A Deep Learning Autoregressive Forecasting Model for Probabilistic Water Pipe Break Prediction

Accurately predicting the likelihood of water pipe breaks is pivotal for proactive maintenance, cost-effective emergency repairs, and mitigating service disruptions. However, crafting a dependable predictive model for water pipeline breaks is formidable. The challenges stem from the sporadic and infrequent occurrences of breaks, irregular intervals between failures, intricate temporal dependencies among pipes with diverse attributes, and the unbalanced distribution of historical data. Although a considerable number of studies in recent years have developed forecasting models using classic statistical techniques, machine learning solutions, and deep learning methods, state-of-the-art models have yet to achieve the predictive power needed to help utilities transform their practices for risk-based proactive maintenance. This study addresses this need by developing and empirically examining the performance of a novel deep learning-based autoregressive forecasting model for probabilistic water pipe break prediction. Notably, the proposed probabilistic forecasting method integrates a multivariate/multidimensional autoregressive model with a recurrent neural network (RNN) in the form of a long short-term memory (LSTM) model to capture complex and irregular temporal patterns, characterizing dependencies and interrelationships among the time series of pipeline attributes over time, and transform the apprehended patterns to a probabilistic pipe failure prediction through a distribution-based mechanism. The proposed method was implemented to predict the likelihood of water pipe breaks in Calgary, Canada. The model was trained and validated using historical data from 1956 to 2019 and tested for its ability to predict breaks from 2020 to 2023. The results demonstrated that the proposed model exhibits strong predictive performance, achieving an area under the curve (AUC) score exceeding 99.96%. The outcomes of this study will help decision makers plan risk-based maintenance operations that prevent service disruptions and safeguard public health.