Authors: Roja Najafi, Daniel P. Ames, Iman Maghami – Department of Civil and Construction Engineering, Brigham Young University, Provo, Utah, USA
Title: Improving the NWM Streamflow Predictions Using a Hybrid LSTM Model Across Climatic Regions
Abstract: In this study, we evaluated the performance of a hybrid model integrating the National Water Model (NWM) and Long Short-Term Memory (LSTM) for streamflow predictions across diverse meteorological conditions. We used NWM short range streamflow forecasts to train an LSTM model using a two-year dataset spanning from April 2022 to April 2023 with USGS streamflow forecasts as target values. The idea is to build a hybrid model where the NWM provides the initial physics-based forecast, and this forecast is then improved by the LSTM model to more accurate represent observed USGS gauge station values. We focused on three distinct climatic regions, Arizona, California, and Michigan, we assessed the hybrid LSTM’s predictive accuracy and reliability through a detailed analysis of key metrics such as Root Mean Square Error (RMSE), Correlation Coefficient (CC), Nash-Sutcliffe Efficiency (NSE), and Percent Bias (PBIAS). The hybrid LSTM model consistently demonstrated a high level of accuracy in predicting streamflow up to 18 hours ahead. Although the NWM initially produced more accurate predictions in the early hours, the hybrid LSTM model outperformed it for the entire forecasting period. Our findings suggest that the hybrid LSTM model is a reliable tool for streamflow prediction across a range of conditions. Furthermore, the results of this research highlight the potential of the hybrid model approach to leverage the strengths of both the NWM and LSTM model, enhancing the accuracy of streamflow predictions in regions characterized by diverse precipitation patterns.