Logo image
Back
A post-processing machine learning framework for bias-correcting National Water Model outputs by accounting for dominant streamflow drivers
Journal article   Open access   Peer reviewed

A post-processing machine learning framework for bias-correcting National Water Model outputs by accounting for dominant streamflow drivers

Savalan Naser Neisary, Ryan C. Johnson, Md Shahabul Alam and Steven J. Burian
Environmental modelling & software : with environment data news, Vol.190, 106459
05/30/2025
DOI: https://doi.org/10.1016/j.envsoft.2025.106459
Web of Science ID: WOS:001475205500001

Metrics

2 Record Views
10 Times Cited - Web of Science

Abstract

While the National Water Model (NWM) provides high-resolution, large-scale streamflow data across the United States, its effectiveness as a key water resources management tool in the drought-prone Western US needs further investigation. Previous studies revealed that the NWM has limitations in controlled basins, impacted by reservoir operations and diversions not explicitly included within the model framework. Responding to the observed reduction in model skill throughout the Western US, we developed a model agnostic post-processing machine learning (PP-ML) framework to account for the impacts of water resources management and regionally dominant hydrological processes on model performance. For our case application of the PP-ML framework, we use daily NWM v2.1 retrospective flow rates as the hydrological model and input upstream reservoir storage, SNOTEL snow water equivalent, and catchment characteristics. Applying the PP-ML framework in the contributing Great Salt Lake watersheds, a key watershed of interest due to its drought-prone nature, we observed a 65%, 335%, and 25% improvement in the median Kling-Gupta Efficiency, Percent Bias, and Root Mean Square Error, respectively, for 30 gauged locations compared to the NWM outputs. Comparing model skills across different flow regimes and station types revealed a substantial (225%) improvement in low-flow estimates at stations with extensive upstream water infrastructure, such as those impacted by reservoir operations, as well as in catchments within negligible water management activities. The research underscores how post-processing hydrological model outputs with ML can account for the effects of water management activities on streamflow estimates, most notably without explicitly incorporating infrastructure rulesets, and demonstrate its capability in bias-correcting streamflow forecasts in response to the regionally dominant streamflow drivers.
pdf
A post-processing machine learning framework for bias-correcting National Water Model outputs by accounting for dominant streamflow drivers2.51 MBDownloadView
Published (Version of record)Article pdfCC BY V4.0,  Open Access
url
A post-processing machine learning framework for bias-correcting National Water Model outputs by accounting for dominant streamflow driversView
Published (Version of record)link to articleCC BY V4.0,  Open

Related links

Details

Logo image