Executive Summary

As a base study for the improvement of long-range forecasts skill and integration of forecast information, we analyzed models’ performance of ENSO and its related atmospheric responses in intra-seasonal scale for the APCC MME individual models during winter month. Although the performance of the models’ Nino index is highly correlated with the observation, models have errors in tropical sea surface temperature and precipitation responses regionally and monthly. The models show similar atmospheric response to ENSO during winter months (DJF) but not show monthly variation such as weakened western Pacific precipitation in January as in observation. The atmospheric responses to ENSO show low skill in the East Asia compared to the Northern hemisphere and the skill is lowest in February regardless of lead time. It is over estimated that the ratio of ENSO related variation to the total variation in model simulation.

We have divided arctic events into four categories on monthly time scale during the winter months and analyzed the frequency of occurrence of cold event in East Asia. Among the four categories of arctic events, we have analyzed the composite patterns of deep arctic warming and shallow arctic warming. It is found that there is high possibility of cold event during January and February when deep arctic warming occurs. Few models can predict the arctic events during the month of January and February, and also reliably reproduce the atmospheric circulation patterns when they predict the occurrence of deep arctic warming event.

For the long range forecast over Korea, the seasonal climate models have capability to simulate the impact from both higher-latitude and lower-latitude. Both Color-Polar (CP) and Warm-Tropical (MT) near Korea are considered as two important indicators that reflecting the impacts from higher and lower latitudes in the forecast. The APCC MME has capability to reasonably simulate the intraseasonal variability of CP and MT, which is also well related to the temperature and precipitation variability in Korea. Therefore, we can select the best model that well simulates the relationship the CP and MT variability with temperature and precipitation in Korea. The best model chosen which reasonably simulates the climatological feature and the relationship with CP and MT with temperature precipitation around Korea, can improve the intraseasonal predictability in Korea. The best model, which is additionally tailored over targeted area (here in Korea), can further improve the intraseaonal predictability.

As to the 1-month long range forecast, the boreal summer intraseasonal oscillation is one of the main prediction sources for the subseasonal forecast over the Asian summer monsoon region. We investigated the predictability of the real-time BSISO index and BSISO impact anomaly in the WMO S2S ECMWF model, currently participating in the 1-month forecast. The ECMWF model exhibited the BSISO prediction skill out to 3-4 weeks, but it has weaker amplitude and slower propagation than the observation. Convective and associated circulation anomaly patterns estimated by the BSISO index can be predicted up to 3 weeks. In addition, predicted BSISO-related anomalies showed higher pattern correlation coefficients in active, warm phase, heatwave phase or dry phase cases in week 3. We confirmed the possible impact of BSISO on Changma onset/withdrawal and mean rainfall in 2022 and therefore, the BSISO impact forecast could be useful information during the monsoon season in Korea.

We developed 1-month temperature prediction model using deep learning technology and established probability prediction system in order to improve the accuracy of 1-month temperature probability prediction around East Asia and including South Korea. A deep learning-based prediction model was designed using bi-directional convolutional long short term memory (ConvLSTM), a channel-attention mechanism like squeeze and excitation block, etc, and was evaluated. Training dataset was built using daily ERA5 2m temperature (T2M) reanalysis data. the European Centre for Medium-Range Weather Forecasts (ECMWF) forecast data of week-1 and week-2, and a rolling prediction method for T2M prediction of week-3 and week–4 were used. After building training and validation dataset using daily ECMWF T2M reanalysis data from 1986 to 2015, we shuffled and trained the deep learning model. The predictability of the developed model was evaluated using daily ECMWF T2M reanalysis data from 2018 to 2021 in two regions, East Asia and South Korea. As evaluation methods, ROC (receiver operating characteristic) score map around East Asia and HSS (heidke skill score) in South Korea were used. As the results of ROC score map around East Asia, the predictability of the developed model was similar to the predictability of ECMWF model. On the other hand, as the results of HSS in South Korea, the probability prediction of the developed model showed slightly higher predictability than that of the ECMWF model. In addition, ECMWF model and the deep learning model showed high predictability of Above Normal and Near Normal, respectively. Therefore, the deep learning-based 1-month T2M probability prediction system developed in this paper can be helpful to improve the accuracy of the 1-month temperature probability prediction. Also, if the advantages and disadvantages of both models are utilized for 1-month T2M prediction, it can be the possible to advance probability prediction.

APCC has been operating a 1-month prediction system since August 2017 to support the KMA’s 1-month forecast. This system has been continuously improved with changes in demand of operator and model characteristics. The forecast produced by APCC showed better skill than the other climate prediction models. These results were fully shared with the KMA and contributed to the improvement of KMA’s 1-month forecast.