Improving Snow to Improve Streamflow: Using Data-derived Snow to Estimate Streamflow in the NextGen National Water Model

Data-driven machine learning (ML) methods such as Long Short-Term Memory Network (LSTM) have shown encouraging results for snow water equivalent (SWE) estimation in mountainous regions. NOAA’s NextGen NWM framework’s modular hydrological modeling structure provides the flexibility to choose from process- and ML-based modeling components. This flexibility makes NextGen NWM suitable for hybrid modeling approaches which can leverage the unique strengths of the data-driven and process-based models, potentially resulting in synergistic advancements in hydrologic prediction.

This study will explore the potential of LSTM-estimated SWE to improve streamflow prediction within NOAA’s NextGen NWM framework. The study first incorporates a LSTM SWE module into the NextGen framework. It then tests the streamflow performance at snow-dominated western US basins using the existing physically motivated NextGen NWM streamflow module. The research aims to answer whether a better SWE estimate can improve streamflow forecasts in NextGen NWM in snow-dominated regions. The LSTM will be trained/validated/tested on a SWE reanalysis dataset, which will be an input to the runoff model. The runoff model will be calibrated to the reanalysis dataset for the streamflow estimates, which will be compared to observed values at the USGS gauges. The accuracy and uncertainty of this hybrid modeling approach will be evaluated.