Automating Energy Management in Green Homes

Nilanjan Banerjee∗

University of ArkansasFayetteville, AR, USAnilanb@uark.edu

Sami Rollins∗

University of San FranciscoSan Francisco, CA, USA

srollins@cs.usfca.edu

Kevin MoranUniversity of San Francisco

San Francisco, CA, USAksmoran@cs.usfca.edu

ABSTRACTHomes powered fully or partially by renewable sources suchas solar are becoming more widely adopted, however energymanagement strategies in these environments are lacking.This paper presents the first results of a study that exploreshome automation techniques for achieving better utilizationof energy generated by renewable technologies. First, usinga network of off-the-shelf sensing devices, we observe thatenergy generation and consumption in an off-grid home isboth variable and predictable. Moreover, we find that reactiveenergy management techniques are insufficient to prevent crit-ical battery situations. We then present a recommendationbased system for helping users to achieve better utilizationof resources. Our study demonstrates the feasibility of threerecommendation components: an early warning system thatallows users of renewable technologies to make more conserva-tive decisions when energy harvested is predicted to be low; atask rescheduling system that advises users when high-powerappliances such as clothes dryers should be run to optimizeoverall energy utilization; and an energy conservation systemthat identifies sources of energy waste and recommends moreconservative usage.

Categories and Subject DescriptorsH.1.2 [Models and Principles]: User/Machine Systems—human factors

General TermsHuman Factors

1. INTRODUCTIONHomes that derive part or all of their energy from renewable

sources, such as solar and wind, are becoming more widelyadopted. Though early adopters of renewable technologies

*lead co-authors listed in alphabetical order

Permission to make digital or hard copies of all or part of this work for

personal or classroom use is granted without fee provided that copies are

not made or distributed for profit or commercial advantage and that copies

bear this notice and the full citation on the first page. To copy otherwise, to

republish, to post on servers or to redistribute to lists, requires prior specific

permission and/or a fee.

HomeNets’11, August 15, 2011, Toronto, Ontario, Canada.

Copyright 2011 ACM 978-1-4503-0798-7/11/08 ...$10.00.

often report that their motivation for adoption was a desireto be more environmentally responsible, new pricing modelssuch as leases have made renewables more affordable. Asa result, renewables have become an alternative for peopleinterested in reducing their energy costs. In the US, thenumber of grid-tied homes with photovoltaic installationsgrew by 40% in 2009 [8,16]. The growth was largely fueled bydecreasing cost of solar technologies and increase in incentivesoffered to consumers who install the technology. These trendsare unlikely to reverse [16].Technology and services that help consumers to manage

the energy generated by their renewables have improvedin recent years. We spoke to one user who installed solarpanels at his home in 2002 and who, in order to log theperformance of the system, manually recorded readings froma control panel on a daily basis over a period of severalmonths. Another user, however, who is in the process ofinstalling solar panels, reports that he will have access to aweb-based service that reports overall energy generation andconsumption at his home. There has also been a significantincrease in technologies, such as the WattsUpMeter [12]and Android@Home [1], that allow users to monitor thefine-grained power consumption of individual appliances, aswell as services, such as Google PowerMeter [4, 7, 10], thatprovide users with convenient access to such information.These technologies, however, do little more than provideaccess to raw data. The user must still attentively monitorreadings, aggregate the information collected by a diverseset of sensing devices, and decide how to use the informationto make smart decisions about energy consumption.

This paper presents the first results of a study that exploresautomated techniques for changing energy consumption pat-terns to achieve better utilization of available resources ingreen homes. Through an empirical case study, we derive fourkey insights that demonstrate the need for and the feasibilityof automated energy management in green homes. First, weexplore how energy is both generated and consumed in anoff-grid home entirely powered by solar panels. Using a net-work of off-the-shelf sensing devices, we observe that energygeneration and consumption in the home is both variableand predictable. Moreover, we observe that the residents inthe home use reactive energy management techniques—theywait until the energy supply is low before taking action toreduce overall energy consumption. This approach, we find,is not sufficient to prevent the critical situation in which agenerator must be used to power the home. Based on ourobservations, we propose a recommendation-based systemthat monitors energy harvested and consumed and provides

19

Charge ControllerMate Remote

Inverter

Primary monitor

Lead Acid Batteries

Solar Panels

Figure 1: The monitoring setup at the off-grid, solar-powered home. The primary monitoring device is aFitPC tethered to a Mate remote control that profiles the inverter. It is possible to access the solar chargingand energy expended data over a standard RS232 interface using custom software.

users with feedback and advice on how to adjust energy con-sumption patterns via a smartphone application. Our studydemonstrates the feasibility of three recommendation compo-nents: an early warning system that allows users of renewabletechnologies to make more conservative decisions when en-ergy harvested is predicted to be low; a task reschedulingsystem that advises users when high-power appliances suchas clothes dryers should be run to optimize overall energyutilization; and an energy conservation system that identifiessources of energy waste and recommends more conservativeusage.

This exploratory study focuses on an off-grid, solar home inArkansas. We recognize that off-grid users are an extremelysmall segment of the population, and our goal is not todevelop technology that will only be useful in an off-gridhome. There are two significant benefits to working in thisspace, however. First, these users are at the extreme end ofthe spectrum when it comes to conservation. Understandingand classifying the manual techniques they have developedfor demand-response can inform the recommendations we canmake to other grid-tied, or even standard grid users. Second,off-grid users are likely to be enthusiastic early adoptersof our technology, helping us to validate and improve ourtechniques without a significant barrier to adoption.

2. AN OFF-GRID SOLAR HOMETo understand the semantics governing energy utilization

in green homes, we have deployed a suite of monitoring toolsin a solar-powered home in Fayetteville, Arkansas. The goal ofthe monitoring infrastructure is to drive measurement-drivenanalysis to understand how users in solar-powered housesinteract with energy. The house is occupied by a couple; oneoccupant works for the University of Arkansas and the otheroccupies the house most of the day.

Our deployment is shown in Figure 1. The house is locatedseveral kilometers from the city and is powered by 8 Kyocera120 W solar panels [11]. The solar panels drive 24 2V 840Ah Lead Acid batteries (two banks of 12 batteries = 24V)through a FlexMax 60 charge controller [13]. A VFX3524Power Inverter is used to transform DC current to AC, whichpowers the appliances at home. We have augmented theabove setup with a monitoring station that consists of alow-power FitPC [6]. A custom software tool on the FitPCperiodically collects data from a controller device (calledthe MATE remote [14]) that collects data from the inverter.

Data at the rate of one sample per minute is captured overthe RS232 interface and written to an append-only log. Eachsample consists of the instantaneous residual battery volt-age (in Volts) and the energy consumed by the house (inKWh). The amount of energy scavenged from the panels canbe calculated using these values and the maximum batterycapacity. At this time the log files are manually transferredby the house occupants. In addition to monitoring the cumu-lative energy consumption for the home, we have deployednetworked WattsUpMeters to monitor individual appliances.We currently monitor a television and a refrigerator. We havecollected data over a period of 55 days (14 days during thesummer of 2010 and 41 days in Nov-Dec 2010) and plan tocontinue collecting data over the next two years.

3. DATA ANALYSISAnalysis of the data collected by our monitoring infras-

tructure highlights four key findings that have informed thedesign of our automated energy management system.

Traditional energy management techniques are in-sufficient in off-grid homes.First, we consider whether established norms of energy

management, for example the 7PM-7AM policy that recom-mends users run important household appliances between7PM and 7AM when the demand for power is low, are ap-propriate for green homes. Figure 2 shows the total energyconsumption in the off-grid house at different times of theday over 41 winter days. While the energy consumptionvaries considerably over 24 hours, we find that a large por-tion of energy is consumed between 10AM-8PM (10-20 onthe x-axis). Conversations with our subject indicate thatconsumption is primarily driven by residual energy in theirbatteries. During the day when the batteries are close tofull, the users maximally use their household appliances likewashers and water heater. This is, in fact, the opposite ofthe 7PM-7AM strategy. These findings demonstrate a needto develop new energy management techniques that considerthe unique energy harvesting conditions in homes poweredby renewable technologies.

20

0 5 10 15 20 250

50

100

150

200

250

300

Hour number (from 12:00 AM)

Ener

gy C

onsu

mpt

ion

(Wh)

Figure 2: Energy consumptionduring different hours of theday averaged over 41 winterdays. The errorbars indicatethe standard deviation across41 days.

1 2 3 4 5 6 7 8 9 10 11 12 13 14

24

25

26

27

28

29

30

Day

Batte

ry V

olta

ge (V

)

Figure 3: Variability in batteryvoltage in the off-grid homeacross a single day over 14 sum-mer days in 2010.

1 5 10 15 20 25 30 35 40

24

26

28

30

Day

Batte

ry V

olta

ge (V

)

Figure 4: Variability in batteryvoltage in the off-grid homeacross a single day over 41 win-ter days in 2010.

Variability in energy generation and consumptionsuggests the need for adaptive management strate-gies.We next consider whether a simple, fixed strategy for en-

ergy management would be appropriate for off-grid homes.Figure 3 and Figure 4 are box-plots of the battery voltage(in the lead-acid batteries in the off-grid home) over 14 sum-mer and 41 winter days in 2010. Each point on the x-axiscorresponds to data collected over a single day. The batteryvoltage is a function of the energy harvested from the panelsand energy consumed by the house. Although the overallenergy consumed per day, shown in Figure 5 varies, a largeproportion of the variability in the battery voltage can beattributed to the energy harvested from the panels. From thefigure, we conclude that energy harvested from the panelsand energy consumed by the house are both highly variable.There is variance across a single day, across several days, andacross seasons. This suggests that fixed energy managementstrategies are insufficient and adapting to variability is a keyelement for green homes.

Reactive techniques do not prevent critical batterysituations.

We further analyze whether manual techniques employedby our green home user are sufficient. Our user reports thathe is extremely conservative; he monitors the voltage of hisbatteries and reduces consumption, for example by switchingoff lights, when the voltage drops below 90%. Even so, thereare periods of time when he must use a generator to power hishome. We observed that he used the generator on 10 days dur-ing the 41-day winter data collection period—approximately25% of the time. To understand what influenced the use ofthe generator, in Figure 10 we plot the battery voltage over24 hours for three instances before the generator was used. Itis clear that the period saw a gradual decrease in the voltageand the AC generator was used when the residual batteryvoltage was close to 24V. Our conversation with our subjectfurther ratified the above observation. While he agreed thatmost of the time the AC generator was used when the batterycapacity was low, on one occasion he had used the generatorto “equalize” the batteries—a process of overcharging thebatteries for a few hours which causes boiling of the elec-trolyte (acid) and helps breakup any calcification on the lead

plates and stratification of the acid1. To understand whetherour subject used any techniques to prevent the use of thegenerator, we plot the total energy consumption of the houseover 24 hours before the generator was used in Figure 6.Although there is a decrease in energy consumption in thelast 10 hours, it did not circumvent the use of the generator.Hence, manual reactive techniques to prevent energy outagesare not sufficient to prevent the system from running out ofcharge.

Energy generation and consumption exhibit predictabil-ity.Finally, we consider whether energy generation and con-

sumption patterns in green homes can be predicted in orderto inform optimal energy management decisions. To quantifythe predictability of the harvester source and energy con-sumed by appliances, in Figure 7 we calculate correlograms(autocorrelation at different lags) from the battery voltagedata and energy consumed by the television set in our solarhome. A lag corresponds to one hour. Hence, if the autocor-relation at lag k is ak, it implies that the correlation betweendata at time t and t+ k hours is ak. For battery voltage, theautocorrelation decreases with increasing lag, hence the bat-tery capacity is predictable at short time spans—for example,around every 5-10 hours. However, the television data showshigh correlation at short lags and lags around 20 hours. Thehigher correlation at 20 hours is due to the periodic use ofthe TV set. Our subject watches the television (located inhis kitchen) when he cooks in the evening. Therefore, whilethere is variability in energy consumed and generated, thereis considerable predictability in the data, pointing to thefeasibility of automated and proactive energy managementschemes.

4. SYSTEM DESIGNOur measurement analysis demonstrates both the need

for and the viability of energy management techniques thatconsider the unique properties of renewable energy sources.We propose a system that exploits these properties and offersthe following contributions:

1We found that on one instance (not shown in the figure)the generator was used when the voltage was close to 29V

21

0 10 20 30 40 500

1

2

3

4

Day

Ener

gy C

onsu

med

(KW

h)

winter

summer

Figure 5: Total energy con-sumption of the off-grid homeduring 14 summer and 41 win-ter days.

0 5 10 15 20 250

500

1000

1500

2000

Hour Before Generator Was Used

Ener

gy C

onsu

mpt

ion

(Wh)

generator was used

Figure 6: Energy consumptionduring 24 hours before the ACgenerator was used. The er-rorbars indicate the standarddeviation over 10 different in-stances.

0 5 10 15 20 250

0.2

0.4

0.6

0.8

1

Lag (hours)

Auto

corre

latio

n co

effic

ient

TV Energy Consumption (hourly)Average Battery Voltage (hourly)

Figure 7: Autocorrelation atdifferent lags (a lag equal to anhour) for the battery voltagedata and energy consumed bya TV set.

• Early Warning: By predicting times when a home’senergy store is likely to be critically low, the user canbe notified in advance and take proactive measures forreducing energy consumption. This both saves the userfrom the manual task of monitoring the state of thehome’s batteries and it also allows for earlier and moreaccurate detection of critical situations.

• Task Rescheduling: By predicting when energy har-vesting is at its peak or when energy storage is full, theuser can be advised of the best times to execute high-power tasks such as running a dishwasher or clotheswasher.

• Energy Conservation: Even the manual energy con-servation techniques employed by our extremely conser-vative off-grid user can be improved. For each devicefrom which we collect power measurements, we iden-tify a risk [9]—the margin of error that is tolerableby the user. We use this information to suggest moreconservative power states, for example adjusting thetemperature of a refrigerator up one or two degrees,that will not have a noticeable impact for the user.

Though our data demonstrates the utility of these recom-mendations in an off-grid environment, our goal is to designa system that will meet the needs of green home users inboth off-grid and grid-tied homes. While an off-grid user maywant to maximally use power during peak generation times, agrid-tied user may actually want to minimize power use andmaximize revenue by selling power back to the grid duringpeak times. By identifying peak and low generation times inadvance, our system allows a user to decide to maximally orminimally use available energy to meet the desired objective.

We present a recommendation-based system that monitorsenergy generation and consumption and provides users withadvice, including warnings in advance of critical battery situ-ations, recommendations for the best times to execute high-power tasks, and opportunities to adjust the power statesof devices to reduce energy consumption. We have chosena recommendation-based model to minimize user irritationand ensure that control of household appliances ultimatelyresides with the user. A smartphone application notifies theuser when the system suggests changes to the power states of

devices, for example suggests that the user turn an applianceon. The user is responsible for implementing the suggestion,and control of the devices is supported via the smartphoneapplication. Figure 8 illustrates the primary components ofthe system.

Monitoring Infrastructure: The monitoring infrastruc-ture collects raw data on energy generation and consumption.Minimally, the system requires measurements of the energygenerated by a renewable source and the overall energy con-sumption of the house. Our system, however, can be improvedby integrating fine-grained power measurements of individualappliances.

Web Application and Scheduling Logic: A central serverimplements the logic of an optimization algorithm and ex-poses recommendations via a web service accessed by a smart-phone application. A profiler aggregates data collected by allcomponents of the monitoring infrastructure and provides thedata to a scheduler. The scheduler implements an optimiza-tion framework that determines appropriate recommenda-tions. Our present optimization framework (work-in-progress)allocates energy from batteries to appliances proportionalto their consumption in the past and their associated risk.When a new recommendation is available, the scheduler noti-fies the web service, which pushes a notification to an iPhoneapplication. The iPhone application can then retrieve furtherinformation, including specifics about the recommendationor raw data produced by the monitoring infrastructure, forexample the current power draw of a specific appliance.

Control System: The smartphone application, shown inFigure 9, both displays the recommendations and other infor-mation about the status of devices in the home and providesthe user with the ability to control supported devices fromthe phone itself. Commands issued in the application, forexample adjusting the brightness of a lamp, generate requeststo the web service. The web service forwards requests to acontrol infrastructure that can directly control individualdevices. The control infrastructure is an application builton top of the HomeOS [3, 5] framework that provides thenecessary drivers and API to control appliances remotely.

22

Client(phone)

Monitoring Infrastructure

(FitPC, Meters)

Device Control (Home OS)

Device Mgr

Profiler

Scheduler

Web Application

Figure 8: The design of a recommendation-based system for improving energy utilizationin a green home.

Figure 9: A smartphone application that dis-plays system recommendations, collects userfeedback, and allows control of selected appli-ances.

5. FEASIBILITY ANALYSISWe are currently in the process of testing our components

and deploying our recommendation system in the off-gridhome. In this section, we demonstrate the feasibility of gen-erating useful recommendations.

Battery voltage can be used to predict when thegenerator will be necessary.The early warning system requires accurate predictions

of when the AC generator is likely to be used. Moreover, itis critical that the prediction occurs long before the actualevent so that remedial actions can be taken. Figure 10 showsthe battery voltage over 24 hours for three instances beforethe generator was used. The trend shows that the generator isused when the voltage either remains close to 24V for a longtime, or has a gradual decrease from a higher voltage to 24V.Such a trend can be captured using simple autoregressive ormoving average models [2], demonstrating the feasibility ofpredicting energy outages long before they occur.

There are instances when the battery voltage is highand overall energy consumption is low.To demonstrate that task rescheduling can help conserve

energy in green homes, we present a scatter plot of batteryvoltage and overall energy consumption of the house in Fig-ure 11. Each point in the plot corresponds to the energyconsumed and the average battery voltage during a particu-lar hour. The figure shows that there are several instanceswhen the battery voltage is low (the AC generator is likely tobe used) and the energy consumption is high. It also demon-strates that it is frequently the case that the voltage is highand energy consumption is low. Such a mismatch points tothe potential for rescheduling delay-tolerant tasks, such asrunning a clothes washer, such that they occur when thereis significant capacity in the batteries that is not being usedfor other critical tasks.

Even our conservative off-grid user could reduce en-ergy usage.

Finally to corroborate the feasibility of energy conservationusing risk associated with appliances, we present results froman experiment using a television set in our solar house inArkansas. Figure 12 shows the power draw of the TV setduring two instances of its use. We observe that on both

occasions, the power draw can be reduced by 3W (quantifiedas risk) without affecting the user.

6. FUTURE WORKThis paper explores the feasibility of using automated

techniques for energy management in homes powered byrenewable technologies. The data collected by a network ofoff-the-shelf components deployed in a single off-grid homedemonstrates the potential for a system that changes energyconsumption patterns to achieve better utilization of availableresources. Our current work focuses on the following areas:

• Improved Monitoring: We are developing auxiliarymonitoring devices that will supplement the data col-lected by the off-the-shelf components with informationabout environmental conditions, such as ambient light,temperature, and humidity, that have a causal rela-tionship with energy consumption (similar to [15,17]).This will help us to derive correlations between energyconsumed and ambient environmental conditions andwill provide insight on how humans duty cycle appli-ances based on environmental conditions. There areseveral challenges to overcome, however, including howto collect data in a non-intrusive way.

• Holistic System: We are building a general-purposeoptimization framework that will adaptively balancethe amount of energy harvested with energy expendedsuch that usability needs are met. This requires identify-ing the importance of individual devices and allocatingthe energy predicted to be harvested accordingly.

• User Evaluation: We plan to deploy the web andiPhone application to collect information about the util-ity of the recommendations generated by the schedul-ing algorithm. The system will collect in situ feedbackabout whether recommendations are useful and cor-relate the recommendations with the power readingsfrom individual devices.

• Wider Deployment: We are in the process of identi-fying other users, in off-grid and grid-tied homes thatare only partially powered by renewable sources, whoare willing to participate in our study.

23

0 5 10 15 20 2524

24.5

25

25.5

26

26.5

27

Hour Before Generator Was Used

Batte

ry V

olta

ge (V

)

Use ofgenerator

Figure 10: Battery voltage over24 hours for three instances be-fore the AC generator was used.Voltages drop from a high orare consistently low before thegenerator is used.

0 100 200 300 400 500 60024

25

26

27

28

29

30

31

Batte

ry V

olta

ge (V

)

Energy Consumption (Wh)

tasks can be rescheduled from these times

task can be rescheduled to these times

Figure 11: Scatter plot show-ing instances when the energyconsumption and battery volt-age are inversely correlated. Re-scheduling tasks can help con-servation.

1.7 1.75 1.8 1.85x 104

0

10

20

30

40

Time (minutes)

Pow

er c

onsu

med

(W)

Risk

Figure 12: Risk associated witha television set—3W of powercan be reduced without affect-ing the user.

AcknowledgmentsWe are thankful to the anonymous reviewers for their in-sightful comments and suggestions. We would also like tothank Ratul Mahajan for sharing the HomeOS platformexecutables with us and our off-grid house subject for volun-teering to have his home monitored. This work was partiallyfunded by National Science Foundation grants CNS-1018112and CNS-1055061. Any opinions, findings, and conclusionsor recommendations expressed in this material are those ofthe authors and do not necessarily reflect the views of theNational Science Foundation.

7. REFERENCES[1] Android @ Home. http://developer.android.com/

guide/topics/usb/adk.html.[2] Autoregressive moving-average models.

http://en.wikipedia.org/wiki/Autoregressive_

moving_average_model.[3] A. B. Brush, B. Lee, R. Mahajan, S. Agarwal, S. Saroiu,

and C. Dixon. Home Automation in the Wild:Challenges and Opportunities. In ACM CHI, 2011.

[4] Cisco Home Energy Management. http://www.cisco.com/web/consumer/products/hem.html.

[5] C. Dixon, R. Mahajan, S. Agarwal, A. Brush, B. Lee,S. Saroiu, and V. Bahl. The Home Needs an OperatingSystem (and an App Store). In ACM Hotnets IX, 2010.

[6] FitPC2. http://www.fit-pc.com/web/.[7] Google Power Meter.

http://www.google.com/powermeter/about/.

[8] Google Trend: Solar Homes. http://www.google.com/trends?q=+solar+homes&ctab=0&geo=all&date=all.

[9] H. Hu. Risk-Conscious Design of Off-Grid Solar EnergyHouses. PhD dissertation, The Georgia Institute ofTechnology, Department of Architecture, Nov. 2009.

[10] Intel Home Energy Meter. http://edc.intel.com/embedded/homeenergy/323157-001US.pdf.

[11] Kyocera Solar Panels.http://www.altersystems.com/catalog/

kyocera-kc120-120-watt-solar-panel-p-569.html.[12] W. meters.

https://www.wattsupmeters.com/secure/index.php.[13] Outback FlexMax 60 Solar Charge Controller.

http://www.altestore.com/store/

Charge-Controllers/Solar-Charge-Controllers/

MPPT-Solar-Charge-Controllers/

Outback-Solar-Charge-Controllers-MPPT/

Outback-Flexmax-60-Solar-Charge-Controller/

p6875/.[14] Outback Mate Remote.

http://www.altestore.com/store/Inverters/

Inverter-Accessories/Remote-Controls/

Outback-Mate-Remote-Monitor-And-Control/p857/.[15] T. Schmid, D. Culler, and P. Dutta. Meter any wire,

anywhere by virtualizing the voltage channel. InBuildSys, 2010.

[16] L. Sherwood. US Solar Market. 2010.[17] J. Taneja, D. Culler, and P. Dutta. Towards

cooperative grids: Sensor/actuator networks forpromoting renewables. In SmarGridComm, 2010.

24