Finding the Optimal Multimodel Averaging Method for Global Hydrological Simulations

Abstract

Global gridded precipitations have been extensively considered as the input of hydrological models for runoff simulations around the world. However, the limitations of hydrologic models and the inaccuracies of the precipitation datasets could result in large uncertainty in hydrological forecasts and water resource estimations. Therefore, it is of great importance to investigate the hydrological value of a weighted combination of hydrological models driven by different precipitation datasets. In addition, due to the diversities of combination members and climate conditions, hydrological simulation for watersheds under different climate conditions may show various sensitivities to the weighted combinations. This study undertakes a comprehensive analysis of various multimodel averaging methods and schemes (i.e., the combination of the members in averaging) to identify the most skillful and reliable multimodel averaging application. To achieve this, four hydrological models driven by six precipitation datasets were used as averaging members. The behaviors of 9 averaging methods and 11 averaging schemes in hydrological simulations were tested over 2277 watersheds distributed in different climate regions in the world. The results show the following: (1) The multi-input averaging schemes (i.e., members consist of one model driven by multiple precipitation datasets) generally perform better than the multimodel averaging schemes (i.e., members consist of multiple models driven by the same precipitation dataset) for each averaging method; (2) The use of multiple members can improve the averaging performances. Six averaging members are found to be necessary and advisable, since using more than six members only imrpoves the estimation results slightly, as compared with using all 24 members; (3) The advantage of using averaging methods for hydrological modeling is region dependent. The averaging methods, in general, produced the best results in the warm temperate region, followed by the snow and equatorial regions, while a large difference among various averaging methods is found in arid and arctic regions. This is mainly due to the different averaging methods being affected to a different extent by the poorly performed members in the arid and arctic regions; (4) the multimodel superensemble method (MMSE) is recommended for its robust and outstanding performance among various climatic regions.