Burst detection in water distribution systems with LSTM

This study presents a data-driven method for detecting pipe bursts in water distribution systems using Long Short-Term Memory (LSTM) neural networks. These types of neural networks are able to process sequential data more effectively than traditional neural networks because they have feedback connections between neurons. The proposed method involves performing one-step ahead predictions about the flow and pressure at different sensor locations in the system, using past time series data along with additional time-related features as inputs. The difference between predictions and actual observations is used to classify bursts and trigger alarms by comparing the errors against a time-varied error threshold. The model is trained using data from burst-free periods in the system. The method was tested using simulated fire hydrant bursts as well as real-world bursts in 8 district metered areas (DMAs) located in the United Kingdom. By harnessing transfer learning, the model can incorporate additional data streams from new sensors, performing well even in data frugal conditions, achieving precision scores of up to 98.1% for the analyzed case studies