Using Deep Learning to Forecast Human-Modified Streamflow at Ungauged Sites
Abstract
In recent years, deep learning models have had notable success in hydrological forecasting. Despite the prevalence of human-modified flows, most of the research using deep learning to model streamflow has focused on unmodified flows. We will present research on using a Long Short Term Memory (LSTM) model to make predictions at both gauged and ungauged sites with flows that are modified by dams and agriculture withdrawals.
Deep learning models perform best when they are trained on large-sample datasets with both static and dynamic inputs. One of the advantages of this approach is that, because the models learn how static variables affect basin hydrology, they can produce accurate predictions even in ungauged basins where no ground truth streamflow data is available [1]. We present preliminary results indicating that similar success is possible for ungauged human-modified flows. Additionally, we demonstrate that ungauged predictions can be improved by allowing the model to learn from data at a gauged location with similar hydrology and human-modification. These findings are based on a study on locations downstream of agricultural withdrawals and dams in California. To simulate ungauged predictions, we made predictions at these locations using a model that had never seen data at any of the locations. As a second step, we allowed the model to train on gauge data from one of the locations, and make predictions at a different location. A second advantage of deep learning is that it is relatively easy to introduce new inputs to the model. We have seen that the model can improve its understanding of human-modification patterns when provided with inputs, such as land-use, that are related to the modification. These findings are important for watershed management, hydropower generation, and meeting environmental objectives in streams with human-modified flows. [1] Kratzert, Frederik, Daniel Klotz, Mathew Herrnegger, Alden K. Sampson, Sepp Hochreiter, and Grey S. Nearing. "Toward improved predictions in ungauged basins: Exploiting the power of machine learning." Water Resources Research 55, no. 12 (2019): 11344-11354.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2020
- Bibcode:
- 2020AGUFMH076...14H
- Keywords:
-
- 0555 Neural networks;
- fuzzy logic;
- machine learning;
- COMPUTATIONAL GEOPHYSICS;
- 1830 Groundwater/surface water interaction;
- HYDROLOGY;
- 1832 Groundwater transport;
- HYDROLOGY;
- 1849 Numerical approximations and analysis;
- HYDROLOGY