MODELING AND PREDICTING PIPED WATER THEFT USING MACHINE LEARNING APPROACH

Abstract

Water theft is a prevalent problem in most countries across the globe which leads to loss of money, contamination of water, poor water supply, pipe bursts, water leakage and unbalanced flow among other problems. The existing works on this topic basically use hardware to detect water theft. Basing on the sophistication and dynamic nature of the problem, intelligent techniques are needed. Machine Learning and data analysis today form an integrative segment of the latest scientific methodology providing intelligent and automated approaches for predicting phenomenon based on past observations, discovering hidden patterns in data and giving insights about the problem. Machine Learning should however not be used as a black box tool, but as a method whose application should be formulated based on the study problem. Understanding the properties, mechanisms and limitations of the Random Forest algorithm was hence necessary prior to its use. This thesis therefore provides in-depth study of Random Forest and the water theft problem in water distribution pipelines and uses the RF algorithm to model a classifier for piped water theft prediction. Data was collected over a three hour period in 10 seconds intervals using the experimental setup of a hardware framework across a distribution pipeline interconnected with flowrate sensors at branch nodes interfaced with Arduino controller. The state of the network was recorded based on the sms alerts received on the mobile phone through GSM modem. The data was then tabulated, cleaned, explored and visualized to understand its pattern. RF model training was done using 80% of the data and likewise for the other benchmark techniques, that is; SVM, KNN and LR. Testing the models utilized 20% of the data and the four models were evaluated on the basis of accuracy, precision, recall and F-score. RF and KNN models achieved the highest accuracy of 97%. Conclusively, the proposed RF model is more advantageous compared to the other techniques in terms of reliable feature importance estimate and efficiency in test error estimation.