Precipitation prediction in Thailand by using regression analysis

Abstract

A methodology is presented for statistical downscaling the general circulation model (GCM) of GFDL-R30 and HadGEM output to predict precipitation at meteorology stations in Thailand under expected future climatic conditions. The meteorology stations that were sampled for testing for prediction were Chiang Mai, Udon Thani, Bangkok, Ubon Ratchathama, Nakhon Ratchasima and Phuket, provinces from each region in Thailand. The approach involves a combination of data preprocessing, data classification and regression analysis. In the preprocessing step, this research used data with GFDL-R30 which was compared between logarithm transformations, square root transformations and no transformations. The results showed the correlation yield of both algorithms was less than no transformations but they can help for negative numbers in predicted data. In the classification step, it applied to the model for cutoff on days that had less rainfall and heavy rainfall. It was varied from 0 mm to 1.5 mm. The results showed the classification cannot increase strong correlation performance. The last section presents about post-processing comparison between multiple linear regressions and Gaussian process which is a non-linear regression. The result showed Gaussian process generally yielded better results both for correlation and RMSE for every station that was tested. Therefore, the Gaussian process is a suitable approach for predicting precipitation in the future without any preprocessing.