Portfolio and Case Studies

Case Study - Contamination Event Detection in Source Water With Anomaly Detection

USE CASE: Drinking Water

BACKGROUND

As part of a comprehensive Source Water Protection program, a large water utility deployed commercially-available source water monitoring sensors at approximately 80 drinking water treatment plants across their footprint. While the water quality parameters supplied by these probes were useful to Treatment Operators, they failed to reliably assist them in determining the presence of possible contamination. When combined with the additional O&M expense, Operators and Water Quality staff were forced to weigh the costs and benefits of having the sensors deployed in the first place, risking the objectives of the enterprise Source Water Protection program.

OUTCOME: A simplified view of source water quality

As part of a 3-year R&D study involving nine Scientists on two teams, Aaron built a model that classified water quality at the intake as simply "normal" or "abnormal", greatly reducing the guesswork in trying to correlate rises and falls in certain parameters to the presence of a harmful chemical or surrogate.

The model was built using supervised learning: contaminants were passed through a pipe loop system in a controlled environment and the data was recorded. By using ensemble learning to combine multiple models together, Aaron's model was able to detect six out of nine true contamination events, including Benzene, the most frequently occuring contaminant of concern. By leveraging data from as many as 80 different Drinking Water Treatment plants, Aaron was able to train and deploy his model on a variety of source waters from geographically distinct watersheds across the United States.

Results of the study were presented at the AWWA Water Quality and Infrastructure Virtual Summit in November of 2020.