Executive Summary

In recent years, artificial intelligence (AI)-based medium-range weather prediction models (within 15 days) have been actively developed by leading companies and research institutes, including NVIDIA, Huawei, Google DeepMind, and the European Centre for Medium-Range Weather Forecasts (ECMWF). Notable examples include Nvidia's Spherical Fourier Neural Operators (SFNO)-based FourCastNet, Huawei's Swin-Transformer-based Pangu-Weather, ECMWF's Graph Neural Network (GNN)-based Artificial Intelligence/Integrated Forecasting System (AIFS), Fudan University's U-Transformer-based FuXi model and FuXi-S2S for subseasonal-to-seasonal (S2S) forecasting, and Google DeepMind's GNN-based GraphCast and diffusion model-based GenCast. These models are predominantly focused on medium-range weather forecasting and have not been widely extended to S2S forecasting. Futhermore, the training of AI models requires large datasets, but the ECMWF ERA-5 dataset-the primary source of climate data for AI training-spans only approximately 75 years. This limitation significantly restricts the amount of data available for training, validation, and testing.

Due to the climate crisis, the importance of high-accuracy subseasonal prediction data in the field of applied climatology is increasing. Such subseasonal prediction data face challenges in forecasting based solely on numerical models, as they rely on physical interactions among various atmospheric and oceanic processes. To overcome these difficulties, efforts have been made to enhance accuracy by expanding datasets and assigning greater weights to important patterns and features in input data. This study utilized the Attention U-Net model and subseasonal prediction models, as well as observational data-based techniques such as Filter, Wrapper, and Embedded methods, to identify the characteristics of variables. Sensitivity analysis was conducted based on various input-output systems. It was confirmed that selecting specific combinations of variables from model/observational data improved accuracy compared to using all variables for subseasonal predictions. The GUI-based input-output system facilitated the generation of input data and sensitivity analysis, contributing to the development of optimized artificial intelligence techniques. This research can serve as foundational data for improving subseasonal prediction accuracy and highlights the need for further advancements in deep learning models in the future.

Recent advances in neural network models and deep learning technologies have gained significant attention across various industries and research fields, including their growing potential in climate prediction. In the context of Sub-seasonal to Seasonal (S2S) forecasting, there is a strong emphasis on the importance of data preprocessing and improving model performance. This study focuses on enhancing the prediction accuracy of daily maximum temperatures (TMAX) in S2S forecasting using ensemble methods such as voting, bagging, boosting, and stacking. The ensemble framework was built using CNN, CNN-LSTM, U-Net, Attention U-Net, and Residual U-Net as base models, and the characteristics and performance of each ensemble approach were systematically analyzed. Results showed that, for ECMWF-S2S TMAX forecasts, all ensemble methods improved the Anomaly Correlation Coefficient (ACC), with bagging achieving the best overall performance. For KMA (GloSea5)-S2S TMAX forecasts, stacking and boosting demonstrated consistent performance improvements across the entire lead time range (1 to 4 weeks), whereas voting and bagging were effective primarily beyond 4 weeks (20 days). Spatial distribution analyses revealed that applying ensemble techniques to ECMWF forecasts enhanced ACC across all lead times (1 to 4 weeks). Notably, the ACC spatial distribution for bagging closely aligned with the pre-ensemble distribution, showing minimal differences. In evaluations targeting specific areas of interest (e.g., four grid points over South Korea), performance improvements were more pronounced for KMA forecasts compared to ECMWF. This highlights that the effectiveness of ensemble techniques can vary depending on the training data and regional characteristics. This study underscores the value of ensemble methods in improving the stability and accuracy of S2S forecasts. The findings are expected to contribute to enhancing the precision and practicality of S2S daily maximum temperature predictions, particularly for 3–6 week lead times, and to improving the reliability of extreme heat forecasts.

In the second year (2023) of this project, we developed an image-based deep learning model for classifying MJO phases using semi-supervised learning (SSL) techniques. For this purpose, we first optimized a supervised learning model by testing various combinations of input variables, model architectures, and training data split methods based on different MJO indices. Then, the optimized model was trained in the SSL framework with both unlabeled and labeled data to classify MJO phases. The semi-supervised model demonstrated its effectiveness in learning critical climate patterns even with reduced labeled data, achieving comparable or superior accuracy to supervised models with all labeled data. Sensitivity analyses revealed that spatial characteristics of climate data should be properly taken into account for data augmentation in SSL. Horizontal flipping reduced accuracy while vertical flipping improved performance for MJO phase classification. Building on this, the study extended the semi-supervised approach for MJO phase classification to RMM prediction in the present year of 2024. A modified ResNet-18 model was developed to predict RMM1 and RMM2 indices for lead times of 0 to 40 days. By integrating anomaly and background climate variables (U200, U850, OLR, VP, TS, TCWV, hadvect), the model achieved skillful RMM predictions up to 22 days based on BCOR 0.5. A custom loss function incorporating MJO amplitude further reduced amplitude-related errors, emphasizing the importance of domain-specific adjustments in deep learning-based MJO research. Interpretable AI techniques highlighted the roles of background wind fields and moisture anomalies in MJO prediction, offering insights for future studies. Finally, probabilistic prediction models were developed using ensemble perturbation methods for initial conditions and model parameters, including ERA5 ensemble data, Perlin noise, and MC dropout. Results suggested that Perlin noise method showed more ensemble spreads over lead times compared to others but was inconsistent across MJO cases, reflecting the challenge of generating ensemble spreads using deep learning models. Future research will explore advanced AI models, high-resolution data for detailed spatial pattern identification to further enhance climate predictability.

To improve the predictability of temperature probabilities on S2S timescales, we developed a deep learning-based 3–6 week temperature probability forecast model for East Asia. The model employs a a U-NET architecture enhanced with attention mechanisms and eXplainable Artificial Intelligence (XAI) to build one-month temperature prediction models. This served as the basis for additional models, incorporating expanded input data from the Atmosphere-Land Surface-Ocean (ALO) system and leveraging state-of-the-art deep learning techniques. A total of 31 models were constructed, including ensemble learning models, N-step forecasting models, and retraing-baed models. These models were evaluated against the ECMWF S2S model using Heidke Skill Score (HSS).

The input variables used in the model design were T2M (atmosphere) and SST (ocean) from the ERA-5 reanalysis dataset, and NDVI (land) from MODIS and AVHRR. These models are 7-day or 14-day inputs and projected 1 day into the future, and for the 3-6 week temperature probability forecasts, a rolling prediction technique was used to build a temperature probability prediction system using the ECMWF S2S model 1-2 week forecast and hindcast data as inputs. Among the input variables of the models, T2M and SST variables are different from the NDVI data, which is only available for 1982-2019, so the evaluation period of the models was divided into 2018-2019 and 2018-2022. For the evaluation period 2018-2022, NDVI data is not available for 2020, so 2019 data was used. The evaluation results show that for the 2018–2019 period, the top-performing model was DL-SE-7+SST+NDVI, followed by DL-SE+FLT2+BTLNCK-7 and NM-SE-7. ECMWF ranked 25th. For the 2018–2022 period, DL-SE-14 ranked first, DL-SE+BIGDATA+FLT2+BTLNCK-14 ranked second, and DL-SE+BIGDATA-14 ranked third, with ECMWF ranking 18th. During El Niño events, the top three models were DL-SE-7+SST+NDVI, DL-SE+FLT2+BTLNCK-7, and DL-SE+BIGDATA-14, with ECMWF ranking 7th. During El Niño events, the top three models were DL-SE-7+SST+NDVI, DL-SE+FLT2+BTLNCK-7, and DL-SE+BIGDATA-14, with ECMWF ranking seventh. Models incorporating the ALO system outperformed others during the 2018–2019 period, suggesting that land (NDVI) and ocean (SST) data positively influenced 4–6 week forecasts. However, for 2020 onwards, the limited availability of NDVI data reduced the predictive power of models utilizing this variable. Overall, SST and NDVI data were found to have a positive impact on the predictability of the models for S2S temperature forecasts.

Improvements of monthly climate predictions were tested by applying a simple data augmentation technique, Cutmix, to a simple convolution neural network (CNN) model. The contirubtion of the data augmentation method were assessed for the predictions of the summer/winter monthly mean temperature and 1-month Standardized Precipitation Index (SPI1). Summer/winter validation accuracy values for both variables increased with data augmentation during the training phase. Epistemic uncertainty was reduced for the test phase when examined by Monte Carlo dropout. Individual monthly predictions, however, only showed improvements for January (LT2), February (LT3) and August (LT3) for monthly mean temperature, and for July (LT2), August (LT3), and December (LT1) for SPI1. Confusion matrices were reviewed; hit rate were improved most of the cases. Although the contribution of simple data augmentation techniques can be limited with overly simplified models even though the cause of poor performance is the lack of training data, some improvements were observed by using a data augmentation technique overall.

To explore AI applications in seasonal climate prediction, this study aims to develop deep learning models for post-processing the Multi-Model Ensemble (MME) forecast of 3-month average precipitation over East Asia. Three different post-processing approaches are applied: ensemble SubSampling (SS), DeBiasing (DB), and Transfer-Learning based Diagnosing (TLD). High-performing architectures and hyperparameters are identified through extensive sensitivity testing. The conventional Quantile Mapping (QM) bias-correction method is used as the baseline for evaluating AI model performance. Among the 12 test seasons in 2023, the DB approach was identified as the best-performing model for 7 seasons, followed by SS for 3 seasons, and TLD and QM each for 1 season. This highlights the effectiveness of AI techniques in seasonal forecast post-processing. However, the significant variability in performance across different post-processing approaches, model architectures, and seasons underscores the need for the MME method to mitigate the uncertainties of individual AI-based post-processing models. Therefore, this study develops codes for calculating both deterministic and probabilistic MMEs of post-processing models and historical performance. It finds that a simple composite mean of 10 post-processing models generally outperforms individual models, with a 40% improvement in ACC compared to the operational APCC-MME. Lastly, an effective visualization layout is designed, potentially introducing a new prototype for an AI-aided APCC-MME.