Data from smart water meters used to predict hotspots in distribution network

To ensure fresh and healthy drinking water in a future of climate change and urbanisation, it’s important to know where drinking water ‘hotspots’ occur in the network, and why. This would allow water companies to be more accurate in assessing potential risks and to take more targeted control measures. Brabant Water has installed at a number of its customers a new type of water meter, which permits the monitoring of drinking water temperature. In this TKI Water Technology project, KWR, Brabant Water and Nelen & Schuurmans studied whether one can predict the temperature at the tap using machine learning, combined with data about the weather and the urban environment. They found that it is possible to predict the minimum drinking water temperature of separate households within an accuracy of about 0.5^oC. Among the environmental factors studied, sunlight, outside temperature, sunshine duration and cloudiness are the key determining factors.

New meter type and machine learning

Fresh drinking water is important for the customer’s comfort and health. Research shows that, if no action is taken, it is likely that climate change and increasing urbanisation will in the future lead to more exceedances of drinking water temperature in distribution networks. Brabant Water has installed at a number of its customers a new type of water meter, which measures the minimum temperature of the drinking water on a daily basis. The large-scale installation of such water meters would make it possible to determine the drinking water temperature in the distribution network and the factors leading to exceedances, as well as detect hotspots and their causes. The research used machine-learning techniques to explore the connections between the water temperatures measured in nine households in Veldhoven and data about the weather, the network and the urban environment.

Accurate predictions

The machine-learning approach generated predictions for minimum water temperatures in separate households of an accuracy of about 0.5^oC. Among the environmental factors studied, sunlight, outside temperature, sunshine duration and cloudiness are the key determining factors. To monitor temperature at the tap for a larger distribution area in detail, the sensor network needs to be extended to include a wider set of water meters, over a variable area and for several years. This project provides a good starting point for the research of a number of follow-up questions using machine-learning techniques. What level of accuracy can be achieved in a dynamic map of drinking water temperature? Can predictions be made for locations where no temperature measurement data are available? What sensor density and measurement period would this require? It is expected that up-scaling to bigger data volumes will improve the quality of the machine-learning predictions: is a visualisation platform, such as Lizard, a suitable means of achieving a high-quality information provision?