Gradual Leak Detection in Water Distribution Networks Based on Multistep Forecasting Strategy

With the availability of real-time monitoring data, leakage detection for water distribution networks (WDNs) based on data-driven methods has received increasing attention in recent years. Accurate forecasts based on historical data could provide valuable information about the condition of the WDN, and abnormal events could be detected if the observed behavior is substantially different from the typical behavior. Therefore, an accurate forecast model is essential for prediction-based leakage detection methods. While most data-driven methods focus on burst detection, it is also important to develop an early warning system for gradual leakage events because they will cause more water loss due to a longer time to awareness. Therefore, a real-time early leakage detection technique based on a multistep forecasting strategy is proposed in this study. A multistep flow forecasting model is introduced to capture the diurnal, weekly, and seasonal patterns in the historical data. The generated multistep forecasting is further compared with the observed measurements, and residuals are calculated based on cosine distance. Based on the analysis of the residual vector, the gradual leakage event could be detected in a timely manner. The proposed method is applied to the L-town datasets containing one year of real-life flow monitoring data. The results prove the superiority of the proposed multistep prediction model-based method over the traditional one-step prediction model for gradual leakage detection. In addition, the results show that the proposed methodology can detect small gradual leakage events within just a few days while generating no false alarms. The method was further applied to a real-life network and showed consistent results.