Distributed Networking in Autonomic Solar

Powered Wireless Sensor Networks

Shusen Yang∗ , Xinyu Yang† , Julie A. McCann∗ , Tong Zhang† , Guozheng Liu† , and Zheng Liu†

*Department of Computing, Imperial College, London, UK,

†Department of Computer Science and Technology, Xi’an Jiaotong University, China,

Abstract—Recent advances in solar harvesting technologies impacts on the design of reliable multi-hop solar-powered
pave the way for sustainable environmental-monitoring applica- WSNs (SP-WSNs). For instance, a routing protocol should
tions in the emerging solar powered wireless sensor networks (SP- adaptively selects and dynamically adjusts end-to-end path
WSNs). The complexities associated with the low-resourced, high-
dynamic, and vulnerable sensor nodes operating in potentially to avoid the time-varying routing hot-spots (i.e. nodes with
unattended or hostile environments require a high degree of low harvesting power). This work therefore aims to provide
self-management and automation. Guided by autonomic com- sustainable data collection capability, while also maximizing
munication principles, this paper presents AutoSP-WSN, a novel end-to-end network performance (e.g. goodput) for SP-WSNs.
distributed framework to achieve sustainable data collection Fortunately, the emerging autonomic computing and com-
while also optimizing end-to-end network performance for SP-
WSNs. Initially, we present the energy-aware support component munication paradigms [5], [6], can help to address the above
that provides reliable energy monitoring and prediction. This issues in the design of SP-WSNs: Firstly, sensor node’s inter-
drives the power management component, which is adaptive to nal power system should be realistically modelled (e.g. solar
time-varying solar power, avoiding battery exhaustion as well as power and battery recharging/discharing process) to provide
maximizing the per-node utility. Finally, to demonstrate the key reliable energy awareness. Secondly, the capacity of self-
design issues of the network protocol component, we propose two
self-adaptive network protocols, a routing protocol SP-BCP and management for each individual node is required to ensure
a rate control scheme PEA-DLEX. We show that the individual sustainable operation while optimizing its long-term solar
components seamlessly integrated as a whole, and the AutoSP- power usage (this is its power management function). Thirdly,
WSN framework exhibits the properties of context-awareness, through adapting to the time-varying and heterogeneous dis-
distributed operation, self-configuration, self-optimization, self- tributed solar harvesting opportunities, distributed and adaptive
protection and self-healing. Through extensive experiments on
a real SP-WSN platform, and hardware-driven simulations, we network protocols should be derived for SP-WSNs to achieve
show that the proposed schemes achieve substantial improve- autonomous multi-hop data collection.
ments over previous work, in terms of reliability, sustainable The current studies on energy harvesting WSNs, per-node
operation, and network utility. power management schemes [4], [7], [8] and network-wide
protocols [9]–[13], aim to address the above issues separately.
I. I NTRODUCTION Furthermore, most of them are theoretical work, and therefore
Environmental monitoring is one of the most important ap- are likely to perform poorly or even fail in real-world SP-
plications in Wireless Sensor Networks (WSNs) [1]. Battery- WSNs. In this paper, we incorporate autonomic communica-
powered sensor nodes have to be manually replaced when they tion principles into the design of SP-WSNs, and present a
are depleted. However, in many outdoor deployments, it is systematical study covering both individual components (i.e.
difficult or even impossible to physically access the sensor energy awareness, power management, and network protocols)
nodes. As a result, the bounded lifetime of WSNs has become and the SP-WSN as a whole. The contributions of this paper
a restriction impacting their use in such applications. are summarized as follows:
Recently, the development of photovoltaic harvesting tech- 1. Guided by autonomic communication principles, we
niques (e.g. [2]), are demonstrating that solar power is indeed develop and implement AutoSP-WSN, to our knowledge,
a viable first step towards autonomous WSNs [3]. However, the first distributed framework for practical SP-WSNs. The
due to the limited sizes of micro solar panels, harvested overarching goal of AutoSP-WSN is to optimize the usage of
solar energy remains scarce [4]. To make best use of this solar power as well as the end-to-end network performance,
resource one needs to know how the node uses power while and more importantly, to achieve Energy Neutral Operation
understanding the dynamics of energy generation. However, (ENO) [7], i.e. to guarantee that no node will run out of energy
hardware power usage is complex and the nature of solar at any point in real-world deployments.
power is highly dynamic. Further complications come about 2. A set of self-managing functions and adaptive algorithms
because of the heterogeneous spatial harvesting capabilities are developed for AutoSP-WSN, including Energy Awareness
across different nodes in a sensing space due to shading or (EA) support, solar prediction, power management, and end-
cloud coverage as shown in Figure 1. These have significant to-end network protocols.
 time-varying solar profile user- defined Distributed Network Protocols
150 objectives
protocol type energy constraint

Power (mW)
100

context Local Power Management

50
information
0 EA solar predictor:WC-EWMA
0:00 6:00 12:00
Time
18:00 24:00
channel support
(a) A sunny day state
120 knowledge base: real-time
the sink off-line profiling monitoring
90
energy software

Power (mW)
60
hardware
30 solar panel MicaZ
topology EA board
heterogeneous harvesting opportunities Li-ion battery
0
0:00 6:00 12:00 18:00 24:00 sensors
Time

solar powered sensor nodes wireless channel (b) A cloudy day ... physical energy flow information flow

Fig. 1. Illustration of a typical multi-hop SP-WSN. Fig. 2. AutoSP-WSN architecture.

3. Through extensive evaluations on both a real SP-WSN and adaptive to other environment context such as channel
platform and the Tossim simulator [14], we show that both condition and topologies. In this paper, we present a routing
the proposed individual schemes and the whole AutoSP-WSN protocol SP-BCP, and a rate control protocol PEA-DLEX to
achieve substantial improvements over previous approaches. demonstrate how autonomic communication principles aid the
The rest of paper is organized as follows. The next section designing of network protocols in SP-WSNs.
presents the overall AutoSP-WSN architecture. Section III B. Time Horizon and Protocol Types
presents EA support component. The energy prediction al- Due to the dynamic nature of solar power, modelling time
gorithm, WC-EWMA, is described in Section IV. Section V is critical to AutoSP-WSN. Continuous time is divided into
presents the details of the LPM component. Two solar-aware discrete durations as shown in Figure 3. We define a slot
network protocols are proposed in Section VI. Evaluations are (of several minutes, dependent on weather and sensor nodes’
presented in Section VII. Section VIII describe the related surroundings) as the atomic duration in which solar power can
work, and finally, Section IX concludes this paper. be considered to be static. Our WC-EWMA algorithm predicts
II. S YSTEM OVERVIEW solar power within a future prediction interval (explain later),
based on the historical solar profile of previous prediction
A. AutoSP-WSN Architecture
intervals, over the current day and multiple previous days.
Figure 2 illustrates our AutoSP-WSN framework. The
system consists of three functional components: a day
The EA support component provides realistic hardware- a prediction interval for S-type protocols
several
driven power models and reliable off-line parameter es-
days a slot
timations (e.g. energy translation efficiency). Besides, EA
component also provides precise online measurements for
a prediction interval for D-type protocols time
power parameters (e.g. real-time solar power). Furthermore, a
lightweight algorithm, WC-EWMA, are developed to forecast Fig. 3. Time horizon and duration terms.
the future solar profile. These self-monitoring and forecasting A prediction interval can consist of either one slot or
functionalities offer solid EA supports for AutoSP-WSNs. multiple slots, depending on the types of network protocols: 1.
The LPM component computes the energy consumption Static type (S-type) protocols. This kind of protocols such as
constraint for each individual node to achieve ENO and rate control algorithms DLEX [10] and Quickfix [15], require
optimal long-term solar power usage, based on the power relatively long-term static local resource constraints, since they
information provided by the EA support component. Since the have relatively heavy control overheads (e.g. the transmission
complex hardware details are encapsulated by LPM, various of end-to-end control information). 2. Dynamic type (D-type)
network protocols can simply use this energy consumption protocols such as routing schemes CTP [16] or BCP [17] often
constraint for their decision makings. Besides, LPM can self- have the light control overheads (e.g. one-hop beacons), and
configure its parameters to provide short-term or long-term therefore they can quickly adapt to network resource change.
energy constraints for different types of network protocols. Hence, we define a prediction interval consists of one slot for
The network protocol component consists of several D-type protocols and multiple slots for S-type protocols.
communication protocols such as Medium Access Control C. Characteristics of Autonomic Communication
(MAC), routing, and sensing rate control, to enable different Besides context awareness, AutoSP-WSN exhibits the fol-
nodes in a multi-hop SP-WSN cooperatively carry out data lowing self-management behaviors:
collection tasks in a distributed manner. Besides energy con- 1. self-optimization. A key objective of AutoSP-WSN is to
straint provided by LPM, network protocols should be aware optimize the usage of solar power. Specifically, LPM optimizes
the per node utility over time, while network protocols aim to 1) Energy Translation Model and Efficiency Parameters:
achieve a near optimal end-to-end network performance (e.g. We first model the battery recharging and discharging process
network goodput and fairness). as follows
2. self-protection and self-healing. A foundation aim of Pload /η1 + Pbattery /η2 = Psolar Pbattery ≥ 0 (1)
AutoSP-WSN is to achieve ENO. Several schemes are also
proposed to detect and recover errors of solar power predic- η1 Psolar + η3 (−Pbattery ) = Pload Pbattery < 0 (2)
tion, resulting in a minimal risk of network failure. In the recharging model (1), part of the solar power
3. self-configuration. LPM can automatically tune its pa- Pload /η1 is provided directly for consumption by the MicaZ,
rameter for different types of network protocols. while the remaining power Pbattery /η2 is stored in the battery.
III. E NERGY-AWARENESS S UPPORT In the discharge model (2), load power is supplied by both
solar (η1 Psolar ) and battery (−η3 Pbattery ). Using these two
This section briefly discusses the EA support component,
models, the three efficiency parameters η1−3 can be easily
including the system hardware, off-line knowledgebase es-
estimated by using least square methods. For instance, η1−3
tablishment, and online measurement. Solar power prediction
of a typical EA board are estimated as 79%, 85%, and 84%
is also an important function of the EA support component,
respectively, by using an oscilloscope.
which will be discussed in detail in Section IV.
2) Off-line Pload Profiling: Most networking protocol op-
A. Brief Description of the Hardware
erations can be translated to the scheduling of the major energy
The objective of the hardware design for AutoSP-WSN is consumers: microcontroller unit (MCU), wireless transceivers
to achieve low cost, high efficiency, and more importantly, en- (e.g. the CC2420 radio), and sensors. Since the power con-
ergy awareness support. It is worth noting that AutoSP-WSN sumption of a given operating state (e.g. MCU idle, CC2420
does not rely on any specific hardware. However, in order to on, sensor on) is stable at runtime, it is feasible to profile such
evaluate real-world performance and demonstrate our design major solar powered sensor nodes by referring to data sheets
methodology, we developed a simple solar powered sensor or using oscilloscopes.
node shown in Figure 4 (a). The hardware system consists 3) Off-line Battery Modeling: The residual energy level
of: a solar panel (4.5V, 50mA, 9 × 3.8cm2 ), a rechargeable of a battery can be measured by estimating the quantitative
Li-ion battery (2.7–4.2V, 800mAh, nearly 100% recharging energy-voltage relation. We have established piecewise linear
efficiency), a MicaZ mote [18], and our own developed EA models for typical Li-ion batteries. However, such models can
board shown in Figure 4 (b). only be used to estimate the initial battery levels off-line,
In Figure 4(a), Psolar , Pload , Pbattery and −Pbattery because the variation of the voltage is too small to be reliably
represent the powers of solar, load, battery recharging, and measured as the residual energy changes at runtime.
discharging, respectively; and 0 < η1 , η2 , η3 < 100% are the C. Online Monitoring
energy translation efficiencies from the solar panel to the load
(direct power), from solar panel to the battery (recharging) and A sensor node can compute its realtime solar power Psolar ,
from battery to the load (discharging) respectively. based on the measured voltage and current of the solar panel.
The realtime battery energy can be computed according to
solar panel MicaZ Eq. (12) in Section V, based on the estimated efficiency
parameters η1−3 , the initial off-line estimated battery level, the
Psolar 1 Pload
online monitored Pload and Psolar . Since all these parameters
input output
2 regulator regulator 3 can be estimated or monitored precisely, AutoSP-WSN can
EA board obtain precise realtime battery level.
Pbattery -Pbattery
Li-ion battery IV. S OLAR P OWER P REDICTION : T HE WC-EWMA
(a) hardware architecture (b) The EA board
A LGORITHM
Our AutoSP-WSN framework requires forecasting solar
Fig. 4. EA support hardware and its architecture.
power of a future prediction interval, i.e. one future slot for D-
The EA board efficiently powers both the load (MicaZ type protocols and multiple future slots for S-type protocols,
mote and sensors) and the battery, as well as providing the with high accuracy and low complexity (Section V will discuss
realtime physical power information. If Psolar > Pload , the how to utilize the predicted solar power). Yet current solar
solar panel will power both the battery and the load; otherwise, prediction algorithms [20]–[22] cannot achieve this. Therefore,
both the solar panel and the battery will power the load; we develop a novel weather-aware solar prediction scheme,
automatically controlled by the input regulator BQ24070 [19]. WC-EWMA, to meet all the requirements of AutoSP-WSN.
WC-EWMA is based on both long-term seasonal and
B. Off-line Modelling and Profiling short-term daily solar profiles. A day is divided into M non-
The knowledgebase of each sensor node is established off- overlapping prediction intervals, each of which consists of L
line by modeling and profiling its power system and corre- slot(s) with a duration T (i.e. M · L · T = 24 hours). We can
sponding parameters, which is then used by other components use the triple (i, l, d) to refer to a slot i in the prediction
at runtime. interval l of the dth day.
A. Reference Power solar power
0.2
wv(d)

power (W)
Let Pre (i, l, d) be the reference solar power in slot (i, l,
d) to reflect the seasonal stable solar pofile governed by the 0.1
long-term geographical climate. The reference power vector is
only updated once at the end of a day as follows: 0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Pre (i, l, d + 1) = d
{
Pre (i, l, d), wv(d) ≥ wvT Fig. 5. Solar profile of 15 days and weather volatility wv(d) for each day
d with αwv = 0.8.
real
αre Psolar (i, l, d) + (1 − αre )Pre (i, l, d) otherwise

where preal
solar (i, l, d) is the real solar power metered in
slot (i, l, d); αre ∈ [0, 1] is the weighting factor; wv(d) C. Adaptive Solar Power Prediction
is weather condition level of the dth day (explained in next At the beginning of slot (1, l, d), WC-EWMA simply
subsection), i.e. the more sunny the dth day is, the smaller predicts the solar power ppred
solar (i, l, d) for all slot(s) (i, l, d),
wv(d) is; wvT is a predefined cloudiness degree threshold. 1 ≤ i ≤ L (i.e. a future prediction interval) as follows
Consequently, Pre (i, l, d) is only updated when the dth day ∑
L
is not quite cloudy, which aims to filter the influence of bad real
Psolar (i, l − 1, d)
pred i=1
weather days (noise) on the seasonal stable reference power. Psolor (i, l, d) = Pre (i, l, d) (7)
∑
L
Pre (i, l − 1, d)
B. Computing The Weather Volatility Value wv(d) i=1
From Figure 5, it is obvious that solar power varies V. L OCAL P OWER M ANAGEMENT
smoothly on sunny days (e.g. the 8th day) and fluctuates on
cloudy days (e.g. the 6th ) day. Based on the above observa- At the beginning of every prediction interval l, LPM
tions, we use two metrics; fluctuation frequency wv0 (d) and computes the maximum feasible load rate (energy budget) for
fluctuation intensity wv1 (d) to profile the solar curve on the all slots in l, i.e. from slot (1, l, d) to slot (L, l, d), based on
dth day. At the end of the dth day, wv0 (d) is calculated as the energy information provided by EA support component.
For S-type protocols, LPM sets L> 1 and enforces maximum
∑
M·L feasible load rate to be equal at every slot during l. For D-type
wv0 (d) = (h(i, d) ⊕ h(i − 1, d)) (3) protocols, LPM sets L=1 so energy budget are provided for
i=3 each future slot. Let Pidle and Pactive (Watts) represent the
where ⊕ is the XOR operator, and for i ∈ [2, M · L], load power when a sensor node is idle (MCU idle, CC2420
off, and sensor off) and active (MCU active, CC2420 on, and
{ sensor on). Since both Pidle and Pactive are nearly constant
real
1 if Psolar real
(i, d) > Psolar (i − 1, d) at runtime [7], [18], we use the maximum feasible duty cycle
h(i, d) = (4)
0 otherwise Dxmax (l) of a sensor node x to represent its maximum feasible
load rate in a prediction interval l. Section VI will show how
It is clear that wv0 (d) is the total number of peaks and troughs Dxmax (l) can be used by network protocols.
of the dth day’s solar curve. wv1 (d) is calculated as
∑
M·L A. LPM for S-type protocols
wv1 (d) = ((h(i, d) ⊕ h(i − 1, d)) ∧ g(i, d)) (5) Let P Sxi and Bxi respectively be the predicted solar power
i=3
of sensor node x at slot i, and x’s battery level at the beginning
where g(i, d), i ∈ [2, M · L], is of slot i, 1 ≤ i ≤ L. Dxmax (l) can be computed by solving
{ the following piecewise linear programming problem:
1 |Psolar
real
(i, d) − Psolar
real
(i − 1, d)| ≥ F IT
g(i, d) = (6)
0 otherwise maximize Dxmax (l) (8)
subject to
where F IT is a predefined threshold. We can see that wv1 (d)
represents the total number of the dth day’s peaks and troughs, BxL+1 ≥ φ (9)
where variation of two successive time slots are no smaller Bmax ≥ Bxi ≥ 0 (10)
than F IT . Combining both the strong and weak weather Dmin ≤ Dxmax (l) ≤ 1 (11)
volatilities, wv1 (d) and wv0 (d) − wv1 (d), we get
Bxi+1 =Bxi + (1 − Dxmax (l))Tη2 |P Sxi
− Pidle /η1 |+
wv(d) = αwv · wv1 (d) + (1 − αwv ) · (wv0 (d) − wv1 (d)) −(1 − Dx (l))T|Pidle − η1 P Sxi |+ /η3
max
(12)
+Dxmax (l)Tη2 |P Sxi − Pactive /η1 |+
where 0 ≤ αwv ≤ 1 is the weighting factor. Figure 5 shows
that wv(d) can reflect the daily weather conditions accurately. −Dxmax (l)T|Pactive − η1 P Sxi |+ /η3 − Eleak
 where Eleak represents the battery leakage in a slot that T is the duration of a slot.
(Eleak ≈ 0 for the Li-ion battery); Dmin > 0 is the user- • case [Link] < η1 P Sxi < Pactive : discharging when active
defined minimum duty cycle; and the operator | · |+ represent and recharging when idle.
max(0, ·) (i.e. for any real number a, |a|+ = a, if a > 0; i
• case 3. η1 P Sx > Pactive : recharging for both active and
|a|+ = 0, otherwise). idle durations.
Due to the limited computational capacity of tiny sensor
Input: Energy Parameters, Dmin , and protocol type. nodes, solving the problem (8) online would be prohibitive.
Therefore, we propose a light-weight algorithm shown in
1: Dxmax (l) ← 1; Figure 6 to compute the optimal Dxmax (l) at runtime.
2: i, j ← 1; E ← Bx1 ; K1 , K2 , K3 ← 0; The LPM algorithm looks a bit complicated due to the
η2
3: P1 ← η3 (Pactive − Pidle )T; P3 ← (Pactive − Pidle )T; three detailed battery update cases. However, its logical flow is
η1
4: while i ≤ L do clear. The main idea is to monotonously and gradually reduce
5: if η1 P Sxi < Pidle then /* case 1 */ Dxmax (l) from the upper bound Dxmax (l) = 1 (line 1), until
6: P ← (η1 P Sxi − Pidle )T/η3 − Eleak ; constraints (9) and (10) are guaranteed. After initialization
7: Bxi+1 ← Bxi + P − Dxmax (l)P1 ; in lines 1-3, Dxmax (l) keeps constant or is reduced in each
8: K1 ← K1 + P1 ; E ← E + P ;
iteration of the while loop (lines 4–32). Lines 5–18 update the
9: else if Pidle ≤ η1 P Sxi < Pactive then /* case 2 */
battery levels from slot i to i + 1 and other energy-relevant
10: P ← (P Sxi − Pidle /η1 )Tη2 − Eleak ;
11: P2 ← (P Sxi − Pidle /η1 )Tη2 variables. Lines 19–22 deal with the battery capacity constraint
−(η1 P Sxi − Pactive )T/η3 ; and record the current slot as the overcharge slot j (note that
12: Bx ← Bx + P − Dx (l)P2 ;
i+1 i max Bxj is always equal to Bx1 or Bmax ). Lines 23–28 process
13: K2 ← K2 + P2 ; E ← E + P ; the two cases: Bxi < 0 and BxL+1 < φ, for which the current
14: else /* case 3 */ Dxmax (l) is reduced and the feasibility of the updated Dxmax (l)
15: P ← (P Sxi − Pidle /η1 )Tη2 − Eleak ; will be rechecked from slot j. Finally, LPM algorithm returns,
16: Bxi+1 ← Bxi + P − Dxmax (l)P3 ; Dxmax (l), the optimal solution to the problem (8), if it exists;
17: K3 ← K3 + P3 ; E ← E + P ;
18: end if
otherwise, LPM algorithms returns Dmin . For brevity, we omit
19: if Bxi+1 > Bmax then /* record overcharging slot as j */ the proof of the optimality of the LPM algorithm.
20: Bxi+1 ← Bmax ; j ← i + 1; In the worst case, LPM algorithm requires O(L2 ) simple
21: E ← Bmax ; K1 ← 0; K2 ← 0; K3 ← 0; arithmetic calculations, when battery Dxmax (l) is updated in
22: i ← i + 1; every slot within prediction interval l. However, this rarely
23: else if Bxi+1 < 0 then /* reduce Dxmax (l) */ occurs in practice. In fact, after the initialization, Dxmax (l) is
30: Dxmax (l) ← E/(K1 + K2 + K3 );
typically updated once to guarantee the final state constraint
24: E ← Bxj ; K1 ← 0; K2 ← 0; K3 ← 0;
25: i ← j; (9). If there exists no battery overcharging and exhaustion,
26: else if Bxi+1 < φ ∧ i = L then /* reduce Dxmax (l) */ the number of simple arithmetic calculations reduces to O(L).
35: Dx (l) ← (E − φ)/(K1 + K2 + K3 );
max Therefore, the computational overhead of the LPM algorithm
27: E ← Bxj ; K1 ← 0; K2 ← 0; K3 ← 0; is similar to a sorting operation for L elements, which is much
28: i ← j; more efficient than directly solving problem (8) online.
29: else /* Dxmax (l) is maximum feasible before i */
30: i ← i + 1; B. LPM for D-type protocols
31: end if Since L=1 for D-type protocols, we directly refer a slot i ∈
32: end while
33: return max(Dmin , Dxmax (l)); [1, 2, ... , M] in the dth day without mentioning a prediction
interval for readability. We rewrite the final state parameter
Fig. 6. Pseudocode of LPM algorithm for each sensor node x.
φ as φix (d) for sensor node x at slot i in the dth day. The
minimum energy consumption of a sensor node in a slot is
The constraint (10) ensures that the battery level should
not exceed the battery capacity Bmax nor be lower than 0. Emin = T(Dmin Pactive + (1 − Dmin )Pidle )
(11) states that Dxmax (l) is x’s duty cycle. Constraint (9) is
the so-called final state constraint [4], [7]. The parameter φ During the night, since η1 P Sxi T < Emin , a sensor node x
ensures enough battery energy remaining for next prediction may exhaust its battery even when it operates in the minimal
interval and the choice of φ influences long-term system required duty cycle Dmin . Therefore, to guarantee ENO, the
performance. In AutoSP-WSN, LPM component sets φ to be node must preserve enough battery energy during the daytime
a constant value for S-type protocols and dynamically adjusts to avoid battery exhaustion at night. To this end, LPM assigns
φ for D-type protocols (next subsection). According to energy φix (d) as shown in Figure 7, where i0 (d) and i1 (d) are the
translation models (1) and (2), the constraint (12) states the first slots of daytime and night in the dth day respectively1 . Let
following three battery updating cases from slot i to i + 1: Bxi (d) be the battery level of sensor node x at the beginning
• case 1. η1 P Sxi < Pidle : discharging for both idle duration 1 The daytime and night are defined as follows: if η P S i ≤ E
1 min /T, slot
x
T(1 − Dxmax (l)) and active duration TDxmax (l). Recall i is in daytime [i0 (d), i1 (d)]; otherwise, i is in night [i1 (d), i0 (d + 1)].
of a time slot i. It can be seen that φix (d) is the lower bound A. A D-type adaptive routing protocol: SP-BCP
x
of battery level Bi+1 (d) for every slot i. The calculation of Recently, a D-type routing protocol, BCP [17], is devel-
x
φi (d) is presented below. oped for data collection in WSNs (please refer to [17] for the
With a duty cycle of Dmin , the battery energy updates details of BCP). BCP is adaptive to sudden link fluctuations,
during night i ∈ [i1 (d), i0 (d + 1)] as queue hot-spots, and topology changes. However, without
Bxi+1 (d) = Bxi (d) − Emin /η3 − Eleak energy awareness, BCP can not provide any guarantee for
efficient solar energy usage or long-term ENO in SP-WSNs.
The minimum battery energy BEN O that a sensor node Therefore, we develop SP-BCP, an energy aware backpressure
should store at slot i1 (d) can be computed as follows routing protocol for sustainable data collections in SP-WSNs.
BEN O = (Emin /η3 + Eleak )(i0 (d) + M (d) − i1 (d)) Before presenting SP-BCP, we first introduce how to use
duty cycle as energy constraints for D-type protocols. To
ensure ENO, the active duration budget of a node x in slot
daytime night i is TDxmax (i), provided by LPM. Let the remaining active
Bxi(d) duration of node x be DRx , initialized as DRx = TDxmax (i)
BENO
i
x (d)
at the beginning of each slot i. Let Nx be the set of x’
i0(d) i1(d) i0(d+1) t all one-hop neighbors. To forward a packet over a link
M(d) slots (x, y), y ∈ Nx , both the transmitter x and the receiver y will
the dth day (M slots) the (d+1)th day spend an active duration of DPx,y 3 , resulting in the reduction
of remaining active durations DRx = DRx − DPx,y and
Fig. 7. Final state parameter φix (d) and real battery level evolutions. DRy = DRy − DPx,y . Therefore, to ensure ENO, a packet
should not be transmitted over link (x, y), if DRx < DPx,y
where M (d) is the number of slots between i0 (d) and or DRy < DPx,y , i.e. both the transmitter and the receiver
i0 (d+1), shown in Figure 7. This normally change seasonally must have enough remaining active durations for a packet
but remain constant (i.e. M (d)=M) for multiple successive transmission. Hence, D-type protocols can easily use DRx
days (e.g. a week). Hence, we can set M (d)=M. Consequently, as a local energy metric for every node x’s decision making.
the following assignment of φix (d), i ∈ [i0 (d), i0 (d + 1)] is The operations of SP-BCP for a packet transmission are
sufficient to ensure ENO for the whole day: quite simple: (1) Energy aware weight calculation: Each
sensor node x computes the weights wx,y for all its neighbors

 (i − i0 (d))BEN O y ∈ Nx . If remaining active durations of x and its neighbor

 , i ∈ [i0 (d), i1 (d))
i1 (d) − i0 (d) y is larger than the estimated packet transmission duration
φix (d) = (i0 (d) + M − i)BEN O (i.e. DRx ≥ DPx,y or DRy ≥ DPx,y ), it sets wx,y =


 , i ∈ [i1 (d), i0 (d + 1)) (Qx − Qy )/DPx,y , where Qx and Qy represents the queue
i0 (d) + M − i1 (d)
backlogs of nodes x and y respectively; otherwise wx,y is set
which can be calculated at i0 (d)2 for each day d. As a as zero. (2) Routing: Each node x selects link (x, y ∗ ) with
result, node x can calculate its maximum feasible duty cycle the maximum weight for optimal potential receiver y ∗ ∈ Nx .
Dxmax (i) at the beginning of every slot i: (3) Forwarding: If wx,y∗ > 0, x forwards the packet to y ∗ .
(4) Remaining activity duration update: When a packet is
Dxmax (i) = max Dx (13) transmitted, both the transmitter x and the receiver y ∗ update
i+1
Dmin ≤Dx ≤1,Bmax ≥Bx (d)≥φix (d) their remaining activity durations as DRx = DRx − DPx,y
The LPM algorithm in Figure 6 can be used to efficiently and DRy∗ = DRy∗ − DPx,y∗ .
solve the problem (13) in one loop, by setting L=1 and The communication overhead of SP-BCP is quite light:
replacing φ and Dxmax (l) as φix and Dxmax (i) respectively. every node x can dynamically obtain Qy and DRy by peri-
odically broadcasting one-hop beacons or using overhearing.
VI. T WO A DAPTIVE N ETWORK P ROTOCOLS Evaluation results presented in Section VII demonstrate that
The per-node energy constraint (i.e. maximum feasible SP-BCP achieves hard ENO guarantee, and therefore prevents
duty cycle) provided by LPM offers an foundational support any node running out of energy.
to develop new network-wide protocol and to extend existing B. A S-type adaptive rate control protocol: PEA-DLEX
battery based WSN protocols to SP-WSNs. To demonstrate
how autonomic communication principles aid the designing of The key practical issue for such protocols is the accumu-
network-wide schemes in the highly dynamic and distributed lated prediction errors of multiple future slots in a prediction
SP-WSNs, this section presents two simple network protocols: interval. For a sensor node, if the predicted solar power
a D-type routing protocol, called SP-BCP; and a S-type rate is larger than the real solar power, LPM would assign a
control protocol, called PEA-DLEX. larger load budget, resulting in a risk of battery exhausting;
otherwise, solar harvesting opportunities can not be fully
2 The value of i (d) can be estimated online based on the reference power
1
vector in WC-EWMA algorithm. 3 DP can be easily obtain based on realtime link estimators (e.g. [17])
x,y
 RD RD RI RI RI RI RI RI
4 3 2 1 S 4 3 2 1 S 4 3 2 1 S 4 3 2 1 S
λ4=9 λ3=3 λ2=6 λ1=8.5 λ4=9 λ3=3 λ2=4 λ1=8.5 λ4=9 λ3=3 λ2=4 λ1=14 λ4=9 λ3=3 λ2=6 λ1=14
r4=r3=1 r2=r1=1.5 r4=r3=r2=0.8 r1=1.5 r4=r3=r2=0.8 r1=6 r4=r3=1 r2=2 r1=6
slot 1 ... slot i1 ... slot i2 ... slot i3
Fig. 8. An example of PEA-DLEX in a line topology SP-WSN in 4 slots 1 < i1 < i2 < i3 ≤ L during a prediction interval l. Saturated and unsaturated
nodes are marked as black and white circles respectively.

utilized, leading to energy waste. For the whole network, this end, i.e. x tracks its state (i.e. saturated or unsaturated)
the prediction errors may be heterogeneous across different and ∆Bx (i) = Bxreal (i) − Bxvirtual (i) in every slot 1 ≤ i ≤ L,
individual nodes in a SP-WSN (e.g. future solar power is over- where Bxreal (i) and Bxvirtual (i) represents the real measured
evaluated by some nodes but under-evaluated by others), which battery level and the expected virtual battery level computed by
would further degrade the end-to-end network performance. LPM algorithm. If |∆Bxi | is larger than a predefined threshold
To show how to use self-management principles to address DBT , x updates Dxmax (l) and λx by using the LPM algorithm.
above issue, we propose an adaptive Lexicographic Max-min Let the ∆λx be the differential between the updated and
(LM) rate control protocol PEA-DLEX, by modifying the previous λx .
static algorithm DLEX (please see details in [10]). PEA-DLEX adjusts the rate assignment if one of the two
We consider a routing tree consisting of a set of sensor events, Rate Decrease (RD) and Rate Increase (RI), is triggered
nodes N , rooted at a sink S in a prediction interval l. Assume by any node x’s state change or ∆λx update as follows
all wireless links have the same capacity C (kbps). Let rx A RD event is triggered if ∆λx <0 or x’s state changes
be the sensing rate at which each node x measures the from unsaturated to saturated. In this case, x has to decrease
environmental data, and STx be the set of all nodes in the AFx to ensure ENO. To this end, x updates rx and ry , y ∈
subtree rooted at x, excluding x. A rate assignment can be STx 5 , then multicasts the updated sensing rates carried by a
represented as a |N |-dimensional vector R = (r1 , r2 , ..., r|N | ) RD packet to the nodes in STx .
where the xth entry represents the sensing rate of node x. A RI event is triggered if x is saturated and ∆λx >0.
Definition 1. Feasible Rate Assignment. A rate assignment In this case, x sends a RI packet that contains the updated
R is feasible if under which every node can achieve ENO. rates ry , y ∈ {x} ∪ STx towards the sink. When x’s parent,
Mathematically, ∑for every node x ∈ N , AFx ≤ λx , where node z, receives the packet, if rz ≤ rx or z is unsaturated,
AFx = rx + 2 y∈STx ry is x’s actual forwarding data rate4 , it add its rate rz into RI packet and forwards the updated RI
and λx = Dxmax (l)C is the maximum allowed data rate. packet to its parent; otherwise, z drops the RI packet. This
Under a feasible rate assignment, every node x must be in process is repeated until the RI packet is received by the sink.
one of the two states: saturated, if AFx = λx ; unsaturated, The sink updates the RI packet and sends it back to x. When
if AFx < λx . a node between the sink and x forwards this RI packet, it
Definition 2. LM Rate Assignment. For two feasible rate updates its rate as assigned by the sink. Upon receiving the
assignment R and R′ in non-descending order, if there exist a RI, x updates its rate and multicasts the RI packet towards its
prefix (r1 ,..., ri ) of R and a prefix (r1′ ,..., ri′ ) of R′ such that upstream node(s). The RI packet sent by x is forwarded by
ri > ri′ , rj = rj′ , ∀1 ≤ j < i, then R is lexicographically unsaturated nodes but dropped by saturated nodes.
′
greater than R . R is LM rate assignment if it is lexicograph- Take Figure 8 for instance, after the initialization in slot 1,
ically greater than all other feasible rate assignments. node 2 triggers a RD event in slot i1 . Since ∆λ2 = −2 < 0,
node 2 assigns the new rates r2 = r3 = r4 = 0.8 and
For instance, consider slot 1 in Figure 8, the LM rate
transmits the updated rates to nodes 3 and 4. In slot i2 , node 1
assignment is (r4 = 1, r3 = 1, r2 = 1.5, r1 = 1.5),
triggers a RI event(∆λ1 = 5.5), it sends RI packet to the sink
which is lexicographically greater than any other feasible rate
which assigns the new rates (1,1,2,6), based on the previous
assignment such as (r4 = 1, r3 = 1, r2 = 1, r2 = 2.5).
information of nodes 2–4 in slot 1 and its current information
Recall that Dxmax (l) provided by the LPM component of
in slot i2 . However, node 2 drops the RI packet sent by the
node x is for the whole prediction interval l. The DLEX
sink since it is saturated. In slot i3 , node 2 triggers a RI
algorithm [10] uses a static LM assignment (calculated in slot
event (∆λ1 = 2). The sink computes the new rate (1,1,2,6)
1) for all slots 1 ≤ i ≤ L. However, due to prediction error,
based on the information from nodes 3 and 4 in slot 1 and
nodes need to adaptively adjust their sensing rate to achieve
nodes 1 and 2 in slot i3 . The updated rates can be sent to
ENO and better rate assignment (in terms of LM fairness). To
every node. It can be seen that the reactions for RD and
4 Due to the half-duplex operation mode of the common commercial
RI events aim to make a trade-off between communication
wireless radios, a node can not transmit and receive synchronously. Each overhead and LM optimality, guided by self-healing and self-
node x has ∑to receive and transmit the data traffic generated by all nodes in
STx (i.e. 2 y∈STx ry ),∑ as well as transmit its own data (i.e. rx ). Therefore, 5 This calculation is based on previous information from nodes in ST
x
we have AFx = rx + 2 y∈STx ry . recorded by x and Eq.(12) in [10].
 TABLE I TABLE II
E XPERIMENT PARAMETERS OF THE THREE PREDICTION ALGORITHMS . E XPERIMENT PARAMETERS OF LPM (S- TYPE PROTOCOLS ).

EWMA WCMA-PDR WC-EWMA L Bmax Bx1 φ Pidle Pactive Dmin

α D α δ γ w Ad αwv αre F IT wvT
6 10656J 1000J 1000J 13.7mW 78.4mW 0
0.5 4 0.7 0.7 0.55 4 1.8 0.8 0.5 5 3

real solar energy battery level φ

EWMA EWMA 0.4 predicted solar energy 2
energy consumption

energy (KJ)
WCMA−PDR WCMA−PDR
60

energy (KJ)
60
pred_err (%)

pred_err (%)
WC−EWMA WC−EWMA 0.3
1.5
40 40
0.2
20 20 1
0.1
0 0
0.5
0
2 2 0 1 2 3 4 5 0 1 2 3 4 5
D
days days
4 A 4
B E
6 C L 6 F (a) solar and load energy evolution (b) battery level evolution
L area area

Fig. 9. Mean prediction errors for three prediction algorithms. Areas A- Fig. 10. Evaluations of LPM for S-type protocols.
F respectively represent Aberdeen, Ashland, Dillon, Hermiston, Moab, and
Madras.
B. LPM (S-type protocols) and PEA-DLEX
The parameter settings of LPM experiments are shown in
optimization principles: adjusting nodes’ rates locally needs Table II. Since the Li-ion battery has a large capacity, we set
no communication but may lead to poor fairness and battery the small initial battery level to show that our LPM algorithm
exhaustion; while globally updating rate assignments in every is still suitable for small-size energy buffers. Figure 10 shows
slot can guarantee strict LM optimality and ENO theoretically, the results of S-type LPM algorithm for five days. With the
but will lead to heavy overheads in practice. maximum duty cycle Dxmax (i) computed by LPM of node x at
slot i, the corresponding energy consumption is computed as
VII. E VALUATION Exmax (i) = Dxmax (i)Pactive +(1−Dxmax (i))Pidle . It is obvious
that both the node’s battery level and energy consumption is
We implemented our AutoSP-WSN framework in TinyOS adaptive to the harvested solar power dynamics. During the
2.1.1 [23] and evaluated them through our aforementioned SP- night, although Dmin was set as zero, the battery level still
WSN platform and the Tossim simulator [14]. All real-world linearly reduces, caused by idle CPU cycles (Pidle ) and battery
experiments used the on-line measured solar power, while all leakage (Eleak ).
simulations were based on the public solar database [24]. For
all experiments, the duration of a slot, T, was set as 30 minutes. 1 (sink)

11 4
2

A. Evaluation of Solar Prediction Scheme WC-EWMA 8

5 9
12 15
We evaluated the performance of WC-EWMA algorithm 3

by comparing it with the classic scheme EWMA [7], [20] and 6 7 10

16
13 14

the state-of-the-art algorithm WCMA-PDR [22], based on real

solar data in six geographical areas [24] with the duration of (a) network deployment and routing tree topology
100 days. Table I shows the simulation settings.
node 4 in DLEX
Figure 9 shows the impact of the prediction interval length
battery energy(KJ)

node 2 in DLEX
L on the mean prediction error pred err (defined in [21], node 4 in PEA−DLEX
rates (kbps)

node 2 in PEA−DLEX
[22]). It can be seen that WC-EWMA is more accurate than 10
1

EWMA and WCMA-PDR (the average pred err for WC- 5 1

0 5
EWMA, EWMA, and WCMA-PDR are 18.7%, 26.7%, and 10 0
25.5% respectively), especially for large L scenarios. We 15 time slot 0 10 20
time slot
can also see that WC-EWMA is relatively insensitive to L,
(b) ranked rate assignment (c) battery level evolutions
but the prediction error of WCMA-PDR increases rapidly as
L increases. Furthermore, WCMA-PDR requires to maintain Fig. 11. Real-world experiment results of DLEX and PEA-DLEX.
solar profile and prediction errors of many days, and a large
number of multiplication and division operations. However, We first evaluated the short-term performance of PEA-
WC-EWMA only requires to maintain two M · L-dimensional DLEX scheme on a real 16-node SP-WSN during a prediction
vectors for reference and real power, and O(M · L) simple interval of L=20 slots (i.e. 10 hours). The network deployment
logical and arithmetic calculations for a whole day, resulting and underlying routing tree are shown in Figure 11 (a). We
in much less computational and storage overheads. set DBT =20 J, and randomly set Bx1 = 1000 ± 50% J and
φx = 1000 ± 50% J for different node x. All other parameters To evaluate the performance of SP-BCP, we first compared
were set as the same in the S-type LPM experiment. Fig- the real-world performance of BCP and SP-BCP in our 16-
ure 11(b) shows the ranked sensing rate of every node in every node SP-WSN for three days. We set Dmin =10%, the initial
time slot assigned by PEA-DLEX. Initially, the rate of every battery level as 1.2 KJ ± 30% for different nodes, and sensing
node is assigned to be equal (8.6 kbps), and there was only (packet generation) rate as one packet per two seconds. As
one saturated node (node 2) in the network. At the 6th slot, shown in Figure 14, for SP-BCP, no nodes run out of energy
node 4 changed from unsaturated to saturated, which triggered during the three days. Therefore, ENO were achieved. The
a RD event, and the network existed two level of max-min sink continuously received data in every slot and achieved
rates from then on. During the whole prediction interval, relatively high packet delivery ratio (about 76.8–98.9%) during
the total events triggered by nodes 2 and 4 are 14 and 12 the daytime. The main reasons of packet loss are limited
respectively. Figure 11(c) shows the battery level evolutions of data buffer size and inelastic sensing rates (a flow controller
the two bottleneck nodes 2 and 4 for PEA- DLEX and DLEX could significantly reduce the packet loss). For BCP, however,
respectively. For DLEX, they did not guarantee their final 10 nodes died during the second day, leading to network
state constraints and the node 2 even exhausted battery energy disconnection and significant degradation of network goodput.
in the 19th slot. In contrast, PEA-DLEX prevented battery
exhausting and maintained enough energy for the future poor 12

goodput (packet/second)
SP−BCP SP−BCP

node failure number

BCP BCP
solar harvesting situations.
6 8
To evaluate the long-term performance of PEA-DLEX of
multiple prediction intervals, we also run a simulation for a 4
4
50-node SP-WSN with a randomly constructed routing tree
during 12 days. The parameter setting was same as that in the 0
0 1 2 3
real-world experiment. Figure 12 shows the ENO performance days
0 1
days
2 3

of PEA-DLEX and DLEX. In PEA-DLEX, no node run out of (a) network goodput (b) node failure evolution
battery energy for the 12 days, while DLEX showed several
bottleneck nodes failing, resulting in the disconnection of the Fig. 14. Real-world experiment results of BCP and SP-BCP.
whole network. Both testbed experiment and simulation results
show that PEA-DLEX scheme manages to ensure ENO and
achieves near LM optimality. 8 SP−BCP BCP

6
goodput
node fealure number

DLEX PEA−DLEX
4
2
2

1 0
0 1 2 3 4 5 6 7 8 9 10 11 12
days
0 (a) network goodput
0 2 4 6 8 10 12
days
node failure number

60 BCP SP−BCP
Fig. 12. Simulation results of DLEX and PEA-DLEX .
40

20
C. LPM (D-type Protocols) and SP-BCP
0
We evaluated the D-type LPM algorithm for three days.
0 1 2 3 4 5 6 7 8 9 10 11 12
We set the Dmin =10%. Figure 13 shows that node’s battery days
level is always above the φix (d) curve, which means that the (b) node failure evolution
φix (d) assignment scheme of LPM for D-type protocols can
Fig. 15. Simulation results of BCP and SP-BCP.
provide hard ENO guarantee.

real solar energy

8 Bix(d) φix(d) Figure 15 shows the simulation results for a random
predicted solar energy
0.4
Emax(i)
deployed 50-node SP-WSNs for 12 days. Sensing rate of every
energy (KJ)
energy (KJ)

x
node was set as one packet per seven seconds, and initial
0.2 4 battery level were set as 200 J ± 30% for different nodes
randomly. The simulation shows similar results to the testbed
experiment, BCP failed quickly in the second day, but SP-
0 0
0 1
days
2 3 0 1
days
2 3 BCP achieved sustainable data collection for the whole 12
(a) solar and load energy (b) battery level and φix (d) days. In summary, both experiment and simulation results
show that SP-BCP can improve the end-to-end performance
Fig. 13. Performance of LPM for D-type protocols. of backpressure routing protocols in SP-WSNs.
 VIII. R ELATED W ORK Beside periodical data collection, event-based applications
such as target tracking or information queries are also im-
Autonomic WSNs. There are several efforts that apply the
portant WSN applications. Such applications normally require
autonomic principles to the design of general WSNs [25]–[27],
hard end-to-end QoS (e.g. delay) guarantees. It is straightfor-
mobile sensor networks [28], and body sensor networks [29].
ward to extend AutoSP-WSN to support such applications, by
However, none of these focus on SP-WSNs, or more generally,
mapping local QoS metric Dmin to end-to-end QoS metrics.
WSNs with dynamic renewable energy sources. Furthermore,
More broadly, for large-scale SP-WSNs with mobile sinks
they are all evaluated based on simulations rather than real-
(or data mules), sensor nodes may store sensor data until a
world platforms.
mobile sink pass by, rather than delivery real-time data through
Energy Harvesting WSNs. For solar prediction, the clas-
end-to-end paths. To support delay-tolerant applications in
sic algorithm EWMA [7], [20] is lightweight but suffer from
such networks (e.g. assisted-reporting garbage bins), new
large prediction error, and the weather-aware scheme WCMA-
functionalities should be added to AutoSP-WSNs, such as
PDR [22], is designed for one-slot solar power prediction. In
classifying opportunistic data muling protocols and extending
contrast, our WC-EWMA can predict multiple-slot solar power
LPM by considering topology and buffer-size awareness for
with lower overheads and higher accuracy. Current power
such protocols.
management schemes [4], [7], [8] focus on optimizing per-
node utility rather than supporting end-to-end network proto- R EFERENCES
cols as our LPM. There also exist several network protocols, [1] J. Yick and B. Mukherjee, “Wireless sensor network survey,” Computer
such as MAC [9], routing [11], [13], localization [30], rate Networks, vol. 52, no. 12, pp. 2292–2330, 2008.
control [10], and cross-layer schemes [15]. However, without [2] V. Raghunathan, A. Kansal, J. Hsu, J. Friedman, and M. Srivastava,
“Design considerations for solar energy harvesting wireless embedded
hardware-driven EA and power management support, none of systems,” in Proc. ACM/IEEE IPSN, 2005.
above is actually implemented in real SP-WSN platforms. In [3] R. Vullers, R. Schaijk, H. Visser, J. Penders, and C. Hoof, “Energy
addition, most of them focus on specific components, while harvesting for autonomous wireless sensor networks,” IEEE Solid State
Circuits Mag., vol. 2, no. 2, pp. 29–38, 2010.
AutoSP-WSN not only improves the performance of individual [4] C. Moser, L. Thiele, D. Brunelli, and L. Benini., “Adaptive power man-
components but also systematically integrates them together in agement for environmentally powered systems,” IEEE Trans. Computers,
a real SP-WSN. For a recent comprehensive survey of energy vol. 59, no. 4, pp. 478–491, 2010.
[5] M. Huebscher and J. A. McCann, “A survey of autonomic computing-
harvesting WSNs, we refer the reader to [31]. degrees, models, and applications,” ACM Computing Surveys, vol. 40,
no. 3, pp. 1–28, 2010.
IX. C ONCLUSION AND D ISCUSSION [6] S. Dobson, S. Denazis, A. Fernndez, D. Gaiti, E. Gelenbe, F. Massacci,
P. Nixon, F. Saffre, N. Schmidt, and F. Zambonelli, “A survey of
In this paper, we develop and implement AutoSP-WSN, autonomic communications,” ACM Trans. Auton. Adapt. Syst., vol. 1,
no. 2, pp. 233–259, 2006.
the first autonomic framework for periodical data collection [7] A. Kansal, J. Hsu, S. Zahedi, and M. Srivastava, “Power management
applications in SP-WSN. AutoSP-WSN achieves sustainable in energy harvesting sensor networks,” ACM Trans. Embedded Comput.
data collection, nearly optimal solar power usage, and high Sys., vol. 6, no. 4, 2007.
[8] C. S. Chen; P. Sinha; N.B. Shroff, “Finite-horizon energy allocation
end-to-end performance. We show that adopting autonomic and routing scheme in rechargeable sensor networks,” in Proc. IEEE
principles, especially context awareness and adaptiveness, ben- INFOCOM, 2011, pp. 2273 – 2281.
efit not only the design of AutoSP-WSN architecture as a [9] [Link], H. Tan, and W. Seah, “Design and performance analysis of
MAC schemes for Wireless Sensor Networks Powered by Ambient
whole, but also the individual components, including reliable Energy Harvesting,” Ad Hoc Networks, 2010.
energy awareness support and solar power prediction (WC- [10] K. Fan, Z. Zheng, and P. Sinha, “Steady and fair rate allocation for
EWMA algorithm), local power management (LPM), as well rechargeable sensors in perpetual sensor networks,” in Proc. ACM
SenSys, 2008, pp. 239–252.
as the network protocols (rate control protocol PEA-DLEX [11] Z. Eu, H. Tan, and W. Seah, “Opportunistic routing in wireless sensor
and the routing protocol SP-BCP). Extensive evaluations based networks powered by ambient energy harvesting,” Computer Networks,
on a real-world SP-WSN platform and the Tossim simulator vol. 54, no. 17, pp. 2943–2966, 2010.
[12] K. Zeng, K. Ren, W. Lou, and P. Moran, “Energy aware efficient
demonstrate the effectiveness of the proposed algorithms. geographic routing in lossy wireless sensor networks with environmental
AutoSP-WSN presents the fundamental tradeoffs between energy supply,” Wireless Networks, vol. 15, no. 1, pp. 39–51, 2009.
network performance and cost, which should be considered [13] L. Lin, N. Shroff, and R. Srikant, “Asymptotically optimal energy-aware
routing for multihop wireless networks with renewable energy sources,”
in developing new SP-WSN schemes. WC-EWMA can be IEEE/ACM Trans. Netw., vol. 15, no. 5, pp. 1021–1034, 2007.
directly used, and LPM can be easily modified for other SP- [14] P. Levis, N. Lee, M. Welsh, and D. Culler, “TOSSIM: accurate and
WSN platforms. Furthermore, with the real-time energy budget scalable simulation of entire tinyOS applications,” in Proc. ACM SenSys,
2003, pp. 126–137.
provided by LPM, various adaptive network protocols can be [15] R. Liu, P. Sinha, and C. Koksal, “Joint energy management and resource
easily developed based on the methodologies of designing D- allocation in rechargeable sensor networks,” in Proc. IEEE INFOCOM,
type and S-type protocols. In addition, beside battery-based 2010, pp. 1–9.
[16] O. Gnawali, R. Fonseca, K. Jamieson, D. Moss, and P. Levis, “Collection
SP-WSNs, AutoSP-WSN can also be used in supercapacitor- tree protocol,” in Proc. ACM SenSys, 2009.
based SP-WSNs, where energy awareness can be better sup- [17] S. Moeller, A. Sridharan, B. Krishnamachari, and O. Gnawali, “Rout-
ported. However, due to the limited capacity and huge leakage ing without routes: the backpressure collection protocol,” in Proc.
ACM/IEEE IPSN, 2010.
of supercapacitors, it could be impossible to guarantee either [18] [Link] [Link].
QoS requirement or ENO during night. [19] [Link]
[20] D. R. Cox, “Prediction by exponentially weighted moving averages and
related methods,” Journal of the Royal Statistical Society, vol. 23, no. 2,
pp. 414–422, 1961.
[21] J. R. Piorno, C. Bergonzini, D. Atienza, and T. S. Rosing, “Prediction
and management in energy harvested wireless sensor nodes,” in Proc.
Wireless Vitae, 2009, pp. 6–10.
[22] C. Bergonzini, D. Brunelli, and L. Benini, “Comparison of energy intake
prediction algorithms for systems powered by photovoltaic harvesters,”
Microelectronics Journal, vol. 41, no. 11, pp. 766–777, 2010.
[23] [Link]
[24] [Link]
[25] D. Marsh, R. Tynan, D. O’Kane, and G. M. O’Hare, “Autonomic
wireless sensor networks,” Engineering Applications of Artificial Intel-
ligence, vol. 17, no. 7, pp. 741–748, 2004.
[26] M. Huebscher, J. McCann, and A. Hoskins, “Context as Autonomic
Intelligence in a Ubiquitous Computing Environment,” International
Journal of Internet Protocol Technology, vol. 2, no. 1, pp. 30–39, 2007.
[27] T. Bourdenas, D. Wood, P. Zerfos, F. Bergamaschi, and M. Sloman,
“Self-adaptive routing in multi-hop sensor networks,” in Proc. IEEE
CNSM, 2011, pp. 1–9.
[28] M. A. K.H. Low, W.K. Leow, “Autonomic mobile sensor network with
self-coordinated task allocation and execution,” IEEE Trans. Syst., Man,
Cybern. C, Appl. Rev., vol. 36, no. 3, pp. 315–327, 2006.
[29] S. Thiemjarus and G. Yang, “An autonomic sensing framework for body
sensor networks,” in Proc. ICST 2nd international conference on Body
area networks, 2007, pp. 61–70.
[30] G. Challen, J. Waterman, and M. Welsh, “IDEA: integrated distributed
energy awareness for wireless sensor networks,” in Proc. ACM MobiSys,
2010, pp. 35–48.
[31] S. Sudevalayam and P. Kulkarni, “Energy harvesting sensor nodes:
Survey and implications,” IEEE Commun. Surveys Tuts.,, vol. 13, no. 3,
pp. 443–461, 2011.