Real-time gradual leakage detection system for water distribution networks based on MIMO-ANN

Leakage detection is a critical issue in water management for water distribution systems (WDSs). With the availability of real-time monitoring data, leakage detection for WDSs based on data-driven methods has received increasing attention in recent years. Current data-driven leakage detection methods are based on a single-step prediction model that only focuses on burst events that are characterized by sudden changes in flow or pressure data in a very short time. However, gradual leakage events that develop from small seeps to noticeable leaks could last for weeks or even months, and these gradual events will cause more water loss and do more harm to the WDS. Furthermore, the gradual leakage events are more challenging to be detected due to its slowly changing pattern. Therefore, this work presents an early warning system for gradual leakage events based on a multistep forecasting strategy. A multi-input multi-output (MIMO) artificial neural network (ANN) is developed to capture the diurnal, weekly and seasonal patterns in the flow monitoring data. The generated forecasting vector is further compared with the observed measurements based on the cosine distance. The residual vector is further analyzed by exponential weighted moving average (EWMA) to smooth the spikes and noises. The final statistics are then used to raise alarms for the monitoring data. The method has been applied to a hypothetical town called L-Town to demonstrate its applicability. The results showed that the proposed method is capable of detecting gradual leakage events with a very small growth rate. In addition, all gradual leakage events are detected with short detection time, high detection accuracy, and low false alarms.