Available online at ScienceDirect. Procedia Engineering 89 (2014 )

Transcription

1 Aville online t ScienceDirect Procedi Engineering 89 (2014 ) th Conference on Wter Distriution System Anlysis, WDSA 2014 A New Indictor for Rel-Time Lek Detection in Wter Distriution Networks: Design nd Simultion Vlidtion Y. Ishido, *, S. Tkhshi Yokohm Reserch Lortory, Hitchi, Ltd., 292, Yoshid-cho, Totsuk-ku, Yokohm, , Jpn Astrct This pper proposes new lgorithm for rel-time lek detection in wter distriution network using rel-time pressure mesurements only. In prticulr, we newly introduce Hed Loss Rtio (HLR), rtio of two differences of pressure mesurements, s n indictor for urst leks in wter distriution network. A detection lgorithm sed on HLRs is then developed. A theoreticl ustifiction is provided for limited clss of wter distriution networks for the proposed lgorithm. Simultion results re lso given to demonstrte tht the proposed lgorithm cn e successfully pplied to wider clss of wter distriution networks The Authors. Pulished y Elsevier Ltd. This is n open ccess rticle under the CC BY-NC-ND license ( Peer-review under responsiility of the Orgnizing Committee of WDSA Peer-review under responsiility of the Orgnizing Committee of WDSA 2014 Keywords: Wter loss; lekge detection; pressure mesurements; dt nlysis 1. Introduction Erly detection of invisile leks in wter distriution network is of gret significnce to most wter utilities. The dely in detection nd repir of filed wter min cn led to lrge mount of wter loss nd serious dmge to infrstructure ner the filure. In densely populted city such s Yokohm, Jpn, the totl dmge cost due to pipe urst mounts to few million US dollrs in some cse [1]. A growing ttention hs een pid in these decdes to reserch on lek detection in which informtion from Supervisory Control And Dt Acquisition (SCADA) is utilized (see e.g., [2-8] nd references therein). Since mny highly-populted cities in the world hve lredy een equipped with SCADA nd some pressure sensors in their * Corresponding uthor. Tel.: ; fx: E-mil ddress: yumiko.ishido.fq@hitchi.com The Authors. Pulished y Elsevier Ltd. This is n open ccess rticle under the CC BY-NC-ND license ( Peer-review under responsiility of the Orgnizing Committee of WDSA 2014 doi: /.proeng

2 412 Y. Ishido nd S. Tkhshi / Procedi Engineering 89 ( 2014 ) wter distriution networks, rel-time lek detection system sed on SCADA, especilly flow nd pressure dt in the distriution network, hs provided mny wter utilities nother option to mnge their wter loss in ddition to the trditionl regulr coustic survey [9], night flow nlysis [10], network clirtion nd simultion [11], trnsient nlysis [12], nd flow sttistics [13, 14]. Mny of the preceding works on lek detection sed on SCADA re long the line of routine ehvior estimtion of flow nd/or pressure sensors from the pst mesurement dt. The system suggested in [2], for exmple, is supposed to detect new urst leks in wter distriution network y detecting sttisticlly significnt difference etween ctul mesurements nd expected ``norml'' vlues for pressure mesurements. Vrious kinds of techniques in dt nlysis nd rtificil intelligence including Artificil Neurl Networks [6], Byesin Networks [7], Support Vector Mchines [8], hve een pplied to improve the estimtion of the routine ehviors of flow nd pressure mesurements so fr. This pper, in contrst, focuses on pre-processing of the ville dt from SCADA for possile improved ccurcy in lek detection in wter distriution networks. In prticulr, we propose Hed Loss Rtios (HLRs) s new lek-detection indictors, nd will develop new detection lgorithm using HLRs. An HLR, which is defined s rtio of two differences of pressure mesurements, is expected to extrct the effect of locl nomlies such s leks nd unusul wter consumption in wter distriution networks. We show y theoreticl result tht the proposed detection lgorithm is expected to e efficient in prticulr clss of wter distriution networks. We then discuss the possile extension to wider clss of wter distriution networks with simultion results. The rest of this pper is orgnized s follows. Section 2 provides the definition of HLRs, new lgorithm for lek detection in wter distriution network, nd theoreticl ustifiction for the proposed lgorithm. Section 3 provides the simultion results. Section 4 concludes the pper. 2. HLRs nd A New Detection Algorithm In this section, we provide the definition of HLRs nd the lek detection lgorithm using HLRs. A theoreticl ustifiction for the proposed lgorithm is lso given for limited clss of wter distriution networks Hed Loss Rtios (HLRs) Definition 1 For given mesurement time, HLR for triple of different pressure mesurements is given y t t t t pi p HLR pi, p, pk, t (1) p p k if when Also, HLR for qudruple of different pressure mesurements t time is given y t t t t pi p HLRp i, p, pk, pl, t (2) p p k if when l When the pressures re mesured t three different loctions with the sme elevtion nd these mesurement loctions re connected y pipes, the right-hnd side of (1) (or (2)) cn e seen s the rtio of hed losses of two different wter flow pths: pth from the node for to the node for nd pth from the node for to the node for (or pth from the node for to the node for ).

3 Y. Ishido nd S. Tkhshi / Procedi Engineering 89 ( 2014 ) A New Lek Detection Algorithm Using HLRs We propose the following lgorithm for rel-time lek detection in given wter distriution network. STEP 1 Strt Lek Detection System, nd set the time. STEP 2 At every mesurement time, clculte HLRs for ll possile triples or qudruples of ville pressure mesurements. Then, plot the vlue for ech HLR. STEP 3 For ny fixed triple or qudruple of sensors, exmine if there is sttisticlly significnt diverge from constnt in the time-series plot of the HLR, i.e., HLR p, p, p, t 0 t T (3) i k STEP 4 If no HLR hs significnt diverge, proceed to STEP 2 for the next time instnt. If significnt diverge is detected in one or more thn one HLRs, the Lek Detection System wrns the network supervisor of possile lek nd proceeds STEP 2 for the next time instnt 2.3. Theoreticl Justifiction for A Limited Clss of Wter Distriution Networks For the lek detection lgorithm proposed in the previous susection, this susection provides theoreticl ustifiction when it is pplied to limited clss of wter distriution networks. Theorem 1 For given wter distriution network nd given time spn, suppose tht the following conditions hold true for ll. i. The wter distriution network hs only one wter source. ii. The wter distriution network hs no wter lek inside. iii. iv. The wter distriution network hs no pump working inside. There exists function such tht ny wter consumption t node in the wter distriution network stisfies t t di i f (4) for some non-negtive constnt. Then, ll HLRs tht re computed y 3 or 4 pressures in the given wter distriution network should e constnt for ll. The proof is omitted due to the spce limittion. Theorem 1 implies tht, for clss of wter distriution networks in which wter is consumed t ll users in the sme wy, if the only one reservoir is working nd ll pumps inside the network re inctive for, then diverge in n HLR from constnt indictes leks t some nodes in the given wter distriution network. This implies tht the proposed lgorithm in the previous susection is expected to e efficient for clss of wter distriution networks in which wter is consumed in similr wy (suppose e.g., district which consists of totlly residentil res), nd tht hs little mount of existing wter leks. 3. Simultion This section is devoted to simultions of the proposed lgorithm. We show the simulted results when the proposed lgorithm is pplied to wider clss of wter distriution networks in which wter is consumed in similr wy, while in the previous section we ssume exctly the sme consumption ptterns in (4). The hydrulic simultions in this section re performed y mens of EPANET2 [15]. We performed simultions for network

4 414 Y. Ishido nd S. Tkhshi / Procedi Engineering 89 ( 2014 ) model with the totl pipe length of 63,088 nd the totl demnd of 2,630 [m 3 /dy]. The network model is produced from rel wter distriution network with the sme size. For this network, we simulted two cses: Lek Incident 1 nd Lek Incident 2. For ll these cses, hydrulic simultion is sequentilly performed for the durtion 5 dys, sy, 13th to 17th, Octoer 2013, with the simultion time step 4 [min]. For this 5-dy simultion, we ssume 4 demnd ptterns t consumers. The four demnds ptterns re ll produced from series of n ctul 5-dy wter distriution dt to rel wter distriution network. Unlike some other hydrulic simultions, though, we develop four different demnd ptterns y dding white noises proportionl to the pressures, to the originl dt. In other words, the four demnd ptterns re similr, ut slightly different ech other in order to express e.g., the vriety in personl life styles of the users. We descrie the two cses in detils elow. I. Lek Incident 1: There is no lek t the eginning time of simultion, i.e., 00:00 on 13th, Octoer A lek hppens t 22:40 on 14th, Octoer The lek is modeled y n emitter in EPANET2 with n emitter coefficient [m 2.5 /hour]. The simulted pressures t three nodes nd for the 5-dys simultion re shown in Figs. 1, 2 nd 3, respectively. The simulted norml pressures, i.e., pressures for the cse of no lek during the whole simultion period, re lso drwn in these figures with roken lines, for comprison. The is given in Fig.4. The HLR for no-lek cse is lso shown y roken line. II. Lek Incident 2: There is no lek t the eginning time of simultion, i.e., 00:00 on 13th, Octoer A lek hppens t 08:00 on 16th, Octoer The lek is modeled y n emitter in EPANET2 with n emitter coefficient [m 2.5 /hour]. The simulted pressures t three nodes nd for the 5-dys simultion re shown in Figs. 5, 6 nd 7, respectively. The simulted norml pressures, i.e., pressures for the cse of no lek during the whole simultion period, re lso drwn in these figures with roken lines, for comprison. The is given in Fig.4. The HLR for no-lek cse is lso shown y roken line. Fig.1 Lek Incident1: - () Zoom-out view; () Zoom-in view. Fig.2 Lek Incident1: - () Zoom-out view; () Zoom-in view.

5 Y. Ishido nd S. Tkhshi / Procedi Engineering 89 ( 2014 ) Fig. 3 Lek Incident 1: () Zoom-out view; () Zoom-in view; Fig. 4 Lek Incident 1: Fig. 5 Lek Incident 2: () Zoom-out view; () Zoom-in view.

6 416 Y. Ishido nd S. Tkhshi / Procedi Engineering 89 ( 2014 ) Fig. 6 Lek Incident 2: () Zoom-out view; () Zoom-in view. Fig. 7 Lek Incident 2: - () Zoom-out view; () Zoom-in view. In either of the cses of Lek Incidents 1 nd 2, we cn oserve tht the selected HLR shows cler divergence from constnt t the lek incident time. The usefulness of this type of indictor is well demonstrted t the oservtion tht pressures themselves, in contrst, hve only very tiny pressure drop mong mny lrge noises. Becuse we cnnot expect exctly the sme routine ehviors in hed mesurements in most of rel cses, this tiny pressure drop is indeed difficult to e detected without n pproprite dt processing. Fig. 8 Lek Incident 2:

7 Y. Ishido nd S. Tkhshi / Procedi Engineering 89 ( 2014 ) While we show in this section only two selected HLRs, mny HLRs hve similr significnt devitions in these simultions. The locliztion of the lek incidents from n ppropritely selected set of HLRs should e one of the future tsks in this reserch. 4. Conclusions In this pper, we hve provided new lgorithm for rel-time lek detection in wter distriution network using pressure mesurements dt only. In prticulr, we hve focused on n online pre-processing of the pressure mesurements nd hve proposed new type of indictor for lek detection. We hve introduced HLRs s new lek incidents indictors, nd hve developed n nomly detection lgorithm using HLRs. We hve provided theoreticl ustifiction for the proposed lgorithm for limited clss of wter distriution networks. Simultion results re lso given to demonstrte tht the proposed lgorithm cn e successfully pplied to wider clss of wter distriution networks. Future tsks in this reserch include possile pplictions of lening intelligence such s Artificil Neurl Networks, Byesin Networks, nd Support Vector Mchines fter the pre-processing proposed in this pper, nd the locliztion of the lek incidents from n ppropritely selected set of HLRs. References [1] JWRC, Technicl report on enhncement of pipe renewl (e-pipe) proect for sustinle wter service, Tech. Rep , Jpn Wter Reserch Center, in Jpnese (2011). URL [2] A. Armon, S. Gutner, A. Rosenerg, H. Scolnicov, Algorithmic network monitoring for modern wter utility: cse study in Jeruslem, Wter Science & Technology 63 (2011) [3] Z. Y.Wu, P. Sge, D. Turtle, Pressure-dependent lek detection model nd its ppliction to district wter system, Journl of Wter Resources, Plnning nd Mngement 136 (2009) [4] M. Romno, Z. Kpeln, D. Svic, Automted detection of pipe ursts nd other events in wter distriution systems, Journl of Wter Resources Plnning nd Mngement 140 (2014) [5] S. R. Mounce, A. Khn, A. S. Wood, A. J. Dy, P. D. Widdop, J. Mchell, Sensor-fusion of hydrulic dt for urst detection nd loction in treted wter distriution system, Informtion Fusion 4 (2003) [6] S. R. Mounce, J. B. Boxll, J. Mchell, Development nd verifiction of n online rtificil intelligence system for detection of ursts nd other norml flows, Journl of Wter Resources Plnning nd Mngement 136 (2010) [7] Z. Poulkis, D. Vlougeorgis, C. Ppdimitriou, Lekge detection in wter pipe networks using Byesin proilistic frmework, Proilistic Engineering Mechnics 18 (2003) [8] S. R. Mounce, R. B. Mounce, J. B. Boxll, Novelty detection for time series dt nlysis in wter distriution systems using support vector mchines, Journl of Hydroinformtics 13 (2011) [9] H. V. Fuchs, R. Riehle, Ten yers of experience with lek detection y coustic signl nlysis, Applied Acoustics 33 (1991) [10] R. McKenzie, C. Sego, Assessment of rel losses in potle wter distriution systems: Some recent developments, Wter Supply 5 (2005) [11] Z. Y. Wu, P. Sge, D. Turtle, Pressure-dependent lek detection modeling nd its ppliction to district wter system, Journl of Wter Resources, Plnning nd Mngement 136 (2010) [12] J. A. Ligget, L.-C. Chen, Inverse trnsient nlysis in pipe networks, Journl of Hydrulic Engineering 120 (1994) [13] S. Bucherger, G. Ndimplli, Lek estimtion in wter distriution systems y sttisticl nlysis of flow redings, Journl of Wter Resources, Plnning nd Mngement 130 (2004) [14] S. R. Mounce, A. J. Dy, A. S. Wood, A. Khn, P. D. Widdop, J. Mchell, A neurl network pproch to urst detection, Wter Science nd Technology 45 (2002) [15] L. A. Rossmn, EPANET2 users mnul, U.S. Environmentl Protection Agency (2000).