USE OF MACHINE LEARNING IN THE AQUATIC SCIENCES & A DATA-DRIVEN FORECAST SYSTEM AS AN OPERATIONAL DECISION TOOL IN A MANAGED RESERVOIR

Abstract

The use of machine learning methods in aquatic sciences has increased greatly over the past few years, at least in part due to the rise of open-source packages that make it easy to apply these complex models in the complex environment we work in. In the current time of generative artificial intelligence and their integrations into coding interfaces, the process of creating machine learning models has become as easy as typing a request in plain language, making a coffee, and reviewing the results. This low barrier to entry has provided a great resource for researchers to better understand their systems and create global models with large datasets or niche models with limited data. This boom, however, has not yet been met with adequate resources for how to rigorously assess machine learning use in aquatic science, especially at the manuscript/journal interface. The first chapter of this thesis presents an essay that draws attention to the democratization of coding tools and how it may impact aquatic sciences and the follow-on benchmarks created with these machine learning techniques. The call to action is that better infrastructure must be provided to the author, reviewer, and editors to be sure that the methods and results described in a manuscript match the code used to build and assess the models. This essay is aimed toward the system we work in as aquatic scientists: it is not aimed at solely authors (who are experts in aquatic science) and reviewers (who volunteer their time to referee manuscripts) that need a framework, but also our journal infrastructure needs an update. The second chapter presents a decision tool built on machine learning best practices as a case study: from conceptual model to stakeholder engagement and trust building. The tool was designed for the Three Lakes System (Grand Lake, Shadow Mountain Reservoir, Lake Granby) managed by Northern Water, who supplies drinking water to about one million people in northern Colorado via municipal utilities and irrigation for ~600,000 acres of agricultural land through the Colorado-Big Thompson Project. In this modeling effort, Northern Water and partner agencies work together to attain water quality and clarity goals in Grand Lake and Shadow Mountain Reservoir. Because the two waterbodies are hydrologically connected, water quality issues in Shadow Mountain Reservoir can directly impact water quality in Grand Lake. This application uses a machine learning tool to forecast water temperature, a primary biological control, in Shadow Mountain Reservoir at a 7-day time horizon that is responsive to operational changes in water delivery that consistently outperforms a persistence model, with a 7-day continuous rank probability skill score of 0.62 at the near-surface depth and 0.78 at the integrated depth relative to a persistence model, where a value of 1 is a perfect forecast at this horizon. Our end users can use this tool to test operational scenarios for real-time forecasting and it can be used to estimate how operations have impacted water temperature within the system. In a hindcast scenario the tool estimates that water temperature was reduced by an average of nearly 3 degrees at the near-surface and 4 degrees at an integrated depth over the course of the adaptive management season July 1 – September 11, 2025.