Leakage has been a major concern for water utilities around the world for a long time. Progressing urbanization and the intensification of climate change are making the loss of water on its way to the customer less and less acceptable. Technologies to identify and localize leakages have also been around for a long time, with methods based on both data analysis and field inspection. As the latter are generally quite labour intensive, the former are preferable as an initial step but unfortunately, despite a plethora of methods published in the academic literature, the holy grail of either leak detection or leak localization has yet to be found.
In an attempt to bring together a number of relevant developments from the scientific literature, KWR created a project called Callisto (Comparison and joint Application of Leak detection and Localization Techniques), in collaboration with Witteveen+Bos Consulting Engineers and water utilities Dunea and Waternet. This project was co-funded by the Topconsortia for Knowledge & Innovation (TKIs) of the Dutch Ministry of Economic Affairs and Climate. The central idea of this project was that if 1) a single method has difficulties detecting all leakages and 2) different methods perform differently under varying conditions, it could be worthwhile to use a number of methods for both detection and localization in parallel, with the working hypothesis that the weaknesses of one method may be covered by another. The input data for this project in is flow and/or pressure data measured in a distribution area or DMA, and a hydraulic model of this area or DMA. What comes out is the timing, duration, magnitude and location of leakages. That’s the theory, at least. A very nice tool was created that implements these ideas and that also includes a data preprocessing module by Witteveen+Bos called Dataprofeet to ensure proper data quality at the start of the analysis.
So how does this work in practice? Well, on synthetic leakages (created in a hydraulic model rather than the real world, resulting in simulated pressure and flow data), it works surprisingly well! Even when you add significant amounts of noise (meant to represent uncertainties in demand patterns, model and measurement errors etc.), both the location and the magnitude of the leakages are recovered quite well, see the tables below.
We also tried to identify leakages in actual measured data from the field. We asked the two participating drinking water utilities to create artificial leaks by opening hydrants. They shared the flow and pressure data that they collected in these DMAs with us to analyze in the Callisto tool. We were able to identify a significant fraction of the leakages (though not all, unfortunately, and in one case, it was also rather the interruptions in an almost contiguous series of flushings that appeared to be detected), but the reconstructed locations were completely WRONG. A closer look at the models and data supplied by the water utilities taught us that the models were not sufficiently representative of the actual situation in the field. This representativeness is a major requirement for successful model based leak-localization. This is, however, a challenge that can be managed.
The most important lessons learnt include the need for a more rigorous validation/calibration of the hydraulic model and also the importance of adequate communication on this issue with the utilities involved. Moreover, the project has demonstrated the validity and the potential of the approach and the tool that was developed. We are aiming for a follow-up project to dot some final i’s, perhaps include further algorithms and some additional clever ideas and to provide successful field demonstrations. And, then, the hunt for leaks will be on!