Abstract

Taiwan is a small island on the east side of the Pacific Ocean near the equator and is vulnerable to extreme climatic conditions caused by global warming. The increase in air temperature (~ 1.4 °C) in Taiwan is twice within the North Hemisphere (~ 0.7 °C), which causes an increase in respiratory ... Show moreTaiwan is a small island on the east side of the Pacific Ocean near the equator and is vulnerable to extreme climatic conditions caused by global warming. The increase in air temperature (~ 1.4 °C) in Taiwan is twice within the North Hemisphere (~ 0.7 °C), which causes an increase in respiratory and cardiovascular mortality. Therefore, the accurate prediction of maximum temperature (Tmax) and understanding of the uncertainty in the forecast of its extremes during the summer season (June through September; JJAS) is of utmost importance. In recent decades, there have been impressive advances in numerical modeling and prediction of weather and climate. However, the numerical weather model's raw products are not skillful beyond five days, and also the uncertainty increases with lead time. Statistical post-processing is required to deal with systematic errors to get more reliable and skillful forecast guidance. In September 2020, NOAA NCEP implemented Global Ensemble Forecast System version 12 (GEFSv12) to support stakeholders for various risk management applications. It has generated consistent reanalysis and reforecast data for 2000-2019. The reforecast was initialized at 00 UTC every day out to 16 days with five ensembles, except on Wednesdays when the integrations were extended to 35 days with 11 members. The success of deep learning techniques over the last few decades has opened up a new avenue of research for weather forecasting. In the present study, we demonstrate how neural networks are often used for probabilistic post-processing ensemble forecasts within the distributional regression framework by using GEFSv12 Reforecast products of Summer (JJAS) Tmax over Taiwan as an input and the ERA5 Tmax reanalysis as a target. The proposed model architecture allows for incorporating various features relevant for correcting systematic deficiencies of ensemble predictions and estimating the network parameters by optimizing the continuous ranked probability score (CRPS) that can be a mathematically principled loss function for probabilistic forecasts. The results of this study will be compared with benchmark post-processing methods for varying training period lengths, such as 2000-2015, 2005-2015, 2010-2015, and 2014-2015, while the validation and testing periods 2016-2017, and 2018-2019 are going to be used respectively. This study may explore the desired optimum length of the data for training and a suitable calibration method for ensemble probabilistic Tmax forecast over Taiwan. Show less