2814 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 14, NO.

5, MAY 2015

Energy Management and Cross Layer Optimization

for Wireless Sensor Network Powered by
Heterogeneous Energy Sources
Weiqiang Xu, Senior Member, IEEE, Yushu Zhang, Qingjiang Shi, Member, IEEE, and
Xiaodong Wang, Fellow, IEEE

Abstract—Recently, utilizing renewable energy for wireless sys- I. I NTRODUCTION

tem has attracted extensive attention. However, due to the instable
energy supply and the limited battery capacity, renewable energy
cannot guarantee to provide the perpetual operation for wireless
sensor networks (WSN). The coexistence of renewable energy and
W IRELESS sensor networks (WSN) consist of a lot of
spatially distributed autonomous sensor nodes with lim-
ited energy, computation and sensing capabilities, to monitor
electricity grid is expected as a promising energy supply manner physical phenomena, and to cooperatively transmit their data to
to remain function of WSN for a potentially infinite lifetime. In
this paper, we propose a new system model suitable for WSN, a sink. WSN have a variety of potential applications, ranging
taking into account multiple energy consumptions due to sensing, from multimedia surveillance, environmental monitoring, and
transmission and reception, heterogeneous energy supplies from advanced health care delivery to industrial process control.
renewable energy, electricity grid and mixed energy, and multi- Traditionally, sensor nodes are powered by a non-rechargeable
dimension stochastic natures due to energy harvesting profile, battery with limited energy storage capacities. However, a lot
electricity price and channel condition. A discrete-time stochastic
cross-layer optimization problem is formulated to achieve the opti- of applications are expected to operate over a virtually infinite
mal trade-off between the time-average rate utility and electricity lifetime. The energy scarcity represents one of the major limita-
cost subject to the data and energy queuing stability constraints. tions of WSN. Indeed, the post-deployment replacement of the
The Lyapunov drift-plus-penalty with perturbation technique and sensors batteries is generally not practical or even impossible.
block coordinate descent method is applied to obtain a fully Thus, a variety of hardware optimizations, energy management
distributed and low-complexity cross-layer algorithm only requir-
ing knowledge of the instantaneous system state. The explicit policies and energy-aware network protocols have been pro-
trade-off between the optimization objective and queue backlog posed to carefully manage the limited energy resources and thus
is theoretically proven. Finally, through extensive simulations, the to prolong the lifetime of a WSN [1]–[3].
theoretic claims are verified, and the impacts of a variety of system Recent advances in hardware design have made energy
parameters on overall objective, rate utility and electricity cost are harvesting (EH) technology possibly applied in wireless sys-
investigated.
tems. Sensor node equipped with EH device replenishes energy
Index Terms—Wireless sensor networks, energy management, from renewable sources with a potentially infinite amount of
energy harvesting, electricity grid, heterogeneous energy, cross- available energy [4]–[6]. Since EH technology is essentially
layer optimization, Lyapunov optimization, drift-plus-penalty,
block coordinate descent. different from the traditional non-rechargeable battery, a new
energy management policy is expected to well-match with the
Manuscript received July 9, 2014; revised November 9, 2014 and January 10, energy replenishment process. As such, a great deal of research
2015; accepted January 15, 2015. Date of publication January 21, 2015; date of efforts have been devoted to investigate the energy management
current version May 7, 2015. This work was supported in part by the National and data transmission in the EH powered scenario. Some ef-
Science Foundation of China (NSFC) under grants 61374020, 61302076,
61329101, 61272311, 61403347, by Key Project of Chinese Ministry of Edu- forts proposed the optimal schemes to achieve the maximum
cation under grant 212066, by Zhejiang Provincial Science Foundation under throughput, the minimum transmission completion time, and/or
grants LR15F010002, LQ13F010008, and by Open Research Fund of the State the minimum information distortion for a single EH node with
Key Laboratory of Integrated Services Networks, Xidian University under
grant ISN14-08. The associate editor coordinating the review of this paper finite or infinite data buffer and finite or infinite battery capacity
and approving it for publication was T. Melodia. (Corresponding author: [7]–[13].
Qingjiang Shi).
W. Xu and Q. Shi are with the School of Information Science and Technol- For wireless multihop network powered by EH, different
ogy, Zhejiang Sci-Tech University, Hangzhou 310018, China, and also with the nodes may have quite different workload requirements and
State Key Laboratory of Integrated Services Networks, Xidian University, Xian available energy sources. Due to the fact that the network
710071, China (e-mail: wqxu@zstu.edu.cn; qing.j.shi@gmail.com).
Y. Zhang is with the School of Information Science and Technology, performance is tightly coupled with energy management policy
Zhejiang Sci-Tech University, Hangzhou 310018, China (e-mail: yszhang@ and mechanisms at the physical, MAC, network, and trans-
zstu.edu.cn).
X. Wang is with the Electrical Engineering Department, Columbia Univer-
port layers, a limited amount of works investigated the cross-
sity, New York, NY 10027 USA, and also with King Abdulaziz University, layer optimization schemes in [14]–[18]. In particular, some
Jeddah 22254, Saudi Arabia (e-mail: wangx@ee.columbia.edu). works of cross-layer optimization leveraged Lyapunov opti-
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. mization techniques. Gatzianas et al. in [19] applied Lyapunov
Digital Object Identifier 10.1109/TWC.2015.2394799 techniques to design an online adaptive transmission scheme

1536-1276 © 2015 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.
XU et al.: MANAGEMENT AND OPTIMIZATION FOR WSN POWERED BY HETEROGENEOUS ENERGY SOURCES 2815

for wireless networks with rechargeable battery to achieve local information only. Some existing works, for instance, [20],
total system utility maximization and the data queue stability. [22], designed the partly distributed optimization solution in
Huang et al. in [20] applied Lyapunov optimization techniques WSN powered by EH. However, a partly distributed optimiza-
with weight perturbation [21] to achieve a close-to-optimal util- tion solution is still impractical, or too costly in large-scale
ity performance in finite energy buffer. The proposed technique networks. Fully distributed optimization solution is particularly
obtains an explicit and controllable tradeoff between optimality attractive.
gap and queue sizes. Similarly, by adopting perturbation-based This motivates us to address a novel energy management and
Lyapunov techniques, Tapparello et al. in [22] proposed the cross-layer optimization for WSN powered by heterogeneous
joint optimization scheme of source coding and transmission energy sources. The key contributions of this paper are summa-
to minimize the reconstruction distortion cost for the corre- rized as follows:
lated sources measurement. All the above-mentioned works
showed that network-wide cross layer optimization is help- 1) We propose a more realistic energy consumption model,
ful for achieving the performance gain. However, the works which takes the energy consumption of sensing, trans-
mentioned above are still not suitable to efficiently deal with mission and reception into account. We propose a new
the energy scarcity limitation of WSN. There are still several heterogenous energy supply model suitable for the node
technical challenges, including: powered by renewable energy or/and electricity grid. We
1) Multiple Energy Consumption: A sensor node is also consider the multi-dimension stochastic natures from
equipped with a sensing module for data measurements and channel condition, energy harvesting profile and electric-
processing, and a communication module for data transmission ity price. For such a model, we formulate a discrete-
and data reception. Almost all of the works mentioned above time stochastic cross-layer optimization problem in WSN
only account for the energy consumed in data transmission. with the goal of maximizing the time-average utility
Traditionally, energy consumption is known to be dominated by of the source rate and the time-average cost of energy
the communication module. However, this is not always true. consumption in electricity grid subject to the data and
In [23], it was shown that communication-related tasks were energy queuing stability constraints.
possibly less energy consumption than intensive processing, 2) To obtain a distributed and low-complexity solution, we
and data transmission is only a slight more energy consumption apply the Lyapunov drift-plus-penalty with perturbation
than data reception. There exists a very limited works in [11], technique [21] to transform the stochastic optimization
[13], [24] to investigate the problem of energy allocation problem into a series of iterations of the determin-
accounting for the energy requirement of data transmission and istic optimization problems. Furthermore, by exploit-
sensing together, only suitable for a single EH nodes. Almost ing the special structure, we design a fully distributed
no works, except [22], studied the joint energy allocation for algorithm—Energy mAnagement and croSs laYer Opti-
communication module and sensing module together in the mization (EASYO) which decomposes the deterministic
multihop scenario. optimization problem into the energy management (in-
2) Hybrid Energy Supply: Due to the low recharging rate cluding energy harvesting and energy purchasing), source
and the time-varying profile of the energy replenishment pro- rate control (implicitly including energy allocation for
cess, sensor nodes solely powered by harvested energy can not sensing/processing), routing selection (implicitly includ-
guarantee to provide reliable services for the perpetual oper- ing energy allocation for data reception), session schedul-
ation. They may currently be suitable only for very-low duty ing and transmission power allocation. EASYO is a fully
cycle devices. Other complementary stable energy supplies distributed algorithm which makes greedy decisions at
should be required to remain a perpetual operation for WSN. each time slot without requiring any statistical knowledge
As the electricity grid (EG) is capable of providing persistent of the channel state, of the harvestable energy state and
power input, the coexistence of renewable energy and electric- of the electricity price state. Note that our proposed
ity grid is considered as a promising technology to tackle the fully distributed algorithm is different from the cross-
problem of simultaneously guaranteeing the network operation layer optimization algorithms in [20], [22], where the
and minimizing the electricity grid energy consumption, which transmission power allocation problem is optimized in
had been confirmed in single-hop wireless system [25], [26]. the centralized manner, leading to the huge challenging
However, as far as we know, no prior work considered cross- in practical implementation.
layer optimization for WSN powered by heterogeneous energy 3) We analyze the performance of the proposed distributed
sources in multihop scenario. algorithm EASYO, and show that a control parameter
3) Fully Distributed Implementation: In WSN, the entire V enables an explicit trade-off between the average ob-
system state is characterized by channel condition, energy jective value and queue backlog. Specifically, EASYO
harvesting profile, electricity price, data queue size and energy can achieve a time average objective value that is within
queue size. Therefore, the centralized solution requiring the O(1/V ) of the optimal objective for any V > 0, while
entire system state will lead to heavy signaling overhead and ensuring that the average queue backlog is O(V ). Finally,
high computational complexity in the central optimizer. Fur- through the extensive simulations, the theoretic claims are
thermore, this information about the entire system state may be verified, and the impacts of a variety of system parameters
hard to obtain or even unattainable in practical implementation. on overall objective, rate utility and electricity cost are
It is desirable to have the distributed optimization based on investigated.
2816 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 14, NO. 5, MAY 2015

TABLE I
S UMMARY OF K EY N OTATIONS

Fig. 1. Diagram of single node system.

Throughout this paper, we use the following notations. The

probability of an event A is denoted by Pr(A). For a random
variable X, its expected value is denoted by E[X] and its
expected value conditioned on event A is denoted by E[X|A].
The indicator function for an event A is denoted by 1A ; it equals
1 if A occurs and is 0 otherwise. [x]+ = max(x, 0).
The remainder of the paper is organized as follows. In
Section II, we give the system model and problem formu-
lation. In Section III, we present the distributed cross-layer
optimization algorithm. In Section IV, we present the perfor-
mance analysis of our proposed algorithm. Simulation results
are given in Section V. Concluding remarks are provided in
Section VI.

A. Source Rate and Utility

II. S YSTEM M ODEL AND P ROBLEM F ORMULATION
At time slot t, the node n measures Fn independent parallel
We consider a general interconnected multi-hop WSN that information sources Fn . The measured samples of the session
implement CDMA-based medium access, and operates over f ∈ Fn is compressed with rate r f (t) before putting into the
time slots t ∈ T = {0, 1, 2, . . .}. WSN is modeled by a direct data queue,2 where r f (t) denotes the source rate of the session
graph G = (N, L). N = NH ∪ NG ∪ NM = {1, 2, 3, . . . , N} de- f at time slot t. We assume that
notes the set of sensor nodes in the network, NH is the set of
nodes powered by EH, called EH nodes, NG is the set of nodes 0 ≤ r f (t) ≤ rmax
f , ∀f ∈ F (1)
powered by EG, called EG nodes, and NM is the set of Mixed
energy (ME) nodes powered by both EH and EG, respectively. where rmax f ≤ Rmax for all f with some finite Rmax at all time. We
Ns ⊂ N denotes the set of all source nodes which measure the assume that each session f is associated with a utility function
information source(s). Each source node n ∈ Ns has multiple U f (r f (t)), which is increasing, continuously differentiable and
sensor interfaces, such that it can measure multiple information strictly concave in r f (t) with a bounded first derivative and
sources Fn = {1, 2, 3, . . . , Fn } at the same time.1 We use F = f
U f (0) = 0. We denote βU = max f ∈F,t∈T βU (t), where βU (t) is
f

Fn = {1, 2, . . . , F} to denote the set of all information the maximal first-order derivative of U f (r f (t)) w.r.t. r f (t).
n∈Ns
sources in the network. The source node transmits the data to
the corresponding destination node through multi-hop routing. B. Data Transmission
L = {(n, m), n, m ∈ N} represents the set of communication
links. O(n) denotes the set of nodes m with (n, m) ∈ L, and I(n) We assume that the links in the network may interfere
denotes the set of nodes m with (m, n) ∈ L. Fig. 1 describes the with each other when they transmit data simultaneously. We
composition of a single node system. The key notations of our define pT (t) = (pTnb (t), (n, b) ∈ L) as the transmission power
system model are shown in Table I. allocation matrix for data transmission at slot t, where pTnb (t)

2 We measure time in unit size slots, for simplicity, and thus we suppress the
1 In the following, we use the terms information source, flow and session implicit multiplication by 1 slot when converting between data rate and data
interchangeably. amount.
XU et al.: MANAGEMENT AND OPTIMIZATION FOR WSN POWERED BY HETEROGENEOUS ENERGY SOURCES 2817

is the transmission power of link (n, b), and then the following for acquiring the data at a particular rate r f (t) of the session f
inequality should be satisfied: at node n. Inspired by [22], we also assume a linear relationship
between the rate r f (t) and pSf (r f (t)), i.e., pSf (r f (t)) = P̃Sf r f (t),
0≤ ∑ pTnb (t) ≤ Pnmax , n ∈ N. (2) where P̃Sf denotes the energy consumption per data of the
b∈O(n)
f -th session for data sensening/processing. Thus, the total
where Pnmax is a finite constant to denote the maximal transmis- energy consumption pTotal
n (t) of node n at slot t is:
sion power limitation at node n. Δ
We use γnb (t) to denote the signal to interference plus noise pTotal
n (t) = ∑ P̃Sf r f (t) + ∑ pTnb (t) + P̃nR ∑ ∑ xanf (t)
f ∈Fn b∈O(n) a∈I(n) f ∈F
ratio (SINR) of link (n, b):
(5)
Δ  
γnb (t) = γnb pT (t), SC (t)
where P̃nR is the energy consumed when node n receives one
SCnb (t)pTnb (t) unit data from the neighbor nodes in the network.
= ,
N0b + ∑a∈Jn,b ∑(a,m)∈L SCab (t)pTam (t)
D. Energy Supply Model
where N0b is the noise spectral density at node b, and SCnb (t)
represents the link fading coefficient from n to b at the slot t. First, we describe the energy supply model of ME node
Jn,b is the set of nodes whose transmission may interfere with shown in Fig. 1. Each ME node is equipped with a battery
the receiver of link (n, b), excluding node n. We assume that having the limited capacity θEn . As depicted in Fig. 1, the
SCnb (t) may be time varying and independent and identically harvested energy en (t) at time t for ME node n is stored in the
distributed (i.i.d.) at every slot. Denote SC (t) = (SCnb (t), (n, b) ∈ battery. On the other hand, the energy supplied by the electricity
L) as the network channel state matrix, taking non-negative grid at time t for ME node n is denoted by gn (t). Different from
values from a finite but arbitrarily large set SC . the ME node, the EH node only stores the harvested energy
The link capacity is defined as en (t) and the EG node only stores the energy gn (t) supplied by
the electricity grid.
C̃nb (t) = log (1 + Knbγnb (t)) . We assume each n knows its own current energy availability
En (t) denoting the energy queue size for n ∈ N at time slot t.
Here, Knb denotes the processing gain of the CDMA system. We define E(t) = (En (t), n ∈ N) over time slots t ∈ T as the
Note that the dependence of C̃nb (t) on pT (t) and SC (t) is vector of the energy queue sizes. The energy queuing dynamic
f
implicit for notational convenience. Let xnb (t) denote the data equation is
transmission rate of the session f over link (n, b), b ∈ O(n).
Because of the total rates of all sessions cannot exceed the link En (t + 1) = En (t) + 1n∈NH ∪NM en (t)
capacity, we have the following constraint:
+ 1n∈NG ∪NM gn (t) − pTotal
n (t) (6)
∑ xnbf (t) ≤ C̃nb (t), ∀n ∈ N, ∀b ∈ O(n). (3)
f ∈F with En (0) = 0. At any time slot t, the total energy consump-
tion at node n must satisfy the following energy-availability
It is well known that this constraint is non-convex in pT (t) [27]. constraint:
So, we make a tightened constraint, which can be transformed
into the convex constraint, shown in Section III-D: En (t) ≥ pTotal (t), ∀n ∈ N. (7)
n

∑ f
xnb (t) ≤ Cnb (t), ∀n ∈ N, ∀b ∈ O(n), (4)
At any time slot t, the total energy volume stored in battery
f ∈F
is limited by the battery capacity, thus the following inequality
with Cnb (t) = log γnb (t) and γnb (t) = Knbγnb (t). Since Knb must be satisfied
is typically very large in CDMA networks, Cnb (t) is very
close to C̃nb (t). So the tightened constraint (4) slightly re- En (t) + 1n∈NH ∪NM en (t) + 1n∈NG ∪NM gn (t) ≤ θEn (8)
duces the feasibility set of the optimization problem with the
constraint (3). We assume the available amount of harvesting energy at slot
t is hn (t) with hn (t) ≤ hmax for all t. The amount of actually
harvested energy en (t) at slot t, should satisfy
C. Energy Consumption Model
At every time slot t, each node n allocates power3 to accom- 0 ≤ en (t) ≤ hn (t), ∀n ∈ NH ∪ NM , (9)
plish its tasks, including data sensening/processing, data trans-
mission and data reception. We define a function pSf (r f (t)) to where hn (t) is randomly varying over time slots in an i.i.d.
denote the energy consumption of sensing/processing module fashion according to a potentially unknown distribution and
taking non-negative values from a finite but arbitrarily large
3 We measure time in unit size slots, for simplicity, and thus we suppress the set SH . We define the harvestable energy state S H (t) = (hn (t),
implicit multiplication by 1 slot when converting between power and energy. n ∈ NH ∪ NM ).
2818 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 14, NO. 5, MAY 2015

The energy supplied by the electricity grid gn (t) of the source rate and the time-average cost of energy consumption in
battery of node n at slot t should satisfy: electricity grid, which subject to all of the constraints described
above. Specifically, we define
0 ≤ gn (t) ≤ gmax
n , ∀n ∈ NG ∪ NM , (10)
O(t) = ϖ1 ∑ U f (r f (t)) − (1 − ϖ1 ) ∑ ϖ2 PnG (t)gn (t)
with some finite gmax
n . f ∈F n∈NG ∪NM
(14)
E. Electricity Price Model
where ϖ1 (0 ≤ ϖ1 ≤ 1) is a weight parameter to combine
The cost per unit of electricity drawn from the electricity grid the objective functions together into a single one, and ϖ2 is
at node n ∈ NG ∪ NM at slot t is denoted by PnG (t). In general, it a mapping parameter to ensure the objective functions at the
may depend on both gn (t), the total amount of electricity from same level.
the electricity grid at slot t, and an electricity price state variable Mathematically, we will address the stochastic optimization
SnG (t), which represents such as both spatial and temporal problem P1 as follows:
variations, etc. For example, the per unit electricity cost may
1 T −1
be higher during daytime, and lower at late night. We assume maximize
{χ(t),t∈T}
O = lim
T →∞ T
∑ E {O(t)}
that SnG (t) is randomly varying over time slots in an i.i.d. fashion t=0
according to a potentially unknown distribution and taking non- subject to (1), (2), (4), (7)–(10), (12), (13) (15)
negative values from a finite but arbitrarily large set SG . Denote with the queuing dynamics (6) for all n ∈ N and (11) for all
S G (t) = (SnG (t), n ∈ NG ∪ NM ) as the electricity price vector. n ∈ N, f ∈ F.
Similarly in [28], we assume that PnG (t) is a function of both Δ
SnG (t) and gn (t), i.e., χ (t) = (ee(t),gg(t), p T (t),rr (t),xx(t)) is the set of the optimal
variables of the problem P1, where e (t), g (t), p T (t), r (t), x (t)
  f
PnG (t) = PnG SnG (t), gn (t) are the vector of en (t), gn (t), pTnb (t), r f (t), xnb (t), respectively.

Note that the dependence of PnG (t) on SnG (t) and gn (t) is implicit
III. D ISTRIBUTED C ROSS -L AYER O PTIMIZATION
for notational convenience in the sequel. For each given SnG (t),
A LGORITHM : EASYO
PnG (t)is assumed to be a increasing and continuous convex func-
tion of gn (t). Let β1G and β2G denote the maximum and minimum In this section, we propose an Energy mAnagement and
unit electricity price in any slot in any node, respectively. croSs laYer Optimization algorithm (EASYO) for the problem
P1. Based on the Lyapunov optimization with weight perturba-
F. Data Queue Model tion technique developed in [21], [29] and [30], EASYO will
f determine the energy harvesting, and the energy purchasing,
For f ∈ F at node n, we use Qn (t) to denote the data backlog source rate control, energy allocation for sensing/processing,
f
of the f -th session at time slot t. We define Q (t) = (Qn (t), n ∈ transmission and reception, routing and scheduling decisions.
N, f ∈ F) over time slots t ∈ T as the data queue backlog vector. EASYO is a fully distributed algorithm which makes greedy
Then the data queuing dynamic equation is decisions at each time slot without requiring any statistical
knowledge of the harvestable energy states, of the electricity
Qnf (t + 1) = Qnf (t) − ∑ f
xnb (t)
price states and of the channel states.
b∈O(n)

+ ∑ f
xan (t) + 1 f ∈ Fn r f (t). (11) A. Lyapunov Optimization
a∈I(n)
First, we introduce the weight perturbation θE = (θEn , n ∈ N).
In (11), the last term is revised as follows: 1 f ∈Fn r f (t) with Note that the weight perturbation θEn is the limited battery
f
Qn (0) = 0. In any time slot t, the total data output at node n capacity of node n defined in Section II-D. Then we define the
Δ
must satisfy the following data-availability constraint: Q(t),E
network state at time slot t as Z (t) = (Q E (t)), and define
the Lyapunov function as
0≤ ∑ f
xnb (t) ≤ Qnf (t), ∀n ∈ N, f ∈ F. (12)
1  2 1  2
b∈O(n)
L(t) = ∑ ∑ Qnf (t) + ∑ En (t) − θEn . (16)
2 n∈N f ∈F 2 n∈N
To ensure the network is strongly stable, the following in-
equality must be satisfied: Remark 3.1: From (16), we can see that when minimizing
T −1  f  the Lyapunov function L(t), the energy queue backlog is pushed
1
lim
T →∞ T
∑ ∑ ∑E Qn (t) < ∞. (13) towards the corresponding perturbed variable value, and the
t=0 n∈N f ∈F data queue backlog is pushed towards zero, which ensure the
strong network stability constraint (13). Furthermore, as long
as we choose appropriate perturbed variables according to (40)
G. Optimization Problem
in Theorem 1 at the next section, the constraint (7) will always
The goal is to design a full distributed algorithm that achieves be satisfied due to (44) in Theorem 1 at the next section. Thus,
the optimal trade-off between the time-average utility of the we can get rid of (13) and (7) in the sequel.
XU et al.: MANAGEMENT AND OPTIMIZATION FOR WSN POWERED BY HETEROGENEOUS ENERGY SOURCES 2819

Now define the drift-plus-penalty as TABLE II

A LGORITHM 1: EASYO
Δ
ΔV (t) = E (L(t + 1) − L(t) −VO(t)|Z(t)) (17)

where V is a non-negative weight, which can be tuned to control

O arbitrarily close to the optimum with a corresponding tradeoff
in average queue size.
We have the following lemma regarding the upper bound of
the drift-plus-penalty ΔV (t):
Lemma 1: Under any feasible energy management, source
rate control, transmission power allocation, routing and
scheduling actions that can be implemented at time t, we have
the upper bound of ΔV (t) as follows

ΔV (t) ≤ B + E Δ V (t)|Z(t) , (18)
C. Components of EASYO
where
At each time slot t, after observing Z (t), all components of
EASYO is iteratively implemented in a distributed manner to
B = NFBQ + ∑ BE (19)
cooperatively solve the problem P2. Next, we describe each
n∈N
component of EASYO in detail.
with BQ = 32 lmax
2 X 2 + R2 , where l
max max max denotes the largest 1) Energy Management: For each node n ∈ N, combining
number of the outgoing/incoming links that any node in the the first term of the RHS of (20) with the constraint (8)–(10),
network can have, and Xmax denotes some finite constant we have the optimization problem of en (t) and gn (t) as follows:
such that Cnb (t) ≤ Xmax , ∀n ∈ N, ∀b ∈ O(n), ∀t ∈ T. BE =  
1 max 2 1 Total 2
2 (1n∈NH ∪NM hmax + 1n∈NG ∪NM gn ) + 2 (Pn,max ) , Pn,max =
Total minimize En (t) − θEn 1n∈NH ∪NM en (t)
en (t),gn (t)
∑ P̃Sf rmax V (t) is given in (20), shown
+ Pnmax + P̃nR lmax Xmax . Δ
f
f ∈Fn + (Dn (t) + En (t) − θen ) 1n∈NG ∪NM gn (t)
at the bottom of the page, where
subject to 0 ≤ en (t) ≤ hn (t) (24)
Δ
Dn (t) =V (1 − ϖ1 )ϖ2 PnG (t), (21) 0 ≤ gn (t) ≤ gmax (25)
n
Δ
An (t) = En (t) − θEn , (22)
1n∈NH ∪NM en (t) + 1n∈NG ∪NM gn (t)
and ≤ θEn − En (t) (26)
f Δ f
Wnb (t) = Qnf (t) − Qb (t) + Ab (t)P̃bR . (23) Remark 3.3: Energy management component is composed
of energy harvesting and energy purchasing. Furthermore, since
Proof: See Appendix A.
PnG (t) is increasing and continuous convex on gn (t) for each
SnG (t), it is easy to verify that energy management component
B. Framework of EASYO is a convex optimization problem in (en (t), gn (t)), which can be
We now present our algorithm EASYO. The main design solved efficiently.
principle of EASYO is to minimize the right hand side (RHS) Remark 3.4: From (26), we can see that all the incoming
of (20) subject to the constraints (1), (2), (4), (8)–(10), and (12). energy is stored if there is enough room in the energy buffer
The framework of EASYO is described in Algorithm 1 according to the limitation imposed by θEn , and otherwise it
summarized in Table II. stores all the energy that it can, filling up the battery size of
Remark 3.2: Note that the algorithm EASYO only requires θEn . Hence, En (t) < θEn for all t, which means that EASYO can
the knowledge of the instant values of Z (t). It does not require be implemented with finite energy storage capacity θEn at node
any knowledge of the statistics of these stochastic processes. n ∈ N.
The remaining challenge is to solve the problem P2, which is 2) Source Rate Control: For each session f ∈ Fn at source
discussed below. node n ∈ Ns , combining the second term of the RHS of (20)

 
V (t) =
Δ ∑ En (t) − θEn 1n∈NH ∪NM en (t) + (Dn (t) + En (t) − θen ) 1n∈NG ∪NM gn (t)
n∈N
 
− ∑ ∑ V ϖ1U f (r f (t)) − Qnf (t)r f (t) + An (t)P̃Sf r f (t) − ∑ ∑ ∑ f f
Wnb (t)xnb (t) + An (t)pTnb (t) (20)
n∈Ns f ∈Fn n∈N b∈O(n) f ∈F
2820 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 14, NO. 5, MAY 2015

with the constraint (1), we have the optimization problem of Remark 3.6: Our proposed EASYO is designed to minimize
r f (t) as follows: the RHS of (20). Each component contributes to minimizing
  the part of the RHS of (20). Taking all components together,
maximize V ϖ1U f (r f (t)) − Qnf (t) − An (t)P̃Sf r f (t)
r f (t) EASYO contributes to minimizing the whole RHS of (20),
subject to 0 ≤ r f (t) ≤ rmax
f (27) and thus to minimize ΔV (t). Because the whole RHS of (20)
r∗f
Let be the unique maximizer. By the Kuhn-Tucker theorem, incorporates the Lyapunov drift, EASYO is stable. Meanwhile,
r∗f is given by since it also incorporates the objective of the problem P1,
  rmax EASYO is optimal.
r∗f = U f−1 Qnf (t) − An (t)P̃Sf
f
(28) Remark 3.7: The former two components of EASYO are
0
−1 computed in closed form or numerically solved through a
where [z]ba
= min {max {z, a} , b}, U f (·) is the inverse of the simple convex optimization problem, only based on the local
derivative of U f (·). information. The unique challenge of distributed implementa-
3) Joint Optimal Transmission Power Allocation, Routing tion of EASYO is to distributedly solve the transmission power
and Scheduling: Combining the third term of the RHS of (20) allocation problem (32). Next, we will develop the distributed
with the constraints (2), (4) and the data-availability constraint algorithm.
(12), we have the optimization problem of x (t) and p T (t) as
follows:
  D. Distributed Implementation of Transmission
maximize ∑ ∑ ∑ Wnb (t)xnb (t) + An (t)pTnb (t)
f f
x (t),ppT (t) n∈N b∈O(n) f ∈F Power Allocation
subject to 0≤ ∑ f
xnb (t) ≤ Cnb (t), ∀n ∈ N, ∀b ∈ O(n) After implementing a variable change p̂Tnm (t) = log(pTnm (t)),
f ∈F and taking the logarithm of both sides of the constraint in
0≤ ∑ pTnb (t) ≤ Pnmax , ∀n ∈ N problem (32), the problem (32) can be equivalently transformed
b∈O(n) into the problem P3

0≤ ∑ f
xnb (t) ≤ Qnf (t), ∀ f ∈ F (29)
max ∑ ∑ W̃nb ∗
  T
(t)Ψnb p̂pT (t) + An (t)e p̂nb (t)
b∈O(n)
p̂pT (t) n∈N b∈O(n)
Next, we will solve the optimization problem (29). Define
the weight of the session f over link (n, b) as:
∑
T
s.t. log e p̂nb (t) − log Pnmax ≤ 0, ∀n ∈ N (33)
f Δ f + b∈O(n)
W̃nb (t) = Wnb (t) − σ , (30)
where p̂pT (t) = ( p̂Tnb (t), n ∈ N, b ∈ O(n)), Ψnb ( p̂T (t)) is defined
where in (34).
σ = lmax Xmax + rmax (31)  Δ
f Ψnb p̂pT (t) = log (γnb (t))
denotes the data amount of the session f which the node n can
receive at most at time slot t. = log SCnb (t) + p̂Tnb (t)
Transmission Power Allocation Component: For each node ⎛ ⎞
∗ ∈ arg max W̃ f (t). Define W̃ ∗ (t) = max W̃ f (t)
n, find any fnb  
f nb nb f nb − log⎝N0b + ∑ ∑ exp p̂Tam (t)+log SCab (t) ⎠.
as the corresponding optimal weight of link (n, b). Observe the a∈Jn,b (a,m)∈L
current channel state SC (t), and select the transmission powers
p T ∗ by solving the following optimization problem: (34)
 ∗ 
maximize ∑ ∑ W̃nb (t)Cnb (t) + An (t)pTnb (t) It is not difficult to prove that Ψnb ( p̂pT ) is a strictly concave
pT n∈N b∈O(n)
function of a logarithmically transformed power vector p̂pT (t)
subject to 0 ≤ ∑ pTnb (t) ≤ Pnmax , ∀n ∈ N (32) [31]. Due to (8) or (26), we have En (t) ≤ θEn , so An (t) ≤ 0, thus
b∈O(n) T
An (t)e p̂nb (t) is a strictly concave function of p̂Tnb (t). To sum up,
Routing and Scheduling Component: The data of session fnb ∗ the objective of P3 is a strictly convex in p̂T (t). Furthermore,
T
f∗ since log ∑ e p̂nb (t) is a strictly concave in p̂T (t), P3 is a
is selected for routing over link (n, b) whenever W̃nb (t) > 0.
b∈O(n)
f∗ f∗
That is, if W̃nb (t) > 0, set xnbnb (t) = Cnb (ppT ∗ ,SSC (t)). strictly convex optimization problem, which has the global
Remark 3.5: If we set σ = 0, the joint transmission power optimum.
allocation, routing and scheduling component is to minimize To distributedly solve P3, we propose a distributed iterative
the third term of the RHS in (20). Inspired by [29] and [30], we algorithm based on block coordinate descent (BCD) method
set a non-zero σ in (30), leading to an easy way to determine whereby, at every iteration, a single block of variables is opti-
the upper bound of all queue sizes shown in Theorem 1. Also mized while the remaining blocks are held fixed. More specifi-
the definition (31) of σ can ensure the constraints (12) will be cally, at iteration ti , which represents the i-th iteration at the time
always satisfied. The detailed proof will be given in Theorem 1. slot t, for each node n ∈ N, the blocks p̂pTn = ( p̂Tnb , b ∈ O(n)) are
Thus, we can get rid of this constraint (12) in (32). updated through solving the following optimization problem
XU et al.: MANAGEMENT AND OPTIMIZATION FOR WSN POWERED BY HETEROGENEOUS ENERGY SOURCES 2821

(35), where p̂T−n (ti ) = ( p̂T1 (ti ), · · · , p̂Tn−1 (ti ), p̂Tn+1 (ti ), · · · , p̂TN (ti ))
are held fixed.
  
maximize ∑ ∑ W̃nb ∗
(t)Ψnb p̂pTn , p̂pT−n (ti )
p̂Tn n∈N b∈O(n)
T
+An (t)e p̂nb

∑
T
subject to log e p̂nb − log Pnmax ≤ 0. (35)
b∈O(n)

The global rate of convergence for BCD-type algorithm has

been studied extensively when the block variables are updated
in both the classic Gauss-Seidel fashion and the randomized
update rule [32], [33]. Since the optimization problem P3 is
strongly convex in p̂pT (t), our proposed BCD-based distributed
iterative algorithm can converge to the global optimum of P3.
Fig. 2. Network topology.

IV. P ERFORMANCE A NALYSIS Proof: Please see Appendix B–E.

Remark 4.1: Theorem 1 shows that a control parameter V
Now, we analyze the performance of our proposed algorithm enables an explicit trade-off between the average objective
EASYO. To start with, we assume that there exists δ > 0 value and queue backlog. Specifically, for any V > 0, the
such that proposed distributed algorithm EASYO can achieve a time
  average objective that is within O(1/V ) of the optimal objective
Cnb p T (t),SSC (t) ≤ δpTnb (t), ∀n ∈ N, ∀b ∈ O(n). (36)
shown in (43), while ensuring that the average data and energy
Theorem 1: Implementing the algorithm EASYO with any queues have upper bounds of O(V ) shown in (41) and (42),
fixed parameter V > 0 for all time slots, we have the following respectively. In the Section V, the simulations will verify the
performance guarantees: theoretic claims.
A) Suppose the initial data queues and the initial energy Remark 4.2: The inequality (44) guarantees that the energy-
queues satisfy: availability constraint (7) is satisfied for all nodes and all times.
Similarly, the inequality (45) ensures that the data availability
0 ≤ Qnf (0) ≤ Qmax , n ∈ N, f ∈ F, (37) constraint (12) is always satisfied.
0 ≤ En (0) ≤ θEn , n ∈ N, (38)

where the queue upper bounds are given as follows: V. S IMULATION R ESULTS
In this section, we provide the simulation results of the
Qmax = ϖ1 βU V + rmax
f , (39)
algorithm EASYO for the network scenario shown in Fig. 2. In
θEn = δϖ1 βU V + Pn,max
Total
. (40) this scenario, we consider a multi-channel WSN with 20 nodes,
78 links, 6 sessions transmitted on 14 different channels. The
Then, the data queues and the energy queues of all nodes
rate utility function is set as U f (r f (t)) = log(1+r f (t)), and thus
for all time slots t are always bounded as
βU = 1. The electricity cost function is set as PnG (t) = SnG (t).
0 ≤ Qnf (t) ≤ Qmax , n ∈ N, f ∈ F, (41) Set several default values as follows: δ = 2; rmax f = 3,
0 ≤ En (t) ≤ θEn , n ∈ N. (42) P̃S=0.1
f , ∀ f ∈ F; g max = 2, ∀n ∈ N ∪ N ; X
n G M max = 2, l max =
6, Pn = 2, P̃n = 0.05, ∀n ∈ N; N0 = 5 × 10 , ϖ1 = 0.6,
max R b −13
B) The objective function value of the problem P1 achieved ϖc = 0.5. We set all the initial queue sizes to be zero.
by the proposed algorithm EASYO satisfies the bound The channel state matrix SC (t) has independent entries that
B̃ are uniformly distributed with interval [SCmin , SCmax ]×d −4 , where
O ≥ O∗ − , (43) SCmin = 0.9 and SCmax = 1.1 as default values and d denotes
V
the distance between transmitter and receiver of the link. The
where O∗ is the optimal value of the problem P1, and energy-harvesting vector S H (t) has independent entries that are
B̃ = B + NFσlmax Xmax . uniformly distributed in [0, hmax ], with hmax =2 as default value.
C) When node n ∈ N allocates nonzero power for data sens- The electricity price vector S G (t) has independent entries that
ing, data transmission and/or data reception, we have: are uniformly distributed in [Smin G , SG ] with SG = 0.5 and
max min
Smax=1 as default values, and thus β1G =Smax
G G and β2 =SG . All
En (t) ≥ Pn,max
Total
, n ∈ N. (44) G min
statistics of SC (t), S H (t), and S G (t) are i.i.d. across time-slots.
D) For node n ∈ N, when any data of the f -th session is We set V = [100, 300, 500, 700, 1000, 1500]. In all simula-
transmitted to other node, we have: tions, the simulation time is 105 time slots. The simulation
results are depicted in Fig. 3. From Fig. 3(a), we see that as V
Qnf (t) ≥ lmax Xmax . (45) increases, the time average optimization objective value keeps
2822 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 14, NO. 5, MAY 2015

Fig. 5. Convergence of BCD-based distributed iterative algorithm.

Fig. 3. Verification of Theorem 1.

Fig. 6. The impact of node power supply manner and maximum available
harvested energy on objective value.

Transmission power allocation problem P3 is the most com-

plex component in our proposed EASYO. We proposed a BCD-
based distributed iterative algorithm to solve the problem P3.
During the implementation of EASYO, we catch four different
snapshots of the iterative procedure of BCD-based algorithm,
Fig. 4. Detailed verification of the queueing bounds. shown in Fig. 5. From Fig. 5, we can see that BCD-based al-
gorithm can quickly converge to the global optimum. Thus, our
increasing and converges quickly to very close to the optimum. proposed EASYO is a low-complexity distributed algorithm.
This confirms the results of (43). From Fig. 3(b), we see that as Next, we investigate the impacts of a variety of system pa-
V increases, the average data queue length keeps increasing. rameters on the objective value, rate utility and electricity cost.
From Fig. 3(c)–(e), we observe that the battery queue size Fig. 6 shows the impact of the node power supply manner and
increases as V increases. A closer inspection of the results also the maximum available harvested energy hmax on the objective
shows a linear increase of the time average data and energy value. From Fig. 6, we can see that the lowest objective value is
queue size with respect to V . This shows a good match between achieved at all EH nodes scenario with hmax = 0.2 much smaller
the simulations and Theorem 1. than gmax = 2 and the highest objective value is achieved at
For better verification of the queueing bounds, we also all EH nodes scenario with hmax = 2 and gmax = 2. Due to
present the data queue process of node 1 for session 1, of node 8 the expense of the highest energy cost, all EG nodes scenario
for session 5 and of node 16 for session 6 under V = 1000 in achieves the objective value lower than all EH nodes scenario
Fig. 4(a), and the energy queue processes for EH node 2, 7, 14, or default node scenario with hmax = 2. In contrast, all EG nodes
for EG node 1, 8, 20 and for ME node 3, 13, 19 under V = 1000 scenario achieves the objective value higher than all EH nodes
in Fig. 4(b)–(d), respectively. It can be verify that all queue sizes scenario or default node scenario with hmax = 0.2, which results
can quickly converge with the upper bound given in Theorem 1. in the low energy supply and low rate utility.
XU et al.: MANAGEMENT AND OPTIMIZATION FOR WSN POWERED BY HETEROGENEOUS ENERGY SOURCES 2823

Fig. 9. The impact of different sensing energy consumption on rate utility and
energy cost.
Fig. 7. The impact of different electricity prices on rate utility and energy
cost. VI. C ONCLUSION
Because of the instable energy supply and the limited battery
capacity in EH node, it is very difficult to ensure the perpetual
operation for WSN. In this paper, we consider heterogeneous
energy supplies from renewable energy and electricity grid,
multiple energy consumptions and multi-dimension stochastic
natures in the system model, and formulate a discrete-time
stochastic cross-layer optimization problem to optimize the
trade-off between the time-average rate utility and electricity
cost. To the end, we propose a fully distributed and low-
complexity cross-layer algorithm only requiring knowledge
of the instantaneous system state. The theoretic proof and
the extensive simulation show that a parameter V enables an
explicit trade-off between the optimization objective and queue
backlog. In the future, we are interested in two aspects of
delay reduction by utilizing the shortest path concept, and by
modifying the queueing disciplines.
Fig. 8. The impact of different weight parameters on rate utility and energy
cost.
A PPENDIX A
P ROOF OF L EMMA 1
We investigate the impact of the electricity price on the rate
utility and energy cost. We set three different electricity prices Though squaring both sides of (11), we have (46), shown at
G = 0.2, SG = 1 and SG = 10, respectively. Fig. 7
as Smax max max
the bottom of the next page. Similarly, we have (47), shown at
shows that the electricity cost increases along with the increase the bottom of the next page, from (6). By plugging (46), (47)
of the electricity price. To reduce the electricity cost, EASYO and (16) into (17), we have (48), shown at the bottom of the
reduce the energy consumption, and thus the corresponding rate next page, with B defined in (19).
utility decreases. Plugging the definition (5) of pTotal
n (t) into (48), and rearrang-
We investigate the impact of the weight parameter on the 
ing all terms of the RHS in (48), ΔV (t) is changed into (20). 
rate utility and energy cost. We set three the weight parameters
as ϖ1 = 0.3, ϖ1 = 0.6 and ϖ1 = 0.9, respectively. When the A PPENDIX B
weight parameter ϖ1 is chosen as a large value, EASYO focuses P ROOF OF PART (A) IN T HEOREM 1
on the rate utility maximization rather than the electricity cost
minimization. The results of Fig. 8 verify this situation, where For t = 0, we can easily have (41), and then we assume (41)
the rate utility increases and the electricity cost also increases is hold at time slot t. Next we will show that it holds at t + 1.
under a large value ϖ1 = 0.9. Case 1: If node n doesn’t receive any data at time t, we have
f f
Fig. 9 shows the impact of P̃Sf on the rate utility and energy Qn (t + 1) ≤ Qn (t) ≤ ϖ1 βU V + rmax
f .
cost. The larger P̃Sf , the more energy is required to supply for Case 2: If node n receives the endogenous data from other
f
data sensing/processing, leading to the less energy used in data nodes a∈I(n), we can get from (30) that Wan (t)−σ≥0. By plug-
f f
transmission, and the lower rate utility. ging (23) and (31), we have Qa (t)−Qn (t)+An (t)P̃nR −(rmax f +
2824 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 14, NO. 5, MAY 2015

f f f
l max Xmax)≥0. Then, Qn (t)≤Qa (t)−(r max
f +l max Xmax )+An (t)P̃nR . Qn (t) ≤ ϖ1 βU V . At every time, the new local data received at
f f
f , and so, Qn (t +1) ≤ Qn (t)+r f
max ≤ ϖ β V +r max .
Due to An (t) ≤ 0 and P̃n > 0, we have
R most is rmax 1 U f
  To sum up the above, we complete the proof of (41).
Qnf (t) ≤ Qaf (t) − rmax
f + lmax Xmax . (49) From Remark 3.5 we can have (42). 

Plugging (41) into (49), we have A PPENDIX C

  P ROOF OF PART (B) IN T HEOREM 1
Qnf (t) ≤ ϖ1 βU V + rmax
f − rmaxf + lmax Xmax
= ϖ1 βU V − lmax Xmax . (50) The proof procedure of Part (B) in Theorem 1 is similar to
that in [20], and hence is omitted for brevity.
At every slot, the node can receive the amount of data at most
rmax
f + lmax Xmax . So A PPENDIX D
P ROOF OF PART (C) IN T HEOREM 1
Qnf (t + 1) ≤ Qnf (t) + lmax Xmax + rmax
f . (51)
It is easy to verify that the following inequality holds accord-
Combing (50) and (51), we have
f
Qn (t + 1) ≤
ϖ1 βU V + rmax ing to the definition of Cab (pT (t), S(t)):
f . 
Case 3: If node n only receives the new local data, accord-  
Cab p T (t),SS(t) ≤ Cab p T (t),SS(t) (52)
ing to (27), the optimal value r∗f will satisfy V ϖ1U f (r∗f ) =
f
Qn (t) − An (t)P̃Sf , where U f (r f (t)) denotes the first derivative where pT (t) obtained by setting pTnm (t) of pT (t) to zero,
f
of U f (r f (t)). So we have Qn (t) − An (t)P̃Sf ≤ V ϖ1 βU , and then (a, b) ∈ L and (a, b) = (n, m).

1  f 2  2
Qn (t + 1) − Qnf (t)
2  2  2
1 1
1 f ∈Fn r f (t) + ∑ xan (t) + ∑ xnb (t)
f
≤ f
2 2 b∈O(n)
 a∈I(n)

+ Qnf (t) 1 f ∈Fn r f (t) + ∑ f
xan (t) − ∑ f
xnb (t)
a∈I(n)
 b∈O(n) 
3
≤ (Rmax ) + (lmax Xmax ) + Qn (t) 1 f ∈Fn r f (t) + ∑ xan (t) − ∑ xnb (t)
2 2 f f f
(46)
2 a∈I(n) b∈O(n)

1  2  2
En (t + 1) − θEn − En (t) − θEn
2  
1  2  2
≤ 1n∈NH ∪NM en (t) + 1n∈NG ∪NM gn (t) + pTotal
n (t)
2 
 
+ En (t) − θEn 1n∈NH ∪NM en (t) + 1n∈NG ∪NM gn (t) − pTotal
n (t)
  
1   2 2
≤ 1n∈NH ∪NM hmax + 1n∈NG ∪NM gmax
n + pTotal
n,max
2 
 
+ En (t) − θEn 1n∈NH ∪NM en (t) + 1n∈NG ∪NM gn (t) − pTotal
n (t) (47)

 
V (t) =
Δ ∑ ∑ Qnf (t) 1 f ∈Fn r f (t) + ∑ f
xan (t) − ∑ f
xnb (t)
n∈N f ∈F

a∈I(n) b∈O(n)

+ ∑ En (t) − θEn 1n∈NH ∪NM en (t) + 1n∈NG ∪NM gn (t) − pTotal
n (t)

n∈N 
−V ϖ1 ∑ U f (r f (t)) − (1 − ϖ1 ) ∑ ϖ2 PnG (t)gn (t) (48)
f ∈F n∈NG ∪NM
XU et al.: MANAGEMENT AND OPTIMIZATION FOR WSN POWERED BY HETEROGENEOUS ENERGY SOURCES 2825

For link (n, m), plugging (23) into (30), we get ACKNOWLEDGMENT
+ We are very grateful to Prof. Michael J. Neely at the
f
W̃nm (t) = Qnf (t) − Qmf (t) + Am (t)P̃mR − σ
+ University of Southern California, Dr. Cristiano Tapparello at
≤ Qnf (t) − σ (53)
University of Padova, and Prof. Osvaldo Simeone at New Jersey
By plugging (31) and (41) into (53), we have Institute of Technology for very helpful discussions.
+
f
W̃nm (t) ≤ ϖ1 βU V + rmax
f − lmax Xmax − rmax
f
R EFERENCES
= [ϖ1 βU V − lmax Xmax ]+ (54)
[1] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci, “Wireless
Since (54) holds for any session f through link (n, m), we have sensor networks: A survey,” Comput. Netw., vol. 38, no. 4, pp. 393–422,
Mar. 2002.
∗ [2] W. Xu et al., “Distributed optimal rate-reliability-lifetime tradeoff in
W̃nm (t) ≤ [ϖ1 βU V − lmax Xmax ]+ (55) time-varying wireless sensor networks,” IEEE Trans. Wireless Commun.,
vol. 13, no. 9, pp. 4836–4847, Sep. 2014.
We assume that En (t) < Pn,max Total , when node n ∈ N allocates [3] J. Chen et al., “Utility-based asynchronous flow control algorithm for
nonzero power for data sensing, compression and transmission. wireless sensor networks,” IEEE J. Sel. Areas Commun., vol. 28, no. 7,
pp. 1116–1126, Sep. 2010.
Furthermore, we assume that the power allocation control vec- [4] S. Sudevalayam and P. Kulkarni, “Energy harvesting sensor nodes: Survey
tor p T ∗ (t) is the optimal solution to (32), and without loss of and implications,” IEEE Commun. Surveys Tuts., vol. 13, no. 3, pp. 443–
∗
generality, there exists some pTnm (t) > 0. By setting pTnm∗ (t) = 0 461, 2011.
[5] S. He et al., “Energy provisioning in wireless rechargeable sensor net-
in p T ∗ (t), we get another power allocation control vector p T (t). works,” IEEE Trans. Mobile Comput., vol. 12, no. 10, pp. 1931–1942,
We denote G(ppT (t),SSC(t)) as the objective function of (32). In Oct. 2013.
this way, we get: [6] H. Ju and R. Zhang, “Throughput maximization in wireless powered
communication networks,” IEEE Trans. Wireless Commun., vol. 13, no. 1,
    pp. 418–428, Jan. 2014.
G p T ∗ (t),SS(t) − G p T (t),SS(t)
 T∗    ∗ [7] V. Sharma, U. Mukherji, V. Joseph, and S. Gupta, “Optimal energy
= ∑ ∑ Cnb p (t),SS(t) −Cnb p T (t),SS(t) W̃nb (t) management policies for energy harvesting sensor nodes,” IEEE Trans.
n∈N Wireless Commun., vol. 9, no. 4, pp. 1326–1336, Apr. 2010.
 b∈O(n)  [8] K. Tutuncuoglu and A. Yener, “Optimum transmission policies for battery
+ En (t) − θEn pTnm∗ (56) limited energy harvesting nodes,” IEEE Trans. Wireless Commun., vol. 11,
no. 3, pp. 1180–1189, Mar. 2012.
∗
From (52), we have Cnb (ppT (t),SS(t)) − Cnb (ppT (t),SSC (t)) ≤ 0 [9] O. Ozel, K. Tutuncuoglu, J. Yang, S. Ulukus, and A. Yener, “Transmission
with energy harvesting nodes in fading wireless channels: Optimal policies,”
for b = m. So IEEE J. Sel. Areas Commun., vol. 29, no. 8, pp. 1732–1743, Sep. 2011.
    [10] R. Srivastava and C. E. Koksal, “Basic performance limits and tradeoffs
G p T ∗ (t),SS(t) − G p T (t),SS (t) in energy harvesting sensor nodes with finite data and energy storage,”
 T∗  ∗
 
≤ Cnm p (t),SSC (t) W̃nb (t) + En (t) − θEn pTnm∗ (57) IEEE/ACM Trans. Netw., vol. 21, no. 4, pp. 1049–1062, Aug. 2013.
[11] P. Castiglione, O. Simeone, E. Erkip, and T. Zemen, “Energy manage-
ment policies for energy-neutral source-channel coding,” IEEE Trans.
Total and the definition
According to our assumption En (t) < Pn,max Commun., vol. 60, no. 9, pp. 2668–2678, Sep. 2012.
of θn in (40), we have
E [12] Y. Zhang, S. He, J. Chen, Y. Sun, and X. Shen, “Distributed sampling rate
control for rechargeable sensor nodes with limited battery capacity,” IEEE
Trans. Wireless Commun., vol. 12, no. 6, pp. 3096–3106, Jun. 2013.
EnH (t) − θeH
n < Pn,max − θn = −δϖ1 βU V
Total eH
(58) [13] S. Mao, M. Cheung, and V. Wong, “Joint energy allocation for sensing
and transmission in rechargeable wireless sensor networks,” IEEE Trans.
Plugging (36), (55) and (58) into (57), we have Veh. Technol., vol. 63, no. 6, pp. 2862–2875, Jul. 2014.
[14] K. Fan, R. Liu, Z. Zheng, and P. Sinha, “Perpetual and fair data collection
G(ppT ∗ (t),SS) − G(ppT ,SSC ) for environmental energy harvesting sensor networks,” IEEE/ACM Trans.
Netw., vol. 19, no. 4, pp. 947–960, Aug. 2011.
≤ δpTnm∗ [ϖ1 βU V − lmax Xmax ]+ − δϖ1 βU V pTnm∗ [15] Z. Mao, C. E. Koksal, and N. B. Shroff, “Near optimal power and rate
<0 control of multi-hop sensor networks with energy replenishment: Basic
limitations with finite energy and data storage,” IEEE Trans. Autom.
Total ,
From the above inequalities, we can see that if En (t) < Pn,max Control, vol. 57, no. 4, pp. 815–829, Apr. 2012.
[16] S. Chen, P. Sinha, N. Shroff, and C. Joo, “A simple asymptotically optimal
pT ∗ is not the optimal solution to (32), which is opposite with energy allocation and routing scheme in rechargeable sensor networks,”
our assumption. So, En (t) ≥ Pn,max
Total , which completes the proof in Proc. IEEE INFOCOM, Orlando, FL, USA, Mar. 2012, pp. 379–387.
[17] S. Sarkar, M. H. R. Khouzani, and K. Kar, “Optimal Routing and Schedul-
of (44).  ing in Multihop Wireless Renewable Energy Networks,” IEEE Trans.
Autom. Control, vol. 58, no. 7, pp. 1792–1798, Jul. 2013.
[18] W. Xu, S.Wen, Q. Shi, J.Wu, and Y.Wang, “Distributed Event-
A PPENDIX E TriggeredFlow Control for Utility Fairness in Energy Harvesting Wireless
P ROOF OF PART (D) IN T HEOREM 1 Sensor Networks,” Int. J. Sensor Netw., [Online]. Available: http://www.
inderscience.com/info/ingeneral/forthcoming.php?jcode=ijsnet
For node n ∈ N, when any data of the f -th session is [19] M. Gatzianas, L. Georgiadis, and L. Tassiulas, “Control of wireless
f networks with rechargeable batteries,” IEEE Trans. Wireless Commun.,
transmitted to other node, we can get W̃nb (t) > 0. From (30),
f f vol. 9, no. 2, pp. 581–593, Feb. 2010.
we have Wnb (t) − σ > 0. By plugging (23), we have Qn (t) − [20] L. Huang and M. J. Neely, “Utility optimal scheduling in energy harvest-
f f f
Qb (t) + Ab (t)P̃bR − σ > 0. Then, Qn (t) > σ + Qb (t) − Ab (t)P̃bR . ing networks,” IEEE/ACM Trans. Netw., vol. 21, no. 4, pp. 1117–1130,
Aug. 2013.
f f
Since Qb (t) ≤ 0, Ab (t) ≤ 0 and P̃bR > 0, we have Qn (t) > σ. [21] L. Huang and M. J. Neely, “Utility optimal scheduling in processing
f networks,” Perform. Eval., vol. 68, no. 11, pp. 1002–1021, Nov. 2011.
By plugging the definition of σ in (31), we have Qn (t) > [22] C. Tapparello, O. Simeone, and M. Rossi, “Dynamic compression-
max
lmax Xmax + r f , which completes the proof of (45). transmission for energy-harvesting multihop networks with correlated
2826 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 14, NO. 5, MAY 2015

sources,” IEEE/ACM Trans. Netw., vol. 22, no. 6, pp. 1729–1741, Yushu Zhang is currently pursuing the master’s de-
Dec. 2014. gree from Zhejiang Sci-Tech University, Hangzhou,
[23] C. B. Margi, V. Petkov, K. Obraczka, and R. Manduchi, “Characterizing China. Her current research interests lie in wireless
energy consumption in a visual sensor network testbed,” in Proc. Trident- communication and network.
Com, Barcelona, Spain, Mar. 2006, pp. 339–346.
[24] X. Liu, O. Simeone, and E. Erkip, “Energy efficient sensing and com-
munication of parallel sources,” IEEE Trans. Commun., vol. 60, no. 12,
pp. 3826–3835, Dec. 2012.
[25] H. Huang and V. K. N. Lau, “Decentralized Delay Optimal Control for
Interference Networks With Limited Renewable Energy Storage,” IEEE
Trans. Signal Process., vol. 60, no. 5, pp. 2552–2561, May 2012.
[26] J. Gong, S. Zhou, and Z. Niu, “Optimal Power Allocation for Energy Har-
vesting and Power Grid Coexisting Wireless Communication Systems,”
IEEE Trans. Commun., vol. 61, no. 7, pp. 3040–3049, 2013.
[27] M. Chiang, “Balancing transport and physical layers in wireless multihop Qingjiang Shi (M’10) received the B.S. degree in
networks: Jointly optimal congestion control and power control,” IEEE J. electronic engineering from the China University of
Sel. Areas Commun., vol. 23, no. 1, pp. 104–116, Jan. 2005. Petroleum, Shangdong, China, in 2003 and the Ph.D.
[28] R. Urgaonkar, B. Urgaonkar, M. J. Neely, and A. Sivasubramaniam, “Op- degree in communication engineering from Shanghai
timal power cost management using stored energy in data centers,” in Jiao Tong University, Shanghai, China, in 2011.
Proc. ACM SIGMETRICS, San Jose, CA, USA, Jun. 2011, pp. 221–232. From September 2009 to September 2010, he visited
[29] M. J. Neely, “Stochastic Network Optimization with Application to Com- the research group of Prof. Z.-Q. (Tom) Luos at the
munication and Queueing Systems,” in Synthesis Lectures on Communi- University of Minnesota, Twin Cities, MN, USA.
cation Networks. San Rafael, CA, USA: Morgan & Claypool, 2010. In 2011, he worked as a Research Scientist at the
[30] L. Georgiadis, M. J. Neely, and L. Tassiulas, “Resource allocation and Research and Innovation Center (Bell Labs China),
cross-layer control in wireless networks,” Found. Trends Netw., vol. 1, Alcatel-Lucent Shanghai Bell Company, Ltd., China.
no. 1, pp. 1–144, 2006. He is currently an Associate Professor with the School of Information and
[31] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, U.K.: Science Technology, Zhejiang Sci-Tech University, Hangzhou, China. He has
Cambridge Univ. Press, 2004. served as a TPC member for IEEE Globecom/ICC 2012–2013, and as a
[32] M. Hong, X. Wang, M. Razaviyayn, and Z.-Q. Luo, Iteration complexity peer reviewer for a variety of IEEE journals and conferences. His current
analysis of block coordinate descent methods, University of Minnesota, research interests lie in algorithm design for signal processing in advanced
Minneapolis, MN, USA, Tech. Rep. [Online]. Available: http://arxiv.org/ MIMO, cooperative communication, physical layer security, energy-efficient
abs/1310.6957 communication, wireless information and power transfer.
[33] A. Beck and L. Tetruashvili, “On the convergence of block coordinate He received the Nomination Award of the National Excellent Doctoral Dis-
descent type methods,” Soc. Ind. Appl. Math. J. Optim., vol. 23, no. 4, sertation of China in 2013, Shanghai Excellent Doctorial Dissertation Award in
pp. 2037–2060, 2013. 2012, and the Best Paper Award from the IEEE PIMRC’09 Conference.
[34] D. P. Bertsekas, A. Nedic, and A. E. Ozdaglar, Convex Analysis and
Optimization. Belmont, MA, USA: Athena Scientific, 2003.

Xiaodong Wang (S’97–M’98–SM’04–F’08) re-

Weiqiang Xu (M’09–SM’13) received the M.Sc. ceived the Ph.D. degree in electrical engineering
degree in communications and information system from Princeton University. He is a Professor of
from Southwest Jiao-Tong University, China, and electrical engineering at Columbia University, New
the Ph.D. degree in control science and engineering York, NY, USA. His research interests fall in the
from Zhejiang University, China, in 2003 and 2006, general areas of computing, signal processing and
respectively. He also was a postdoc at Zhejiang Uni- communications, and has published extensively in
versity, and a Visiting Scholar at Columbia Univer- these areas. Among his publications is a book en-
sity. He is currently a Professor with the School of titled Wireless Communication Systems: Advanced
Information Science and Technology, Zhejiang Sci- Techniques for Signal Reception, (Prentice Hall:
Tech University, Hangzhou, China. He is currently a 2003). His current research interests include wireless
member of the Youth Committee of China Institute communications, statistical signal processing, and genomic signal processing.
of Communications and the Committee of the Chinese Institute of Electronics He received the 1999 NSF CAREER Award, the 2001 IEEE Communications
on Information Theory. His research interests include multi-cell networks, Ad Society and Information Theory Society Joint Paper Award, and the 2011 IEEE
Hoc networks, wireless sensor networks, wireless optical networks, congestion Communication Society Award for Outstanding Paper on New Communication
control, and smart grid, etc. He has served as publication chair for China- Topics. He has served as an Associate Editor for the IEEE T RANSACTIONS ON
com 2014 and a TPC member for IEEE ICC 2013–2014, IEEE Globecom C OMMUNICATIONS, the IEEE T RANSACTIONS ON W IRELESS C OMMUNI -
2012–2013, WCSP 2013, IWCMC 2009, 2010 and 2014, PMSN 2009, IHMSC CATIONS , the IEEE T RANSACTIONS ON S IGNAL P ROCESSING , and the IEEE
2009–2013, and CCCC 2012–2013. He has also served as a peer reviewer for a T RANSACTIONS ON I NFORMATION T HEORY. He is a Fellow of the IEEE and
variety of IEEE journals and conferences. listed as an ISI Highly-cited Author.