Partial Least Squares Regression Model to Predict Water Quality in Urban Water Distribution Systems

Trans. Tianin Univ. 2009, 15: 140-144DOI 10. 1007/s12209-009-0025-2。Tianin University and Springer-Verag 2009Partial Least Squares Regression Model to PredictWater Quality in Urban Water Distribution Systems*LUO Bijun (骆碧君) , ZHAO Yuan(赵元)1.2, CHEN Kai(陈凯), ZHAO Xinhua (赵新华) 1(1. School of Environmental Science and Engincering, Tianjin University, Tianjin 300072, China;2. Architectural Design and Rcscarch Institute, Southeast University, Nanjing 210096, China;3. Tianjin Huashui Water Supply Construction Co., Ltd, Tianjin 300122, China)Abstract: The water distribution system of one residential district in Tianjin is taken as an example to analyze thechanges of water quality. Partial least squares (PLS) regression model, in which the turbidity and Fe are regarded as con-trol objcctives, is used to establish the statistical model. The experimental results indicate that the PLS regression modelhas good predicted results of water quality compared with the monitored data. The percentages of absolute relative error(bclow 15%, 20%, 30%) are 44.4%, 66.7%, 100% (turbidity) and 33.3%, 44.4%, 77.8% (Fe) on the 4th sampling point;77.8%, 8.9%, 8.9% (urbidity) and 44.4%, 55.6%, 66.7% (Fe) on the 5th sampling point.Keywords: water distribution systems: water quality; turbidity; Fe; partial least squares regressionMany factors in space and period can change the ur- value of toxicology and bacteriology target.ban water quality, such as water age, physical, chemical 1.2 Fand biological reactions occuring in the distribution sys-The ferrous tubing is used extensively in the watertems 0. It is impossible to detect water quality of everydistribution systems in China. According to the statisticalsegment and node on the spot because of the complexity result in 1996，the proportion of cast iron pipe is 51.67%of urban water distribution systems. To this end, com- and the steel pipe is 23.85% (41. Serious corrosion and theputer technology can be used in simulating the law of release of Fe exist in ferrous tubing water distributionwater quality in the distribution systemsl-l.systems, which is harmful to human health since the FeAt present, the analysis of water quality in the waterconcentration exceeds the normal standard seriously.distribution systems is to simulate the law of some qual-Studies show that Fe is positively related to turbid-ity parameters on the basis of data collection and assis- ity. The release of Fe is the major factor that influencestance of computer technology. In this paper, turbidity and the turbidity in the distribution systems. Fe gets into wa-Fe (total iron) are regarded as dependent variables in the ter in the form of compounds of Fe*+ and Fe't, by whichmulti-variable water quality model with partial least the sensitive targets of the water body will be influencedsquares (PLS) regression method.greatlys. As a result, Fe in the water can be measured byturbidity.1 Turbidity and Fe in water distributionChemical equilibrium of Fe is occuring all the timesystemsin the microenvironment of the tube wall which consistsof ferrous tubing (metallic iron), furring (iron com-1.1 Turbiditypound) and the water environment of the distributionStudies show that if the turbidity in water is reduced systerms. The chemical reactions involve corrosion, re-to 0.1 nephelometric turbidity units (NTU) or even lease and sediment of Fe. The main reactionsol are aslower, most of the organic pollutants will be removed!3!. follows.Lower turbidity in the water can not only fulfill the re-Second order reactions:quest of sensitive characters, but also help to reduce the中国煤化工Accepted date: 2008-09-10.YHCNMHG*Supported by National Natural Science Foundation of China (No. 50478086) and Tianjin Special Scienific Innovation Foundation(No. 06/2Z5H0900)0LUO Bijun, born in 1983, female, doctorate student.Correspondence to ZHAO Xinhua, E-mail: zxh@tju.edu.cn.LUO Bijun et al: Partial Least Squares Regression Model to Predict Water Quality in Urban Water Distribution SystemsFe2*+20H- >Fe(OH)zinto different parts such as residential area, public build-4Fe2* +O2+8OH*- >4FeOOH+2H2Oing, and campus. Five sampling points are shown inThird order reactions:Fig.1. .4FeCO3+O2+2H2O- +4Fe0OH+4CO2At present, water supply in the district is provided6FeCO3+O2-→2Fe;O4+6CO2by municipal waterworks through 6 water pipes shown inResidual chorine' ”in water distribution systems Fig.1. The tubing of all the pipes in the systemis of castparticipates in the corrosion reaction of the tube wall that iron. In the district, especially in the residential area, themakes the release of Fe,pipes of the distribution systems are collocated so tightly6HCIO+5Fe-→>3FeCl2+2Fe (OH)3that the flow velocity is very low in many pipes accord-The theoretical relationship between other targets in ing to the result of hydraulic calculation, and some arewater such as UV2s4, ammonia nitrogen and Fe is not even close to zero. The change of water quality is cor-clear till now. Therefre, this paper uses the statistical related with the corrosion degree of the distribution sys-theory to set up a quality prediction model.tems because when water enters the residential area, itwill continuously settle in this part of the distribution2 Experimentsystems.The water samples were collected from the hydrantsIn this paper, the water distribution system of one in the buildings constructed along the pipeline of the dis-district in Tianjin is studied considering its location, area tribution systems. The period of sampling time wasand experimental condition. The district can be divided 7:30- 9:30 a.m., from April 12th to May 24th, 2006.= Inlet2-0 Sampling pointFig,1 Sketch map of water distribution system in studied areasults are listed in Tab.l.3 Water quality modelAccording to Tab.1, the total chlorine and tempera-ture have ill correlation with turbidity and Fe. As a result,The PLS regression method supplied by the PLS variables of PLS regression model in this paper are tur-course of SAS/STAT is used. The cross validation bidity, Fe, pH, residual chlorine, UV2s4 and ammoniamethod is used to validate the results.nitrogen.3.1 Independent and dependent variablesThe dependent variables are the turbidity and Fe ofThe relationship between Fe and other targets could the survillance point at time t+1. Here t is the samplingnot be described theoretically. To determine the inder . time, a中国煤化Ie.pendent and dependent variables, this paper chooses theYHCNMHGthe turbidit, pH,4th sampling point as an example, using Excel to analyze UV2s4, uuua uuugui, ivsiuuas uilorineandFeof thethe statstical correlation between eight targets. The re- surveillance point at time I.-141-Transactions of Tanjin University Vol.I5 No.22009Tab.1 Correlation of targets on 4th pointTempera-ResidualAmmoniaTurbidityFeHTotal chlorineUV254turechlorinenitrogen0.645 477Temperature0.300 1880.354 8020.341 028-0.093110.040 7330.053 8760.469 8570.487 3511Residual chlorine-0.229 19-0.288 090.154 941-0.060 790.328 599UV2s40.575 490.698 7080.540 1930.457 7220.413 0480.145 1161Ammonia nitrogen0.467 540.44 4180.483 6570.254 5240.330 5040.317513 0.719906 13.2 PLS model.0 rThe PLS in SAS/STAT system[8,91 is used. According.8 tto the data analysis and the locations in the district of all.66the sampling points, the 4th and 5th sampling points arechosen.+ Monitored value (Ist group)The laboratory data will be removed from the singu-.2 t+ Filing vialuelar value while the individual omissions interpolation willbe flled according to the data to make a total of 30 sets035791315119Timeldof data. The data are divided into two groups, in whichthe former 20 sets (1st group) is to establish and checkFig.2 Comparison of ftting and monitored values ofturbidity on the 4th pointthe model and the latter 10 sets (2nd group) to test the0.25model results.The regression equation of turbidity (41) and Fe:20 t(y42) on the 4th point isC 0.15y4= -0.2809+0.069 6x41+0.189 76x42+号0.10/0.025 8x43+0.076 8xa4+- + Monitored value (Ist group)Fitting value0.05 t0.103 47x4s+2.121 1x46y42=- 0.45119+0.050 46x41+0.137 51x42+05579市519Time/d1.375 1x43+0.055 66x4+0.074 97x45+1.558 7x46 .(2)Fig.3 Comparison of ftting and monitored values of Feon the 4th pointThe regression equation of turbidity (ys1) and FeThe comparisons of the predicted and monitored(ys2) on the 5th point isvalues of turbidity and Fe on the 4th point are shown inys= -0.1655+0.108 59xs1+0.375 3xs2+Fig.4 and Fig.5. Note that the models in this paper use the0.023 52xs3+0.044 33xs4+former sets of data to predict the latter sets of data.(3) Therefore, there are 9 sets of predicted results in total as0.118 2xss+1.426 25xs6shown in Fig.4 and Fig.5. .ys=-0.094 6+0.038 76xs+0.133 9Xxs2+0.7厂0.008 4x3+0.015 82xs4+0.042 18x5s+0.509 1xs6(4)0.5t.4where归is turbidity on the ith point at time t+1, NTU;0.3上yn is Fe on the ith point at time t+1, mg/L; Xxi is turbidit,2卜+ Monitored value (2nd group)NTU; xn is Fe, mg/L; X3 is residual chlorine, mg/L; xi4 is- t - Predicted value0.1十pH; xis is ammonia nitrogen, mg/L; Xxi6 is UV2s4.3.3 Fitting and using of PLS model中国煤化工The comparisons of the fiting and monitored valuesof turbidity and Fe on the 4th point are shown in Fig.2.MYHC N M H G monitored values ofand Fig.3.-142一LUO Bijun et al: Parial Least Squares Regression Model to Predict Water Quality in Urban Water Distribution Systems0.30Monitored value (2nd group)0.250.25-一- Predicted values 0.20云0.20曾0.1.5出0.100.05一Predicted value.0sF“H之34567宫9°T2346789Time/dFig.5 Comparison of predicted and monitored values ofFig.9 Comparison of predicted and monitored values ofFe on the 4th pointFe on the 5th pointThe comparison of ftting and monitored values ofturbidity and Fe on the 5th point are shown in Fig.6 and3.4 Analysis of resultsFig.7.The results of PLS model in the water ditributionsystems are shown in Tabs.2 and 3.Tab.2 Absolute value of error4th point5th pointErorTurbidity/ FelTurbidity Fe/NTU(mg°L l)(mg.L')_Maximum0.110.11370.170.1286十Monilored value (Ist group)Minimum00.00730.010.000 3-Fiting value0.0.049 10.035 I013579113151719Tab.3 Absolute relative errorFig.6 Comparison of ftting and monitored values o[th pointturbidity on the 5th pointErrorTurbidity/(mg.L)___ NTU (mg.L')_十- - Moniored value (Ist group)Maximum 23.04%41.23%32.69% 53.07%# - . Fiting vaulur5.31%1.30%0.32%Mean12.21%21.54%11.68%20.62%! 0.15≤15%44.4%33.3%77.8%0.10≤20%66.7%88.9% .55.6%≤30%100%88.9%According to Tabs.2, and 3, considering the influ-91113151719Timeldence of the experimental errors, the results of PLS modelFig.7 Comparison of ftting and monitored values of Fehas basically achieve the request of the precision.on the 5th pointThe comparison of the predicted and monitored val-4 Conclusionsues of turbidity and Fe on the 5th point is shown in Fig.8To deal with the multi-variable problems, such asand Fig.9.the il-ffct of multiplex correlation between variables,PLS regression method is better than many other multipleregression methods.2 0.言0.3才The multi-variable water quality model in this paperexpands the application of water quality model of distri-0.2→- Monitored valur (2nd group)bution systems to predict double targets of water quality.十Predicted valueReferences02-54s中国煤化工[1]; and modeling variationsFig.8 Comparison of predicted and monitored values ofPYHCN M H Giny [1. Amerten wuer .Works Association, 1990, 82: 46-53.一143- -Transactions of Tianjin University Vol.15 No.2 2009[2] Fang Zheng, Ma Lili, Yuan Jianping. Research on mathe-2(in Chinese) ,matical model of water quality change in water tank com- [6] Sontheimer H, Kolle w, Snoeyink V L er al. The sideriteputer simulation [J]. Engineering Journal of Wuhan Uni-model of the formation of corrosion resistant scales [J].versiy, 2005, 38(5): 99-103 (in Chinese).American Water Works Association, 1980, 72(11): 62-63.[3] Zhang Shifang, Zheng Feng. Lowering the turbidity of the [7] Cantor Abigail F, Park Jae K, Vaiyavatjanmi Prasit et al.drinking water to protect the safety of the drinking waterEffect of chlorine on corrosion in drinking water systems[J]. Water Purification Technology, 2005, 24(2): 39-41[J]. American Water Works Association, 2003,.95(5): 112-( in Chinese).[4] Zhang Xiaojian, Niu Zhangbin. Progress in the study of the[8] Bastien Philippe, Vinzi Vincenzo Esposito. PLS general-iron stability problem in drinking water distribution systemised linear regression [J]. Computational Statistics and[J]. China Water & Waste Water, 2006, 22(2); 13-16(inData Analysis, 2005, 48(1): 63-67.Chinese).[9] Tenenhaus Michel, Vinzi Vincenzo Esposito. PLS path[ 5 ] Niu Zhangbin, Zhang Xiaojian, Han Hongda et al. Analy-modeling [J]. Computational Slatistics and Data Analysis,sis on chemical stability of metal ion in water distribution2005, 48(1): 159-205.system [J]. China Water & Waste Water, 2005, 21(5): 18-中国煤化工MYHCNMHG-144-