DETECÇÃO DE FRAUDES EM SISTEMAS DE ABASTECIMENTO DE ÁGUA POR MÁQUINAS DE VETORES DE SUPORTE - SVM

Abstract

There are several factors related to the causes of water losses in Water Supply Systems and, consequently, the reduction of their energy efficiency. In 2014, the second item of expenses with explorations of the Sanitation utilities, was for the costs with electric energy, having totaled in that year the value of R$ 3,471.0 billion, that is 11.2% of the total. The relevance in this process is due to the need to move water against the action of gravity in pressurized pipes using motor pump sets. Therefore, the reduction of these costs is an objective to be pursued, starting from the previous implementation planning and the routine steps of operation and maintenance. An impacting factor on energy costs for sanitation companies are water the losses in a supply system. In public water supply systems, water losses correspond to unaccounted volumes, which can be physical or real, representing the unconsumed portion, as well as non-physical or apparent losses, which correspond to consumed and unaccounted water. Specifically, regarding apparent losses, one of the main causes is due to fraud, or unauthorized consumption of treated water by the population. One way to reduce such losses is to increase the companies' assertiveness in researching these irregular connections, which are currently done practically at random, using slow detection methods. For this purpose, an analysis of temporal consumption data series and correlation with historical data obtained from the Commercial Information System of a Sanitation Company in Brazil is proposed, to develop a forecast model and test its margin of error, based on methods of Data Mining, Outliers and Clustering. The hypothesis developed in this study is that it is possible to trace, through the methodology of artificial intelligence, specifically support vector machines - SVM, through time series, that are, historical data of consumption of the connections of a certain location, a profile that is common to the share of consumers who practice unauthorized consumption. Through the analysis of this historical data, automated mechanisms can be created to identify fraudulent connections, and thus combat this practice, reducing unauthorized consumption and energy costs linked to the loss of treated water.

Keywords: Data Mining, Apparent Losses, Treated Water, Fraud Detection, Artificial Intelligence, Support Vector Machines.