A hybrid machine learning-based model for predicting failure of water mains under climatic variations: A Hong Kong case study

Abstract

Effective functioning of water systems is critical to ensure the quality of human life. Therefore, failure prediction of water mains under climatic variations is necessary to avoid socio-economic and environmental losses. This paper aims to propose a hybrid model named STL-GC-LSTM for an accurate failure prediction of water mains under the impact of climatic variations. Firstly, the seasonal-trend decomposition based on Loess (STL) method is employed to decompose the failure time series. Next, significant climate variables are selected from the Granger causality (GC) test. Lastly, the final predicted failure of water mains is acquired by adding up the predictive results of the three components which are learned by Long Short-Term Memory (LSTM) models. Several evaluation metrics are used to assess the prediction performance. The results from a case study in Hong Kong imply that STL decomposition is promising for fully mining intrinsic properties of failure series. The developed hybrid models are effective in specifically identifying which component climatic variations exert influence on, and the final failure predictions show satisfactory agreement compared with peer models. This paper could provide an accurate estimation for failures of water mains ahead of time and be used as an essential complement to other numerical prediction models.