Prediction of odor complaints at a large composite reservoir in a highly urbanized area: A machine learning approach

Odorous compound emissions and odor complaints from the public are rising concerns for agricultural, industrial, and water resource recovery facilities (WRRFs) near urban areas. Many facilities are deploying sensors that measure malodorous compounds and other factors related to odor creation and dispersion. Focusing on the Metropolitan Water Reclamation District of Greater Chicago's (MWRDGCs) Thornton Composite Reservoir (7.9 billion gallon capacity), we used meteorological, operational, and H2S sensor data to train a 3‐day advance‐warning predictor of local odor complaints, so as to implement targeted odor prevention measures. Using a machine learning approach, we bypassed difficulties in modeling both physical dispersion and human perception of odors. Utilizing random forest algorithms with varied settings and input attributes, we find that a small network of H2S sensors, meteorological data, and operational data are able to predict odor complaints three days in advance with greater than 60% accuracy and less than 25% false‐positive rates, exceeding MWRDGC’s standards required for full‐scale deployment. Practitioner points A random forest algorithm trained on H2S, weather, and operations data successfully predicted odor complaints surrounding a large composite reservoir. Thirty‐two data attribute combinations were tested. It was found that H2S sensor data alone are insufficient for predicting odor complaints. The best predictor was a Random Forest Classifier trained on weather, operational, and H2S readings from the reservoir corner locations. This study demonstrates odor complaint prediction capability utilizing a limited set of data sources and open‐source machine learning techniques. Given a small network of H2S sensors and organized data management, WRRFs and similar facilities can conduct advance‐warning odor complaint prediction. Pictured: The Thornton Composite Reservoir (TCR) and various data collection points. This attribute data was used to train a random forest algorithm for odor complaint prediction.