A MONITORING METHOD COMBINING MULTIPLE FLOW PREDICTION MODELS TO DETECT A BURST IN A WATER DISTRIBUTION NETWORK

Abstract

 As the ageing of pipe networks progresses, a quick response to a pipe burst is gaining importance. Online monitoring methods have been proposed to detect a burst, but these methods tend to fail or take too long to detect a gradually developing burst. This study proposes a monitoring method to detect both step-shaped and gradually developing bursts earlier by combining multiple flow prediction models with various prediction horizons. A prediction model with a shorter horizon contributes to detect a step-shaped burst earlier based on a smaller prediction error. On the other hand, another model with a longer horizon is expected to detect a gradually developing bursts earlier since its prediction is more robust compared with the former. The proposed method is composed to take the advantage of each model and complement each other. In a case study, the proposed method was evaluated by applying to a set of synthetic data based on an open dataset of actual flow readings. The set of synthetic data was generated by adding actual flow readings and a sigmoid shaped burst flow with various combinations of size, rate of development, and datetime of a simulated burst. It was found that the proposed method shortened the time by 1‒2 hours to detect gradually developing bursts compared with a method with a single flow prediction model in the case study area of this study. The proposed method is expected to contribute to prevent secondary accidents and expansion of supply suspension by the saving the time to detect.