| Abstrakt: |
This study presents the results of an effort to improve the forecast of precipitation (> 0.1 mm/hr or > 0.1 mm/3hr) in the Local Data Assimilation and Prediction System (LDAPS) and the Global Data Assimilation and Prediction System (GDAPS) by applying the Random Forest (RF) model in South Korea. LDAPS and GDAPS are Numerical Weather Prediction (NWP) models operated by the Korea Meteorological Administration (KMA) for weather forecasting. GDAPS operates the Unified Model (UM) and the Korean Integrated Model (KIM). This study used weather forecast data from LDAPS, GDAPS/KIM, and GDAPS/UM. Precipitation forecasts from LDAPS and GDAPS were corrected by RF training with rain gauge observations from about 685 stations. Approximately 35 selected NWP model output variables were used as inputs to the RF training. To reflect recent trends in biases between observations and NWP, the precipitation probability prediction model was designed for real-time learning using a sliding window technique. In addition, the precipitation data had a data imbalance problem with more precipitation cases than non-precipitation cases, so an undersampling method was applied to solve this problem. Comparing the performance of the proposed method with NWP in predicting precipitation, the CSI was improved by 14.7-23.1% (LDAPS), 33.9% (GDAPS/KIM), and 6.7%-38% (GDAPS/UM) over NWP, and the accuracy was also better. In future research, automating the sampling rate selection to reflect recent weather trends when undersampling is likely to improve forecast performance. [ABSTRACT FROM AUTHOR] |
