Page 1Page 2Page 3Page 4Page 5Page 6Page 7Page 8Page 9Page 10Page 11Page 12Page 13Page 14Page 15Page 16Page 17Page 18Page 19Page 20Page 21Page 22Page 23Page 241989-01 Spring 1989 A statistical Analysis of Water SOUTHERN METHODIST UNP Main Breaks for the Dallas Water Utilities Jeff Muscarella, Jorge Cueto andMatthew McGrath ;tical Analysis of Water Main Bre for the Dallas Water Utilities OREM 4390 May 13, 1989 DEPARTMENT OF OPERATIONS RESEARCH AND ENGINEERING MANAGEMENTSCHOOL OF ENGINEERING AND APPLIED SCIENCE DALLAS, TEXAS 75275I 1• I I IA Statistical Analysis of Water Main Breaksfor the Dallas Water Utilities I.OREM 4390 May 13, 1989 Ip I I U IJeff Muscarella Jorge CuetoMatthew McGrath I I I. In IExecutive Summary Public works such as the Dallas Water Utilities need to stretch Itheir capital resource budgets as far as possible in order to give the best service to the community. In order to keep their costs down, Ithe Dallas Water Utilities has a need to know when worn out pipes need replacing. The purpose of our project was to find a scientific and reliable method of predicting areas of the city that will need water main Isystem replacement in the future. By analyzing data provided on past water main breaks, we attempted to identify any trends in previous pipe breaks which could tell us about where pipes will break in the coming years. Although our analysis clearly defined certain trends in the past water main breaks, our study found that in order to obtain results Ibeneficial to the DWU (Dallas Water Utilities) it would be necessary to conduct a more in-depth analysis. This second analysis involves Idividing the city into sectors and gathering information on all of the pipes in that sector, both pipes with breaks and those without breaks. The experiment then analyzes the new data and determines the pipes in certain areas most likely to break which is the desired Iresult. I I.ki ITable of Contents Executive Summary.3 Project Description .......................................5 Experimental Design ....................................6 Analysis of Results .......................................8 Conclusions and Recommendations ......11 Technical Appendix .....................................12 4I•Project Description The purpose of our experiment was to study the effects of Ivarious factors on water main breakage in the area served by the Dallas Water Utilities.Different factors such as pipe type, depth Ibelowground, temperature range, cause of break, etc. were things Ithatgave us information into the conditions when the pipe broke. Each time that a main breaks and is repaired in the DWU 1system, the repair crew fills out a Break Card.These break cards contain the information on the pipes that needed repair.Given on these Break Cards are the factors discussed above as well as location, Isize of pipe, soil type, and type of break.The Break Cards were then transposed by the use of a scanner to a daily ledger and stored on file by the DWU. To begin the experiment, four years of water main breaks Iwerecompiled into a data base.The DWU then obtained the dailytemperatures and the water pumpage in millions of gallons for each Iday of the four years. Although other factors are important in the Iwater main breakage such as water temperature, soil resistivity, and most importantly pipe age, we did not have any access to this data Iwhich left us looking for trends without these factors. Now that we had a data base of information with which to Iwork, a statistical model was needed that could approximate the real Iworld system. This model needed the ability to accept large amounts of data from thousands of main breaks. I I.5 II•Experimental Design IIn the design of this experiment, certain variables were Iobviouslyjointly affected, such as the type of pipe and the cause of Ibreak.Whenever a pipe breaks, the factors have a certain interaction with each break.Therefore, we were looking for Icombinationsof factors which were present most often which would give us information on other places to look in the system having 1these same factor interactions.Because of this, we needed to design the experiment and choose a model in a way to allow for analysis of Ithesejoint effects. IInorder to model the experiment correctly, a Log Linear Model was chosen.A Log Linear Model uses nonlinear data and tries to find correlations between the variables which determine a qualitative response variable.Since we were looking at an analysis which looks Iatwhether the pipe was broken (meaning that the result is not a Iquantitativeor numeric answer), the Log Linear Model was our best choice.Also, Log Linear models are used when the data merely Icontainfrequency counts of certain variables at different levels.For instance, the depth variable has different levels (2", 4", 6", etc) which Iwecan count in interaction with other variables and levels of variables. IWhen analyzing large amounts of data, it is common to use a statistical computer package such as SAS.SAS allows one to input data in one of a variety of forms and then operate on it as needed to Igetmany different forms of output.We used SAS because it allowed us to use the Log Linear model on the large scale of the mainframe 6 IIIBM on the SMU campus. In the first step of the experiment, we entered the two data sets into a SAS data set. This was accomplished by using two different SAS program files. Once the two sets were Iordered by date and merged into one data file, we were able to use Ithe SAS Catmod procedure which employs the Log Linear model. The major drawback to our analysis using SAS was that the Imemory available to us at the Bradfield Computer Center was much too small. Our full analysis needed more than 15 megs of computer Ispace to run the Catmod Procedure using all seven variables. We Itherefore needed to reduce the number of levels in each variable which was accomplished by taking frequency counts on the original Idata and finding the levels of each variable which occured most often. We reduced the levels of each variable to no more than six per variable. Even after reducing the levels though, the procedure required more data space than permissible. Eventually, we needed Ito reduce the number of variables that we could interact in each run Ifrom seven to three. In the Catmod procedure, the log linear model was specified by Ithree