Improvement of Surface Temperature Prediction Using SVR with MOGREPS Data for Short and Medium range over South Korea

Friday, 19 December 2014
Sung-Joon Lim, Reno K Choi, Kwang-Deuk Ahn, Jong-Chul Ha and Chun-Ho Cho, National Institute of Meteorological Research, Seogwipo-si, South Korea
As the Korea Meteorology Administration (KMA) has operated Met Office Global and Regional Ensemble Prediction System (MOGREPS) with introduction of Unified Model (UM), many attempts have been made to improve predictability in temperature forecast in last years. In this study, post-processing method of MOGREPS for surface temperature prediction is developed with machine learning over 52 locations in South Korea. Past 60-day lag time was used as a training phase of Support Vector Regression (SVR) method for surface temperature forecast model. The selected inputs for SVR are followings: date and surface temperatures from Numerical Weather prediction (NWP), such as GDAPS, individual 24 ensemble members, mean and median of ensemble members for every 3hours for 12 days.
To verify the reliability of SVR-based ensemble prediction (SVR-EP), 93 days are used (from March 1 to May 31, 2014). The result yielded improvement of SVR-EP by RMSE value of 16 % throughout entire prediction period against conventional ensemble prediction (EP). In particular, short range predictability of SVR-EP resulted in 18.7% better RMSE for 1~3 day forecast. The mean temperature bias between SVR-EP and EP at all test locations showed around 0.36°C and 1.36°C, respectively. SVR-EP is currently extending for more vigorous sensitivity test, such as increasing training phase and optimizing machine learning model.