Ensemble projection and uncertainty attribution of potential evapotranspiration in northwest China in the future

Abstract: Potential Evapotranspiration (PET) has been one of the most important factors on terrestrial water availability. It is a high demand to predict the PET changes for the robust responsive strategies to the future climate scenarios in northwest China. However, the projected outcomes cannot fully match the decision-making on the planning, due to the uncertainties originating from different sources. Therefore, it is crucial to develop a large-scale ensemble of PET projections, considering the multiple sources of uncertainty to quantify the contribution of each uncertainty source. In this research, a monthly regression correction was carried out on the simplified PET models (temperature- and machine learning-based PET models) using historical meteorological data. A three-dimensional evaluation framework was then constructed to predict the PET changes using the meteorological data of 6 Global Climate Models (GCMs) in the sixth Phase of the Coupled Model Intercomparison Project (CMIP6), 3 future scenarios (Shared Socioeconomic Pathways, SSPs), and 6 PET models (2 combined and 4 simplified PET models). A large-scale ensemble projection was obtained, including 108 projections of annual and seasonal PET changes (△PET). The △PETs in the mid-term future and long-term future were obtained to subtract the average PET in the historical period from 2041 to 2070, and from 2071 to 2100, respectively. The different forecasts of climate scenarios (SSPs), PET models (PETMs), and GCMs were quantified to determine the multiple sources of uncertainty on the projection. Three-factor Analysis of Variance (ANOVA) was selected to quantify the contribution rate of each uncertainty source and the interactions to the total uncertainty. The result showed that: 1) The correction coefficients in the PET models after the monthly regression were represented the influence of meteorological elements (except for the temperature), indicating an improved simulation accuracy of the simplified PET models with the reduced systematic error. It infers that the regional modified and combined PET models were more reasonable than ever to evaluate the PET changes. 2) In the mid-term future (2041-2070) and the long-term future (2071-2100), the ensemble projected average values of the annual △PET were 67.8 and 95.3 mm, respectively, where the variances of projection were 17.6 and 21.4 mm, respectively. The SSP5-8.5 scenario, random forest, and ACCESS-ESM1-5 model were tended to produce the larger values in the ensemble projection, while it was smaller using SSP1-2.6 scenario, PMCO2 and MIROC6 model. In addition, there was the increased variation in the forecast outcomes caused by the uncertainties from the SSP and PETM over time. 3) The importance of various sources of uncertainty in the PET ensemble projection was ranked in the descending order of PETM, SSP, and GCM in the mid-term future. The SSP was the most important source of uncertainty with a contribution rate of 65.3% in the long-term future. Meanwhile, the contribution rates of PETM and GCM were reduced to 19.1% and 7.7%, respectively. In terms of seasonal distribution, the PETM and SSP presented a higher contribution rate to the total uncertainty in the warm seasons (summer, autumn) and cold seasons (winter, spring), respectively. Therefore, the large-scale ensemble projection with enough samples and the statistical characteristics can be widely expected to serve a more comprehensive evaluation than before, particularly for climate change. The findings can provide a potential direction to reduce the uncertainty of future PET assessments.