Leveraging LSTM for Seasonal Streamflow Prediction Using Post-Processed NWM Retrospective Data

Seasonal streamflow prediction is crucial for water management in the drought-prone western US, influencing hydropower, water allocation, and agriculture planning and management. We analyzed the use of National Water Model version 3.0 (NWM) daily retrospective data as a possible source for water supply forecasts to bypass common long-term water supply forecasting limitations, such as hydrological modeling limitations and the impacts of water infrastructure, and explore the power of machine learning to address identified limitations. Complementing the streamflow forecasting analysis, we evaluated several hydrological model inputs to assess the impact on model prediction skill, including the Analysis of Record for Calibration (AORC), North American Land Data Assimilation System (NLDAS), and Airborne Snow Observatory (ASO) datasets. Using the Great Salt Lake watershed as a case application, our analysis indicates that the NWM has low accuracy, particularly concerning drought and low-flow prediction in watersheds with high human activity. To address the limitations in model skill, we applied the Long-Short Term Memory (LSTM) as a post-processing mechanism to the NWM and Ensemble Streamflow Prediction (ESP) diagrams to enhance seasonal prediction accuracy. The results demonstrate that LSTM effectively accounts for the impacts of human water management activities on streamflow estimation, assisting water managers with seasonal planning and mitigating the consequences of hydrological drought.