Tracking Wet Weather Pollution Control Performance Using Logistic Regression: A Comparison of Statistical and Machine Learning Approaches
TL;DR
It is shown how annual calculation of rainfall thresholds could have identified a malfunctioning CSO tank that caused a 4-year spill and nearly $3 million in fines, and statistical logistic regression is preferable to the more commonly used machine learning approach in this application.
Abstract
Wet weather pollution includes combined sewer overflows (CSOs) and wastewater treatment plant (WWTP) bypasses. Logistic regression can estimate the daily rainfall threshold required to trigger a pollution event, providing a useful metric. Calculating this rainfall threshold only requires the event dates and rainfall. The rainfall threshold can be used to evaluate the effectiveness of upgrades. For example, the rainfall required to trigger a WWTP bypass increased significantly after upgrades to the headworks. Furthermore, we show how annual calculation of rainfall thresholds could have identified a malfunctioning CSO tank that caused a 4-year spill and nearly $3 million in fines. Annual regression can identify changes in wet weather infrastructure. If the rainfall threshold increases, this suggests improved ability to handle excess water. This approach can be used to evaluate the effectiveness of low-impact development or other upgrades. Finally, statistical logistic regression is preferable to the more commonly used machine learning approach in this application because it clearly quantifies the threshold uncertainty required for hypothesis testing.