An Interactive Context-aware Power Management Technique for

Optimizing Sensor Network Lifetime

Jinseok Yang

, Sameer Tilak

and Tajana S. Rosing

Department of Electrical and Computer Engineering, UCSD, La Jolla, CA, U.S.A.

CalIT2, UCSD, La Jolla, CA, U.S.A.

Department of Computer Science and Engineering, UCSD, La Jolla, CA, U.S.A.

Keywords:

Wireless Sensor Network, Environment Monitoring, Power Management.

Abstract:

A key problem in sensor networks equipped with renewable energy sources is deciding how to allocate energy

to various tasks (sensing, communication etc.) over time so that the deployed network continues to gather

high-quality data. The state-of-the-art energy allocation algorithm takes into account current battery level and

harvesting energy and fairly allocates as much energy as possible along the time dimension. In this paper we

show that by not considering application-context this approach leads to very high and uniform sampling rates.

However, sampling the environment at ﬁxed predeﬁned intervals is neither possible (need to accommodate

system failures) nor desirable (sampling rate might not capture an important event with desired ﬁdelity). To

that end, in this paper we propose a novel interactive power management technique that adapts sampling rate as

a function of both application-level context (e.g., user request) and system-level context (e.g harvesting energy

availability). We vary several key parameters including application request patterns, geographic locations,

time slot length, battery end point voltage and evaluate the performance of our approach in terms of energy

efﬁciency and accuracy. Our simulations use sensor data and system speciﬁcations (battery and solar panel

specs, sensing and communication costs) from a real sensor network deployment. Our results show that the

proposed approach saves signiﬁcant amounts of energy by avoiding oversampling when application does not

need it while using this saved energy to support sampling at high rates to capture events with necessary ﬁdelity

when needed. The computational complexity of our approach is lower (O(n)) than the state-of-the-art non-

interactive energy allocation algorithm (O(n

)).

1 INTRODUCTION

Sensor networks are revolutionizing the scientiﬁc

applications by gathering data about the natural

(Mainwaring et al., 2002)(Cerpa et al., 2001)(CRE,

2013)(GRE, 2015) and built environment data at un-

precedented spatio-temporal granularity. A key prob-

lem in sensor networks is deciding how to allocate

energy to various tasks (sensing, communication etc.)

over time so that the deployed network continues to

gather high-quality data. There has been extensive re-

search in the area of power management and resource

allocation algorithms in sensor networks. However,

new challenges arise for sensors equipped with re-

newable energy sources (Chang and Bonnet, 2010).

Fair energy allocation along time dimension in sen-

sor networks with renewable energy is important in

order to support constant operation. It has two im-

plications: allocating as much energy as possible re-

sults in high frequency sampling and fair energy allo-

cation over time results in sampling the environment

at ﬁxed rate. Our experience with real-world sensor

network deployments in collaboration with limnolo-

gists and coral reef ecologists shows that the afore-

mentioned high frequency ﬁxed rate sampling tech-

nique does not work well in practice because of the

following reasons.

Real-world Deployments Depend on Periodic Inter-

action to Maintain Optimal Sampling Regime: Sen-

sor networks need periodic interaction primarily for

the following two reasons. (A) Early identiﬁcation

of system failures: Sensor networks embedded in in-

hospitable environment are prone to fail for a vari-

ety of reasons such as biofouling, exposure to ex-

treme temperature or humidity etc. (B) Identiﬁcation

of interesting trends: Both anticipated (nightly tem-

perature drops) and unanticipated episodic events (ty-

phoon, hurricanes etc.). At present the interaction is

Yang, J., Tilak, S. and Rosing, T.

An Interactive Context-aware Power Management Technique for Optimizing Sensor Network Lifetime.

DOI: 10.5220/0005728600690076

In Proceedings of the 5th International Confererence on Sensor Networks (SENSORNETS 2016), pages 69-76

ISBN: 978-989-758-169-4

Figure 1: Buoy deployed in a lake located in northern Wis-

consin, USA measuring several key limnological variables.

manual, where the domain scientists periodically look

at the incoming data to ensure that it is generating

science-quality data (Chang and Bonnet, 2010). Sci-

entists also often explore the data to see if something

interesting happened in last day or two and whether

the current sampling rate is sufﬁciently capturing the

events with necessary ﬁdelity. At present, it is too

complex to automate this process. This is both due

to lack of priori knowledge of the all possible events

and system failures and specifying and capturing all

the interesting events and system failures. In addi-

tion, even if the events are known, programming and

detecting all possible events makes the system pro-

hibitively complex. In future, as machine learning al-

gorithms will be more sophisticated and sensor net-

works become equipped with more computing power,

we believe that this manual approach will be replaced

by an automated system that requires no human in-

teraction. Never the less, either a real end user or an

automated system will interact with the deployed net-

work on a periodic basis. In this paper we use the term

”user request” to denote both the request generated by

human beings as well as automated systems.

Periodic Sampling at Fixed Rate Is Not Sufﬁcient:

Sampling the environment at ﬁxed predeﬁned inter-

vals is neither reliable (need to accommodate system

failures) nor desirable (current sampling rate might

not capture an important event with desired ﬁdelity)

(Chang and Bonnet, 2010). We now explain this in

detail. Failure: Suppose monitoring system monitors

both temperature and humidity level at every 1 min.

If either measurement is missing the other is useless.

This indicates that the missing value should be com-

pensated by either repeating a measurement within a

few seconds. Interesting events: Consider an applica-

tion that requires sampling a sensor at a high rate (i.e.

10 samples/seconds) when rain is detected and other-

wise a much lower sample rate (i.e. 1 sample/minute).

Scientists therefore want systems that can adapt sam-

pling rates and meet their science requirements. Peri-

odic sampling can often result in either oversampling

(thereby wasting energy) or under sampling (thereby

not capturing an event with necessary ﬁdelity).

Setting Sampling Regime Is Often an Exploratory and

Iterative Process: Scientists are often operating in

unexplored territory and therefore setting up sampling

rate is not a one-time process, but is an iterative and

exploratory process. Scientists typically set the sam-

pling rate to the best of their knowledge and then use

the gathered data to adjust it. This process can take

anywhere from few days to few months.

Based on these observations, we propose a novel in-

teractive power management technique that adapts

sampling rate as a function of both application-level

context (e.g., user request) and system-level context

(e.g harvesting energy availability and stored energy).

We prove the energy-efﬁciency and accuracy of the

proposed approach using data and infrastructure de-

tails (battery levels, sensing every consumption etc.)

from a real-world deployment.

2 RELATED WORK

There has been considerable work in the area of

sensor network reprogramming (Hui and Culler,

2004)(Levis and Culler, 2004)(Naik et al., 2007).

These approaches are mainly designed for real

network-wide software updates and are not suitable

for more frequent sampling rate updates. The in-

dustrial automation systems or building management

systems integrated with control system require guar-

anties for real-timeliness, functional safety, security,

energy efﬁciency, etc (Chen et al., 2010). In these

sensor-actuator networks resource allocation deci-

sions are typically done in a centralized manner (at

the plan data center). In contrast, we propose a fully

distributed approach for energy allocation.

Context has been used extensively for efﬁcient

sensor network protocol design in the area of rout-

ing (Koo et al., 2009)(Zhou and Hou, 2007), clus-

ter formation (Haque et al., 2009), and power man-

agement (Wood et al., 2008). Context aware power

management protocol (Wood et al., 2008) proposed a

context aware power management protocol considers

heterogeneous energy sources in which some nodes

are powered by batteries and others are plugged into

wall. However, they do not consider green energy

sources in their research. Solar energy allocation al-

gorithm (Gorlatova et al., 2011) determines fair en-

ergy allocation along time dimension in systems with

predictable as well as stochastic renewable energy in-

puts. Their energy allocation algorithm – Progressive

Filling (PF) fairly allocates energy over time dimen-

sion and it has O(n

) computational complexity. PF

algorithm starts from time slot 0 and increments its

allocated energy by α until it reaches the target bat-

SENSORNETS 2016 - 5th International Conference on Sensor Networks

Table 1: Power speciﬁcation for our deployment.

Device Power consumption(W)

3G cost+Processing 5

Vaisala Weather Station 0.168

Sonde 3.372

Templine 0.42

Sensing cost 3.96

tery level. Since PF is the state-of-the-art energy al-

gorithm, we use it to compare with the proposed inter-

active technique. In this paper we use interchangebly

use PF algorithm and non-interactive technique. To

the best of our knowledge, this is the ﬁrst work that

proposes a novel interactive power management tech-

nique that adapts sampling rate as a function of both

application-level context (e.g., user request) and the

system-level context (e.g harvesting energy availabil-

ity).

Solar panels are frequently used in sensor net-

works because they can theoretically provide quite

a bit of harvested energy. However, they are not a

reliable, consistent source of energy because of the

Sun’s cycles and the everchanging weather condi-

tions. To that end, short-term solar prediction al-

gorithms (Kansal et al., 2007)(Piorno et al., 2009)

have been studied. Exponentially Weighted Moving-

Average (EWMA) (Kansal et al., 2007) algorithm re-

lies on the assumption that the energy available at a

given time of the day is similar to the energy genera-

tion observed at the same time on the previous days.

The high prediction errors shown by EWMA when

sunny and cloudy days alternate is due to the high im-

pact that the weather conditions of the previous day

have when estimating the energy generation for the

current day. Weather-Conditioned Moving Average

(WCMA) (Piorno et al., 2009) prediction algorithm

avoids this effect by effectively taking into account

both the current and past-days weather conditions.

3 SYSTEM MODEL

Our system consists of two components (1) ﬁeld de-

ployed sensor network (2) data center. The sensor net-

work consists of a network of platforms (e.g. buoys or

towers), which are large enough to house large solar

panels and bulky batteries and an embedded computer

to which multiple sensors (order of 30) are connected

either via serial or Bluetooth link. The computer runs

a low-power operating system and is equipped with

one or more network modalities (e.g, WiFi, cellular,

and satellite). Figure 1 shows our latest deployment

of an instrumented buoy for a lake monitoring ap-

plication. This buoys hosts a variety of sensor for

Figure 2: Interactive Context-Aware Power Management

System Architecture.

monitoring lake processes, including temperature at

twenty seven depths, dissolved oxygen, conductivity,

pH/ORP, ﬂourescence sensors (Chloraphyll a, Blue-

green Algae, and Rhodamine WT) and voltage. These

sensors are connected to an Android Cell phone via

IOIO board. The phone runs the data acquisition pro-

gram and sends data back to a data center over the

cellular network. We use one Instapark 80W Mono-

crystalline solar panel as our green energy source (Ins,

2015). It has following power speciﬁcations: Maxi-

mum Power Voltage: 17.39V; Open Circuit Voltage:

21.97V; Maximum Power Current: 4.61A. Table 1

summarizes the sensing and communication and pro-

cessing power consumption, which we use in our sim-

ulations.

4 INTERACTIVE-POWER

MANAGEMENT

Our approach uses application-context (e.g., feedback

from domain scientists or an automated system run-

ning user-speciﬁed rules) to optimally set sensor sam-

pling rates. Figure 2 describes the proposed power

management framework that runs at each sensor node.

It consists of two major subsystems, namely, power

manager (PM) and and the Interactive Resource Al-

locator (IRA) subsystem. The PM subsystem makes

resource allocation decisions based on the current bat-

tery level and predicted harvesting level (ref. Eq (1)).

The IRA subsystem then adapts the aforementioned

sampling rate in an interactive manner (ref. Algo-

rithm 2). We now describe the details of PF and IRA

algorithms.

Each sensor node divides a time into K slots (Gor-

latova et al., 2011). We denote S = [s

,...,s

] as a set

of allocated energy to K time slots and s

is the allo-

cated energy to slot i. The embedded power manager

allocates energy to each time slot based on the current

batter level and predicted battery level (ref. Eq. (1)).

B(i) is battery level and H(i) is the predicted harvest

level at time slot i. We use B

max

to denote total Battery

capacity and B

min

to denote minimum battery capac-

An Interactive Context-aware Power Management Technique for Optimizing Sensor Network Lifetime

Algorithm 1: Advanced Progressive Filling (APF).

1 avgHarvstEnergy =

∑

i=1

H(i)/K ;

2 s(1 : K) := avgHarvestEnergy ;

3 for i = K; i ≥ 1;i = i − 1 do

4 [over, amount] ← check validity(s(i)) ;

5 if over == TRUE then

6 s(i) = s(i) − amount ;

7 end

8 end

10 Function [over, amount] = check validity(s)

11 B

current

= current battery level

12 for i = 1; i ≤ K; i = i + 1 do

13 B

current

← min{B(i) + Q(i) −s(i), B

max

} ;

14 if B

current

< s(i + 1) then

15 return [T RU E,s(i + 1) − B

current

] ;

16 end

17 end

18 return [FALSE,0] ;

ity. U(·) is utility function that calculates sampling

rate given allocated energy. It is a non-negative, in-

creasing, strictly concave function (Tutuncuoglu and

Yener, 2012). The ﬁrst constraint of Eq. (1) obeys the

energy neutral operation in energy harvested wireless

sensor network(Kansal et al., 2007).

max

∑

i=1

U(s

)

s.t.B(i − 1) + H(i − 1) − s(i − 1) ≤ B(i)

0 ≤ s

≤ B

max

, 0 ≤ H(i), B

min

≤ B(i) ≤ B

max

(1)

In order to solve Eq. (1), PM allocates constant

energy over K time slots as described in Algorithm

1. The validity of this approach has been proved

by (Gorlatova et al., 2011)(Tutuncuoglu and Yener,

2012). This reduces the computational complexity of

PM to O(n) compared to that of PF O(n

However, constant energy allocation do not con-

sider a situation that system needs to consume more

energy than harvested one. Interactive resource allo-

cator (IRA) considere those scenario. We now de-

scribe the IRA subsystem. The sensor node virtu-

ally divides its battery into two parts, B

current

and

saved

. The IRA subsystem (ref. Algorithm 2) inter-

acts with the application (user or automated system)

and then calculates the energy required to meet the

requested sampling rate. Intuitively, when the energy

needed to satisfy the user request (s

req

) is less than

the energy allocated (s(i)) by the APF algorithm (sys-

tem is currently oversampling), IRA turns down the

current sampling rate and saves this extra energy to

saved

while achieving the necessary ﬁdelity. How-

ever, when user requires sampling at higher rate than

the current sampling rate (the system is currently un-

der sampling), the sensor node augments B

current

with

Algorithm 2: Interactive Resource Allocator.

1 s ← Algorithm 1

2 for i = 1; i ≤ K; i = i + 1 do

3 s

req

← sampling rate given s(i) ;

4 if s

req

< s(i) then

5 // Support user request

6 s(i) ← s

req

;

7 B

saved

← B

saved

+ {s(i) −s

req

} ;

8 else

9 if s

req

< s(i) + B

saved

then

10 s(i) ← s

req

;

11 B

saved

= B

saved

− {s

req

− s(i)};

12 else

13 Aggressive : s(i) ← s

req

;

14 Conservative : s(i) ← s(i) ;

15 Hybrid : s(i) ← s(i) + B

saved

;

16 end

17 end

18 end

saved

to support it. When (s(i) + B

saved

< s

req

given

s(i) < s

req

), the system is under sampling because it

does not have enough energy to support the requested

sampling rate. In this case, we consider three policies

(1) Aggressive: The ongoing event is so critical that

the user sees beneﬁt in capturing that even at the cost

of reduced network lifetime. In this case, the IRA

algorithm increases the sampling rate for the given

slot to the requested rate. (2) Conservative: IRA al-

gorithm decides to continue sampling at the current

sampling rate at the cost of reduced ﬁdelity. (3) Hy-

brid policy: The system selects the best sampling rate

it can support in a greedy manner. This happens in

the case where although the requested rate is not fea-

sible due to energy constraints, but there is still bene-

ﬁt in increasing the sampling rate to the level that can

be supported. A user monitors collected data and de-

termines current optimal sampling rate that meets the

science requirements. Sensor nodes receive the re-

quest and adjust their conﬁguration based on onboard

power management technique (described later).

5 MATHEMATICAL ANALYSIS

In this section, we theoretically compare the perfor-

mance of interactive and non-interactive power man-

agement approaches in terms of user satisfaction. The

interactive approach makes its decision based on bat-

tery level, predicted harvesting energy, and user re-

quest, while the non-interactive approach typically

considers only the ﬁrst two as its inputs. At a given

slot i, when the allocated energy (s(i)) is larger than

energy required to meet the user request (s

req

), a sen-

sor node can satisfy user request at that slot. However,

SENSORNETS 2016 - 5th International Conference on Sensor Networks

in this case, the node is oversampling and wasting its

energy.



1 i f s(i) ≥ s

req

0 i f s(i) < s

req

(2)

As shown in equation (2), z

deﬁnes the condition of

i’s time slot. When s(i) is larger than s

req

the z

has

1 which indicates that the system is oversampling and

spending extra energy. Otherwise z

has 0 as shown in

equation (2).

Let us deﬁne the probability p = Pr(s(i) ≥ s

req

The average number of slots in which a node spends

oversamples and wastes energy is given by equa-

tion (3). Thus, in this case, the average number of

time slots in which the non-interactive approach over-

spends energy in K · Pr (s

≥ x

E[Z

] = K · p = K · Pr(s(i) ≥ s

req

) (3)

The proposed interactive power manager (ref. Al-

gorithm 2) saves energy when the energy needed to

satisfy user request (s

req

) is less than allocated energy

(s(i)), and uses the saved energy (B

saved

) as a boost

when the energy needed to satisfy a user request is

more than allocated energy. The interactive approach

with hybrid policy fails to satisfy user requests only

when the sum of allocated and saved energy is lower

than the amount of user request, s(i) + B

saved

< s

req

We describe this in equation (4).



1 i f s(i) + B

saved

≥ s

req

0 i f s(i) + B

saved

< s

req

(4)

The s(i) + B

saved

≥ s

req

includes both s

≥ x

and

s(i)+ B

saved

≥ s

req

situations. Thus, the average num-

ber of time slots that satisfy user request with in-

teractive approach is K · {Pr(s(i) ≥ s

req

) + Pr(s(i) +

saved

≥ s

req

)}. This result means that user-interactive

power management always satisﬁes more user re-

quests than the non-interactive mechanism because

Pr(s(i) + B

saved

≥ s

req

) ≥ 0. The interactive approach

with conservative policy will show same performance

with non-interactive one and aggressive policy always

satisﬁes user satisfaction.

6 RESULTS

US climate Reference Network(USCRN), maintains

a database of environmental data collected from vari-

ous monitoring stations across the US. For our simu-

lations, for solar energy prediction, we use data from

USCRN database for Necedah, Wisconsin location

since it is the closest location to our deployment. Our

past research has shown that the state-of-the-art en-

ergy predictors such as Weather-Conditioned Moving

Average, WCMA can be used to accurately predict

the amount of harvesting energy (Piorno et al., 2009).

Therefore, in this paper we use WCMA algorithm for

solar energy prediction. To calculate accuracy, we

use one week worth of sensor data (Wind speed data)

from our deployment. We use Matlab to conduct sim-

ulations.

6.1 Study of Impact of Time Slot Length

Variations on Energy Efﬁciency

In this study, we consider 24 hours duration and vary

the time slot length from 1 (24 slots/day) hour to 24

hours (1 slot/day). We ﬁx 1 sample per 10 min as de-

fault sampling rate. We consider user request pattern

from 10% to 100%. In the case of 10% request pat-

tern, among all time slots, 10% time slot support high

request rate which requires 1 sample every 1 min. Re-

maining request pattern requires 1 sample every 5min.

We use end point battery level as 11.1V. Thus, the

capacity is (12-11.1)*55 = 49.5Wh. Table 2 shows

percentage of energy consumed for each approach for

different time slot lengths and request patterns. As

expected (ref. Table 2), when we decrease request

frequency, the overall energy consumption decreases.

However, we observe an interesting patten when time

slot length is varied. When time slot length is be-

tween 1 hour to 6 hours, the environmental conditions

(for solar energy production) do not vary consider-

ably and the overall energy consumption goes up as a

function of slot length. However, for lengths greater

than 6 hours the environmental conditions within a

slot can vary signiﬁcantly thereby changing the har-

vesting energy production (solar energy availability

during day-night shifts). This results in lower energy

consumption for 12 hours and 24 hours slot lengths as

compared to slots of 1, 2, 3, and 4 hours duration.

6.2 Study of Impact of EPV Variations

on Energy-efﬁciency

Our deployment uses Interstate DCM0055 Lead-Acid

battery (Int, ) with 55Ah capacity with Initial Bat-

tery Level (IBL) as 12V. The technical speciﬁcation

for this battery mentions that there are ﬁve different

End Point Voltage (EPV) levels : 9.6V, 10.2V, 10.5V,

10.8V, 11.1V for this battery. When the battery level

reaches the EPV, it stops working until the recharge

process starts. We then calculate the available/ tar-

get battery capacity for each of the discharge levels

as: (IBL - EPV) * battery capacity. For example, for

9.5V EPL, the target battery capacity is: (12 - 9.6)* 55

= 132 Wh. It can be seen that the energy efﬁciency de-

creases as the application request ratio increases since

An Interactive Context-aware Power Management Technique for Optimizing Sensor Network Lifetime

Table 2: Impact of time slot length and request pattern variations on energy efﬁciency.

Request pattern 1 hr 2 hr 3 hr 4 hr 6 hr 12 hr 24 hr

10% 0.956276 0.834421 0.830564 1.293239 0.899995 0.468826 0.375603

20% 1.457325 1.494519 1.417344 1.342128 1.193348 1.186236 0.448982

30% 1.677017 1.690152 1.821041 1.929334 2.073806 1.251442 0.889197

40% 2.178127 2.179191 2.187712 2.173487 2.440535 1.512372 1.109292

50% 2.508317 2.790583 2.554353 2.662568 2.513921 1.512403 1.402832

60% 3.106887 2.961605 2.884691 3.200636 3.174275 1.903787 1.622956

70% 3.302188 3.279281 2.994695 3.542866 3.541096 2.36027 2.063197

80% 3.607844 3.474945 3.654829 3.983248 4.201438 2.555879 1.916398

90% 3.950051 3.695036 3.801447 4.080857 4.054915 2.686218 2.430062

100% 4.206684 4.086216 4.241561 4.374419 4.275124 2.882064 2.28323

Figure 3: Impact of End Point Voltage variations on energy-

efﬁciency.

in our case each request needs higher sampling rate

(sampling every 1 minute). We can see that the pro-

posed interactive approach is signiﬁcantly more en-

ergy efﬁcient than the non-interactive approach. This

is because the later one allocates as much energy as it

can in a fair manner, which leads to oversampling and

wastage of energy. Figure. 3 also shows that higher

discharging rate cannot use the total capacity, 55Ah

because it draws high current. This situation is ex-

plained by Peukert’s Equation (Doerffel and Sharkh,

2006).

6.3 Study of Impact of Harvesting

Energy Variations on Energy

Efﬁciency and Accuracy

In this paper we employ WCMA (Piorno et al., 2009)

algorithm for solar energy prediction for the inter-

active and non-interactive approaches. Solar energy

availability varies signiﬁcantly as a function of geo-

graphic location and season. To understand its im-

pact on the performance of our approach in this study

we consider solar energy variations during the win-

ter (2012/01/4 - 2012/01/10) season at three different

geographic locations in the United States, namely, 1)

Necedah, Wisconsin (44.0262, -90.0737), 2) Austin,

Texas (30.25, -97.75), and 3) Santa Barbara, CA

(34.425833, -119.714167). We set default sampling

interval to be 10 minutes and high request sampling

interval to be 1 minute.

Our results indicate that the proposed interactive

approaches are orders of magnitude more energy ef-

ﬁcient than the non-interactive approach. In partic-

ular, Table 3 shows the the percentage of remaining

battery level after one week of operation. In case

of non-interactive approach it is just 0.8152% for the

Wisconsin winter case. The accuracy is expressed in

terms of Root Mean Square Error (RMSE). Note that

as shown Table 4 error of this approach is quite high

(RMSE = 2.3057). We calculated that PF algorithm

allocated energy to sample sensors every 17 seconds.

This is counterintuitive because the approach sam-

ples data at very high frequency (default sampling rate

for the interactive approach is every 10 minutes), but

it still its error is higher than all the interactive ap-

proaches. A careful investigation shows that PF ap-

proach sets it Target Battery Level (TBL) to the bat-

tery end point voltage (9.6V). It will try to allocate

maximum energy during each time slot in a fair man-

ner. However, this includes the stored and harvesting

energy. They assume an ideal solar prediction algo-

rithm that always predicts the harvesting energy ac-

curately. However, WCMA, the state-of-the-art solar

energy prediction algorithm has a relative mean error

of only 10%. When we plug-in this realistic solar en-

ergy prediction algorithm with the non-interactive al-

gorithm, we see that during the one week of operation,

the batter level goes below the end point level (target

battery level) for approximately 11% of the slots. The

system then stops operating thereby completely miss-

ing the sampling opportunities in those slots. In con-

trast, the interactive approaches avoid oversampling

when not needed thereby saving the energy to allow

higher sampling rates upon request. We also observe

that the geographic locations did not have any major

impact on the energy efﬁciency or the accuracy of the

studied protocols.

SENSORNETS 2016 - 5th International Conference on Sensor Networks

Table 3: Percentage of remaining battery after 1 week.

WI CA TX

Non-interactive 0.815 0.981 1.795

Interactive-conservative 99.92 99.92 99.94

Interactive-aggressive 99.83 99.83 99.86

Interactive-hybrid 99.89 99.89 99.91

Table 4: Impact of harvesting energy variations on system

accuracy.

WI CA TX

Non-interactive 2.305 2.498 2.132

Interactive-conservative 1.08 1.08 0.08

Interactive-aggressive 0.01 0.015 0.01

Interactive-hybrid 1.100 1.100 1.100

7 CONCLUSION

The state-of-the-art energy allocation algorithm that

takes into account current battery level and harvest-

ing energy strives to fairly allocate as much energy

as possible along the time dimension. This approach

by not considering application-context leads to very

high and uniform sampling rates. However, sampling

the environment at ﬁxed predeﬁned intervals is nei-

ther possible (need to accommodate system failures)

nor desirable (sampling rate might not capture an im-

portant event with desired ﬁdelity). To that end, in this

paper we propose a novel interactive power manage-

ment technique that adapts sampling rate as a function

of both application-level context (e.g., user request)

and system-level context (e.g harvesting energy avail-

ability). Our simulations use sensor data and system

speciﬁcations (battery and solar panel specs, sensing

and communication costs) for a real sensor network

deployment. Existing interactive algorithm considers

an ideal solar energy prediction algorithm that makes

no prediction errors. However, by plugging-in a real-

istic solar energy prediction algorithm, we show that

the existing approach often leads to draining the bat-

tery below the end point voltage thereby resulting in

lower accuracy while spending high energy (due to

high sampling rate). Our results show that the pro-

posed approach saves signiﬁcant amounts of energy

compared by avoiding oversampling when applica-

tion does not need it and uses this saved energy to

support sampling at high rates to capture event with

necessary ﬁdelity when needed. The computational

complexity of our approach is lower (O(n)) than the

state-of-the-art non-interactive energy allocation al-

gorithm (O(n

)).

ACKNOWLEDGEMENTS

This work was supported in part by the TerraSwarm

Research Center, one of six centers supported by the

STARnet phase of the Focus Center Research Pro-

gram (FCRP) a Semiconductor Research Corpora-

tion program sponsored by MARCO and DARPA.

This work also has been funded by NSF OCI Award

1219504 and a grant from the Gordon and Betty

Moore Foundation.

REFERENCES

Dcm005 batter speciﬁcation. https://www.

interstatebatteries.com/content/product info/specs/

dcm0055.%pdf.

(2013). Creon, the coral reef environmental observatory

network. http://www.coralreefeon.org.

(2015). Gleon, the global lake ecological observatory net-

work. http://www.gleon.org.

(2015). Instapark solar panel. http://www.instapark.com/

solar-power-panels.

Cerpa, A., Elson, J., Estrin, D., Girod, L., Hamilton, M.,

and Zhao, J. (2001). Habitat monitoring: Application

driver for wireless communications technology. vol-

ume 31, pages 20–41, New York, NY, USA. ACM.

Chang, M. and Bonnet, P. (2010). Meeting ecologists’ re-

quirements with adaptive data acquisition. In Pro-

ceedings of the 8th ACM Conference on Embedded

Networked Sensor Systems, SenSys, pages 141–154,

New York, NY, USA. ACM.

Chen, J., Cao, X., Cheng, P., Xiao, Y., and Sun, Y.

(2010). Distributed collaborative control for indus-

trial automation with wireless sensor and actuator net-

works. volume 57, pages 4219–4230.

Doerffel, D. and Sharkh, S. A. (2006). A critical review

of using the peukert equation for determining the re-

maining capacity of lead-acid and lithium-ion batter-

ies. volume 155, pages 395 – 400.

Gorlatova, M., Wallwater, A., and Zussman, G. (2011). Net-

working low-power energy harvesting devices: Mea-

surements and algorithms. In INFOCOM, Proceed-

ings IEEE, pages 1602–1610.

Haque, M., Matsumoto, N., and Yoshida, N. (2009).

Context-aware multilayer hierarchical protocol for

wireless sensor network. In Sensor Technologies and

Applications, SENSORCOMM. Third International

Conference on, pages 277–283.

Hui, J. W. and Culler, D. (2004). The dynamic behavior of

a data dissemination protocol for network program-

ming at scale. In Proceedings of the 2nd International

Conference on Embedded Networked Sensor Systems,

SenSys, pages 81–94, New York, NY, USA. ACM.

Kansal, A., Hsu, J., Zahedi, S., and Srivastava, M. B.

(2007). Power management in energy harvesting sen-

sor networks. volume 6, New York, NY, USA. ACM.

An Interactive Context-aware Power Management Technique for Optimizing Sensor Network Lifetime

Koo, B., Won, J., Park, S., and Eom, H. (2009). Paar: A

routing protocol for context-aware services in wireless

sensor-actuator networks. In Internet, AH-ICI. First

Asian Himalayas International Conference on, pages

1–7.

Levis, P. and Culler, D. (2004). The ﬁrecracker protocol. In

Proceedings of the 11th Workshop on ACM SIGOPS

European Workshop, EW 11, New York, NY, USA.

ACM.

Mainwaring, A., Culler, D., Polastre, J., Szewczyk, R., and

Anderson, J. (2002). Wireless sensor networks for

habitat monitoring. In Proceedings of the 1st ACM

International Workshop on Wireless Sensor Networks

and Applications, WSNA, pages 88–97, New York,

NY, USA. ACM.

Naik, V., Arora, A., Sinha, P., and Zhang, H. (2007). Sprin-

kler: A reliable and energy efﬁcient data dissemina-

tion service for extreme scale wireless networks of

embedded devices. volume 6, pages 777–789.

Piorno, J., Bergonzini, C., Atienza, D., and Rosing, T.

(2009). Prediction and management in energy har-

vested wireless sensor nodes. In Wireless Commu-

nication, Vehicular Technology, Information Theory

and Aerospace Electronic Systems Technology, Wire-

less VITA. 1st International Conference on, pages 6–

10.

Tutuncuoglu, K. and Yener, A. (2012). Optimum trans-

mission policies for battery limited energy harvesting

nodes. volume 11, pages 1180–1189.

Wood, A., Stankovic, J., Virone, G., Selavo, L., He, Z., Cao,

Q., Doan, T., Wu, Y., Fang, L., and Stoleru, R. (2008).

Context-aware wireless sensor networks for assisted

living and residential monitoring. volume 22, pages

26–33.

Zhou, H.-Y. and Hou, K.-M. (2007). Civic: An power-

and context-aware routing protocol for wireless sensor

networks. In Wireless Communications, Networking

and Mobile Computing, 2007. WiCom 2007. Interna-

tional Conference on, pages 2771–2774.

SENSORNETS 2016 - 5th International Conference on Sensor Networks