Transferability of hydrological models and ensemble averaging methods between contrasting climatic periods

Abstract

Understanding hydrological model predictive capabilities under contrasting climate conditions enables more robust decision making. Using Differential Split Sample Testing (DSST), we analyze the performance of six hydrological models for 37 Irish catchments under climate conditions unlike those used for model training. Additionally, we consider four ensemble averaging techniques when examining interperiod transferability. DSST is conducted using 2/3 year noncontinuous blocks of (i) the wettest/driest years on record based on precipitation totals and (ii) years with a more/less pronounced seasonal precipitation regime. Model transferability between contrasting regimes was found to vary depending on the testing scenario, catchment, and evaluation criteria considered. As expected, the ensemble average outperformed most individual ensemble members. However, averaging techniques differed considerably in the number of times they surpassed the best individual model member. Bayesian Model Averaging (BMA) and the Granger‐Ramanathan Averaging (GRA) method were found to outperform the simple arithmetic mean (SAM) and Akaike Information Criteria Averaging (AICA). Here GRA performed better than the best individual model in 51%–86% of cases (according to the Nash‐Sutcliffe criterion). When assessing model predictive skill under climate change conditions we recommend (i) setting up DSST to select the best available analogues of expected annual mean and seasonal climate conditions; (ii) applying multiple performance criteria; (iii) testing transferability using a diverse set of catchments; and (iv) using a multimodel ensemble in conjunction with an appropriate averaging technique. Given the computational efficiency and performance of GRA relative to BMA, the former is recommended as the preferred ensemble averaging technique for climate assessment.

Robustness of hydroclimate metrics for climate change impact research

“…They showed that models with similar skill could lead to very different projections, highlighting the need to consider parameter uncertainty when estimating climate change impacts on water resources. Broderick et al (2016) reached similar conclusions based on a multimodel and multicatchment study. They demonstrated the need to test the transferability of parameter sets between contrasting climate conditions and catchment types.…”

Possible Increases in Flood Frequency Due to the Loss of Eastern Hemlock in the Northeastern United States: Observational Insights and Predicted Impacts

“…Matheny, Mirfenderesgi, & Bohrer, 2017;Mursinna et al, 2018;). Mendoza et al (2015) argue that a priori reduction of model complexity can limit the relevance of model predictions, particularly when model-climate transferability must be acknowledged as in Broderick et al (2016). While this is arguably true, we also recognize that increased model complexity involving highly parameterized processes, as are common in numerical representations of ecohydrologic plant dynamics (e.g., Maneta & Silverman, 2013;Tague & Band, 2004), impose greater computational demand, larger model development, calibration, and validation data requirements, and more robust calibration efforts (e.g., Kuppel et al, 2018a).…”

Benchmarking Ensemble Streamflow Prediction skill in the UK

“…It is a daily lumped catchment rainfall-runoff model 25 with a parsimonious structure consisting of four free parameters that require calibration against steamflow observations using daily P and ETp as input. GR4J has been shown to reliably simulate the hydrology of a diverse set of catchments (Perrin et al, 2003) including temporal transition between wet and dry periods (Broderick et al, 2016), and for the generation of ESP forecasts (e.g. Pagano et al, 2010 (Coron et al, 2016(Coron et al, , 2017 with the inbuilt calibration optimisation algorithm based on a steepest descent local search procedure and default parameter ranges.…”