Accurately estimating and predicting solar photovoltaic (PV) power generation is essential for ensuring a stable renewable energy supply. This study proposes a machine learning-based approach to estimate and predict solar PV power generation. A linear model and three machine learning models were trained using hourly datasets from the Yeongam F1 solar power plant from 2019 to 2022. To ensure model stability, a year-based 4-fold cross-validation was performed. Of the tested models, XGBoost model performed the best, achieving a normalized mean absolute error (nMAE) of 4.16% with a standard deviation of 0.07%. Furthermore, in a practical forecasting scenario where models were trained on data from 2019 to 2021, XGBoost model successfully predicted next-day solar PV power generation for 2022 using LDAPS forecasts initialized every 2100 KST, achieving a nMAE of 7.89%. These results demonstrate that combining short-term numerical weather forecasts with machine learning algorithms can enable reliable prediction of solar PV power generation one day in advance.
This study develops and evaluates an optimized subdivision of severe weather alert zones for Busan and Ulsan, using meteorological, high-impact weather, socioeconomic, and land surface data. Based on these variables, a multivariate cluster analysis was conducted at the census output area level. The optimal number of subdivision zones was determined to be three for Busan and two for Ulsan by several statistical methods. From multiple zoning combinations, three candidate plans were selected, and their operational effectiveness was evaluated based on recent occurrences of heatwaves and heavy rainfall. Among them, the combination of high-impact weather and socioeconomic characteristics showed the highest efficiency. This approach is expected to reduce unnecessary alerts, improve the spatial focus of warnings, and it also aligns well with administrative boundaries. The results demonstrate that a single large severe weather alert zone cannot adequately capture local risk patterns. This work provides a scientific basis that supports the redesign of severe weather alert systems and offers a practical framework that could be adapted for other metropolitan areas. In the context of increasing localized extreme weather events, these findings may help optimize resource allocation, and strengthen public responsiveness.
This study evaluates the subseasonal-to-seasonal (S2S) prediction skill of Madden- Julian Oscillation (MJO) and its East Asian teleconnections using the Korea Meteorological Administration's Global Seasonal Forecasting System version 6 (GloSea6) hindcasts. The MJO prediction skill of the GloSea6 model, in terms of the bivariate correlation of the Real-time Multivariate MJO (RMM) index, is about three weeks in winter, extending to four weeks for MJO phases 2-3. However, the prediction skill of MJO teleconnections to East Asia is limited to about two weeks. This discrepancy arises from prediction errors in MJO spatial structure and magnitude, which drastically increase after forecast week two. In particular, the centers of divergence and convergence become latitudinally biased with weaker magnitudes. These errors are attenuated by latitudinal averaging when computing the RMM index. The bivariate correlation metric also underrepresents magnitude errors. Together, they lead to an overestimation of MJO prediction skills. These results highlight the importance of considering the detailed structure and magnitude of MJO circulation and convection when assessing the prediction skills of MJO and its teleconnections.
Within the quality control (QC) procedures, this study reassesses the threshold of the persistence test from the perspective of sub-daily timescale analysis, using 1-min mean windspeed time series from seven Automatic Weather System (AWS) sites in the Ulsan region in 2024. First, we performed preprocessing steps to prevent distortions in spectral analysis arising from temporal discontinuities or nonphysical signals. Next, to quantify the impact of flat-line segments on the spectrum, we randomly selected 21 out of 286 days with no missing values and flat lines shorter than 5 min, and inserted and moved a 0 m s(-1) segment of varying length (6, 10, 15, 20, 30, 120, 240 min) within each time series. As a result, the spectral relative error increased with segment length; in particular, the increase in error was markedly larger for 10 similar to 15 min than for 6 similar to 10 min. In the 240-min case, the largest error appeared at a timescale of about 8 h, and consistency with the spectrum of the original time series deteriorated substantially across the entire frequency range. In addition, the longer the flat-line segment, the more the daily mean wind speed was underestimated and the more the variance tended to be overestimated. These results suggest that the 240-min threshold used in the KMA persistence test imposes limitations on analyses at sub-daily timescales. Considering the need to minimize spectral and statistical distortions, this study proposes 10 min as a realistic and physically sound threshold for the persistence test applied to 1-min mean wind-speed data.
This study examined net biome exchange (NBE) data from various institutions to evaluate the reliability and characteristics of NBE estimates on a global, continental, and national scale from 2015 to 2020. All three NBE datasets-CT2022, CMS-Flux, and CAMS- consistently showed that the global terrestrial biosphere acts as a net carbon sink. The mid-to high-latitude regions of the Northern Hemisphere, which are rich in vegetation, mainly absorb CO2. However, the datasets show substantial discrepancies in the regional magnitude and variability of NBE. Large discrepancies in NBE exist among the datasets in the low-latitude Northern and Southern Hemispheres, where observational data are scarce. At the national scale, China, Canada, the United States, and Russia, while significant anthropogenic CO2 emitters, are consistently identified as carbon sinks by the terrestrial biosphere across all datasets, indicating relatively low uncertainty in NBE estimates. However, countries such as India and Brazil show large inconsistencies in NBE estimates, highlighting significant uncertainties among the data-sets. South Korea, Japan, and the United Kingdom exhibit near-zero NBE values across all data-sets. These results underscore the necessity of multi-model comparisons, cross-validation, and expanded observational networks in underrepresented regions, as well as the importance of integrating biosphere carbon flux considerations into national carbon neutrality strategies.
