HAPpy Hour Seminar: Development of Streamflow Predictions through Entity-Aware LSTM Models, Comparison to NOAA’s National Water Model, and Evaluation of Information Content of Several Meteorological Forcing Products for AI-based Hydrologic Applications

Abstract: 

Machine learning approaches have been successfully used to map meteorological input datasets to hydrologic response variables with high accuracy. In this study, a specific type of Recurrent Neural Network named Entity-Aware Long Short-Term Memory (EA-LSTM) is used to develop daily streamflow predictions for more than a thousand river basins across the contiguous United States (CONUS). The predictive skill is then compared to NOAA’s physics-based National Water Model (NWM). Our findings clearly show that EA-LSTM achieves a better Nash-Sutcliffe Efficiency score at over 75% of the study sites compared to NWM. In the second phase of this study, we explore the use of different meteorological forcing products in hydrologic modeling and assess their relative strengths and weaknesses. The gridded meteorological forcing products included are: (1) NOAA's Analysis of Record for Calibration (AORC) dataset, the base forcing dataset used in NWM; (2) Gridded Surface Meteorological (GridMET) dataset; (3) Oak Ridge National Laboratory’s Daymet dataset; (4) NCAR’s CONUS404 dataset, a 40-year reanalysis at 4km resolution developed using the Weather Research and Forecasting (WRF) model; and (5) bias-adjusted CONUS404. These forcing products were separately used to train individual EA-LSTM models. Our findings suggest that EA-LSTMs, when forced with observation-based products (i.e., AORC-, GridMET-, and Daymet-forced EALSTM models), outperform CONUS404-forced models across the majority of basins in the country. However, in the Northwest and Rocky Mountain regions, CONUS404-forced EA-LSTM models consistently show better performance, highlighting the strength of WRF model outputs over regions with low weather station network density. Additionally, we conducted a feature importance analysis to quantify the sensitivity of the EA-LSTM model to different meteorological variables. It is found that precipitation and maximum daily temperature are the primary influential variables to the EA-LSTM model.

VIEW RECORDINGS OF PREVIOUS MEETINGS HERE