into equality,
solves the problem in (11) if and only if it solves we introduce variable
t t t
max Rnt (Px,n
t
)−η G t (Pnet
t
)+ Cmt (PG,m
t
) VG,m = PG,m − P̃G,m , (16)
A∈S
t,n t t,m
which denotes the CG output variation in two consecutive time
+ H jt (|tj |) (12) slots. Hence, constraints (5), (7), and (11f) can be rewritten as
t, j
follows, respectively,
up
with η = η given by −PG,m ≤ VG,m
t
≤ PG,m
down
, m ∈ M,t ∈ T (17)
 t )
Rnt (Px,n Q tj = Q̃ tj +tj , j ∈ J, t ∈ T (18)
 t,n
η    . tj = j − PB, j , j ∈ J, t ∈ T .
(13) t t
t  )+
G t (Pnet t  )+
Cmt (PG,m H jt (|tj |) P̂ch, (19)
t t,m t, j t
By virtue of the introduced auxiliary variables P̃G,m and Q̃ tj ,
Lemma 1 can be proven by a similar approach as in [25].
It is clear that η is just the maximum CE. Let us denote the time coupling constraints (5) and (7) are now decoupled
over each time slot. For the power supply-demand balance
the optimal value of problem (12) by F(η). Then, as implied
by Lemma 1, the maximum CE, η , satisfies F(η ) = 0. constraint (11k), we introduce several copies that are stored at
the MGEM, i.e.
In fact, the Dinkelbach method is an iterative algorithm that
generates a sequence of values of η iteratively converging to t = Pt ,
P̂x,n x,n
t
P̂G,m = PG,m
t ˆ t = t .
,  (20)
j j
the maximum CE monotonically such that F(η ) = 0. The
As a result, we can rewrite (11k) as
Dinkelbach method for solving problem (11) is summarized
Nc M J
in Table II. ˆ tj + Pnet
( P̂x,n
t
+ Pc,n ) ≤ t
P̂G,m −  t
, t ∈T. (21)
n=1 m=1 j =1
B. Low-Complexity ADMM-Based Algorithm for Solving (12)
Besides, to facilitate the analysis of the convergence
In this part, we propose the low-complexity ADMM-based (cf. Proposition 1), we also introduce the following auxiliary
algorithm for solving problem (12). For a brief introduction variables:
to the ADMM method, the readers can refer to Appendix A. t t t t
We emphasize that the difficulty in designing the V̂G,m = VG,m , P̂B, j = PB, j (22a)
distributed algorithm comes from the variable coupling t
P̂net,s = t
Pnet , t
P̂net,G = t
Pnet , t
P̂net,x = t
Pnet . (22b)
6052 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 16, NO. 9, SEPTEMBER 2017

Then, by employing these new variables, problem (12) can Moreover, E = [ I, · · · , I , 0] and F = [0, I, · · · , I ] select
     
be equivalently expressed as T −1 T −1
the first T −1 and the last T −1 blocks of vectors, respectively.
max Rnt (Px,n
t
)−η G t t
(Pnet ) Besides, S f = [ I, 0, · · · , 0] selects the first block of vectors.
 t,n t
  
T −1
Therefore, problem (23) can be rewritten as the standard form
+ Cmt (PG,m
t
)+ H jt (|tj |)
t,m
(cf. (33)):
t, j
s.t. (11b)−(11e), (11g)−(11 j ), (14)−(21), max Rnt (Px,n
t
)−η G t (Pnet
t
)
xa ,xb ,za ,zb
(22a), (22b). (23) t,n t

where the quantity  = { A, t

P̂ch, j,
t
P̃G,m , Q̃ tj , VG,m
t
, + Cmt (PG,m
t
)+ H jt (|tj |)
t ˆ
P̂x,n , P̂G,m ,  j , V̂G,m , P̂B, j , P̂net,s , P̂net,G , P̂net,x }.
t t t t t t
It is t t,m t, j
important to point out that the constraints (11b) − (11e), s.t. Ca xa = Da ẑ a , Cb xb = Db ẑ b , xa , xb ∈ C1 , z a , z b ∈ C2
(11g) − (11j), (17) and (21) in problem (23) are now (28)
separable between two optimization variable sets:
where C1 and C2 are feasible sets corresponding to the
1) The tuple [xaT , xbT ]T , where xa = [[Pch, t , j ∈ J],
j set of constraints of (11c)-(11e), (11j), (21) and (11b),
t ∈ T ], [[PB, j , j ∈ J], t ∈ T ], [Pnet , t ∈ T ],
t t
 (11g)-(11i), (17), respectively. Now, we can apply ADMM
[[ ˆ t , j ∈ J], t ∈ T ], [[ Q̃ t , j ∈ J], t ∈ T ] T , and method to problem (28).
j j
xb = [[ P̂G,m t , m ∈ M ], t ∈ T ], [[ P̃ t , m ∈ M ], t ∈ Next, we construct the partial augmented
G,m 
t , n ∈ N ], t ∈ T ] T . Lagrangian (AL) function for problem (28).
T ], [[V̂G,m
t
, m ∈ M ], t ∈ T ], [[ P̂x,n c
Let us use {δch,
t
, θ̃ t
, δ̃ t
, , θ t
, δ t
, δ t
V ,m Q, j , j ,
2) The tuple [z aT , z bT ]T , where z a = [ P̂net,x t , t ∈ T ], j G,m Q, j

[ P̂net,G , t ∈ T ], [ P̂net,s , t ∈ T ], [[ P̂ch, j , j ∈ J],

t t t
t , θ̂ t
θx,n G,m , δ̂, j , θ̂ V ,m , θ B, j , δnet,s , δnet,G , δnet,x } to denote the
t t t t t t

t ∈ T ], [[ P̂B, Lagrange multipliers corresponding to the equality constraints

j , j ∈ J], t ∈ T ], [[ j , j ∈ J],
t t
 T in (14)−(20), (22a) and (22b), respectively. By collecting
t ∈ T ], [[Q tj , j ∈ J], t ∈ T ] , and z b = [[PG,m t
,m ∈ {δch, j , δ̃ Q, j , δ Q, j , δ, j , δ̂, j , δnet,s , δnet,G , δnet,x } into vector
t t t t t t t t
M ], t ∈ T ], [[VG,mt , m ∈ M ], t ∈ T ], [[P t , n ∈ N ],
T x,n c
δ, and {θ̃G,m t , θVt ,m , θx,n
t , θ̂ t
G,m , θ̂ V ,m , θ B, j } into vector θ ,
t t
t ∈T ] . we can express the partial AL function for (28) as follows:
The rest (14)-(16), (18)-(20),(22a) and (22b) are all linear L ρ1 ,ρ2 (xa , xb , z a , z b ; δ, θ )
equality constraints. Furthermore, these equality constraints
can be written into Ca xa = Da ẑ a and Cb xb = Db ẑ b , with = Rnt (Px,n
t
)
t,n
ẑ a = [z aT , [Q 0j , j ∈ J]]T , ẑ b = [z bT , [PG,m
0
, m ∈ M ]]T
⎡ ⎤T −η G t (Pnet
t
)+ Cmt (PG,m
t
)+ H jt (|tj |)
I 0 0 0 0 0 0 0 0 0 t t,m t, j
⎢0 I 0 0 0 0 0 I 0 0 ⎥
⎢ ⎥ ρ1
Ca = ⎢ ⎥ + δ (Ca xa − Da ẑ a ) − Ca xa − Da ẑ a 2
T
⎢0 0 I I I 0 0 0 0 0 ⎥ (24)
2
⎣0 0 0 0 0 I 0 0 0 0 ⎦ ρ2
0 0 0 0 0 0 I 0 F Sf + θ (Cb xb − Db ẑ b ) − Cb xb − Db ẑ b 2 ,
T
(29)
2
⎡ ⎤T
I 0 0 0 0 0 where ρ1 > 0 and ρ2 > 0 denote constant AL parameters.
⎢ 0 0 0 I F Sf ⎥ The key step of ADMM applied to problem (28) is to run
Cb = ⎢ ⎣0 I 0 0 0 0 ⎦
⎥ (25) one-iteration alternating optimization to address the following
0 0 I 0 0 0 AL problem:
⎡ ⎤T
0 0 I 0 0 0 0 0 0 0 max L ρ1 ,ρ2 (xa , xb , z a , z b ; δ, θ )
xa ,xb ,za ,zb
⎢0 0 0 I 0 0 0 0 0 0⎥
⎢ ⎥ s.t. xa , xb ∈ C1 , z a , z b ∈ C2 . (30)
⎢0 0 0 0 I 0 0 0 0 0⎥
⎢ ⎥
⎢I 0 0 0 0 0 0 I 0 0⎥ The proposed algorithm is summarized in Table III, which
Da = ⎢ ⎢ ⎥ (26)
⎥ is referred to as the low-complexity ADMM-based algo-
⎢0 I 0 0 0 0 0 0 0 0⎥
⎢ 0 0 0 0 0 I −I −I 0 0 ⎥ rithm (LCABA). Particularly, the proposed LCABA schedules
⎢ ⎥
⎣0 0 0 0 0 0 I 0 E 0⎦ the energy generation in the MG and allocates the trans-
0 0 0 0 0 0 0 0 0 I mit power at the BSs ahead of time based on the pre-
⎡ ⎤T dicted wireless channel realizations, the uncertainty regions
I 0 0 I I 0 of the renewable energy generation amount, and the energy
⎢ 0 I 0 −I 0 0 ⎥
Db = ⎢ ⎣0 0 I 0 0 0⎦
⎥ (27) prices, and thus it is an offline algorithm. The detailed
derivation of each updating step of LCABA is given in
0 0 0 0 0 I Appendix B. Also note that the convergence of our pro-
where I and 0 denote the identity matrix and the posed LCABA to the optimal solutions of problem (23) is
zero matrix with appropriate dimension, respectively. readily implied by Proposition 1 in Appendix A (it can be
ZHANG et al.: COST-EFFICIENT CELLULAR NETWORKS POWERED BY MICRO-GRIDS 6053

TABLE III
S UMMARY OF THE P ROPOSED L OW-C OMPLEXITY ADMM-BASED A LGORITHM (LCABA).

checked that CaT Ca , CbT Cb , DaT Da and DbT Db are all full-rank in preparation for calculating output change VG,m t and the
matrices.) t
output at the present time slot PG,m (cf. Step (v) in B-B).
Moreover, for each BS n, it updates its transmit power Px,n t

via Step (iii) in Appendix B-B. Up to this time, BSs’ energy

C. Information Exchange Mechanism of LCABA consumption {Px,n t , n ∈ N }, ESDs’ discharging {P t , j ∈ J},
c B, j
In this section, we elaborate how the LCABA is imple- and net charging {tj , j ∈ J}, REGs’ total direct energy supply
mented in a parallel and completely distributed fashion. I 
J
By referring to Appendix C, the LCABA is decoupled across E it − t
Pch, j , and CGs’ generation {PG,m , m ∈ M } have
t
i=1 j =1
time slots. For simplicity, we consider the implementation in all been obtained. To maintain the supply-demand balance, this
time slot t and assume that the algorithm is performed by the information need to be broadcast to the MGEM for computing
MGEM and the LCs. the transaction amount traded with the external market. Note
Let us first take a look at the implementation for REGs that, based on the supply-demand balance constraint in (11k),
and ESDs. Consider the update of the optimization variable the MGEM only requires the knowledge of {tj , j ∈ J} and
set {Pch,t
j , j ∈ J} (cf. Step (i) in Appendix B-A). This I
updating step is in fact performed by the LC of REGs instead E it without knowing specifically about {Pch,
t
j , j ∈ J} and
of ESDs since the LC of REGs decides that how much i=1
{PB,
t ,
j j ∈ J}. Once such information is conveyed to the
renewable energy generated should be charged to each ESD
MGEM, it makes local copies { P̂x,n t , n ∈ N }, {
c ˆ t , j ∈ J} and
and how much should be delivered to the MG. Then the j
updated Pch, t
j is transmitted to the LC of each ESD j . After
{ P̂G,m , m ∈ M }, and utilize these local copies to compute the
t

receiving Pch, t
j , each ESD j makes a local copy of it, that t −
I
t . To this point, ESDs have been aware of the amount
transaction amount, which can be represented as Pnet E it
is P̂ch, j i=1
of charging. Besides, each ESD schedules the discharging (cf. Step (iii) in Appendix B-A).
amount PB, t Based on the above description of the information exchange
j so that the energy consumption can be satisfied.
Thus optimization variable PB, t mechanism, LCABA can be implemented in a parallel manner
j is updated (cf. Step (i) in
t t for BSs part, REGs part, CGs part, ESDs part, and the
Appendix B-A). After P̂ch, j and PB, j updated, each ESD j MGEM. Within each part, it can be performed in a completely
can compute the ESD net changing tj (cf. Step (iv) in
distributed way among the LCs.
Appendix B-B). It is worth pointing out that, at the start of
each time slot, each ESD j would store the energy level from IV. S IMULATION R ESULTS
the last time slot, namely Q tj−1 , as Q̃ tj in order to calculate In this section, we provide simulation results to evaluate
the energy level at the end of time slot t according to (18). the performance of LCABA for the MG-powered multi-cell
Next, we discuss the implementation for CGs, i.e. the system. For comparison, we also demonstrate the performance
update of CG output {PG,m t , m ∈ M } as well as the CG
of the dual decomposition algorithms [32], [33], i.e., the
output change in two successive time slots {VG,m t
, m ∈ M }. dual decomposition algorithm for CE maximization referred
To this end, at the beginning of each time slot t, each CG to as DDA-1 and the dual decomposition algorithm for cost
t −1 t
m will store the output in the last time slot PG,m as P̃G,m minimization referred to as DDA-2 that solve our considered
6054 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 16, NO. 9, SEPTEMBER 2017

Fig. 2. Convergence Performance of (a) CE; (b) r.

TABLE IV 5 iterations to almost the same level in terms of CE, with5.88

R ELEVANT PARAMETERS OF THE CG S bits/dollar and 5.89 bits/dollar, respectively. However, the con-
stant CE for DDA-2 only attains 0.91 bits/dollar, which is more
than 4.97 bits/dollar lower than the LCABA and DDA-1.
To get an insight into this huge discrepancy, we list the
optimal CEs, total costs, transmit power Px,n t , and sum-rate for

these three algorithms in Table VIII. From the table, DDA-2

focuses on cost minimization and cuts down the total cost
from more than 600 dollars to 588.35 dollars at the expense
CE maximization problem in (11) and the cost minimization of a significant decrease in sum-rate. In particular, the transmit
problem (cf. problem (31)), respectively. The dual decom- power, Px,nt , for the LCABA and DDA-1 always attains the
position algorithms, DDA-1 and DDA-2, are based on the maximum value for all n ∈ Nc and all t ∈ T resulting in
Lagrange dual decomposition technique and the optimization the sum-rate of 3534.50 bits. However, DDA-2 always attains
software package, e.g., CVX. Particularly, the cost minimiza- the lowest allowable QoS target leading to the sum-rate of only
tion problem considered in our simulation is given by 532.66 bits. That is to say, the cost minimization problem
min G t (Pnet
t
)+ Cmt (PG,m
t
)+ H jt (|tj |) focuses on a low budget but fails to make full use of each
A dollar. For this reason, we use CE in the context of “green
t t,m t, j
s.t. (5), (7), (11b) − (11 j ). (31) communications” to achieve higher MT users’ utility with less
money.
We assume that all BSs and MTs have the same set of The convergence performance of the time-averaged resid-
max = P max = 48 dBm, P
parameters, i.e., Px,n c,n = Pc = ual r corresponding to constraint (11k) is depicted in Fig. 2b.
70 dBm, γn = γ = 20 dB and σn2 = σ 2 = −85 dBm. It is observed that r approaches 0 after around 700 iterations
Moreover, we adopt the complex Gaussian distributed channel for ρ1 = ρ2 = 0.01 and after about 200 iterations for
with zero mean and unit variance. The considered MG com- ρ1 = ρ2 = 0.1. Obviously, as the values of ρ1 and ρ2 getting
prises M = 3 CGs, J = 2 ESDs and I = 5 REGs. The CG larger, the convergence speed of the LCABA accelerates.
cost function is assumed to be Cmt (PG,m t ) = a Pt2 + b
m G,m m Next, we present the optimal power scheduling for PG,m t
,
and the ESD cost is set to be H j (| j |) = c j |tj |2 + d j . The
t t
Pch, j , PB, j ,  j , Q j and the transaction amount with the
t t t t
relevant parameters of the CGs, ESDs, and REGs as well as the t − Ẽ t , (except P t , the scheduling for which
market, i.e., Pnet i x,n
time-varying purchasing and selling market prices are listed has been discussed above). Specifically, the optimal power
in Tables IV, V, VI and VII.
 I Furthermore, for REGs,
 I we set scheduling for PG,m t
and the transaction amount are depicted
t t
E i = 10E ti , E min
t
= 1.2 i=1 E ti , E max
t = 0.9 i=1 E i and in Fig. 3a. Note that we take CG 3 as an example and the
π E = 0.1$/kWh. In particular, the time horizon considered other CGs’ outputs will be discussed later. Besides, the optimal
spans T = 8 time slots. For LCABA, we set ρ1 = ρ2 = 0.01. power scheduling for PB, t , t and Q t is depicted in Fig. 3b.
j j j
First, we examine the convergence performance of LCABA Moreover, we only depict the power scheduling results for
by investigating the convergence of the CE and the time- the first ESD in Fig. 4a since the two ESDs incorporated in
averaged residual r of constraint (11k), which is given by our model have the same parameters. Particularly, the power
T
1  Nc t
M
t
J
 scheduling results for the LCABA and DDA-1 are both
r= (Px,n + Pc,n )− PG,m + tj − Pnet
t
. (32) depicted in Figures. 3-4a. Note that the optimal solutions of
T
t =1 n=1 m=1 j =1 these two algorithms may be different, but they change with
Particularly, the convergence performance of the CE is the same trend. Specifically, from Fig. 3a, the highest levels of
depicted in Fig. 2a for the LCABA, DDA-1, and DDA-2. t
PG,3 occur from time slots 4 to 6. This is because that this time
From the figure, the LCABA and DDA-1 both converge within period is just the electricity usage peak with relatively higher
ZHANG et al.: COST-EFFICIENT CELLULAR NETWORKS POWERED BY MICRO-GRIDS 6055

TABLE V
R ELEVANT PARAMETERS OF THE ESD S

Fig. 3. Power scheduling (a) for CG output and transaction amount and (b) for net charging, energy level and discharging.

Fig. 4. (a) Power scheduling for charging; (b) Effects of cost function parameters on CG outputs.

TABLE VI TABLE VIII

L IMITS OF THE F ORECASTED R ENEWABLE T HE O PTIMAL C OST E FFICIENCIES , C OSTS , T RANSMIT
E NERGY G ENERATION (100W/ H ) t AND S UM -R ATE
POWER Px,n

TABLE VII
is seen in Fig. 3a, the transaction with the market achieves
T HE T IME - VARYING M ARKET P RICES ($/ K W H )
the lowest amount during this time period, i.e. time slots 4-6.
In Fig. 3b, a relative large amount of ESD discharging also
occurs from time slots 4 to 6 in response to the relatively
higher market prices. As a consequence, there is a drop of the
ESD energy level from time slot 4. Also note that, in Fig. 4a,
the charging amount reaches the minimum in time slot 4 for
purchasing and selling prices. As a result, CG 3 generates as the reason that much of the generated renewable energy is
much electricity as possible so as to minimize the transaction delivered directly to the MG so as to reduce the transaction
cost that is induced by trading with the market. Accordingly, as amount. From Fig. 4a, the charging amount gradually increases
6056 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 16, NO. 9, SEPTEMBER 2017

Fig. 5. (a) Time-varying price; (b) CE versus time slots.

Fig. 6. (a) CG generation cost and total cost; (b) ESD cost and transaction cost for different price settings.

TABLE IX
C OMPUTATIONAL C OMPLEXITY C OMPARISON

after time slot 4 in order to keep the ESD energy level above CG 1, 2, and 3. Particularly, in order to elaborate the effects of
the minimum allowable value. From the results, the proposed spatially varying prices on power scheduling, we investigate
algorithm is capable of scheduling the energy resources in the the optimal power scheduling for the geographically spread
MG in response to the variations in electricity prices so as to CGs, which have different cost function parameter settings
enhance the system CE as much as possible. from each other. From the cost function parameter settings
The complexity of the LCABA and DDA-1 can be evaluated in Table IV, CG 1 has the highest generation cost while CG 2
by the execution time when implemented by the software, has the lowest generation cost. This can also be verified by
e.g., CVX, and is shown in Table IX, where K and L denote Fig. 4b. Note that CG 1 yields the least output over all time
the iteration number for the Dinkelbach method and the iter- slots due to too high generation cost while the cheapest CG 2
ation number within the LCABA or the DDA-1, respectively. generates the most electricity during time slots 1, 2, 7 and 8.
From the table, with the same values of K and L, the pro- However, from time slots 3 to 6, CG 2 can only attain the
posed LCABA saves about 95% of execution time compared upper limits of generation, i.e. 20 kWh due to the limits of
with DDA-1. Therefore, the LCABA not only has almost the maximum output. As a result, CG 2 generates the second
the same performance with DDA-1 in terms of the maxi- most electricity from time slots 3 to 6 only next to CG 3 that
mum CE and the optimal solutions, but also can reduce the has higher output upper-limits, i.e., 30 kWh.
computational complexity greatly. The significant reduction of To investigate how the time-varying prices can affect
execution time makes a big difference in practical applications, the CE, we consider two cases with different time-varying
especially when the energy scheduling is expected to keep price settings as shown in Fig. 5a. Fig. 5b shows the CE
up with the fast changes of wireless channels, arrival rate of of each price setting versus time slots. Since the transaction
renewable energy resources and variations in electricity prices. amount always keeps positive in our simulation (cf. Fig. 3a),
Finally, the effects of the spatially and temporally varying we can only investigate the effects of purchasing price. As can
prices on power scheduling and CE are shown in Fig. 4b for be observed, the CE corresponding to Case A is higher than
ZHANG et al.: COST-EFFICIENT CELLULAR NETWORKS POWERED BY MICRO-GRIDS 6057

 A ∈ R , B ∈m R , 
that corresponding to Case B over all time slots, since the where l×n l×m and c ∈ Rl ; f : Rn →

purchasing price in Case A is much lower than that in Case B R {+∞} and g : R → R {+∞} are convex func-
in each time slot. It is also seen that, in Case A, the CE tions; C1 and C2 are non-empty convex sets. The partial
varies in the opposite direction to the purchasing price on the augmented Lagrangian function for problem (33) can be
whole, however, in Case B, it follows the same trend with formulated as
the purchasing price. This is contrary to our common sense
L ρ (x, y, z) = f (x) + g(z) + yT ( Ax + B z − c)
that the higher purchasing price the lower CE. To gain an
ρ
insight into this abnormality, we depict the CG generation cost +  Ax + B z − c22 , (34)
and total cost in Fig. 6a and ESD cost and transaction cost 2
in Fig. 6b. As is shown in Fig. 6a, there is a big fall in the total where y ∈ R is the Lagrangian dual variable vector associated
l

cost for Case B from time slot 4 to 5. This is because CGs with the linear equality constraint, and ρ > 0 is a prede-
contribute a lot to the electricity generation during these time fined constant. The ADMM algorithm solves problem (33)
slots with remarkable increasing CG cost shown in Fig. 6a. by sequentially performing the following three steps at each
Accordingly, from time slot 4 to 5, the transaction cost in iteration k:
 
Case B is significantly reduced to almost the same level as x (k+1) = arg min L ρ x, z (k), y(k) (35a)
that in Case A as depicted in Fig. 6b. That is to say, too high x∈C1
purchasing price during electricity usage peak, i.e. from time  
z (k+1) = arg min L ρ x (k+1), z, y(k) (35b)
slots 4 to 5, inspires the potential of electricity generation z∈C2
in the MG and therefore greatly improves the CE during that  
y(k+1) = y(k) + ρ Ax (k+1) + B z (k+1) − c . (35c)
period of time. In summary, we would like to point out that the
proposed CE maximization algorithm responds to the spatial Note that the dual variable update uses a step size equal
and temporal variations in the electricity prices timely, and it is to the augmented Lagrangian parameter ρ. The subprob-
able to adjust the power scheduling results accordingly in order lems (35a) and (35b) are solvable in closed-form when both
to achieve the optimal CE. Therefore, by designing appro- the constraints sets and the objective function are simple. The
priate pricing mechanisms, the system CE can be improved convergence and the optimality of the ADMM algorithm is
further. summarized in the following proposition [26]–[28]:
Proposition 1: Assume that the optimal solution set
V. C ONCLUSION AND F UTURE W ORK of problem (33) is non-empty, and AT A and B T B are
In this paper, we have studied the efficiency aspect invertible. Then the sequence of {x (k), z (k) , y(k) } generated
of the MG-powered multi-cell system from the economic by (35a), (35b) and (35c) is bounded, and every limit point of
perspective. Specifically, a concept of efficiency in the context {x (k), z (k) } is an optimal solution of problem (33).
of “green communications”, i.e. CE, has been introduced to
characterize the information bits delivered among all BSs at A PPENDIX B
each dollar. We have aimed to maximize the CE by joint D ERIVATION OF U PDATING S TEPS IN LCABA
energy scheduling in the MG and transmit power optimization
In this part, we derive each of the update procedure of
at BSs. In order to solve the formulated CE maximization
LCABA.
problem, we have first applied the Dinkelbach method and then
proposed the LCABA. Particularly, our proposed algorithm
can be implemented in a parallel and completely distributed A. Solving step 2(1) for LCABA
fashion. Simulation results show the convergence performance
and optimal power scheduling results for our proposed algo- We solve the problem in step 2(1) to optimize the first
rithm in comparison with the dual decomposition algorithms. block of variables xa , xb by fixing the second block of
Furthermore, the proposed algorithm provides offline ahead- variables z a , z b and the Lagrange dual variable vectors δ, θ .
of-time energy scheduling in the MG-powered multi-cell sys- Particularly, the variables {PB, t
j }, {Pch, j }, { V̂G,m }, { Q̃ j } and
t t t

tem based on the predicted data. Development of online energy { P̃G,m } can be solved separately in closed-form solutions
t

scheduling will be an interesting research topic to explore in by utilizing the first-order optimality condition, while the
the future work. Besides, integration of renewable energy and coupling variables {Pnet,t }, { P̂x,n t }, {ˆt } and { P̂ t } can
j G,m
smart-grid technologies into the communication system will be first decoupled with the Lagrange dual decomposition
be a promising direction to pursue. method. Afterwards, the variables {Pnet,t } can be solved
via bisection search, while the rest variables { P̂x,n t }, {ˆt }
j
A PPENDIX A and { P̂G,m } can be solved separately by using the first-
t

B RIEF R EVIEW OF ADMM A PPROACH order optimality condition. Detailed derivations are as
The algorithm of ADMM solves problems with the follow- follows:
ing structure [27]: (i) Subproblems for {PB, t }, {P t } and { Q̃ t }: The subprob-
j ch, j j

min f (x) + g(z) lems for {PB, j }, {Pch, j } and { Q̃ tj } can all be decomposed
t t

x∈Rn ,z∈Rm over each ESD j and each time slot t. As a result, for each
subject to Ax + B z = c, x ∈ C1 , z ∈ C2 , (33) j ∈ J and t ∈ T , the following optimization problems
6058 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 16, NO. 9, SEPTEMBER 2017

are solved: t  to (43) can be obtained as

while the optimal solution P̃G,m
ρ1 t t −1
max δ tB, j (PB,
t
j − P̂B, j )−
t
|P − P̂B,
t
j|
2
ρ2 (PG,m
t −VG,m
t + PG,m )− θ̃G,m
t −θVt ,m
−Pch,
max≤P t ≤P max
j B, j B, j
2 B, j t
P̃G,m = . (45)
ρ1 t 2ρ2
+ δ,
t
j ( P̂ch, j − PB, j − j )−
t t t
| P̂ch, j − PB,
t
j − j | ,
t 2
ˆ t , P̂ t }: For this subset
2 (iii) Subproblem for {Pnet t , P̂ t , 
x,n j G,m
(36) of variables, the corresponding updating procedure only
J
 t ρ2 t  relates to power constraint (21) and is decoupled over
max δch, j (Pch,
t
j − P̂ch, j )−
t
|Pch, j − P̂ch,
t
j|
2
each time slot. Particularly, for each time slot, the fol-
0≤Pch, j≤Pch, j
t max 2
j =1 lowing optimization problem is solved
J I t t
max δnet,x (Pnet − P̂net,x )
j≤ E it ,
t
s.t. Pch, (37) ˆ t , P̂ t
min ≤ P t ≤ P max , P̂ t ,
Pnet net net x,n j G,m
j =1 i=1 ρ1
ρ1 t − t
|Pnet − P̂net,x |2 +δnet,G
t
(Pnet − P̂net,Gt
)
max δ tQ, j (Q tj − Q̃ tj − tj ) − |Q j − Q̃ tj − tj |2 2
Q̃ tj 2 ρ1
− |Pnet − P̂net,Gt
|2 +δnet,s
t
(Pnet − P̂net,s
t
)
ρ1 t −1 2
+ δ̃ tQ, j (Q tj−1 − Q̃ tj ) −
|Q j − Q̃ tj |2 . (38) ρ1
2 − |Pnet − P̂net,st
|2 −ηG t (Pnet t
)
The optimization problems (36)-(38) can all be solved by 2
J
 t 
utilizing the first-order optimality condition. Specifically,
+ ˆ tj )− ρ1 |tj − 
δ̂, j (tj −  ˆ tj |2
the optimization problem (36) is a box-constrained uni- 2
j =1
variat quadratic optimization problem, with the optimal
solution PB,t M
 ρ2 t 
t t t t
j
+ θ̂G,m (PG,m − P̂G,m )− |PG,m − P̂G,m |2
2
t
 δ tB, j − δ,
t
j +ρ1 ( P̂B, j + P̂ch, j − j )
t t t m=1
PB, = min max{ , Nc
j
2ρ1  t ρ2 t 
 + θx,n (Px,n
t
− P̂x,n
t
)− |Px,n − P̂x,n
t 2
|
− Pch,
max
j }, PB, j .
max
(39) 2
n=1
The optimal solution {Pch, t  } for the optimization prob- Nc M J
j
lem (37) can be obtained as follows s.t. ( P̂x,n
t
+Pc,n ) ≤ t
P̂G,m− ˆ tj+Pnet
 t
.
δch,  j =1
j −λ +
t n=1 m=1
t t max
Pch, j = min{[ P̂ch, j + ] , Pch, j }, (40) (46)
ρ1
where λ is the optimal Lagrange dual variable with Note that, given t ,
Pnet G t (Pnet
t )
can be explicitly obtained,
respect to the coupling constraint, and can be chosen as shown in Appendix C. Let us denote the optimal
via bisection method to ensure the complementarity and Lagrange dual variable corresponding to the coupling
feasibility. As with the optimization problem (38), its constraint in (46) as λt , then both the optimal solution
t  of problem (46) and λ can be obtained via bisec-
optimal solution Q̃ tj can be obtained as Pnet t
tion search. To be specific, λt is chosen such that the
ρ1 (Q tj − tj + Q tj−1 ) − δ tQ, j − δ̃ tQ, j complementarity and feasibility condition for the cou-
Q̃ tj = . (41) pling constraint in (46) are satisfied. For each given λt ,
2ρ1
we search for the optimal solution Pnet t  (λ ) in the range
t
(ii) Subproblems for {V̂G,m t
} and { P̃G,m
t
}: The subproblems of [Pnet , Pnet ] using the bisection method in order that
min max
for {V̂G,m } and { P̃G,m } can both be decoupled over each
t t
the gradient equals 0. Furthermore, given the optimal λt ,
CG m and each time slot t. Therefore, the following opti- the rest of the optimal solutions, i.e., { P̂x,n ˆ t  , P̂ t  },
t , 
t t j G,m
mization problems can be solved for each V̂G,m and P̃G,m , can be obtained by the first-order optimality condition:
respectively: t
ρ2 t P̂x,n = Px,n
t
− (θx,n
t
+ λt )/ρ2 (47)
max θ̂Vt ,m (VG,m
t
− V̂G,m
t
)− |VG,m − V̂G,m
t
|2 , (42) ˆ tj = tj − (δ̂, 
t
V̂G,m 2  t
j + λt )/ρ1 (48)
t
t −1 ρ2 t −1 P̂G,m = t
PG,m − t
(θ̂G,m + λt )/ρ2 . (49)
t t t
max θ̃G,m (PG,m − P̃G,m )− |P − P̃G,m |2
t
P̃G,m 2 G,m

+ θVt ,m (PG,m
t
− P̃G,m
t
−VG,m
t
) B. Solving step 2(2) for LCABA
ρ2 t In this step, we solve the problem in step 2(2) to
|P − P̃G,m
− t
− VG,m
t
|2 . (43) optimize the second block of variables z a , z b by fix-
2 G,m
ing the first block of variables xa , xb and the Lagrange
By employing the first-order optimality condition,
dual variable vectors δ and θ . Particularly, the variables
the optimal solution for (42) is given by t }, { P̂
{ P̂B, j net,x }, { P̂net,G }, { P̂net,s }, and {Px,n } can be solved
t
t
V̂G,m = VG,m
t
− θ̂Vt ,m /ρ2 , (44) in closed-form solutions separately by taking advantage
ZHANG et al.: COST-EFFICIENT CELLULAR NETWORKS POWERED BY MICRO-GRIDS 6059

ρ2 t
of the first-order optimality condition, while the vari- + δch,
t
j (Pch, j − P̂ch, j ) −
t t
|P − P̂ch,
t
j| )
2

able sets {tj , P̂ch, 2 ch, j

j , Q j } and {PG,m , VG,m } can be solved
t t t t
ρ1
iteratively by utilizing the alternating optimization algo- + δ tQ, j (Q tj − Q̃ tj − tj ) − |Q tj − Q̃ tj − tj |2
2
rithm, respectively. Detailed derivations are demonstrated as t t t t ρ1 t t t 2
+ δ, ( P̂
j ch, j − P B, j − j )− | P̂ − PB, j − j |
below: 2 ch, j
(i) Subproblem for { P̂B, t }: This subproblem is decoupled
j
− ηH jt (|tj |). (59)
across each ESD j and each time slot t. Then the t t
t Note that variables P̂ch, j and Q j are separable, although
following problem is solved for each P̂B, j: they both relate to tj . Thus, problem (59) can be solved
ρ1
max δ tB, j (PB,t
j − P̂B, j ) −
t
|P t − P̂B,
t
j | . (50)
2
via the alternating optimization algorithm. Specifically,
t
P̂B, 2 B, j for any given tj , P̂ch,t t
j
j (resp. Q j ) can be obtained while
t  can be obtained by applying t t
The optimal solution P̂B, Q j (resp. P̂ch, j ) is held fixed. Particularly, these optimal
j
the first-order optimality condition: solutions can be expressed as follows via the first-order
t optimality condition:
j = PB, j − δ B, j /ρ1 .
t t
P̂B, (51)
t
δ,
t
j −δch, j +ρ1 (Pch, j + PB, j + j )
t t t t

(ii) Subproblems for { P̂net,x

t }, { P̂net,G
t }, { P̂net,s
t }: This set of P̂ch, j ( j ) =
t
(60)
2ρ1
variables can be solved separately. Let us first consider +1 t +1
 δ tQ, j+δ̃ tQ, j +ρ1 ( Q̃ j + Q̃ j+ j )
t t
the subproblem corresponding to the variable set { P̂net,x t }, Q tj (tj ) = min max{ ,
2ρ1
which can be expressed as follows for each t: 
j }, Q j
Q min .
max
ρ1 (61)
max δnet,x
t
(Pnet − P̂net,x
t
)− |Pnet − P̂net,x t
|2 . (52)
t 2 Then, when P̂ch,t t
P̂net,x j and Q j are fixed, the problem for
t updating  j is univariate and can be solved by the first-
t
The optimal solution P̂net,x can be obtained by using the
order optimality condition:
first-order optimality condition:
t tj ( P̂ch,
t
j, Q j)
t
P̂net,x = Pnet
t
− δnet,x
t
/ρ1 . (53)
= (δ̂,
t t t ˆt
j −δ, j −δ Q, j +ρ1  j + ρ1 (Pch, j − PB, j )
t t
By following the same approach, we can obtain the
t
optimal solutions { P̂net,G t  }:
} and { P̂net,s + ρ1 (Q tj − Q̃ tj ))/(3ρ1 +2a j η). (62)
t By iteratively computing (60), (61) and (62) until the
P̂net,G = Pnet
t
− δnet,G
t
/ρ1 (54)
algorithm is convergent, then we can obtain the optimal
t
P̂net,s = Pnet
t
− δnet,s
t
/ρ1 . (55) solutions tj , P̂ch,
t t
j, Qj .
(iii) Subproblem for {Px,n t }: This subproblem is decoupled (v) Subproblem for {PG,m , VG,m
t t
}: The corresponding sub-
over each BS n and each time slot t. Then the following problem for this variable set is decoupled over each CG
subproblem can be solved to optimize each Px,n t : m and each time slot t, and then it is given by
t t t
ρ2 t max θ̂G,m (PG,m − P̂G,m )
max θx,n
t
(Px,n
t
− P̂x,n
t
)− |Px,n − P̂x,n
t 2
| t ≤P max ,−P u p ≤V t ≤P down
0≤PG,m G,m G,m G,m G,m
γn σn /gn≤Px,n≤Px,n
2 t t max 2 ρ2 t t +1 t +1
− |P − P̂G,m
t
|2 + θ̃G,m (PG,mt
−P̃G,m )
+ Rnt (Px,n
t
). (56) 2 G,m
ρ2 t t +1 2
By using the first-order optimality condition, the optimal − |PG,m − P̃G,m | + θVt ,m (PG,m t t
− P̃G,m t
−VG,m )
t  can be obtained by solving the following 2
solution Px,n ρ2 t
− |PG,m − P̃G,m
t
−VG,m
t
|2 + θ̂Vt ,m (VG,m
t
− V̂G,m
t
)
quadratic equation: 2
ρ2 t
t
a(Px,n )2 + b Px,n
t
+ c = 0, (57) − |VG,m − V̂G,m
t
|2 − ηCmt (PG,m t
). (63)
2
where a = −ρ2 , b = θx,n t + ρ P̂ t − ρ σ 2 /g t and c =
2 x,n 2 n n
Problem (63) can also be solved by the alternating
optimization algorithm. Specifically, we first fix VG,m t and
(θx,n + ρ2 P̂x,n )σn /gn + 1/ log 2. Therefore, the optimal
t t 2 t
t
t  can be obtained as
solution Px,n obtain the optimal solution PG,m (VG,m ) by applying the
t

√ first-order optimality condition as follows:

t −b ± b2 −4ac σn2 γn max t
Px,n = min{max{ , t }, Px,n }. (58) PG,m (VG,m
t
)
2a gn  t t +1 t +1
= min{[ θ̂G,m + θ̃G,m +θVt ,m +ρ2 ( P̂G,m
t
+ P̃G,m + P̃G,m
t
(iv) Subproblem for {tj , P̂ch, t
j , Q j }: The subproblem cor-
t
 +
responding to this set of variables can be decoupled over + VG,m ) /(3ρ2 +2ηam )] , PG,m }.
t max
(64)
each ESD j and each time slot t. Thus the optimization t
Then we fix PG,m and solve problem (63) for the optimal
problem for each ESD j and each time slot t is given by t
solution VG,m (PG,m
t
) via the first-order optimality condi-
ˆt ρ1 t ˆ tj |2 tion:
max δ̂,
t
j ( j −  j ) −
t
| j − 
 j , P̂ch, j ,Q j ≤Q j≤Q j
t t min t max 2 t t

VG,m (PG,m ) = min max{(θ̂Vt ,m −θVt ,m +ρ2 (PG,m
t t
−P̃G,m
+1 t +1 ρ1 t 
+δ̃ tQ, t
j (Q j − Q̃ j ) − |Q j − Q̃ tj+1 |2 +V̂G,m
t up
))/(2ρ2 ),−PG,m }, PG,m
down
. (65)
2
6060 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 16, NO. 9, SEPTEMBER 2017

TABLE X
A LTERNATING O PTIMIZATION A LGORITHMS FOR S OLVING S UBPROBLEMS (59) AND (63)

By iteratively computing (64) and (65) until the algo- [10] N. Hatziargyriou, H. Asano, R. Iravani, and C. Marnay, “Microgrids,”
rithm is convergent, then we obtain the optimal solutions IEEE Power Energy Mag., vol. 5, no. 4, pp. 78–94, Jul./Aug. 2007.
t  and V t  . [11] E. Dall’Anese, H. Zhu, and G. B. Giannakis, “Distributed optimal power
PG,m G,m flow for smart microgrids,” IEEE Trans. Smart Grid, vol. 4, no. 3,
Note that the algorithms for solving subproblems (59) and (63) pp. 1464–1475, Sep. 2013.
are listed in Table X. [12] Y. Zhang, N. Gatsis, and G. B. Giannakis, “Robust energy management
for microgrids with high-penetration renewables,” IEEE Trans. Sustain.
Energy, vol. 4, no. 4, pp. 944–953, Oct. 2013.
A PPENDIX C [13] S. Sun, M. Dong, and B. Liang, “Distributed real-time power balancing
S OLVING FOR W ORST-C ASE Ẽ it in renewable-integrated power grids with storage and flexible loads,”
IEEE Trans. Smart Grid, vol. 7, no. 5, pp. 2337–2349, Sep. 2015.
Recall (10), given Pnett , the optimal solution for the worst- [14] K. Rahbar, J. Xu, and R. Zhang, “Real-time energy storage manage-
case total renewable energy harvested during each time slot, ment for renewable integration in microgrid: An off-line optimization
approach,” IEEE Trans. Smart Grid, vol. 6, no. 1, pp. 124–134,

I
i.e. Ẽ t = E it , can be obtained via solving the following Jan. 2015.
i=1 [15] J. Xu and R. Zhang, “Cooperative energy trading in CoMP systems
optimization problem: powered by smart grids,” IEEE Trans. Veh. Technol., vol. 65, no. 4,
 pp. 2142–2153, Apr. 2015.
max (π p −πs )[Pnett − Ẽ t ]+ +π (P t − Ẽ t )+π Ẽ t . (66) [16] S. Zhou, J. Gong, Z. Yang, Z. Niu, and P. Yang, “Green mobile
s net E access network with dynamic base station energy saving,” in Proc. ACM
et ∈E t
MobiCom, Sep. 2009, pp. 10–12.
It can be readily checked that the objective function [17] L. Chiaraviglio, D. Ciullo, M. Meo, and M. A. Marsan, “Energy-
of (66) is non-increasing and concave since π p > πs aware UMTS access networks,” in Proc. 11th Int. Symp. Wireless Pers.
Multimedia Commun., Sep. 2008, pp. 8–11.
and π E < π p . Besides, considering the polyhedron uncertainty [18] A. P. Jardosh, K. Papagiannaki, E. M. Belding, K. C. Almeroth,
set E t , the globally optimal solution must be attained at the G. Iannaccone, and B. Vinnakota, “Green WLANs: On-demand WLAN
extreme points of E t . We assume that E min t is larger than infrastructures,” Mobile Netw. Appl., vol. 14, pp. 798–814, Dec. 2009.
I [19] S. Bu, F. R. Yu, Y. Cai, and X. P. Liu, “When the smart grid meets
t t
the sum of the lower bound of each E i , i.e. E min ≤ E ti , energy-efficient communications: Green wireless cellular networks pow-
i=1 ered by the smart grid,” IEEE Trans. Wireless Commun., vol. 11, no. 8,
as a consequence, the optimal solution Ẽ t  can be obtained pp. 3014–3024, Aug. 2012.
[20] G. Y. Li et al., “Energy-efficient wireless communications: Tutorial,
as Ẽ t  = E min
t
. Therefore, G t (Pnet
t ) for given P t
net can be survey, and open issues,” IEEE Wireless Commun. Mag., vol. 18, no. 6,
obtained herein. pp. 28–35, Dec. 2011.
[21] Q. Shi, C. Peng, W. Xu, M. Hong, and Y. Cai, “Energy efficiency opti-
mization for MISO SWIPT systems with zero-forcing beamforming,”
R EFERENCES IEEE Trans. Signal Process., vol. 64, no. 4, pp. 842–854, Feb. 2016.
[1] Z. Hasan, H. Boostanimehr, and V. K. Bhargava, “Green cellular net- [22] Y. Chen et al., “Fundamental trade-offs on green wireless networks,”
works: A survey, some research issues and challenges,” IEEE Commun. IEEE Commun. Mag., vol. 49, no. 6, pp. 30–37, Jun. 2011.
Surveys Tuts., vol. 13, no. 4, pp. 524–540, Nov. 2011. [23] J. Ma, H. H. Chen, L. Song, and Y. Li, “Residential load scheduling
[2] F. R. Yu, X. Zhang, and V. C. M. Leung, Green Communications and in smart grid: A cost efficiency perspective,” IEEE Trans. Smart Grid,
Networking. Boca Raton, FL, USA: CRC Press, 2012. vol. 7, no. 2, pp. 771–784, Mar. 2016.
[3] X. Wang, Y. Zhang, G. B. Giannakis, and S. Hu, “Robust smart- [24] Q. H. Spencer, A. L. Swindlehurst, and M. Haardt, “Zero-forcing meth-
grid-powered cooperative multipoint systems,” IEEE Trans. Wireless ods for downlink spatial multiplexing in multiuser MIMO channels,”
Commun., vol. 14, no. 11, pp. 6188–6199, Nov. 2015. IEEE Trans. Signal Process., vol. 52, no. 2, pp. 461–471, Feb. 2004.
[4] J. Xu and R. Zhang, “CoMP meets smart grid: A new communication [25] W. Dinkelbach, “On nonlinear fractional programming,” Manage. Sci.,
and energy cooperation paradigm,” IEEE Trans. Veh. Technol., vol. 64, vol. 13, no. 7, pp. 492–498, Mar. 1967.
no. 6, pp. 2476–2488, Jun. 2015. [26] D. P. Bertsekas and J. N. Tsitsiklis, Parallel and Distributed Computa-
[5] E. Oh, B. Krishnamachari, X. Liu, and Z. Niu, “Toward dynamic energy- tion: Numerical Methods. Belmont, MA, USA: Athena Scientific, 1997.
efficient operation of cellular network infrastructure,” IEEE Commun. [27] S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein, “Distributed
Mag., vol. 49, no. 6, pp. 56–61, Jun. 2011. optimization and statistical learning via the alternating direction method
[6] J. Xu, L. Duan, and R. Zhang, “Cost-aware green cellular networks with of multipliers,” Found. Trends Mach. Learn., vol. 3, no. 1, pp. 1–122,
energy and communication cooperation,” IEEE Commun. Mag., vol. 53, Jan. 2011.
no. 5, pp. 257–263, May 2015. [28] W.-C. Liao, M. Hong, H. Farmanbar, X. Li, Z.-Q. Luo, and H. Zhang,
[7] J. Xu, L. Duan, and R. Zhang, “Energy group buying with loading “Min flow rate maximization for software defined radio access net-
sharing for green cellular networks,” IEEE J. Sel. Areas Commun., works,” IEEE J. Sel. Areas Commun., vol. 32, no. 6, pp. 1282–1294,
vol. 34, no. 4, pp. 786–799, Apr. 2016. Jun. 2014.
[8] J. Gong, S. Zhou, and Z. Niu, “Optimal power allocation for [29] Y. Zhang and G. B. Giannakis, “Distributed stochastic market clearing
energy harvesting and power grid coexisting wireless communica- with high-penetration wind power,” IEEE Trans. Power Syst., vol. 31,
tion systems,” IEEE Commun. Mag., vol. 61, no. 7, pp. 3040–3049, no. 2, pp. 895–906, Mar. 2015.
Jul. 2013. [30] T. Li and M. Dong, “Real-time energy storage management with
[9] T. Han and N. Ansari, “Powering mobile networks with green energy,” renewable integration: Finite-time horizon approach,” IEEE J. Sel. Areas
IEEE Wireless Commun., vol. 21, no. 1, pp. 90–96, Feb. 2014. Commun., vol. 33, no. 12, pp. 2524–2539, Dec. 2015.
ZHANG et al.: COST-EFFICIENT CELLULAR NETWORKS POWERED BY MICRO-GRIDS 6061

[31] C. R. Inc. (Sep. 2012). CVX: Matlab Software for Disciplined Convex Guanding Yu (S’05–M’07–SM’13) received the
Programming. [Online]. Available: http://cvxr.com/cvx B.E. and Ph.D. degrees in communication engi-
[32] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge U.K.: neering from Zhejiang University, Hangzhou, China,
Cambridge Univ. Press, 2004. in 2001 and 2006, respectively. He joined Zhejiang
[33] L. Zhang, Y. Cai, Q. Shi, and G. Yu, “Cost efficiency optimization for University, where he is an Associate Professor with
multi-cell systems powered by micro-grids,” in Proc. IEEE GLOBE- the College of Information and Electronic Engi-
COM, Washington, DC, USA, Dec. 2016, pp. 1–6. neering. From 2013 to 2015, he was also a Vis-
iting Professor with the School of Electrical and
Computer Engineering, Georgia Institute of Tech-
nology, Atlanta, GA, USA. His current research
interests include energy-efficient communication and
green networks, device-to-device communications, full-duplex communica-
tions, LTE in unlicensed spectrum, and mobile edge computing.
Ling Zhang received the B.S. degree in electronic He received the Exemplary Reviewer Award from the IEEE T RANSACTIONS
engineering from the Ocean University of China ON C OMMUNICATIONS of the IEEE Communications Society (ComSoc) in
in 2015. She is currently pursuing the M.S. degree 2016 and 2017, and the IEEE ComSoc Asia-Pacific Outstanding Young
with the College of Information Science and Elec- Researcher Award in 2016. He regularly sits on the technical program
tronic Engineering, Zhejiang University, Hangzhou, committee (TPC) boards of prominent IEEE conferences, such as ICC,
China. Her interests include distributed adaptive GLOBECOM, and VTC. He has also served as the General Co-Chair for the
signal processing, energy efficient wireless commu- CRNet 2010 Workshop, and TPC Co-Chair for the Green Communications
nication, convex optimization, and smart grid. and Computing Symposium of Chinacom 2013. He has served as a Guest
Editor of the IEEE Communications Magazine Special Issue on Full-Duplex
Communications, an Editor of the IEEE J OURNAL ON S ELECTED A REAS IN
C OMMUNICATIONS Series on Green Communications and Networking, and
a Lead Guest Editor of the IEEE Wireless Communications Magazine Special
Issue on LTE in Unlicensed Spectrum. He is currently serving as an Associate
Editor of the IEEE W IRELESS C OMMUNICATIONS L ETTERS , an Editor of the
IEEE T RANSACTIONS ON G REEN C OMMUNICATIONS AND N ETWORKING,
Yunlong Cai (S’07–M’10–SM’16) received the B.S. and an Associate Editor of the IEEE ACCESS .
degree in computer science from Beijing Jiaotong
University, Beijing, China, in 2004, the M.Sc. degree
in electronic engineering from the University of
Surrey, Guildford, U.K., in 2006, and the Ph.D.
degree in electronic engineering from the University
Geoffrey Ye Li (S’93–M’95–SM’97–F’06) received
of York, York, U.K., in 2010. From 2010 to 2011, he
his B.S.E. and M.S.E. degrees in 1983 and 1986,
was a Post-Doctoral Fellow with the Electronics and
respectively, from the Department of Wireless Engi-
Communications Laboratory of the Conservatoire
neering, Nanjing Institute of Technology, Nanjing,
National des Arts et Metiers, Paris, France. Since
China, and his Ph.D. degree in 1994 from the
2011, he has been with the College of Information
Department of Electrical Engineering, Auburn Uni-
Science and Electronic Engineering, Zhejiang University, Hangzhou, China,
versity, Alabama.
where he is currently an Associate Professor. From 2016 to 2017, he was a
He was a Teaching Assistant and then a Lecturer
Visiting Scholar with the School of Electrical and Computer Engineering,
with Southeast University, Nanjing, China, from
Georgia Institute of Technology, Atlanta, GA, USA. His current research
1986 to 1991, a Research and Teaching Assistant
interests include transceiver design for multiple-antenna systems, sensor array
with Auburn University, Alabama, from 1991 to
processing, adaptive filtering, full-duplex communications, cooperative and
1994, and a Post-Doctoral Research Associate with the University of Maryland
relay communications, and wireless information and energy transfer.
at College Park, Maryland, from 1994 to 1996. He was with AT&T Labs -
Research at Red Bank, New Jersey, as a Senior and then a Principal Technical
Staff Member from 1996 to 2000. Since 2000, he has been with the School of
Electrical and Computer Engineering at the Georgia Institute of Technology as
an Associate Professor and then a Full Professor. He is also holding a Cheung
Kong Scholar title at the University of Electronic Science and Technology of
Qingjiang Shi received the Ph.D. degree in com- China since 2006.
munication engineering from Shanghai Jiao Tong His general research interests include statistical signal processing and
University, Shanghai, China, in 2011. From 2009 communications, with emphasis on cross-layer optimization for spectral- and
to 2010, he visited the Prof. Z.-Q. (Tom) Luo’s energy-efficient networks, cognitive radios and opportunistic spectrum access,
Research Group with the University of Minnesota, and practical issues in LTE systems. In these areas, he has published around
Twin Cities, MN, USA. In 2011, he was a Research 200 journal papers in addition to around 40 granted patents and numerous
Scientist with the Research and Innovation Center conference papers. His publications have been cited over 28,000 times and
(Bell Labs, China), Alcatel-Lucent, Shanghai. He he has been recognized as the World’s Most Influential Scientific Mind, also
was an Associate Professor with the School of known as a Highly-Cited Researcher, by Thomson Reuters. He has been
Information and Science Technology, Zhejiang Sci- involved in editorial activities for over 20 technical journals for the IEEE,
Tech University, Hangzhou, China, in 2013. He is including founding Editor-in-Chief of IEEE 5G Tech Focus. He has organized
currently a Professor with the College of Electronic and Information Engi- and chaired many international conferences, including technical program
neering, Nanjing University of Aeronautics and Astronautics, Nanjing, China. vice-chair of IEEE ICC’03, technical program co-chair of IEEE SPAWC’11,
His current research interests lie in algorithm design for signal processing in general chair of IEEE GlobalSIP’14, and technical program co-chair of IEEE
advanced MIMO, cooperative communication, physical layer security, energy- VTC’16 (Spring). He was awarded IEEE Fellow for his contributions to
efficient communication, wireless information, and power transfer. signal processing for wireless communications in 2005. He won 2010 Stephen
He received the National Excellent Doctoral Dissertation Nomination Award O. Rice Prize Paper Award, 2013 WTC Wireless Recognition Award, and
in 2013, the Shanghai Excellent Doctoral Dissertation Award in 2012, and the 2017 Award for Advances in Communication from the IEEE Communications
Best Paper Award at the IEEE PIMRC Conference in 2009. He is an Associate Society and 2013 James Evans Avant Garde Award and 2014 Jack Neubauer
Editor of the IEEE T RANSACTIONS ON S IGNAL P ROCESSING, and has served Memorial Award from the IEEE Vehicular Technology Society. He also
as a TPC member of the IEEE Globecom/ICC and peer reviewer for a variety received 2015 Distinguished Faculty Achievement Award from the School of
of IEEE journals and conferences. Electrical and Computer Engineering, Georgia Tech.