Performance of Multi-Model Combinations in Reproducing Hydrological Signatures Relevant for Climate Change Impact Studies in High Latitudes

Abstract

Assessing the hydrological impacts of climate change is rather challenging, particularly due to the discrepancies between climate and hydrological models used for such assessments. The reliability and robustness of climate models, downscaling techniques, and bias-adjustment procedures are typically judged based on their ability to reproduce distributions of climate variables. On the other hand, hydrological models are developed to reproduce complete series of hydrological variables, primarily flows. Although climate change impact studies focus exclusively on pertinent hydrological signatures, such as mean flows or annual maxima/minima of varying durations, these aspects of hydrological models’ performance are generally limited to investigative research studies. Since hydrological models are neither developed nor evaluated according to their performance in reproducing hydrological signatures and their distributions, they often yield poor performance in this regard, especially when it comes to signatures related to extreme flows. One approach to improving model performance is the application of multi-model combination methods (MMCMs). This study is aimed at evaluating the effects of the application of MMCMs in improving performance in reproducing numerous signatures relevant for climate change impact assessments in high-latitude catchments. To this end, ten MMCMs are applied with an ensemble of 29 spatially-lumped, bucket-style models in 50 Swedish catchments that span a wide range of hydroclimatic regimes. All selected MMCMs are point-estimate methods (i.e., they result in a single flow estimate, referred to as a model combination), and they are mainly based on the information criteria. The selected methods also include the equal weights method, Bates-Granger- and Granger-Ramanathan methods, the Mallows method, and its simplex version. The MMCMs outputs are used to compute numerous commonly used performance indicators and distributions of the selected signatures (following Todorović et al. (2022)), which are compared to the results of a reference model. The reference model is selected as the on-average best-performing individual model across the 50 selected catchments. Additional computations are performed to infer whether (1) the selection of the candidate models, or (2) targeting specific signatures, such as annual maxima or minima, can improve the performance of the model combinations. The results suggest that the application of MMCMs can improve efficiency in terms of traditionally used performance indicators; however, no improvement is obtained when it comes to the distributions of the signatures. Neither omitting the poor-performing candidate models from the ensemble, nor applying the MMCMs with the series of targeted signatures, can improve this aspect of performance. These results clearly reveal the need for further improvement of hydrological models so that they can properly reproduce distributions of hydrologic signatures, which is crucial for climate change impact studies.