Tuesday, 23 January 2024
Heatwaves can cause more fatalities than any other weather-related disasters in East Asia due to heat exhaustion and heat strokes. The number of people with heat related illness also increased over the years as the intensity and frequency of heatwaves increased. Increasing of heatwave frequency and severity suggest that the prediction of heatwaves and tropical nights are becoming more important. The Korean Meteorological Administration (KMA) operates Local Ensemble Prediction System (LENS) to stochastically provide weather information at 2.2 km resolution. Research shows the LENS model underestimates daytime temperature and overestimates temperature at nighttime. The statistics indicates difficulty in predicting heatwaves and tropical nights over the Korean Peninsula (K.P.) by LENS model. In this study, we conducted Convolutional Neural Network (CNN) deep learning method to create heatwave and tropical night prediction model using LENS and observation data (Automated Weather Stations; AWS, Automatic Synoptic Observation System; ASOS). In order to obtain distribution of temperature prediction on K.P,, we choose U-Net model has up-sampling process with various activation functions (Rectified Linear Units; ReLU, LeakyReLU, hyperbolic tangent). Based on the three years of test datasets, U-Net model with hyperbolic tangent function (Tanh) showed 0.85 R2 and 1.60 oC RMSE in the next day heatwave prediction higher than LENS in 0.51 R2 and 2.80 oC RMSE. In next day tropical night prediction, the Tanh model showed 0.87 R2 and 1.18 oC RMSE higher than LENS in 0.84 R2 and 1.32 oC RMSE. In addition, we calculated heatwave severity index (HSI) could reflect heatwave illness on the K.P. and deep learning model showed more similar distribution of AWS observation data compared to LENS and Model Output Statistics (MOS). We constructed heatwave and tropical night prediction model can forecast up to 3 days. Among prediction models, the Tanh model showed best results.
Currently we are developing Long-Short Term Memory (LSTM) model, which can apply past heatwave and topical night data to forecast.
Currently we are developing Long-Short Term Memory (LSTM) model, which can apply past heatwave and topical night data to forecast.
※ This work was funded by the Korea Meteorological Administration (KMA) Research and Development Program under grant KMI2017-02410.

