Search

Abstract This study presents a novel application of machine learning to deliver optimised, multi-model combinations (MMCs) of Global Hydrological Model (GHM) simulations. We exemplify the approach using runoff simulations from five GHMs across 40 large global catchments. The benchmarked, median performance gain of the MMC solutions is 45% compared to the best performing GHM and exceeds 100% when compared to the ensemble mean (EM). The performance gain offered by MMC suggests that future multi-model applications consider reporting MMCs, alongside the EM and intermodal range, to provide end-users of GHM ensembles with a better contextualised estimate of runoff. Importantly, the study highlights the difficulty of interpreting complex, non-linear MMC solutions in physical terms. This indicates that a pragmatic approach to future MMC studies based on machine learning methods is required, in which the allowable solution complexity is carefully constrained. Highlights • We present the first use of machine learning-based model combination applied to a global hydrological model ensemble. • The multi-model combination (MMC) performs in most cases better than any individual input model and the ensemble mean. • MMC is not always able to out-perform model combination based on multiple linear regression. • The physical interpretation of the MMC solutions is limited by the complexity of their non-linear weighting schemes. [ABSTRACT FROM AUTHOR]

• Comparably lower performance of GHMs in simulating monthly runoff-deficits. • CHMs and GHMs were similar in estimating SRI. • Ensemble models performed better than the individual models. • Ensemble models simulated drought frequency well. • Differences in ability to simulate the occurrence of observed drought event. Although global- and catchment-scale hydrological models are often shown to accurately simulate long-term runoff time-series, far less is known about their suitability for capturing hydrological extremes, such as droughts. Here we evaluated simulations of hydrological droughts from nine catchment scale hydrological models (CHMs) and eight global scale hydrological models (GHMs) for eight large catchments: Upper Amazon, Lena, Upper Mississippi, Upper Niger, Rhine, Tagus, Upper Yangtze and Upper Yellow. The simulations were conducted within the framework of phase 2a of the Inter-Sectoral Impact Model Intercomparison Project (ISIMIP2a). We evaluated the ability of the CHMs, GHMs and their respective ensemble means (Ens-CHM and Ens-GHM) to simulate observed hydrological droughts of at least one month duration, over 31 years (1971–2001). Hydrological drought events were identified from runoff-deficits and the Standardised Runoff Index (SRI). In all catchments, the CHMs performed relatively better than the GHMs, for simulating monthly runoff-deficits. The number of drought events identified under different drought categories (i.e. SRI values of -1 to -1.49, -1.5 to -1.99, and ≤-2) varied significantly between models. All the models, as well as the two ensemble means, have limited abilities to accurately simulate drought events in all eight catchments, in terms of their occurrence and magnitude. Overall, there are opportunities to improve both CHMs and GHMs for better characterisation of hydrological droughts. [ABSTRACT FROM AUTHOR]