Proactive exfiltration severity management in sewer networks: A hyperparameter optimization for two-tiered machine learning prediction

Abstract

Effective management of aging sewer pipelines requires accurate analysis of sewer pipeline exfiltration. Previous research studies have not paid attention to proposing an index to represent the exfiltration severity. To this end, this study introduces a novel Exfiltration Severity Index (ESI) by considering the frequency and severity of defects captured from Closed-Circuit Television (CCTV) reports. To address the need for a proactive tool that automates the exfiltration estimation and reduces dependence on CCTV reports, we leverage Machine Learning (ML) and Deep Learning (DL) models to predict sewer exfiltration occurrence and severity. In this regard, our proposed methodology comprises a series of steps that start by computing the ESI of pipeline segments considering the frequency, type, and severity of defects. After that, physical, environmental, and climatic factors influencing pipeline exfiltration are gathered and aggregated to build predictive models. We compare the performance of six ML models and two DL models, developed in two tiers to predict exfiltration occurrence and severity, respectively. The hyperparameters of each model are optimized using GridSearchCV to enhance prediction accuracy. Among the eight algorithms, the light gradient-boosting machine performs best, with 71% and 85% accuracy in the first and second tiers, respectively. Furthermore, our study investigates the influence of various factors on pipeline exfiltration and reveals that pipe diameter and population have the most significant impact on exfiltration occurrence and severity. Our method provides a valuable tool for managing sewer pipeline exfiltration and can be utilized to prioritize sewer network maintenance and repair efforts.