Authors: John Kemper – Utah State University
Title: Forecasting water quality in gaged and ungaged watersheds using the National Water Model
Presentation Type: Poster
Abstract: As water quality datasets grow in scale and scope, large scale hydrometeorological forecasts like the National Water Model have become increasingly widely available, rendering the time ripe to use machine learning models to transform forecasts of streamflow into forecasts of concentration and load. Here we undertake this approach in the Lake Champlain basin of the northeastern US, where discrete, low-frequency sampling has been used to monitor water quality in 18 tributary watersheds across 30+ years. Using the LightGBM implementation of gradient-boosted decision trees, we build models of several constituents (total & dissolved phosphorus, total nitrogen, suspended sediment, and chloride) predicated on various dynamic hydrological drivers and static watershed attributes that are then fed discharge forecasts from National Water Model operational output. Model performance is best in low-gradient, highly agricultural catchments and weakest in large basins with steep, forested headwaters. Despite this gradient in performance, model capabilities are relatively robust, especially at short lead times, suggesting that the forecasting framework presented herein represents a promising path forward for water quality prediction. Results from this research may help to improve proactive management of water resources, which has tangible implications for partner organizations such as the Lake Champlain Basin Program.