Improving U.S. National Water Model Streamflow with Long Short-Term Memory Networks

This is a Preprint and has not been peer reviewed. This is version 5 of this Preprint.


Download Preprint

Supplementary Files

Jonathan Frame , Grey Nearing , Frederik Kratzert, Mashrekur Rahman 


Long short-term memory (LSTM) deep learning networks were used to post-process the U.S. National Water Model (NWM) outputs for improved daily averaged streamflow predictions at 531 basins across the continental United States (CONUS). We compared post-processed streamflow against the NWM and a baseline LSTM without NWM outputs. The LSTM post-processors perform better, on average, than the NWM. Overall median NSE scores are 0.62 for the NWM, 0.74 for the standalone LSTM and 0.73 and 0.75 for the two post-processors. The LSTM with NWM inputs was not significantly better than a standalone LSTM, indicating that the NWM provides only situational benefit for LSTM streamflow prediction. Accuracy of predictions in 2 of the 531 basins was severely reduced by post-processing during tests on ungauged basins, and we found no way to identify ahead of time (without streamflow observations and predictions for comparison) basins where this might occur. The baseline LSTM performs well in ungauged basins. The post-processor improves NWM streamflow predictions in all regions within CONUS. A sensitivity analysis was used to diagnose the land surface component of the NWM as the source of mass bias error and the channel router as a source of simulation timing error. Our assessment indicates that the NWM routing scheme should be considered a priority for NWM improvement.



Earth Sciences, Hydrology, Physical Sciences and Mathematics


Long short-term memory, National Water Model, theory-guided machine learning


Published: 2020-07-02 10:27

Last Updated: 2021-04-05 00:39

Older Versions

Academic Free License (AFL) 3.0

Add a Comment

You must log in to post a comment.


There are no comments or no comments have been made public for this article.