A multivariate bias correction algorithm for climate model predictions and projections

Model output statistics (MOS) are generally tailored for single variables, despite their application to multivariate time series being constantly required in many problems. For instance, distributed, physical-based hydrological models often require as input meteorological variables (e.g. precipitation, land temperature, evapotranspiration, etc) that are strongly correlated. Preserving the spatio-temporal variability of single variables as well as the inter-variables dependence structure is thus of fundamental importance in climate model outputs to enhance, for instance, reliable hydrological predictions. In this context, we extend the multivariate bias correction algorithm (MBCn) proposed by Cannon (2018) through pre-filtering the input data and improving the orthogonal rotation matrix. We finally evaluate different bias correction algorithms. Our proposed approach is grounded on the multivariate techniques of principal component analysis (PCA) and sparse principal component analysis (SPCA). It seeks to promote bias correction while preserving spatial and inter-variable dependencies. We apply and test our algorithm using S2S predictions provided by the C3S multi-system seasonal forecast service, which includes climate models such as ECMWF, NCEP and canCM4i. The ERA 5 reanalysis data are used as reference meteorological data. We particularly explore the application of the proposed methodology to daily S2S forecasts of precipitation, temperature, wind field and surface solar radiation, which are notably valuable as input to hydrological models and to estimate evapotranspiration, droughts indices and renewable energy yields.

 

Acknowledgment

This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (No. 2019R1A2C2087944).