A Statistical Downscaling Study On The Daily Extreme Temperature In South China

Global climate model(GCM) is one of the most important models to forecast and describe the global climate change. It can simulate well the large-scale average climatic characteristics for future. However, because of its low spatial resolution, it is hard for GCM to make a detailed prediction for the local weather variation. Through establishing the statistical relationship between the low resolution large-scale GCM outputs and local weather variables, Statistical Downscaling(SD) can describe well the variation of weather condition at high resolution, even at a certain point. SD is now one of the most important techniques in predicting local weather variations by using large scale global climate model outputs.Based on the principle of Statistical Downscaling, this study proposed several methods: stepwise regression, principal component regression, and BP neural networks combined with genetic algorithm, to characterize the variability of daily extreme temperatures at different sites in South China from 2000 to 2013. The data of 2000-2010 were used for model calibration and the data of 2011 to 2013 were used for model validation. The daily temperature extremes, daily maximum temperature and daily minimum temperature, at 21 weather stations in South China were applied as predictands, and the large scale GCM output variables significantly correlated to the local weather conditions were screened out as the predictors. The large scale GCM output variables were 1 ° × 1 ° Final Operational Global Reanalysis data provided by National Centers for Environmental Prediction(NCEP) and National Center for Atmospheric Research(NCAR) in the United States.The performance of the three models in simulating the daily temperature extremes at the 21 stations in South China were compared in terms of coefficient of determination(R2), Root Mean Square Error(RMSE), and Mean Absolute Error(MAE). The following conclusions are obtained.(1) The proposed three regression downscaling models can provide accurate estimates of fundamental statistical and physical properties of Tmax and Tmin. Among them, BP neural network combined with genetic algorithm model works best with the average R2 about 0.85 and average RMSE about 1.5 o C. However, it is more time consuming for model computation than the other two models.(2) Generally, SD models can provide more accurate estimates of the local Tmax and Tmin characteristics at the coastal stations than the inland stations in South China.(3) Comparing the models performance in different seasons, the three models can catch the variation of the daily temperature extremes better in winter than in summer in terms of R2. Due to the influence of severe weather such as tropical cyclones in summer, it is hard for the three models to simulate well the daily temperature variation, but the simulation error is small in terms of MAE and RMSE.