Combining a statistical model with machine learning to predict groundwater flooding (or infiltration) into sewer networks

Abstract

Groundwater flooding (or infiltration) in sewer systems leads to significant negative consequences such as discharge of untreated sewage, reduction of system capacity, structural deterioration, and dilution of the wastewater stream delivered to a treatment plant causing malfunction. Cities with aging networks along coastal areas, where aquifers are shallow, are particularly vulnerable. Rehabilitation is necessary to mitigate the negative impact of infiltration but costly. Therefore, a prioritization strategy of intervention is required. This paper presents a decision-support model to identify the probability of infiltration into aging sewer when observations of infiltration and sewer conditions are sparse and time-limited. The model is based on logistic regression, where the variables are: material, soil, water table, and pipe size and shape. As a proof-of-concept, the method was applied to the city of Hoboken, NJ. Machine learning was used to calibrate, validate, and test the model using infiltration measurements, provided by the water authority. Upon calibration, model predictions agree well with the measurements with an accuracy of 82%. Sensitivity analysis of the model was carried out and shows that the most important parameter is the water table of the shallow aquifer. Overall, the proposed approach can be a valuable tool for strategic intervention of sewer repair and flood mitigation in urban areas.