Oyindamola Abatan
Research Mentor: Dani Jones
Mentor Department: Cooperative Institute for Great Lakes Research (CIGLR), SNRE
Author(s): Oyindamola Abatan, Lindsay Fitzpatrick, Dani Jones
Session: Session 1 (9:00 AM – 9:50 AM)
Presentation Type: Poster 2
Abstract
Accurate prediction of the Great Lakes water levels carries both environmental and commercial significance for the 34 million people in the United States and Canada. The United States Army Corps of Engineers currently utilizes physical models to create 6 month forecasts that struggle to predict extreme levels due to the difficulty in measuring components such as evaporation that are used to track net basin supply. Extreme high or low water levels threaten wildlife ecosystems, commercial shipping capabilities, tourism, and flooding in shoreline communities. By utilizing machine learning models to improve accuracy of water level predictions for extremes over 12-month forecasts, we can better prepare various federal, local, and commercial entities to implement proactive strategies. The methodology of this project involves fine-tuning parameters of a custom cost function used within a Gaussian Process Regression model trained using historical water level data, to accurately account for extreme high and low water levels while maintaining accurate readings for average level. Previous experimentation has produced effective models and we seek to conclude with a model that is further optimized for extreme predictions. Successful fine-tuning will not only help mitigate environmental and commercial losses but provide machine learning solutions for other environmental data sets that require extreme data point prediction.
