Machine Learning Reveals New Climate Zones for South Korean Building Energy


A meteorological investigation using machine learning clustering across ninety South Korean weather stations has revealed four distinct climate zones, uncovering a previously unrecognized coastal classification and shifting the regulatory boundaries for twenty-seven regional administrative districts under national building energy codes. The research re-examined South Korea’s Building Energy Conservation Design Standards by applying unsupervised learning algorithms to ten years of weather records spanning 2015 to 2024. In the existing regulatory structure, South Korean provinces are segmented into four broad administrative zones—Central 1, Central 2, Southern, and Jeju—to assign minimum envelope insulation standards and evaluate building code compliance. However, prior code divisions relied heavily on heating degree days rather than multidimensional meteorological factors. By applying multivariate statistical learning to comprehensive weather parameters, the study demonstrated that existing heating-centric boundaries overlap substantially, grouping regions with sharply contrasting heating and cooling degree day patterns into identical regulatory categories while failing to distinguish unique maritime climate conditions along the peninsula’s perimeter.

South Korea governs total building energy performance primarily through two regulatory frameworks: the Building Energy Conservation Design Standards and the Zero Energy Building certification system. The design standards mandate minimum thermal transmittance requirements for building envelopes, mechanical equipment specifications, and renewable energy adoption. Compliance historically operated through a prescriptive pathway based on the Energy Performance Index, but since 2010 authorities implemented a computational simulation-based performance pathway, known as the Total Building Energy Consumption Regulation System. This evaluates annual primary energy consumption across heating, cooling, hot water, lighting, and ventilation. In 2017, South Korea introduced the Zero Energy Building certification framework to advance national carbon neutrality goals, rating buildings from Level 5 to Plus based on code-compliant energy simulation results and energy self-sufficiency rates. Because building envelope criteria and baseline simulation inputs derive directly from climate zone designations, regional zoning accuracy directly influences thermal design loads, compliance approvals, equipment sizing, and certification ratings.

The current building code divisions trace back to a 2018 revision that expanded the national classification from three regions into four. Despite this administrative adjustment, official documentation did not detail which meteorological parameters governed the reclassification. Over recent decades, observed climate data across the Korean peninsula reflect clear warming trends. Comparing long-term thirty-year historical averages from 1991 to 2020 against recent ten-year records from 2015 to 2024 across sixty-five meteorological stations revealed a distinct leftward shift in heating degree days and a rightward shift in cooling degree days. Heating degree days decreased while cooling demand steadily rose nationwide. Furthermore, an evaluation of sixty-five stations using K-means clustering across both periods indicated that nine percent of regions shifted boundaries between thirty-year and ten-year records, while mean heating degree days across zones dropped by two to three percent, confirming that warming weather patterns alter regional building thermal demands.

To establish a systematic data-driven classification, researchers compiled ten-year average meteorological observations from ninety weather stations operated by the Korea Meteorological Administration. The initial data pool encompassed sixteen candidate parameters reflecting thermal, hygric, solar, and geographical characteristics: annual average temperature, maximum air temperature, minimum air temperature, summer maximum temperature, winter minimum temperature, annual temperature standard deviation, daily air temperature range, heating degree days with an eighteen-degree-Celsius base, cooling degree days with a ten-degree-Celsius base, precipitation, relative humidity, minimum relative humidity, humidity standard deviation, latitude, longitude, and elevation. To rank variable significance, all sixteen parameters were standardized using z-scores, and K-means clustering with four clusters was executed across two hundred random initializations. Feature importance, defined as the variance of standardized cluster centroid coordinates across clusters, showed that winter minimum temperature, minimum air temperature, annual temperature standard deviation, and heating degree days ranked in the top five in at least ninety-nine percent of runs, while annual average temperature ranked in the top five in fifty-four percent of runs.

Because the five highest-ranking parameters exhibited strong collinearity—with absolute correlation coefficients between 0.75 and 0.98, variance inflation factors between 22.6 and 142.4, and the first principal component accounting for 92.2 percent of total variance—researchers conducted extensive sensitivity assessments. Diagnostic tests confirmed the correlation matrix had a condition number of 1157 and a mean absolute correlation of 0.90. Sensitivity evaluations included leave-one-out clustering, replacing annual average temperature with alternative candidates such as minimum humidity or daily temperature range, clustering in principal component analysis space, clustering with weakly correlated subsets, and executing two hundred subsampling iterations using eighty percent of the stations. When individual variables were removed, only 2.2 to 5.6 percent of stations changed zone assignments under K-means, and clustering along the first two principal components yielded identical zone boundaries. K-means and Ward hierarchical clustering demonstrated strong mutual consistency, achieving an adjusted Rand index of 0.90 with divergent assignments occurring in only three stations: Yeongdeok, Gochang, and Jangheung.

The optimal number of climate clusters was evaluated across cluster counts ranging from three to ten using three internal validation metrics: the silhouette coefficient, the Davies-Bouldin index, and the Calinski-Harabasz index. The silhouette coefficient reached its highest value of 0.447 at four clusters for K-means, slightly exceeding 0.446 at three clusters, while Ward hierarchical clustering peaked at 0.437 for four clusters. The Davies-Bouldin index for K-means reached its minimum of 0.648 at four clusters, and elbow evaluation of within-cluster sum of squares demonstrated reductions of forty-three percent at three clusters, twenty-seven percent at four clusters, and twenty-six percent at five clusters, flattening below eighteen percent beyond six clusters. Based on these validation metrics and alignment with the four existing regulatory tiers of the 2025 standards, researchers designated four data-driven zones: CZ 1 (Jeju Region), CZ 2 (Coastal Region), CZ 3 (Central Plain Region), and CZ 4 (Central Inland Mountainous Region), formally establishing a dedicated coastal classification that does not exist in current national building codes.

Under the proposed four-zone K-means structure, twenty-seven of the ninety analyzed weather stations were reclassified into different climate zones compared to the 2025 building standards. Specifically, twelve stations moved from Central 2 to Central 1, eleven stations moved from Southern to Central 2, three stations shifted from Central 2 to Southern, and one station shifted from Central 1 to Central 2, while the Jeju region remained unchanged. The boundary separating the Central Plain Region and the Central Inland Mountainous Region reflected a continuous thermal gradient without a sharp geographic demarcation, yielding a mean silhouette coefficient of 0.43 for both clusters, with four transition stations—Geochang, Chupungnyeong, Imsil, and Yeonpyeong—registering silhouette scores below 0.2. Because national codes currently lack dedicated insulation specifications for the newly defined clusters, researchers mapped each proposed zone to its closest counterpart in the 2025 standards based on regional geographical composition and heating degree day similarity, matching CZ 4 to Central 1, CZ 3 to Central 2, CZ 2 to Southern, and CZ 1 to Jeju.

To evaluate how reclassification alters code-compliance assessments, researchers conducted computational thermal simulations using EnergyPlus version 22.1.0 on two modified U.S. Department of Energy prototype buildings adapted to South Korean regulations: a ten-floor high-rise residential apartment covering 7836.5 square meters and a three-floor medium commercial office covering 4982.2 square meters. Four representative cities capturing all four reclassification trajectories were modeled using typical meteorological year weather files spanning 2011 to 2025: Wonju (reclassified Central 1 to Central 2), Cheonan (Central 2 to Central 1), Gangneung (Central 2 to Southern), and Daegu (Southern to Central 2). Simulation parameters incorporated envelope thermal transmittance limits, window-to-wall ratios of 29.8 percent for residential and 33.0 percent for commercial models, ASHRAE Standard 90.1-2022 operating schedules, internal lighting and equipment loads, and automated heating, ventilation, and air-conditioning system sizing with water-source heat pumps and packaged variable-air-volume central systems.

The building simulation results demonstrated that altering climate zone designations produces region-specific and building-specific shifts in annual and peak energy demands rather than uniform changes. In the medium office prototype, annual total heating and cooling consumption rose by 4.5 percent in Wonju (+0.92 kWh/m²) and 6.6 percent in Gangneung (+1.34 kWh/m²), but dropped by 4.0 percent in Cheonan (−0.87 kWh/m²) and 6.0 percent in Daegu (−1.30 kWh/m²). For high-rise residential models, total annual energy consumption changed by +0.4 percent in Wonju, −0.8 percent in Cheonan, +6.9 percent in Gangneung, and −6.0 percent in Daegu. These differences stemmed from envelope thermal properties and solar heat gain coefficients interacting with internal building loads. In the high-rise apartment, switching to Central 2 criteria in Wonju increased window solar heat gain coefficients by forty-three percent from 0.127 to 0.181, allowing greater daytime solar heat absorption that reduced winter heating demands by 0.59 kWh/m² while elevating summer cooling loads by 0.66 kWh/m². Conversely, office building windows experienced negligible solar coefficient changes from 0.487 to 0.484, causing weakened wall insulation to increase winter heating by 1.06 kWh/m². Peak-day heating loads also showed pronounced shifts, swinging between −20.6 percent and +20.2 percent in offices, where morning heating startup at 08:00 accounted for fourteen to fifteen percent of daily heating demand, compared to a tighter variation of −7.1 percent to +6.9 percent in residential structures.

The study authors identified several methodological boundaries and practical limitations governing their findings. First, because the selected clustering parameters are strongly correlated along a single thermal axis, weighting principal component axes equally caused thirty-four percent of stations to shift zones. Second, annual average temperature exhibited lower stability across initializations, and hierarchical clustering displayed greater sensitivity to variable composition than K-means, shifting up to forty percent of stations when individual variables were dropped. Third, internal validation indices indicated that four clusters represent an effective grouping rather than a unique statistical optimum. Fourth, applying existing building standard insulation levels served to measure compliance discrepancies under current rules rather than evaluating zone-specific optimal insulation thresholds. Fifth, the simulations evaluated four representative cities and two building types using typical meteorological year data, excluding extreme weather anomalies, airborne salinity effects on coastal glazing, and future climate projections. Consequently, the researchers concluded that while data-driven machine learning provides an objective, scientifically robust methodology for identifying regional climatic realities, their findings serve as a foundational reference for future regulatory revisions rather than direct justification for immediate statutory amendments.

Subject of Research: Data-driven climate zoning methodology and building energy performance evaluation for South Korean building codes

Article Title: Climate-data-driven zoning methodology for building energy code-compliance: A case study of Korean standards

Article References: Choi, B., Kim, H., & Kim, C. (2026). Climate-data-driven zoning methodology for building energy code-compliance: A case study of Korean standards. Results in Engineering, 32, Article 114207. https://doi.org/10.1016/j.rineng.2026.114207

Image Credits: AI Generated

DOI: 10.1016/j.rineng.2026.114207

Keywords: climate zoning, building energy codes, machine learning, K-means clustering, hierarchical clustering, heating degree days, cooling degree days, building energy simulation, EnergyPlus, Energy Conservation Design Standards, zero energy buildings, thermal insulation, building envelope, medium office prototype, high-rise residential, collinearity analysis, Korea Meteorological Administration

एक उत्तर छोड़ें

आपका ईमेल पता प्रकाशित नहीं किया जाएगा। आवश्यक फ़ील्ड को * से चिह्नित किया गया है