Automated Decision Support for Model Selection in the Nextgen National Water Model

The modular design of the Next Generation Water Resources Modeling Framework ([Nextgen]; Ogden et al., 2021) provides users a model agnostic platform to mosaic multiple hydrologic models for one modeling task and evaluate their performance using a unifying structure and standard. However, there is no existing methodology for choosing an optimal model (or models) for a given catchment in Nextgen, but may be one of the best investments to improve hydrologic predictions, especially in ungauged basins. Therefore, we developed a model selection system to predict model performance based on catchment attributes. We trained separate random forest regressors [RF] to predict the performance (Normalized Nash-Sutcliffe Efficiency [NNSE]) of three rainfall-runoff models: Long short-term memory [LSTM], Conceptual Functional Equivalent [CFE], and version 2.0 of the National Water Model [NWM]. LSTM and CFE are already implemented in Nextgen. NWM is currently under deployment. The study domain included 495 catchments in the Catchment Attributes and Meteorology for Large-sample Studies [CAMELS] dataset. The RF predicted NNSE reasonably well and correctly identified the model with the highest NNSE in most cases. An ensemble of multiple models with simulated streamflow weighted by the predicted NNSE of each model by RF further improved the model selection results to be similar to the single best model. The results from this study suggest that predicting performance metrics based on catchment attributes is a suitable pathway for model selection, and that the resulting performance metrics can be used to inform a model ensemble. This methodology was designed to be adaptable as new hydrologic models are implemented into Nextgen, as well as unrestricted to the regressor(s) (e.g., RF) and the performance metric(s) (e.g., NNSE or hydrologic signatures) used, therefore providing a flexible and robust workflow for taking full advantage of the modeling capabilities offered by the framework, particularly in ungauged basins.

Ogden, F. L. et al. (2021), The next generation water resources modeling framework: Open source, standards based, community accessible, model interoperability for large scale water prediction, Abstract (H43D-01) presented at 2021 AGU Fall Meeting, 13-17 Dec.