Leveraging Long Short-Term Memory for Individual Water Main Failure Prediction Using Pipe Intrinsic Variables and Climate Data

The optimal annual replacement of deteriorating water mains is a critical task for water utilities. Pipe failure prediction models serve as essential tools in the strategic planning of rehabilitation efforts for urban water distribution infrastructure. Numerous studies highlight that climatic factors have a significant impact on the integrity of water mains. However, existing failure prediction methods often struggle to effectively account for the complex interactions between static variables, such as pipe material and diameter, and dynamic factors, including historical failure trends and climate-related variables, thereby limiting their accuracy and reliability. This study aims to develop a novel deep learning approach for predicting the probability of failure for individual pipes within a water distribution system. The research utilizes three key datasets from Saskatoon, Canada: pipe inventory data, historical break records, and weather information. A long short-term memory (LSTM) neural network model is developed to capture the temporal dependencies and non-linear interactions in the data. The results demonstrated that incorporating invariant variables including pipe characteristics, improved the model’s predictive performance, while the inclusion of climate data further enhanced the model’s performance, highlighting the importance of combining static and dynamic factors in failure prediction.