Executive Summary

Since 2017, the Asia-Pacific Economic Cooperation Climate Center(APCC) has established a robust collaborative framework to support the Korea Meteorological Administration (KMA)'s operational 1-month forecast by collecting, processing, and providing the latest subseasonal prediction information. This study describes the stable operation of the subseasonal prediction system and the deep learning-based probabilistic temperature forecasting system tailored for this purpose. In addition, the migration and improvement of data collection and forecast production servers were carried out to enhance operational efficiency. Through a case study of August 2025, which recorded the second-highest temperatures on record, the system's performance in predicting detailed weather variability—which cannot be identified through monthly mean values alone—was verified. The results demonstrate the high utility of subseasonal prediction information in capturing intra-monthly variability. To address the limitation of seasonal prediction information in a rapidly changing climate, APCC pursued the development of a subseasonal prediction system designed to enhance the value of utilization.

To develop the APCC subseasonal prediction system, a wide range of global subseasonal-to-seasonal(S2S) prediction datasets were systematically collected and standardized. Using these datasets, it is confirmed that the Multi-Model Ensemble (MME) provides stable and superior performance within the subseasonal prediction range. In addition, prediction skill of SubC and S2S models, together with APCC’s in-house model, SCoPS, was rigorously compared and validated. The group comprising the largest number of models exhibited the highest prediction skill, demonstrating that including a broad set of models can enhance MME performance. We also confirmed that the SubC project model due to the various component model composition can further increase the overall contribution to the MME. Also, we evaluated parametric probabilistic prediction methods for temperature and precipitation in order to identify methodologies optimized for the statistical characteristics of each variable. The results indicate that parametric methods outperform non-parametric approaches in terms of quantitative predictive skill and can effectively reduce spatial noise arising from an insufficient ensemble size. For precipitation in particular, we taken into account a hybrid gamma algorithm for the high frequency of dry events that automatically switches to a nonparametric method when the effective ensemble size in insufficient.

Building on these research findings, we established the model configuration and probabilistic methods required for a APCC subseasonal prediction system. The system ingests weekly forecast data from 10 models (including APCC's SCoPS, BOM, NCEP, HMC, etc.), to generate subseasonal prediction products. Data acquisition employs various methods, such as FTP, the ECMWF API, and the IRI Data Library, depending on the provider. A critical component of the MME system is the automated pre-processing system. All individual model data is standardized to a 1-degree resolution, transformed into various mean fields (weekly, 1–4 week mean, etc.), and converted into the NetCDF format. For the final MME production, the Simple Composite Method (SCM) is used for deterministic forecasts, while the Hybrid Gamma (Mixed Gamma + Quantile) method is employed for probabilistic forecasts. This integrated system allows for the production and display of MME prediction results, initialized to start every Monday.

To diversify and enhance the usability of subseasonal information, the study explored and identified seamless content that integrates subseasonal (weekly) and monthly forecast information. We identified products that can be delivered immediately such as integrated monthly/weekly probability distributions and weekly variability information, as well as more advance products requiring additional technical development including probabilities of intramonthly extreme events. By distinguishing these contents, we provides a clear roadmap for APCC to become a leading provider of integrated subseasonal-seasonal climate information. The approaches and findings presented herein are expected to enhance APCC’s competitiveness and expand its role in next-generation climate services.

To include SCoPS as one of the contributing models, we developed a real-time subseasonal forecast system that produces 60-day forecasts every week, based on seasonal forecasts every month. To facilitate this, initial fields are produced by automatically collecting CFS analysis fields and ARGO ocean observation data every Thursday for the preceding week (Wednesday to Tuesday). SCoPS performs the 60-day forecast with 10-ensemble member from every Tuesday, and the processes including initial, forecast, post-processing production are completed by Friday. This forecast result is then utilized as an input for the MME subseasonal forecast performed the following Monday. However, given the relatively old development baseline of SCoPS and its lower skill compared with other MME participating models, there is clear need to reduce its systematic errors. This study therefore focused on reducing errors in the initial conditions first for SCoPS and, in turn, decreasing subseasonal forecast errors.

To reduce land-related initial errors, which are critical for subseasonal prediction, we developed a soil moisture initialization technique for SCoPS and quantitatively evaluated its impact. Using the Ensemble Adjustment Kalman Filter (EAKF) algorithm with ERA5-Land reanalysis data, we conducted hindcast experiments for 2003-2016. Initialization achieved 99.3% global RMSE reduction within 10-12 days, with soil moisture memory persisting 30-40 days. Land-atmosphere coupling strengthened significantly over mid-latitude continental regions, particularly Central US and Southeastern Europe. Temperature forecast skill improved 10-60% regionally for 16-30 day lead times, with sustained improvements up to 60 days over regions with strong coupling. Precipitation showed no significant response, consistent with its primary control by atmospheric circulation rather than land surface conditions. The causal chain from initialization → memory → coupling → predictability was quantitatively established, demonstrating that soil moisture initialization effectively contributes to subseasonal prediction through physically consistent mechanisms.

We also examined how improvements in oceanic and atmospheric initial conditions influence prediction performance on both subseasonal and seasonal time scales. The ocean-initialized experiment effectively reduced the long-standing warm biases and overly deep thermocline in the eastern Pacific, but it has only a limited and weakly time-evolving impact on a atmospheric variables and subseasonal prediction skill. In contrast, the atmosphere-initialized experiment shows a more rapid and direct response: global, tropical, and East Asian mean near-surface temperature errors are consistently reduced across lead times, and the large-scale lower- and upper-tropospheric circulation is improved. These changes in temperature and wind fields enhance lower-tropospheric humidity and precipitation forecasts associated with the East Asian summer monsoon.

On the basis of these results, we have initiated pilot operations that newly established APCC subseasonal prediction to provide enhanced forecast information. We plan to develop display systems for weekly prediction products, select highly useful contents for operational dissemination. New products such as forecasts based on improved SCoPS initial conditions and probabilities of intramonthly extreme events will be incorporated on a continuing basis. Through these enhancements and expansions, the subseasonal prediction information developed in this study is expected to strengthen forecast support for KMA and improve climate prediction services across the Asia-Pacific region.