Abstract:
In 2012, the National Non-Revenue Water assessment revealed that South Africa has 37% of
non-revenue water. With the steadily growing demand for this scarce resource, the detection
of leaks in pipe networks is becoming more important. Currently, in South Africa the primary
method of detecting leaks is to install pressure management systems and monitoring
minimum night time
ows [1].
The pressure-
ow deviation method, can be used to formulate an inverse analysis model
based leak detection problem. This problem can then be solved using Arti cial Neural Networks,
Support Vector Machines and other optimization methods.
With EPANET, di erent networks were tested to compare these methods to nding leaks,
using an inverse analysis formulated problem. Four di erent numerical networks were modeled
and tested, a simple single pipe network, a small agricultural site, a distribution network
proposed and investigated by Poulakis et al. [2] and the simulated model of the experimental
network that was designed and commissioned during the study in our laboratory.
From the numerical investigation, it was found that the optimization methods struggled
to nd solutions for simple networks with in nite number of solutions for the problem. For
more complex numerical networks, it was seen that the Support Vector machine and the
Arti cial Neural Networks trained to the averages of their respective data sets.
Errors to ensure an accurate solution found by these algorithms were calculated as 2:6%
for the numerical experimental network. The experimental network consisted of six possible
leaking pipes, each having a length of 3m and a diameter of 10mm. Three leak cases were
tested with diameters of 3mm and 2mm. Overall, the Support Vector machine could locate
the leaking pipe with the best accuracy, while the minimizing of non-regularized error could
calculate the size and location of the leak the most accurately.
Multiple leak cases were measured with the experimental network. The Support Vector
machine was tested on these measurements, where it was found that two of the three leak
cases could be solved with relative accuracies. Sensor usage optimization was completed on the measurements for the experimental network, where it was found that the leaks could be
classi ed correctly with probabilities higher than 98% if only two sensors were used in the
training of the SVM instead of all twelve.
Overall this method of leak detection shows promise for certain applications in the future.
With practical applications on water distribution, transportation, and agricultural networks.
