A Hybrid Statistical Downscaling Framework Based on Nonstationary Time Series Decomposition and Machine Learning

Downscaling techniques are effective to bridge the scale gap between global circulation models and regional studies. Statistical downscaling methods are prevalent due to their advantages in high computational efficiency and accuracy. However, an implicit assumption of most statistical techniques is that time series should be relatively stationary after certain function transformations. Otherwise, statistics of nonstationary time series may be meaningless for describing future behavior. In this study, a hybrid statistical downscaling framework was developed through integrating bivariate empirical mode decomposition (BEMD) and a machine learning method to extract multi‐timescale features from nonstationary data to enhance downscaling performance. The proposed framework can reduce the effects of non‐stationarity in data‐driven models by using the BEMD method, which can decompose time series into independent and stationary components at multiple time‐frequency resolutions. It was applied to downscale monthly precipitation and temperature of Canadian Earth System Model in multiple stations with different climate types in the Central Valley of California, USA, to verify its accuracy and generalization ability. The performance of downscaling maximum and minimum temperatures ( R 2  > 0.9) was more accurate than that of precipitation. The potential reason is that precipitation is more sensitive to transient weather phenomena, which can only be extracted from data with higher temporal resolution. The proposed model was further compared with models based on discrete wavelet transform and models without time series decomposition. The results showed that a decomposition strategy of the proposed framework can improve the downscaling accuracy, potentially providing a viable option to deal with the nonstationary of data in statistical downscaling models.