Non-Intrusive Appliance Load Identification with the Ensemble of Classifiers

Transcription

1 Non-Intruve Applane Load Identfaton wth the Enemble of Clafer Potr Blk Inttute of Radoeletron and Multmeda Tehnologe Waraw Unverty of Tehnology Waraw, Poland Weław Wnek Inttute of Radoeletron and Multmeda Tehnologe Waraw Unverty of Tehnology Waraw, Poland Abtrat The paper preent the applaton of the lafer fuon to dentfy eletral applane preent n the houehold. The analy baed on the proeng of feature extrated from the urrent gnal, reorded by the dedated data aquton hardware, ntalled n the vnty of the energy meter. The eleted feature are baed on the medum frequeny meaurement. The propoed dentfaton module ue three lafer of the ame type: deon tree, rule and random foret. Expermental reult prove the effetvene of ombnng varou lafer to the ame tak and how ther advantage and drawbak. Keyword NIALM; artfal ntellgene; lafaton; medum frequeny meaurement I. INTRODUCTION The Non-Intruve Applane Load Montorng (NIALM) [] urrently the tandard tool ondered for the analy of the power onumpton n partular deve workng n houehold. The omplete ytem are propoed to determne the energy profle for eah type of applane, dentfyng the mot expenve one. A only one meaurement module (loated n the vnty of the meter) ued here, the man problem the dentfaton oftware, operatng on the et of eleted voltage and urrent feature. Both ue are evaluated, by propong ombnaton of omputatonal algorthm and data vetor for proeng. Multple Artfal Intellgene (AI) approahe were mplemented o far, nludng Artfal Neural Network [2], Support Vetor Mahne [3], rule [4] or Hdden Markov Model [5]. Ther trength le n the ablty to automatally tran and lafy event ourrng n the power network. The mot preng ue the maxmzaton of the applane dentfaton auray, whh the ba for the further energy onumpton analy. To aheve th am, more omplex method may be mplemented, uh a the deep learnng or the enemble of lafer. Alo, omparon of the partular method outome requred to elet the bet one for the partular problem. The paper preent the applaton of the lafer fuon n the NIALM approah to dentfy applane workng n the houehold. It the method of mplementng multple deon makng algorthm to ombne ther deon n attempt to nreae the dentfaton auray ompared to the eparate lafer. Two apet are mportant n the preented approah: eleton of feature extrated from the urrent and voltage gnal (aqured by the pealzed hardware) and mplementaton and onfguraton of lafaton algorthm, reponble for dentfyng all applane operatng n the apartment. The mplemented algorthm belong to the rulebaed approah: deon tree, rule nduton and random foret. The ombnaton of th group of lafer hould be more effetve n the tak than ung any ndvdual method. The paper truture a follow. The derpton of the mplemented NIALM arhteture n Seton 2. The vetor of feature (further alled example) and detal of t generaton are preented n Seton 3. The algorthm mplemented n the enemble are ntrodued n Seton 4. Expermental reult are preented n Seton 5, whle Seton 6 ontan onluon and future propet for the propoed oluton. II. PROPOSED NIALM ARCHITECTURE The ytem mplemented n the preented reearh (Fg. ) degned to work on-lne n the envronment wth multple two-tate applane (workng n the on-off tate, uh a the kettle or the lghtbulb) and fnte tate mahne (uh a wahng mahne or dryer), workng n parallel. It able to detet hange n the deve onfguraton, leadng to the modfed energy onumpton. Th nlude turnng the deve on or off, or hangng t tate (lke the wahng program). The arhteture ont of two part. The frt one the DAQ hardware, loated near the energy meter, typally meaurng the overall energy onumpton n every apartment. It aumed that the aurate applane dentfaton not poble ung only the tandard nfratruture urrently ntalled n the houehold (uh a preented n [6]). Therefore the pealzed module ued to aqure urrent and voltage gnal wth the 2kHz amplng frequeny. It ondered by the Author a the medum range, n the mddle between the ngle [7] and hundred of Hertz (lke n the EMI analy [8]). The eond part of the arhteture the oftware module, proeng the aqured gnal. It degned to perform the analy on-lne, detetng the hange n the onfguraton of the operatng applane. It operaton baed on exeutng three ubequent tage (Fg. 2):. Deteton of hange n the meaured urrent. It the man oure of nformaton about hange n the deve

2 onfguraton (further alled the event ). It deteton performed twe a eond, baed on the analy of the one thouand ample vetor, for whh the maxmum value alulated. Th way t poble to detet the hange n every applane aumng uh event do not our fater than one a half-eond. The event deteted f the dfferene n two ubequent maxmum value are above the deteton threhold θ. The operaton then performed on the envelope of the urrent gnal (Fg. 3), reated from the orgnal urrent waveform by eletng the maxmum value from vetor of ample. Every ndex value (from the x-ax) repreent the half a eond perod. Seleton of the θ value the omprome between the ablty to detet hange even n the energy effent deve (uh a lghtbulb) and the fale alarm (uggetng that the hange took plae although n fat t dd not). Beaue turnng the deve on ometme related wth the tranent tate (hgh pth n the urrent, qukly returnng to the new level), the begnnng of the teady tate mut be ndated (ee rle n Fg. 3). Only then the operaton of the feature extraton from ample may ommene. The blak rle are the value dfferng from the prevou one above the threhold. The red one ndate tme ntant when the applane tate hange ompleted, trggerng t dentfaton phae. The event derbed ether by the ngle red rle (for the hange wthout the tranent tate) or the equene of blak rle followed by the red one (for the onfguraton hange wth the tranent tate). 2. After the deteton, when the urrent gnal already n the teady tate, feature ued for the dentfaton are alulated. They nlude urrent and voltage parameter extrated from the ample vetor (derbed n Seton 3). 3. The feature are the nput to the fuon of lafer module, whh make ther deon ndvdually. After the votng (where repone of all lafer are ombned), the fnal deon d made. predefned amount of tme (for ntane one hour), n whh the urrent and voltage were meaured. All other applane remaned, o the ubequent proeng referred only to th deve. The ample vetor were ued to alulate the feature repreentng the ngle applane and upplement them wth the applane dentfer, formng the tranng example e. Th way, multple vetor ould be extrated, n varou tme ntant of the deve operaton perod. In the preented reearh eah of k applane wa repreented n the et L by 00 labeled example (.e. wth known dentfer of applane, whh requred for the ubequent uperved learnng of the AI approahe [9]). u, Change deteton Fg. 2. The oftware part of the NIALM arhteture e e 2 2 L en n Feature alulaton k 2 n Deon tree Rule nduton Random foret Clafer fuon m 2m nm k + d () Fg. 3. The urrent envelope for the event deteton Fg.. The propoed NIALM arhteture All mplemented lafer are traned n the off-lne mode on the prevouly prepared data et L (). The tranng example e,, e n were obtaned for eah applane eparately durng ther operaton montorng. The partular deve wa workng n the olated envronment for the After detetng the event, all feature are alulated for the extrated vetor of ample n the teady tate. The obtaned value refer to the ombnaton of urrently operatng applane. To dentfy the deve aung the event, the feature mut be obtaned a the dfferene between the feature alulated after the prevou event (n the tme ntant t j-) and after the urrent event (n the tme ntant t j) ee (2). Suh an operaton requred beaue n the teted envronment multple deve work multaneouly, therefore the dentfaton n the partular moment not amed at determnng all urrently workng applane. Intead, only the hange wth repet to the prevou onfguraton deteted and the applane reponble for t dentfed.

3 t t (2) j j III. DATA VECTORS Seleton of the feature requred to dentfy the partular applane the mportant part of the employed mahne learnng heme [0]. It tll unknown, whh parameter of the urrent and voltage gnal are the mot ueful to dtnguh between varou deve []. Therefore t reaonable to ollet ther maxmum et, a durng the tranng of any mplemented lafer only part of them wll be ondered (although eah algorthm mght elet dfferent ubet). The alulaton were performed for 024 ample. The followng haratert were extrated from the urrent and voltage pattern n the teady tate: Ampltude of the frt xteen harmon omponent of the urrent (where the frt omponent at 50Hz) Phae hft of the frt xteen harmon omponent of the urrent Root Mean Square (RMS) of the urrent Mean value of the urrent Maxmum urrent value DC omponent n the urrent Mean power Value of the atve power n the frt xteen harmon Value of the reatve power n the frt xteen harmon The et of overall 69 real-valued feature wa obtaned for the tranng example and alulated after detetng the hange n the urrent level. IV. ENSEMBLE CONSTRUCTION Th eton brefly ntrodue the mplemented lafaton algorthm and ther onfguraton for the enemble. All the eleted method are rule-baed. Knowledge they repreent after the tranng legble for the human, whh faltate undertandng of the enemble operaton and poble modfaton. They dffer n the generalzaton ablty (.e. the poblty to orretly dentfy example not preent durng the tranng). The odd number of algorthm eleted for the fuon mnmze the threat of the deon unertanty when outome of the ubequent module are equally dtrbuted among dfferent applane dentfer [2]. The mplemented algorthm are traned on the ame et L, but rule ontruted n the proe may be dfferent for eah algorthm. Although there are multple approahe to ombne the partal dentfaton nto the fnal deon [3], n the preented reearh the mplet ae wa ondered, where eah lafer ha equal vote and the deon made baed on the majorty. Below eah method preented, nludng the adjuted parameter to maxmze the dentfaton auray. All three method were mplemented n the Matlab envronment by the Author. Th allowed for ontrollng the ode and adjutng the parameter of every approah. Modfaton of the orgnal approahe are brefly derbed n the followng ubeton. A. Deon tree (DT) Th the memory-effent method [4], preentng the tree-lke truture, tartng from the hghet node (root), onneted wth node of lower level and endng wth the termnal node (leave). All node exept the latter, tore tet that ompare the value of the eleted feature wth the threhold. The reult of the tet lead to the node on the lower level, untl the leaf reahed. Seleton of the tet to the node durng the tree ontruton the key element of tranng. The mnmum entropy rteron wa ued for th purpoe. The parameter ntrodued by the Author of the DT wa the method of eletng feature for the tet when multple feature have dental mnmum entropy. B. Rule Induton (RI) The algorthm mplemented by the Author the modfed veron of the la AQ algorthm [5], adjuted to proe the ontnuou data (for detal ee [6]). The generated rule are not onneted to one another, whh allow for frng multple rule at the ame tme. Eah rule ha the IF THEN form, wth the preme (ondton to meet) and the onluon (dentfaton of the applane) part. Ther generaton ont n the teratve generaton of multple preme wth the ame onluon, allowng to over the ame example. Among them, the bet one n the ene of generalzaton eleted and all example overed by t are elmnated from the et L. The proe onluded when there are no more example to over. The parameter of the method the number of anddate for the rule for eah example. C. Random foret (RF) Th the newet approah to the rule-baed ytem, ung multple DT a part of the lafer [7]. Eah tree make the ndvdual deon baed on the nput vetor of feature and the fnal deon made baed on votng, where the majorty of tree pontng at the dental applane dede about the overall outome. Therefore the RF the fuon of tree. Contrary to the DT preented n eton 4A, th tme eah tree bult onderng not only the bet attrbute for the node, but alo the wore one. Th way the traned truture effent from the tattal pont of vew and hould have hgh generalzaton ablte. The adjuted parameter of the RT nlude the number of ontruted tree and the et of attrbute value ued for ontrutng the node. V. EXPERIMENTAL RESULTS The experment were performed n the olated laboratory ontanng all applane workng n varou ombnaton. The data aquton and proeng hardware wa mplemented on the peronal omputer equpped wth the DAQ ard reponble for meaurement. The analyzed deve nluded (n braket the number of the dentfer gven): power-avng bulb (), dryer (2), vauum leaner (3), mxer (4), juer (5) and kettle (6). They were turned on or off n varou moment of tme, leadng to dfferent onfguraton of deve workng at the partular moment. The example of the half-hour experment preented n Fg. 4, where the upper dagram ontan the meaured urrent, whle the lower one how whh applane were atually turned on. Eah olor repreent the ndvdual

4 deve, havng one of two value: 0 (when t turned off) and t dentfer (when t turned on). auray on the tetng et. The olumn d how the auray of the enemble, whh greater than any ndvdual lafer. TABLE II. IDENTIFICATION ACCURACY [%] OF INDIVIDUAL CLASSIFIERS Algorthm DT RI RF d Overall auray Fale alarm auray Applane dentfaton auray The followng onluon an be drawn from the experment: Fg. 4. The example of the operatng equene for x applane In the equene from Fg. 4, 28 event were deteted. Beaue the event deteton threhold θ wa et to 00mA, among the atual hange n the applane onfguraton, the fale alarm were dentfed a well. For eah event the dentfaton proedure wa ommened by every lafer n the enemble. The fragment of the equene proeng hown n Tab.. The frt olumn ontan the ndex of the event, whle the eond one () ontan the atual dentfer of the applane (aordng to prevou agnment of the dentfer to the partular deve). Three ubequent olumn (d DT, d RI, d RF) ontan ndvdual deon made by the lafer, whle the lat olumn (d) ontan the deon made by the enemble. The 0 value for the eght event for the fale alarm (no atual hange n the deve onfguraton ourred). The - value for the RI method mean no rule wa fred - the lafer unable to make any deon. The overall enemble auray hgher than t partal aurae (a dfferent lafer are the bet n detetng fale alarm and dentfyng applane). The mot dffult event to detet the hange of the power avng bulb tate. It epeally hard to dentfy when deve wth hgh power onumpton operate n the bakground. Sometme t wthng on or off med, whh hould be orreted by the montorng ytem durng the next event deteton. Beaue of the large number of fale alarm, ther orret deteton mut be performed by the lafer. In mot ae, t poble to aheve by the enemble. The DT the mot effent n detetng the fale alarm, although not o effent n other ae. Only two (out of 92) of the former were norretly dentfed. The RF ha overally the bet performane, but alo make more mtake durng the fale alarm dentfaton. TABLE I. RESULTS OF THE SEQUENCE PROCESSING No d DT d RI d RF d ? Reult how the nput of every lafer to the enemble. Ther ombnaton allow for mnmzng the norret deon, made by eah module ndvdually. In ome ae even the ombnaton of lafer unable to make deon (event No. 2), a eah produe dfferent output. To olve the problem, the weghtng to the votng hould be ntrodued, upportng the lafer wth the mallet number of norret deon on the tetng data. The omparon of the ndvdual performane of eleted algorthm n Tab. 2. Aordng to them, the outome of the enemble would be the (orret) dentfer 5, a ponted by the RF, whh ha the maxmal VI. CONCLUSIONS The propoed heme of the enemble proved t effeny when workng wth a et of applane operatng n the houehold. Although eah lafer fal n dentfyng multple event, ther ombnaton allow for detetng mot of tate hange. The mot problemat for the module are fale alarm and deteton of the power-avng deve. Identfyng all thee event may need the ophtated votng mehanm or mplementng more lafer to the heme. The problem may be ung all preented method n the embedded mroproeor ytem, requrng low level ode optmzaton. Alo, the module hould be teted on the greater number of deve workng at the ame tme, alo outde the laboratory envronment, where the nfluene of other apartment may be preent. Thee apet wll be nvetgated n the nearet future. ACKNOWLEDGMENT Th work ha been aomplhed wthn the projet PBS2/A4/0/203 The Non-nvave Sytem for Montorng and Analy of Eletrty Conumpton n the Area of the End-uer fnanally upported by the Natonal Centre for Reearh and Development.

5 REFERENCES [] M. Zefman, and K. Roth, Nonntruve Applane Load Montor-ng: Revew and Outlook, IEEE Tran. Conumer Eletron, vol. 57, No., pp , February 20. [2] H.-T. Yang, H.-H. Chang, C.-L. Ln, "Degn a Neural Network for Feature Seleton n Non-ntruve Montorng of Indutral Eletral Load," Pro. th Internatonal Conferene on Computer Supported Cooperatve Work n Degn, Aprl 2007, Melbourne, Autrala, pp [3] T. Onoda, H. Murata, G. Rath, and K.-R. Muller, "Expermental Analy of Support Vetor Mahne wth Dfferent Kernel baed on Non-Intruve Montorng Data," Pro. of the Internatonal Jont Conferene on Neural Network, 2-7 May 2002, Honolulu, pp [4] J. Power, B. Margoan, and B. Smth, "Ung a Rule-Baed Algorthm to Daggregate End-Ue Load Profle from Preme-Level Data," IEEE Computer Applaton n Power, 99, pp [5] J. Z. Kolter and T. Jaakkola Approxmate Inferene n Addtve Fatoral HMM wth Applaton to Energy Daggregaton, Proeedng of the 5th Internatonal Conferene on Artfal Intellgene and Statt (AISTATS) 202, La Palma, Canary Iland. [6] G.W. Hart, Nonntruve Applane Load Montorng, Proeedng of the IEEE, Vol. 80, No. 2, De. 992, pp [7] J. Lao, G. Elafoud, L. Stankov and V. Stankov, "Power Daggregaton for Low-amplng Rate Data," NILM Workhop, Autn, Texa, 3. June 204. [8] S. Gupta, M. S. Reynold and S. N. Patel, EletrSene: Sngle-Pont Senng Ung EMI for Eletral Event Deteton and Clafaton n the Home, Pro. Ubomp '0 Proeedng of the 2th ACM nternatonal onferene on Ubqutou omputng, 200, pp [9] J. Kelly, W. Knottenbelt, Neural NILM: Deep Neural Network Appled to Energy Daggregaton, 205, onlne: arxv.org/pdf/ [0] A. Zoha, A. Gluhak, M. A. Imran, and S. Rajaegarar, "Non-Intruve Load Montorng Approahe for Daggregated Energy Senng: A Survey," Senor (Bael). 202 De., 2(2), pp [] L. Jang, S. Luo and J. L, "Automat power load event deteton and applane lafaton baed on power harmon feature n nonntruve applane load montorng," Pro. 8th IEEE Conferene on Indutral Eletron and Applaton 9-2 June 203, pp [2] P. Blk, Analy of the lafer fuon effeny n the dagnot of the aelerometer, Meaurement, Vol. 67, May 205, pp [3] L. Kunheva, "Combnng Pattern Clafer. Method and Algorthm," Wley and Son, New Jerey, [4] N.-T. Nguyen and H.-H. Lee, "Deon Tree wth Optmal Feature Seleton for Bearng Fault Deteton," Journal of Power Eletron, vol. 8, nº., pp. 0-07, January [5] R. S. Mhalk, I. Mozet, J. Hong, and N. Lavra, The AQ5 ndutve learnng ytem: An overvew and experment, Report 260, Department of Computer Sene, Unverty of Illno at Urbana- Champagn, 986. [6] P. Blk, W.Wnek, The Rule-Baed Method for the Non-Intruve Eletral Applane Identfaton, Pro. IDAACS 205 Conferene, Sept. 205, Waraw, Poland, pp [7] A.-L. Bouletex, S. Jantza, J. Kruppa and I.R. Köng, "Overvew of random foret methodology and pratal gudane wth empha on omputatonal bology and bonformat," Wley Interdplnary Revew: Data Mnng and Knowledge Dovery, vol. 2, nº. 6, pp , November/Deember 202.