Post-Processing the U.S. National Water Model with a Long Short-Term Memory Network
Abstract
Hydrological science is in the midst of an exciting revolution with rapid improvement coming in two ways: 1) The domain of hydrology models is expanding to a continental scale and beyond ( e.g. , Salas et al. 2018), and 2) the predictive accuracy of hydrology models is improving at an accelerating rate with the integration of machine learning (ML) and, more specifically, deep learning (DL) ( e.g. , Kratzert et al. 2018). We present results post-processing daily streamflow predictions in the U.S. National Water Model (NWM), a large domain hydrologic forecasting system, with long short-term memory (LSTM) networks, a DL method that is particularly well suited to model hydrologic processes.
We trained an LSTM to ingest dynamic NWM output (states and fluxes) to improve streamflow simulations, and tested performance at 531 basins across the continental United States. Relative to a stand-alone NWM benchmark, the LSTM post-processor provided a significant benefit to nearly all aspects of NWM streamflow predictions. The LSTM post-processor improved the NSE score of the NWM mean daily streamflow at a total of 488 basins (92%), improved the total bias at a total of 331 basins (62%) and improved the peak timing error at a total of 494 basins (93%). Relative to a stand-alone LSTM benchmark, adding NWM states and fluxes as dynamic inputs to the LSTM improved the representation of physically-motivated hydrologic signatures, but not the overall performance metrics. The hydrologic signature representations were analyzed by the correlation coefficient (r 2 ) of the models compared to observations. The signatures best represented by the NWM are similarly the best represented by the LSTM post-processor, and the same is true for the poorly represented hydrologic signatures. The most significant improvement is the representation of the baseflow index. Results were further explored by region, in calibrated vs. uncalibrated basins, and for different dynamic training input sets for the LSTM network. Salas, F. R., et al. , 2018. Towards Real-Time Continental Scale Streamflow Simulation in Continuous and Discrete Space. Journal of the American Water Resources Association 54, no. 1 Kratzert, F., et al. , 2018. Rainfall-Runoff Modelling Using Long Short-Term Memory (LSTM) Networks. Hydrology and Earth System Sciences 22, no. 11- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2020
- Bibcode:
- 2020AGUFMH121...06F
- Keywords:
-
- 1805 Computational hydrology;
- HYDROLOGY;
- 1816 Estimation and forecasting;
- HYDROLOGY;
- 1839 Hydrologic scaling;
- HYDROLOGY;
- 1847 Modeling;
- HYDROLOGY