Application of Time-Scale Decomposition Statistical Method in Climatic Prediction of Summer Extreme High-Temperature Events in South China

A time-scale decomposition (TSD) method to statistically downscale the predictand and predictors is used for seasonal forecast of summer extreme high temperature events (hot days) in South China. The hot days present a significant variability that is associated with distinct possible predictors. Both the hot days and the possible predictors are decomposed into inter-decadal and inter-annual components by fast flourier transformation filtering. Three downscaling regression models are then separately set up for the total hot days and the inter-decadal and inter-annual components of hot days. The downscaling regression model of the total hot days is named as direct regression model, while the downscaled inter-decadal and inter-annual regression models are combined together and named as TSD statistical regression model to obtain the total hot days. The fitting results of the direct regression model and TSD statistical regression model are tested by 10-fold cross-validation. The results show that compared to the direct regression model, the TSD statistical regression model decreases the root-mean-square error (RMSE) from 2.6 d to 2.3 d and increases the correlation coefficient with observations from 0.69 to 0.73 for the inter-decadal component; the TSD statistical regression model also decreases the RMSE from 3.2 d to 2.9 d and increases the correlation coefficient from 0.4 to 0.48 for the inter-annual component; for total hot days, the TSD statistical regression model decreases the RMSE from 4.1 d to 3.7 d and increases the correlation coefficient from 0.48 to 0.68. The hindcast results of hot days during 1979-2010 show that the correlation coefficient between observations and outputs of the direct regression model is 0.57, while the value is improved to 0.72 by the TSD statistical regression model. The forecast results of hot days during the independent validation period (2011-2013) show that the relative RMSE is 26.4% by the direct regression model, and it is 12.3% by the TSD statistical regression model. Compared with observations, both of the direct regression model and the TSD statistical regression model can predict the hot days to some extent in South China, and the TSD statistical regression model performs better for forecasts during 1979-2013.