Real-Time Recognition and Proﬁling of Appliances through a …sclab.yonsei.ac.kr/courses/12TAI/termproject/keypaper... · 2012. 6. 15. · Real-Time Recognition and Proﬁling of

Real-Time Recognition and Profiling of Appliancesthrough a Single Electricity Sensor

A.G. Ruzzelli, C. Nicolas†, A. Schoofs, and G.M.P. O’HareCLARITY: Centre for Sensor Web Technologies

School of Computer Science and Informatics, University College Dublin, Ireland†Institute of Higher Education in Computer Science and Communication,

University of Rennes, France{ruzzelli, anthony.schoofs, gregory.ohare}@ucd.ie

[email protected]

Abstract—Sensing, monitoring and actuating systems are ex-pected to play a key role in reducing buildings overall energyconsumption. Leveraging sensor systems to support energy effi-ciency in buildings poses novel research challenges in monitoringspace usage, controlling devices, interfacing with smart energymeters and communicating with the energy grid. In the attemptof reducing electricity consumption in buildings, identifyingindividual sources of energy consumption is key to generateenergy awareness and improve efficiency of available energyresources usage. Previous work studied several non-intrusiveload monitoring techniques to classify appliances; however, theliterature lacks of an comprehensive system that can be eas-ily installed in existing buildings to empower users profiling,benchmarking and recognizing loads in real-time. This hasbeen a major reason holding back the practice adoption ofload monitoring techniques. In this paper we present RECAP:RECognition of electrical Appliances and Profiling in real-time.RECAP uses a single wireless energy monitoring sensor easilyclipped to the main electrical unit. The energy monitoring unittransmits energy data wirelessly to a local machine for dataprocessing and storage. The RECAP system consists of threeparts: (1) Guiding the user for profiling electrical applianceswithin premises and generating a database of unique appliancesignatures; (2) Using those signatures to train an artificialneural network that is then employed to recognize applianceactivities (3) Providing a Load descriptor to allow peer appliancebenchmarking. RECAP addresses the need of an integrated andintuitive tool to empower building owners with energy awareness.Enabling real-time appliance recognition is a stepping-stonetowards reducing energy consumption and allowing a numberof major applications including load-shifting techniques, energyexpenditure breakdown per appliance, detection of power hungryand faulty appliances, and recognition of occupant activity. Thispaper describes the system design and performance evaluationin domestic environment.

I. INTRODUCTION

Electricity represents 41% of the total energy used inAmerican homes [1]. The delivered energy use per householddeclines at an average annual rate of 0.6 percent, mostly dueto technological progress in power efficiency [2]. To furtherincrease that trend, smart energy grids are being promoted toaddress optimal management and improved control of energy,by introducing intelligence into the electricity grid. The recentmomentum for Smart Grid meters, visible in a number ofgovernment driven large-scale pilot deployments such as in

Italy [11] and in the US [12], intends to accelerate theirintroduction into households.

Fig. 1. Typical energy consumption in domestic premises

Within this context, embedded sensor networks and actu-ating systems are expected to play a key role in monitoringand reducing building’s overall energy consumption. Recentstandardization efforts have generated a push towards theintegration of sensor systems in building automation systemsand home environments. IEEE 802.15.4, ZigBee [16], andIETF 6LoWPAN/ROLL [5], [6] are enabling technologies thatfacilitates the connections of low-cost sensing and monitoringunits and gather energy consumption information in real-time.Low-power wireless networking has enabled easy access tohouseholds meter readings, making them available to energyutilities for monitoring and control, and to building ownersfor direct feedback on their energy consumption e.g. Tedenergy detective [14]. Fine-grained energy decomposition isnevertheless not available. Key is to process the energy datato finally provide meaningful information to empower buildingowners with hints for reducing their energy cost. To this end,providing a breakdown of the energy expenditure per applianceis of particular interest to identify energy hungry devices andprovide other interesting services to the homeowner (e.g. homeactivity patterns).

Figure 1 shows typical domestic energy consumption overa time period with some appliance activity annotations. The

978-1-4244-7151-5/10/$26.00 ©2010 IEEE

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE Secon 2010 proceedings.

main aim of this work is to develop a low-cost system thatcan attribute names to appliances contributing to each of theenergy spikes in real time with a single energy monitor. In thispaper, we present Real-time Electrical Appliance Recognition(RECAP), a system that provides fine-grained recognition ofappliances in real-time, based on one single Zigbee-basedbuilding energy monitor attached to the main electrical unit.RECAP system components are appliance signature profiling,real-time signature recognition, and intuitive user feedback. Upto now, several works studied non-intrusive load monitoringtechniques to classify appliances. However, the market showsa lack of practical adoption of such techniques due to practicalissues when deploying the systems into into existing buildings.In contrast, this paper focuses on devising a testing a compre-hensive system to allow system plug-and play capability andempower users profiling, benchmarking and recognizing loadsin real-time, which was holding back the deployment of suchsystems into practice.

The rest of the paper is organized as follows: Section IIrefers to existing appliance load monitoring techniques. InSection III, we describe the system challenges and motivateour design choice. In Section IV, we detail the design of theRECAP system including appliance profiling, recognition, thedatabase of signatures and user interface. Section V presentsthe experimental results from a real deployment. Finally,Section VI discusses the system and future work beforeconcluding the paper in Section VII.

II. RELATED WORK

Many approaches to appliance load monitoring have beeninvestigated. Hart paved the way with the Nonintrusive Ap-pliance Load Monitoring (NALM) [4]. NALM segments nor-malized power values, to characterize the power signal insuccessive steps or events, and match them to appliancesignatures. The technique has achieved an average error of6.3% for total household energy consumption. Remainingchallenges to NALM, which are addressed by RECAP, werethe ability to decompose a power signal made of overlappingon/off events on multiple appliances, and to recognize complexappliance patterns.

The load disaggregation algorithm [9] takes a very similarapproach of comparing each change in the total power signal toeach appliance operating range. In order to differentiate trickycases where observed patterns may fit multiple appliances, aclassification of appliances according to their frequency of usebalances the decision making to the frequently used device.

With ViridiScope, Kim et Al. [15] use indirect sensing toevaluate the power consumption of home appliances. Ambientsignals placed near appliances estimate power consumptionby measuring sound and magnetic field variations when ap-pliances are on or off. Even though sound sensors may becheaper than a home energy monitor, one sensor and onetransmitter per appliance are needed. Furthermore, more thanthe unaesthetic aspect, inaccessible or outdoor appliances aswell as the addition of new appliances make the installationand correct operation of sound sensors difficult. In contrast,

RECAP aims at achieving appliance recognition by deployinga single energy monitor clipped around the live wire of themain electrical unit.

Patel et Al. [10] detect the electrical noise on residentialpower lines created by the abrupt switching of electricaldevices and the noise created by certain devices while in op-eration. The approach relies on the fact that abruptly switchedelectrical loads produce broadband electrical noise either inthe form of a transient or continuous noise. The deploymentphase consists in collecting and recording noise signaturesfrom appliances in the on, off and idle states. Aforementionedproblems of variable power drawn by some appliances as wellas concurrent on/off events affect similarly this approach.

Quantum Consulting Inc. developed an algorithm with rulesbased on pattern recognition. The input is the premise levelload data, information about standard appliances and assump-tions about the customer’s behavior [7]. Forty houses wereevaluated during four summer months. Disaggregated loadprofiles have differed by less than 10%. Unfortunately, thissystem requires at least one sensor per appliance deployed forseveral days for the setting of initial operating characteristics. This is not a cost-effective solution and makes the systemhardly applicable in real scenarios.

Farinaccio et Al. [8] use a pattern recognition approachto disaggregate the total electricity consumption in a houseinto the major end-uses. However, this work does notaddress appliance profiling and assumes a constant appliancesignature, which in reality varies with the house/room loadand the way the appliance is set. Other techniques use wiredsolutions or employ smart sockets. This requires retrofittingthe whole building, which is not cost-effective and mayapply only to new structures. In contrast RECAP is basedon a single wireless and low-cost low-power solution thatintegrates profiling of appliances, namely Unique ApplianceSignatures (UAS), storing of signatures for further use,autonomous recognition through machine learning techniqueand a simple user interface.

Overall, although the literature shows some existing re-search activities on this domain, existing systems addressrequirements in a disconnected manner, target specific casesand fail to meet system usability requirements. Up to now thereis a lack of a low-cost tool that addresses system usabilityto empower the user with a system that integrates applianceprofiling, generation of unique signatures, relational signaturestorage, and a basic user interface for appliance activityrecognition, which are the main focus of this contribution toextend the current literature.

III. CHALLENGES

The main challenges in recognizing appliance activity aremainly due to the following:

• Appliances with similar current draw: The systemshould be able to discriminate between two applianceswith similar or same energy consumption;


• Appliances with multiple settings: Some appliances canbe either tuned according to user needs or have differentphases with different associated consumption, e.g. stand-by mode or washing cycles. The system should eitherunderstand the various appliance settings or recognizeappliances based on additional data independent from thechosen setting;

• Parallel appliances activity: The system should disag-gregate appliances activity identifying each constituentaccounting for the total power consumption;

• Environment noise: The system should be resilient toexternal factors such as not-profiled appliances that canbe turned on unexpectedly;

• Load variation: The energy provider can deploy devicesat substation level for power factor correction, which candestabilize the matching with the appliance profile;

• Long appliance cycles: The system should be able tocope with appliance with long working cycle, which mayresult in long profiling periods.

In order to address these challenges, system adaptivityand resilience to dynamic and unpredictable environmentare needed. To this end, the properties of existing machinelearning techniques represent a suitable solution to reach thegoal. In attempting identify an appropriate machine learningtechnique for recognizing appliances, we initially consideredthe following classifiers:

• Markov Chain classifier: Although Markov Chains(MC) are employed in many classification and patternrecognition algorithms a negative aspect is that simpleMarkov Chains can merely handle one state at time. Thismeans that the number of states could grow greatly if wemap each state with a possible combination of appliancesactive at the same time. Although MC can be a suitablesolution for monitoring a limited number of appliances,the system may not scale well to handle appliances inthe order of tens via a single energy meter, which is amajor objective of RECAP. Multistate Markov chains area possible solution to address this issue but they maygreatly increase the complexity of the system when alarge number of appliances is profiled. Another limitationof this solution is its flexibility. If the user wants to adda new appliance, the Markov chain requires a number ofparameters to be set, which can obstruct system usabilityin light of the fact that the system may be used by non-IT experts. In view of such drawbacks, we opted for theinvestigation of a more scalable classifier.

• Bayesian classifier: A main advantage of this solutionis the simplicity of the algorithm. Despite its apparentminimalism, Bayesian classifiers can give appropriateresults with only limited data. A limitation of this type ofclassifiers is the resistance to parameter variations such aspower variation and duration. Since parameter variationsare a key elements due to power factor correction by theenergy providers and signature aging control existing cir-cuit breakersThis factor is key for the scope of appliance

recognition

In contrast to previous techniques, Artificial Neural Net-work (ANN) to perform appliance recognition are manifoldincluding: (1) the ability to handle any type of data (2) theunnecessary prior understanding of appliance behaviour; (3)the easy extensibility to higher number of inputs, many typesof values or dissimilar kind of data; (4) the learning processthat can be automated for example through additional profilingsensors that can turn on/off appliances remotely; (5) the abilityto learn while running through mechanisms of error feedbackfrom the user; (6) the ability to handle multiple simultaneousappliance states. In contrast, a drawback of the ANN solutionis the lengthy training process that may take few minutes, e.g.in the presence of more than 15 appliances to profile or ifsome appliances have long signatures e.g. a washing machinewith a multi-state signature.

IV. SYSTEM DESIGN

A. Data Acquisition System

Recent standardization efforts have generated an increasingtrend towards the integration of sensor systems in building au-tomation systems, allowing the connection of low-cost sensingand monitoring units and the gathering of energy consumptioninformation in real-time.

Fig. 2. Energy Monitoring Data Acquisition System

Although the RECAP system is independent from the com-munication protocol used by the energy monitor, the unit usedfor testing transfers data via a ZigBee-based acquisition systemto a gateway connected to a local machine, which connectsto either a local or remote relational database for storage,as shown in Figure 2. The RECAP system resides on thelocal machine and processes energy data as they arrive fromthe network. In particular RECAP is able to firstly generateappliance signatures and then train an ANN to recognizeappliance activities on the spot. This starts with the applianceprofiling phase, a one-off procedure that allows RECAP tocharacterize appliances that the user wants to recognise. Theprofiling will create a set of unique appliance signatures thatwill then be used for the real-time activity recognition. Tokeep record of appliance activity times, once an appliance isturned on/off, the system records this into a dedicated table ina remote database.


B. Appliance Profiling

A crucial aspect to consider is what parameters will con-tribute to the generation of a given signature. For examplethe real power consumption can discriminate between appli-ances with dissimilar power consumption but may fail whenappliance consumption is similar.

In order to identify the important constituents for a uniqueappliance signature, we now highlight the main electricalparameters for an appliance working on alternate current (AC).According to its internal circuit, an appliance can be ofresistive, inductive, or capacitive, predominance. For examplea kettle is almost purely resistive while a fan can be predom-inantly inductive. Inductors and capacitors affect the powerconsumption by shifting the alternate current with respect tothe alternate voltage. In particular, capacitors delay the currentwith respect to the voltage while the opposite happens forinductors. Considering that the power is the multiplicationof voltage and current, if voltage and current are shifted,the power transferred to the appliance is less. This effect iscaptured by the active and reactive power components, which,in mathematical terms, correspond to real and imaginary partrespectively, as shown in Figure 3. In general, applianceswork through the real power (active), while the reactive power(passive) is due to the presence of storage elements in theappliance circuit (inductors or capacitors), does not work atthe load and heats wires. Pure resistive appliances show noshift of current and voltage, the reactive part is null and allthe power is transferred to the load. In contrast, the largerthe current/voltage shift the greater the imaginary component.Reactive and active powers are key parameters to calculatethe power factor, which is captured by the energy meter.Equation 1 reports the relation between the active, reactiveand power factor.

S = P + jQPf = P/|S| (1)

where S = Apparent/Complex power, Q = Reactive Power,P = Active Power, Pf = Power Factor, and |S|= real part ofthe apparent power.

Fig. 3. Relation between reactive and active power

C. Unique Appliance Signature

Based on the relations between the power components andhow they map to appliance types, this section introduces theconstituents of a unique appliance signature. The real power isthe first important constituent that can discriminate appliancesof dissimilar consumption. To address appliance with similarconsumption, the power factor can discriminate between ap-pliances of resistive, capacitive and inductive types. Following,the peak current relates to the appliance circuit specifics, asit represents the maximum amount of energy the applianceallows before reacting. RECAP collects also RMS currentthat provides consumption information independently from thevoltage given by the energy provider. Finally, peak voltageand RMS voltage relate to the specific voltage providedwhen the signature is made. Overall, the system identifies 6constituents to generate a unique appliance signature, which isthe base to discriminate between multiple appliances activity.Additional factors captured when profiling appliances are thesignature length and the meter sampling frequency. Theseparameters are key to translate signatures from dissimilartypes of energy meters into a standard signature. In fact,when profiling appliances, users may generate signatures ofdissimilar duration in order to capture diverse appliance powermodes. For example, an electric oven presents an initialperiod of almost constant current draw followed by periodicdeactivations when the set temperature is reached a shown inFigure 4. Finally, to avoid inconsistencies between signaturesgenerated with meters at dissimilar sampling frequencies, RE-CAP implements a simple function that translates signaturesinto a standard frequency before storing it in the relationaldatabase. Figure 4 shows power signatures for 4 appliances ofdifferent lengths and standard sampling frequency of 1 valueper minute.

Electric heater

Fig. 4. Active Power Signatures for 4 appliances

D. Signature Database

Once a new appliance is profiled, the signature togetherwith some metadata relative to the appliance model and


location are stored locally and duplicated in a remote database.The duplication allows the creation of a common repositoryof signatures used to train the ANN and share signatureswith other users, namely Unique Signature State Informa-tion (USSI) database. In fact, providing a common signaturerepository for multiple homes can progressively reduce theinitial training phase required by RECAP. Figure 5 showsthe USSI database associated to RECAP . USSI consists of6 main relational tables. 3 main tables, namely CapturedParameters, Physical and Environmental relate directly to thesignature. Since certain types of appliance type with eitherdissimilar models or from various manufacturers are likely tohave dissimilar signatures, the table ”Physical” captures thespecifics of the appliance.

As the USSI database grows, it is necessary to provide tech-niques to present to the user an initial standard set of relevantsignatures. Through the Environmental table the system canprovide a common list of appliance signatures based on userlocation (password protected). It is in fact common to havesame appliance models concentrated within the same area orregion (e.g. electric showers are very common in Ireland andUK while a certain HVAC model are more common in warmercountries). By using the RECAP interface, the user can thenbrowse the list of appliance models in the area or searchfor other signatures should the appliance be not in the list.Currently, the USSI system in RECAP is implemented in anSQL-based relational database.

Furthermore, the Environmental table provides informationof surrounding conditions during measurements as this mayaffect the signature accuracy. USSI was designed with a broaduse in mind such as a large number of signatures generatedby contributors. To address multiple contributors for the samesignature, the database implements a Contributor table thatincludes a confidence rate, which increases according to thereputation of the contributor. We envision that a reputationwould increase based on collection of opinions from otherusers. The USSI system can handle multiple signatures of thesame appliance ID according to the Signature Property table.Similar to contributor reputation, the Energy Meter table storesthe accuracy of the energy meter, which can be used to tunethe appliance recognition algorithm. For example, in RECAPthis would enable testing the meter accuracy and associateaccuracy levels to different activation functions.

E. Training and Recognition

Following the profiling phase, the generated signatures areused to train an ANN for the recognition of appliances.The basic element of an ANN is a neuron, which can berepresented as a simple succession of mathematical operations,such as weight balancing, sum and an activation function asshown in Figure 6.

Each input of a neuron is balanced by a different weightand is then aggregated into an activation function that can beas simple as a step function, or a more complex function suchas hyperbolic tangent. An ANN consists of several neuronsinterconnected. Figure 7 shows a common type of ANN, 3-

Fig. 5. Integration of unique signature state information

Fig. 6. Single neuron showing input weigths, weighted sum and activationfunction

layer ANN, which is in fact the type adopted for RECAP, asprovides a judicious balance between complexity and responsetime. The first layer consists of Inputs Neurons i.e., neuronswith one or more inputs connected to external or internal data.The second layer consists of Hidden Neurons that have inputsconnected to the outputs of the first layer and are not in directrelation with inputs and outputs of the ANN. The Third layerconsists of Output Neurons that have inputs connected to theoutputs of the hidden neurons. Output neurons represent thedirect outputs of the ANN. The connections between the layersand neurons can vary. For example, the input layer can beconnected to the output of the ANN in order to provide afeedback informing the new input state about the previous


output. RECAP uses a similar feedback mechanism to improvethe accuracy by allowing the user to notify the system of anincorrect guess.

Fig. 7. An example of 3-layer ANN showing input, hidden and outputneurons

We now describe the ANN learning process. At the begin-ning of the learning phase, all weights are random. Weight co-efficients of neurons are modified using the data from a train-ing data set, based on combination of appliance signatures.The data and corresponding weight modification propagatedthrough the hidden layer and then the output layer. Duringtraining phase, the output signal generated by the ANN isthen compared against the desired value, namely the target,as given in the training data set. The difference between thetarget and the output layer of the network is called error δ. Thiserror is then back propagated to the hidden neurons and theinput neurons and the weight of each neuron may be modifiedaccordingly. Equation 2 provides the calculation used to tunethe weights.

Wn = Wo − (δ ∗ Lr) (2)

where Wn is the new Weight, Wo is the old weight andLr is the learning rate sets to a constant value to avoidinconsistencies due to rapid weight changes.

The back propagation ends when all the layer weights areadjusted and the ANN is ready for another ”wave” of data.Naturally, the more sophisticated the training data set, the finerthe weights will be tuned to recognize any combination ofappliance activation. To improve system usability, RECAP im-plements an automatic training program (ALP) that allowsautonomous training of the system. ALP uses the generatedsignatures and creates a training data set with all possiblecombinations of appliance activity, which is then used to tunethe neuron weights autonomously.

It is important to determine the correct number of neuronsand the configuration of the network. In general a large numberof hidden neurons may result in long training times and a

system that may perform extremely well on the training dataset but cannot handle unseen data. In contrast, an inadequatenumber of neurons cause the inability to handle complexcombination of appliances and poor results. Although theliterature offers several programs to find the optimal number ofneurons, we opted for tuning the ANN empirically. A numberof trials allowed the identification a saturation point afterwhich the network showed little further improvement. Thecurrent version of RECAP adopts 6 input neurons matchedwith 6 hidden neurons, which performed adequately for the 6inputed parameters used to generate the signatures.

Finally, an important aspect of the ANN is the ActivationFunction, which shapes the output of the neurons. Consideringthat the output of the network should correspond to one of theprofiled appliances, after several empirical trials we adopted ageneral sigmoid function, which is close to a threshold functionwith smooth angles to allow some neuron uncertainty in caseof errors within the order of 5% or more e.g., to account forsmall variation of power delivered by the energy provider.

F. Graphical User Interface

In order to facilitate user interaction, RECAP provides a ba-sic user interface, which integrates the profiling of appliances,the network training and the final displaying of applianceactivity. These functionalities are divided in 4 separate panelsas follows:

• Add An Appliance: As shown in Figure 8, this panelallows the user entering a JPEG image, name/model andtype of appliance to profile. The panel provides also apredetermined list of appliance types already available inthe database for the convenience of the user. Once theinformation is entered, the user can press on ”save andgo to next step”, which is the appliance profiling phase.

Fig. 8. The ”Add Appliance” panel used for inserting the appliances to beprofiled

• Generate A Signature: This panel, shown in Figure 9,is used to record appliances profile in order to generate


signatures. The panel is provided with a list of addedappliances; by clicking on the button ”Start Recording”,the user will be prompted with ”please, turn on theappliance to be profiled”. The system will automaticallyrecognize the delta variation relative to the appliancebeing turned on and will start recording the parameters.For the duration of the profiling period, the button willstay red and display ”recording in progress”. Duringthis period, the system requires that no other appliancesbe turned on or off. The button will turn green afterreceiving a sufficient number of values per each energyparameter. Based on some initial experiments, the mini-mum number of values to build an accurate signature hasbeen set to 4, which correspond to 4 packets from theenergy meter. Obviously, the time needed to generate asignature depends on the meter transmission rate. In ourexperiments, the meter was transmitting at a rate of 1packet/min as preset by the manufacturer. Should systemdetect some anomalies during this recording period, theappliance profiling is aborted and the user may restart theprocess.

Fig. 9. The ”Record Values” panel used for generating appliance signatures

• Training: This panel, as shown in Figure 10, consists ofa list of profiled appliances, on the left hand side of thefigure, which will be used to train the network. In case ofsome unutilized appliances in the list, their removal fromthe training list is done through the Remove button on theright-hand side of the panel. The training of the ANNis performed by pressing the ”Start learning” button,while the expected recognition performance and trainingprogress bar are on the lower part of the panel. Over-lapping appliance activities during profiling may causejagged signatures and poor performing training periods.In case the recognition performance does not meet therequirements dictated by the application, the user cango back to the ”Record Values” panel and re-recordthe uncertain signatures. We also identified that longer

and steady signatures increase the system confidence andprovide a better performance forecast.

Fig. 10. The ”Train system” panel used for training the ANN with profiledappliances

• Recognition: As shown in Picture 11, this panel displaysthe real-time activity of appliances. Active appliancesare displayed in colour while inactive appliances aredisplayed in black and white. At the bottom of the panel,a statistics window allows following the historical datarelative to appliance activity and energy cost associatedto each appliance.

Fig. 11. The ”Watt’s On” paned used for displaying appliances activity andstatistics

V. EXPERIMENTATIONS

The experimentations of the RECAP system have been con-ducted in the kitchen area of the CLARITY centre at Univer-sity College Dublin. We connected the energy monitor to the


miniature circuit breaker in the electrical unit, which controlsthe whole kitchen area and the corridor including sockets andlights. For the experiments, we used the off-the shelf ZEM-30ZigBee-based Energy Monitor from Episensor [3], as shownin Figure 12. This energy monitor is equipped with a currenttransformer (CT) clipped around the live wire of the consumerelectrical unit.

Fig. 12. Episensor ZEM-30 ZigBee Energy Monitor

The first objective of the experiment was to recognize3 main appliances present in the kitchen area: a kettle, amicrowave and a fridge. Initially, 3 appliances were profiledand signatures produced for them. We tested the system forone week and instructed people using appliances and socketsin the area to annotate time of usage. During the week, tehannotation reported that the 3 main appliances were activatedseveral times per day together with other lower consumingdevices, such as lights and a small fan. Occasionally peopleused the sockets to plug in their laptops and phone rechargers.This initial trial showed accuracy over 95% due to a muchhigher power consumption of the 3 appliances than the lightsand other low consuming appliances.

Given this initial encouraging performance, we decidedto test the system in a more demanding scenario in whicha user profiles an appliance that shows both similar powerconsumption and power factor characteristics as other existingsignatures. We identified an electric fire as a resistive appliancewith similar consumption of microwave and kettle. The electricfire was introduced in the kitchen and its signature generated.The objective of this experiment was to test whether RECAPcould still distinguish appliances based on the remainingsignature parameters. Figure 13a shows the real appliance ac-tivity, which was manually annotated and used for comparisonagainst the output from RECAP. Figure 13b shows the rawoutput from the ANN in response to the input energy data. Anappliance is considered active if the output from ANN is equalor greater than 1. The output coming from the system is thenpassed through a filter module, as shown in Figure 13c, whichconverts the raw output into actual appliance activity, displaythe results on the user interface and annotate the activity intoa dedicated table in the database. We collected informationover 65 minutes while many low consuming un-profiled ap-

pliances where also activated occasionally. In contrast, wenever activated the electric fire for the entire duration of theexperiment to simulate the worst-case scenario of having anunknown high consuming appliance. By doing this, the systemcould not reinforce its neurons by comparing the heater againstother appliances activity and learn the difference. In spite ofthis adverse condition, the experiment demonstrated how thesystem guessed correctly 55 times over a total of 65, whichcorresponded to an accuracy greater than 84%.

In order to test system scalability, we performed testsregarding the time needed to train the ANN with a variousnumber of appliance signatures. Average training time showeda large dependence on the machine used and the length ofsignatures. However, we can report with confidence that, fora regular pentium 4 machine, the training remained below1minute for up to 15 appliances while showing a muchlarger increase when more signatures where introduced. Basedon this, a limit of RECAP is the inadequacy to recognize agreat number of appliances in a building. Instead, it shouldbe trained to recognise a subset of appliances activity ata time in the order of 15 or less. Finally, the presentedevaluation demonstrated how RECAP can guess with highlevel of accuracy appliances that have an almost constantlevel of energy consumption, while further tests are neededto understand the performance for appliances with dissimilarpower and changing power consumption due to power cycles.

Fig. 13. Tests performed in the kitchen area at the CLARITY centre. (a)real appliance activity (b) ANN raw output; (c) output filter.


VI. DISCUSSION AND FUTURE WORK

We foresee 3 major user cases for this platform: (1) creatingreal-time energy awareness for home owners to pinpointappliance consumption useful to understand the reasons ofexpenditures to curb energy cost; (2) Enabling load shifting,by autonomously understanding activity, to alter the pattern ofenergy usage so that on-peak energy use is shifted to off-peakperiods; (3) Providing a personalised energy bill includingpersonalised hints as to how to reduce cost by using energy innon peak hours. By informing end users, e.g. that their localsubstation is subjected to repetitive energy peaks, the elec-tricity provider can also give adjunct services such as adviceof how to implement electricity usage shifting techniques forcertain appliances. The advantage for the energy provider isthat it can employ peak leveling techniques aiming at a moreefficient energy distribution by temporary deactivation of somepower stations.

The results presented in this paper are encouraging asobtained in real conditions. However, we acknowledge thatfurther experimentation is needed to test the system formore complex appliances, such as appliances with differentpower cycles that periodically go into stand-by, e.g. ironand electric oven, when the temperature reaches a certainthreshold. Following the experimentations presented in thispaper, currently the RECAP system is undertaking major testsin 20 homes, each of them provided with a ZigBee basedsmart meter around the area of Dublin City. This large scaletest aims at providing substantial scalability tests for a largerange of appliances and scenarios. The RECAP system can bedownloaded from the CLARITY centre Wiki page [13].

Future work will include a more extensive performanceevaluation and user feedback in regards to the applianceprofiling and user interface. Future work will address the fullautomation of signature creation and energy data annotation.Although the RECAP user interface guides the user throughthis process, the system still requires human supervision forprofiling and calibrating appliance load monitoring systems,which delays performance testing in large-scale. A majorfuture work will also include the creation of a Wikipedia forappliance signatures connected to an open database. Appliancerecognition is a powerful tool and affords possible inputs intovaried application domains such as Ambient Assisted Living(AAL) where activity patterns could be determined throughappliance usage.

VII. CONCLUSIONS

This paper presents a detailed description and results ofan intelligent system for Recognition of Electrical Applianceactivities in Real-time, namely RECAP. The objective wasto provide a plug-and-play tool to create energy awarenessby identifying consumption of individual appliances. Thisrepresents a basis for related load shifting techniques aimedat reducing energy expenditure in buildings. The system in-tegrates (1) appliance profiling to generate unique appliancessignatures; (2) a database for storing appliance signatures; (3)a neural network-based technique to recognise the appliances

profiled; (4) a simple graphical interface to guide the userprofiling appliances and displaying their disaggregated con-sumption.

RECAP was tested in a prototypical real kitchen scenarioand showed accuracy greater than 84% for all the casesstudied. The system is based on recent advances on energymonitoring through ZigBee-based low-power wireless com-munication. RECAP is amenable to quick installation withinan existing building and it is suitable to general off-the shelfenergy meters. The system uses a single electrical energysensor clipped around the live wire of the main electricalunit. RECAP is a versatile system with potential applicationsthat include: decomposing the electricity bill, recognisingenergy hungry devices and electricity leaks, and identifyingpossible improvements in terms of energy saving. Ancillaryapplications include recognition of occupant behaviour forambient-assisted living. The RECAP system is available todownload from the Wikipedia of the CLARITY centre [13].

VIII. ACKNOWLEDGMENTS

This work is supported by Science Foundation Irelandunder grant 07/CE/I1147 and through the SFI-funded NationalAccess Programme (NAP).

REFERENCES

[1] EIA. Energy information administration. Use of Energy in the UnitedStates Explained, 2009.

[2] EIA. Energy information administration. In Annual Energy Outlook2009 with Projections to 2030, 2009.

[3] Episensor. Zem30 energy monitor. http://www.episensor.com, 2009.[4] G. Hart. Nonintrusive appliance load monitoring. Proceedings of the

IEEE, 80(12):1870–1891, Dec 1992.[5] IETF. Internet Engineering Task Force, IPv6 for low-power PANs

Working Group (6LoWPAN), 2004.[6] IETF. Internet Engineering Task Force, Routing over low-power and

lossy networks Working Group (ROLL), 2004.[7] B. S. J. Powers, B. Margossian. Using a rule-based algorithm to disag-

gregate end-use load profiles from premise-level data. In Transactionsof the IEEE Computer Applications in Power, 1991.

[8] R. Z. L. Farinaccio. Using a pattern recognition approach to disaggregatethe total electricity consumption in a house into the major end-uses. InEnergy and Buildings, 1999.

[9] M. L. Marceau and R. Zmeureanu. Nonintrusive load disaggregationcomputer program to estimate the energy consumption of major enduses in residential buildings. In Energy Conversion and Management,volume 41, pages 1389 – 1403, 2000.

[10] S. Patel, T. Robertson, J. Kientz, M. Reynolds, and G. Abowd. At theflick of a switch: Detecting and classifying unique electrical events onthe residential power line. In UbiComp 2007: Ubiquitous Computing,LNCS 4717, Springer, pages 271–288, 2007.

[11] P. Petroni, M. Cotti, and O. Bono. The new edge for the enel telegestore:An integrated solution for the remote management of electricity and gasdistribution allowing a total management of the energy consumptions.In Electricity Distribution, 2009 20th International Conference andExhibition on, pages 1–5, June 2009.

[12] PG&E. Smartmeter gas and electric meters installation.http://www.pge.com/smartmeter/.

[13] RECAP. Recognition of appliance and profiling in real-time.http://www.clarity-centre.org/claritywiki/index.php/REAR Tutorial.

[14] Ted. The energy detective. http://www.theenergydetective.com/index.html.[15] Z. C. Younghun Kim, Thomas Schmid and M. B. Srivastava. Viridis-

cope: design and implementation of a fine grained power monitoringsystem for homes. In UbiComp, pages 245–254, 2009.

[16] Zigbee. Zigbee alliance working group web page for rf-lite.http://www.zigbee.org/, 2002.


Documents

Real-Time Recognition and Proﬁling of Appliances through a …sclab.yonsei.ac.kr/courses/12TAI/termproject/keypaper... · 2012. 6. 15. · Real-Time Recognition and Proﬁling of