CCWI2017: F20 'Online Burst Detection in Water Networks With an Ensemble of Flow Prediction Models'

Pipe bursts in water distribution networks have to be dealt with quickly so as to prevent secondary accidents and expansion of supply suspension. Online burst detection methods have been actively studied from within the framework that an anomaly is detected by assessing the deviation of actual sensor readings from the corresponding predictions. However, these methods might fail or take too long to detect a gradually developing burst because their predictions are based on sensor readings from recent periods that would be immediately affected by the burst. In this study, we adopt an ensemble of flow prediction models to detect both step-shaped and gradually developing bursts earlier. One prediction model using recent flow readings has a smaller prediction range, which enables earlier detection of step-shaped bursts, while another model using only the older flow readings can robustly detect gradually developing bursts. The proposed method forms an ensemble of the models with different usages of flow readings and takes advantage of the strengths of each so that it can detect bursts of various development rates earlier. As a case study, the proposed method was applied to a set of synthetic data of a DMA with 20m 3 /h inflow on average. It was found that half of the randomly generated gradually developing bursts were detected around when the development of the burst flow slowed down. It was also demonstrated that adopting the ensemble models actually reduced the amount of time it took to detect gradually developing bursts. A future task is the evaluation of the proposed method utilized for a real burst event.<br>