The detector who cried event

Or the importance of advanced flow tracking.

Water distribution network event detection is a game-changer enabled by computational methods, essential in ensuring fast reaction as well as effective water loss management. Multiple factors are at play in the competent detection of network events.

The key features that enable a software to pick up true events, but stop it from crying wolf too frequently, can be hard to spot at first glance. It all begins by being good at establishing normality, i..e, at knowing what the zone's consumption at the next time step is going to be with a sufficiently narrow dispersion band, as in the figure.

If that band is not narrow enough, it's much harder to tell with an acceptable degree of confidence that this value we are seeing is normal, or not normal – and therefore, potentially indicative of an event. For the event management algorythm, the key decision at that point is to know when not to issue a notification, if it cannot reliably ensure something is escaping normality. Just a hunch is not good enough and it quickly loses credibility among the operational teams.

Detecting events on wide dispersion bands is likely to generate too many false positives. Of course, wide dispersion is often unavoidable on flow predictions, and there is little that can be done – except to avoid generating false notifications that will only drive users' confidence away.

It is for this reason that our researchers and datascientists at Baseform work hard to leverage our Big Data trove (over 40 thousand years of real-time flow monitoring experience and over 1 million events detected) as our predictive engines learn daily how to not cry wolf in vain.

In NRW management, working hard to get predictions right saves time, money and water.