Data-Driven Analytics Tools to Support Prioritized Management of Stormwater Infrastructure

Abstract

Maintenance of aging underground infrastructure compounded with urbanization encroachment and economic pressures have caused a number of challenges for municipal agencies that are tasked with managing and operating these infrastructure systems. However, the budget and time constraints that municipalities often face present challenges for managing culverts. Reliable prediction of infrastructure conditions can help these municipal agencies manage this burden for culverts by providing decision support regarding optimal renewal, replacement, and maintenance, which offers reduced overall costs and improved system reliability. As opposed to traditional physics-based models, data-driven models (e.g., machine learning) offer benefits in processing large datasets where some data is missing or there is “noisy” data. Considering these compounding issues, our study aimed to develop predictive models for culvert condition assessment to better inform repair and replacement schedules. The overall goal is to complement and bolster existing toolsets for prioritization & asset management. The project objectives were to 1) estimate current conditions based on measured data and infrastructure characterization, 2) design models to predict conditions via data analytics, 3) train and validate the models by evaluating each model and verifying results, and 4) identify at-risk pipelines and culverts within the study area. This report describes the data-driven modeling approach via machine learning for predicting culvert conditions to identify at-risk culverts using Charlotte-Mecklenburg Stormwater Services existing data inventory. Several supervised machine learning models were considered and their performance evaluated, respectively. Simulation test results and discussion are included to showcase the key contributions and findings of this work. The results show that one model, the random forest classifier (RFC) gives the prediction performance. The main contributions highlighted in this report include: 1) the performance of four different machine learning models which are compared for the target application to predict culvert conditions, 2) the application of the model for stormwater infrastructure asset management, 3) the viewpoint that the machine learning approach offers economic savings for utilities, 4) the machine learning approach provides potential for improved reliability of stormwater infrastructure.