Improving Operational Hydrologic Prediction Using Mosaiced Model Formulations with the Next Generation Water Resources Modeling Framework

The fact that there is no single best hydrologic model appropriate for all hydroclimatic regimes presents a challenge for operational water resources forecasting at a continental scale. To tackle this problem and advance the capabilities of the National Water Model (NWM), the NOAA-NWS Office of Water Prediction (OWP) is developing the Next Generation Water Resources Modeling Framework (NextGen). This standards-based, model-agnostic interoperability software facilitates scientific evaluation of different model formulations by executing them in a given location using standardized geospatial and forcing data and a single computational architecture. One advantage of this framework is its ability to link discrete process modules together to mimic existing or create new model formulations with the desired processes represented (e.g., the coupling of a snowmelt module to an infiltration-runoff module in a snowy basin).

This capability to deploy optimized, novel model combinations—tailored to a specific domain, time, and purpose—stands to benefit both the research and operations communities. To ease uptake and adoption, one of our key development tenets has been the use of open-source software and standard libraries. To that end, we employ the Basic Model Interface (BMI), developed by the Community Surface Dynamics Modeling System group at the University of Colorado Boulder, in the NextGen framework and formulations to standardize model execution and coupling. The BMI allows the user to easily exchange one method of modeling a process for another without affecting the other processes or modules, regardless of the programming language, configuration, or paradigm.

We encourage the adaptation of community models into NextGen and also provide a set of formulations implemented for and tested in the NextGen framework. The latter includes a set of conceptual, physical, and machine learning modules to represent various components of the hydrologic cycle. To date, we have configured a total of 12 models and process modules, including those for snow accumulation and melt, soil freeze-thaw dynamics, interception, infiltration-runoff partitioning, soil and groundwater reservoirs, baseflow, and evapotranspiration. To run or test your own model in the NextGen Framework, program it in C, C++, Fortran, or Python and use a compliant BMI implementation.

In this presentation we overview the modules developed for use in NextGen, the processes they represent, and their output. We also provide preliminary evidence supporting our guiding hypothesis that a mosaic of hydrologic models provides improved streamflow predictions relative to a single model deployed over a large domain. To perform an initial evaluation, we first pre-selected an evapotranspiration module using simulated aridity index values at a selection of Catchment Attributes and Meteorology for Large-sample Studies (CAMELS) basins. We then calibrated NextGen-based formulations using the dynamically dimensioned search algorithm in the model-agnostic ngen_cal calibration package. At 64% of these CAMELS basins we found NextGen-based formulations outperformed the NWM version 2.1 benchmark when compared using the Kling Gupta Efficiency (KGE). Importantly, we also discovered that selecting the best formulation in each basin led to a marked increase in overall performance. The heterogeneous blend of formulations had a median KGE greater than the 75th percentile value from the NWM v2.1 benchmark.

NOAA intentionally selected open-source software and workflows to encourage community involvement through collaboration with the university and federal research communities, improving model outcomes and water resources predictions. We especially encourage collaboration through our GitHub repositories and the testing of additional models and modules in NextGen.