Simplifying a hydrological ensemble prediction system with a backward greedy selection of members – Part 1: Optimization criteria

Abstract

Hydrological Ensemble Prediction System (HEPS), obtained by forcing rainfall-runoff models with Meteorological Ensemble Prediction Systems (MEPS), have been recognized as useful approaches to quantify uncertainties of hydrological forecasting systems. This task is complex both in terms of the coupling of information and computational time, which may create an operational barrier. The main objective of the current work is to assess the degree of simplification (reduction of members) of a HEPS configured with 16 lumped hydrological models driven by the 50 weather ensemble forecasts from the European Center for Medium-range Weather Forecasts (ECMWF). Here, the selection of the most relevant members is proposed using a Backward greedy technique with <i>k</i>-fold cross-validation, allowing an optimal use of the information. The methodology draws from a multi-criterion score that represents the combination of resolution, reliability, consistency, and diversity. Results show that the degree of reduction of members can be established in terms of maximum number of members required (complexity of the HEPS) or the maximization of the relationship between the different scores (performance)

A semi‐empirical wind set‐up forecasting model for Lake Champlain

“…Probabilistic forecasting can involve ensemble forecasts from a single NWP system, but it can also make use of forecasts generated by multiple NWPs to cover the uncertainty of the NWP selection in the process, known as a multi‐model forecast. The idea of multi‐model forecasts has been used in hydrological modelling to create a GESP (Grand Ensemble Streamflow prediction) that allows a more complete representation of uncertainty (Brochero et al, 2011; Demirel et al, 2013; Velázquez et al, 2011). This principle could also be used for wind set‐up forecasting to create a GEF (Grand Ensemble Forecast).…”

Generating Ensemble Streamflow Forecasts: A Review of Methods and Approaches Over the Past 40 Years

“…Fusing is typically achieved by pooling the predictions from the multiple models (weather and/or hydrological models) and extracting a single prediction value corresponding to the “best” forecast from the multi‐model ensemble. The multi‐model combination can be performed by fusing individual streamflow forecasts from multiple hydrological models (e.g., Brochero et al., 2011; Demirel et al., 2015; Mahanama et al., 2012), combining multiple NWP model ensembles of weather forecasts (e.g., Bao et al., 2011; Bogner et al., 2011; Kim et al., 2017), and ultimately, merging the different approaches (e.g., Bourdin & Stull, 2013; Dutta et al., 2012; Slater et al., 2017). However, even though the combination methods can improve some of the statistical properties of the forecasts, they are little used in operational EFSs, where they are put into practice in a decision‐making context.…”