Benefits of modeling interdependent environmental variables, streamflow and stream temperature, with deep learning

Abstract:

Accurately predicting streamflow and stream temperature in a river network can have major environmental and economic benefits. Deep learning has proven to be proficient at accurately predicting these and other hydrologic variables, however, these variables have typically been modeled independently. This presentation will explore the benefits of modeling streamflow and stream temperature together using a physics-guided deep learning model. Changes to streamflow and stream temperature are both part of the stream energy exchange process. Leveraging this process connection in our deep learning model can provide at least two benefits. First, by training the deep learning model using both temperature and streamflow, the model has two sets of independent observations from which to learn the interdependent behavior of the two variables. This gives the model more data from which to learn because there are places and times where one variable is measured, and the other is not. Second, modeling stream temperature and streamflow together may encourage physical realism of the model. For example, if the streamflow predictions are mostly baseflow, the temperature predictions should be colder and less variable than temperature predictions under other streamflow conditions. This presentation will compare the results of a physics-guided deep learning model trained using varying amounts of stream temperature and streamflow data in the model training data including: 1) only streamflow data, 2) only temperature data, and 3) both streamflow and temperature data together. The benefits of modeling streamflow and temperature together can have implications broadly and would be especially important for our case study area, the Delaware River Basin, where management decisions are made to regulate both streamflow and temperature.