Abstract
An evaluation is presented of the utilization of change-point methods for the detection of anomalies in water consumption time series, and their applicability to water loss detection in water distribution networks. Special attention is given to the relative unconstrained least-squares importance fitting (RuLSIF) change-point detection method [106,173], whose suitability to WDN streaming data is examined by use of a two-month-long hourly water consumption signal. The RuLSIF method successfully detects unusual fluctuations in the water consumption patterns, and classifies them as anomalies. The first water consumption anomaly type examined relates to a discontinuity in the signal (a break in the consumer's water consumption patterns), whereas the second type relates to an unusual increase in the signal (water loss incidents). Even though the proposed analysis does not predict future anomalies, it is suitable for past and near-real-time anomaly detection; an attribute that is sufficient for water loss management as it allows for a timely detection of anomalies is streaming water flow data. Further, the method dynamically assigns anomaly scores to the detected changes in the signal, thus easing water loss detection and appraising the severity of each detected incident.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call