Development of a PCA-based climatic similarity index to enhance weather file selection criteria for climate-based daylight modelling simulations in tropical climates Authors: Siew Bee Aw, Pau Chung Leng, Yaik Wah Lim Published: 2026 ABSTRACT Distance-based selection has been the conventional criterion for weather file selections, although this topographically- and meteorologically-agnostic decision may lead to unreliable Climate-Based Daylight Modelling (CBDM) simulation results. To mitigate this risk, this study explores alternative criteria that may be better predictors of weather file proxies for locations within the tropical climate. Principal Component Analysis (PCA) and k-means clustering were conducted on screened tropical climate files to determine the primary attributes and meteorological clusters. Four Principal Components (PCs) explained 89.1% of the variance in weather files. Sensitivity analysis of Malaysian files determined that the four PCs provided a 78% probability of simulating Spatial Daylight Autonomy (sDA) and Useful Daylight Illuminance (UDI) within a 10% tolerance of the Petaling Jaya (PJ) benchmark. Screening for files within weighted PCA distance (PCAd) outperformed pure geometric proximity selections when evaluated at both Malaysian and Class A levels. Overall, the Climatic Similarity Index (CSI) protocol achieved a 86.67% predictive selection accuracy rate across a proportionate random sampling of all Class A TMYx files. This study proves that a combination of atmospheric and geometric feature matching has the potential to overcome the limitations of local weather file availability for a target site.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1038/s41598-026-58112-x. Introduction In recent decades, computational simulations of daylight performance in buildings have emerged as a viable alternative to traditional experimental models and expert intuition 1,2. Virtual prototyping circumvents spatial, financial, and temporal constraints of physical prototypes 2. Accurate simulation inputs are crucial to minimise errors that result in discrepancies between the simulated and actual daylight performance1,3. Generally, daylight simulations require a virtual model, an identified sky type or weather file, a simulation grid, the height of the simulation plane, and an identification of desired daylight metrics. Out of the inputs, the choice of sky type or weather file is the determinant of external daylight parameters that will be imposed on the model4. Weather files emerged as an alternative to static CIE sky types, in part due to a combination of lower computational cost5, a growing awareness of CIE sky shortcomings6, the use of the Perez All-Weather Sky model to complement the Radiance simulation engine, and the formalisation of Climate-Based Daylight Metrics (CBDM) in the IES LM-83-12 standard4. Downloadable weather files and other gridded datasets have become the standard input source for CBDM simulations3,7,8. Daylight simulations consider both the sun and sky as distant light sources, although transient cloud patterns complicate the determination of sky brightness2. Conventionally, users select the weather file within the closest approximate distance from the target site. This proximity-led decision has several drawbacks, given that terrain variation, which affects micro-climatic conditions, can occur within short geographical distances. Micro-climate shifts may occur even within 10 km or 50 km variation9. Weather station scarcity may also impact the accuracy of the simulation results7. These issues pose a problem for daylight simulations, which are sensitive to fluctuations in local sky luminance and cloud cover2,9,10. Past studies have studied the impact of the granularity of weather files11, the implications of using different weather file types4,11, and the impact of different Radiance sky models12. Impacts of weather files have been studied through stratified or purposive sampling of specific cities within one or more Köppen-Geiger climate classifications for energy simulations7,11,13, daylight simulations4,6,11,12, or both3,14. Li et al. (2026) suggested that environmental similarity may be weighted with spatial proximity for improved global solar radiation accuracy15. Another study suggested that mesoscale, high-resolution Weather Research and Forecasting (WRF) data may be superior to Typical Meteorological Year (TMY) files for building energy simulations16. Others even suggested micro-climatic data would be most suitable due to urban heat island and terrain variations3,17,17; however, weather station scarcity and the limited availability of on-site weather data impede such exercises15,18. In a Malaysian study, even intra-state stations demonstrated high degree of error in solar radiation estimations between coastal and inland locations 19. Terrain differences may exhibit variations in precipitative water, cloud patterns, and atmospheric turbidity that increase the risk of inaccurate simulations. Given the impact of spatial variability on daylight simulation results, we hypothesise that attribute-based selection based on commonalities in topo-morphological characteristics should be the primary determinant of weather file selection, rather than simple geographical proximity bias. As daylight-related simulations are sensitive to cloud cover20, an identification of proxy metrics derived from core weather file data was necessary to facilitate its analysis. This serves to improve the applicability of as many weather files as possible for the purpose of this study. Out of these metrics, a combination of clearness index (kT) and diffuse fraction index (kd) have been used as primary complementary indices for sky model generation, including the Perez All-Weather Sky Model that is coupled with Radiance for CBDM simulations21,22. As they are directly derivable from core weather file data, this paper uses these indices as radiometric proxies for cloudiness. It is important to note that kT and kd are proxies for atmospheric transmissivity and cloud cover, indicative of radiative attenuation for CBDM, rather than as a descriptor of luminance distribution or sky type. This study aims to explore the inter-changeability of weather files for daylighting simulations in Class A tropical climates by identifying the degree of micro-climatic representativeness afforded by geographic (latitude and terrain), atmospheric (clearness and diffuse fraction indices), and anthropogenic (urban density) variables. Common CBDM metrics, namely Spatial Daylight Autonomy (sDA), Useful Daylight Illuminance (UDI), and Annual Sunlight Exposure (ASE) were selected for simulation results analyses. Methodology To identify meteorologically-coherent archetypes, unsupervised machine learning techniques consistent with recent protocols in environmental and energy informatics were used23,24. Principal Component Analysis (PCA) was conducted to reduce dimensionality while preserving the global tropical climate variance within the dataset23,24. The selection of primary Principal Components (PCs) was guided by the explained variance24. Transformation and scaling were performed on features with high skewness before being standardised. Loadings were derived by calculating the eigenvectors of the correlation matrix of the standardised features. PCA was conducted for both the overall and unique Class A climates. Variables were considered to be contributors to the loadings based on an absolute coefficient of |λ|≥ 0.40. Upon achieving dimensionality reduction via PCA, k-means clustering was utilised to identify the optimal number of clusters and centroid per cluster. Starting with k = 2, the number of clusters was gradually increased until the silhouette score peaked and inertia stabilisation was achieved. The centroids were considered to be statistically significant representatives of the mean behaviour in their clusters25. This approach preserves the topographical and meteorological nuances of the weather files. Feature creation and weather file selection framework As PCA requires numerical data, clearness index (kT), diffuse fraction index (kd), latitude, elevation, distance of weather station from the coastline, and urban density were identified as PCA features. A stratified strategy based on latitude bands and terrain categories was employed to ensure balanced representation of tropical radiative regimes. Climate datasets and individual weather files were sourced from Climate.OneBuilding (COB), a comprehensive repository for global weather files26. The datasets were screened for stations located within the tropical climate to identify the latitude, longitude, elevation, country, city, and file type. The most common weather file type by quantity was selected for further analysis. The tropical climate is a tropical belt encompassing countries within 23.5˚ north or south of the equator. Based on the Köppen-Geiger classification, this encompasses a range of tropical climates such as tropical rainforests (Af), tropical monsoons (Am), and tropical savanna (Aw or As). The weather files were stratified into three latitude bands, namely |0˚–8˚|, |8˚–16˚|, and |17˚–23.5˚| away from the equatorial line. The present version of the Köppen-Geiger classification map was used to classify the weather files. This was achieved through the Python kgcpy library, which uses the 1981–2010 climatological normal27,28. Atmospheric clarity Candidate files that did not contain DNI, DHI or GHI were excluded from the study. Climate features kT and kd were derived from the weather file data. The formula of kT and Io are as follows, respectively:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${k}_{T}= GHI/{I}_{o}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${I}_{o}=SC \times \left[1 +0.033 \times cos\left(360n/365\right)\right]\times cos\left({\theta}_{z}\right)$$\end{document}where: kT = clearness index, GHI = global horizontal irradiance (W/m2), Io = extra-terrestrial solar radiation (W/m2), θ z = zenith angle. kT is a dimensionless ratio of global solar radiation to extra-terrestrial solar radiation. It is the inverse of the cloud index (Cd), which represents the fraction of extra-terrestrial radiation attenuated by clouds. kd assesses sky structure. It provides a radiometric measure of the amount of solar radiation that is scattered by clouds, aerosols, and water vapour in the atmosphere. For this paper, the Erbs model was used to compute kd, as follows:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${k}_{d}= DHI/GHI$$\end{document}where: kd = diffuse fraction index, DHI = diffuse horizontal irradiance (W/m2), GHI = global horizontal irradiance (W/m2) To ensure data integrity, weather files with a mean kT that do not fall between 0.12 and 0.82 were excluded from the dataset in line with PVsyst documentation recommendations, as they fall outside typical atmospheric attenuation29. The inverse relationship between kT and kd can be described by the Hollands and Huget model for solar energy engineering, which estimates kd based on the monthly average kT30. The selection of a more generic model promotes consistency across a dataset where high-resolution atmospheric data is not uniformly available. Therefore, the dataset was also screened for high kT, high kd scenarios, which may be indicative of poor-quality data. The formula for the relationship for 0.35 < kT < 0.75 is as follows:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${k}_{d}=1.391-3.560{k}_{T}+4.189{k}_{T}^{2}-2.137{k}_{T}^{3}$$\end{document} Terrain categorisation As radiative diffusivity may vary by terrain, the weather files were also categorised accordingly into coastal, inland, or highland stations. In climatology and meteorology, 20 km inland is typically considered the Internal Boundary Layer (IBL), beyond which terrestrial conditions exert stronger influence31. This is also in line with information from the Metrological Department of Malaysia32 and other precedent studies9,10 regarding the inland reach of sea breezes. Thus, weather files that were sourced from sites within 20 km from the coastline were considered to be “coastal”, sites beyond 20 km were considered to be “inland”. The threshold for lifting condensation level (LCL), which indicates the base height of convective clouds, was used to categorise “highland” sites, as the different cloud conditions may affect daylight simulations. Hirsch et al.33 determined 500 m as the threshold, in line with the methodology used by University of Wyoming. However, different regions may have different LCL thresholds, with coastal sites having potentially lower LCL33. Therefore, the mean LCL of all tropical TMYx files was used to define the lower threshold of “highland” sites. The Python MetPy package by the American Meteorological Society, which implements the LCL equation defined by Romps34, was used for this exercise. Python version 3.12 along with the pandas, geopandas, and shapely libraries were used to categorise the weather station locations. A global coastline shapefile was downloaded from Natural Earth Data to identify vectorised coastline boundaries. Using the “nearest neighbour” logic available in the shapely library, the Euclidean distance of each station from the shoreline was calculated using coordinate values to categorise them as “coastal” or “inland”. Highland categorisation was carried out using purely the elevation data from weather files. Airport stations were identified based on explicit nomenclature via the “.AP”, “Airport”, “Intl”, “Aeroport”, “Field”, or “AFB” suffixes in station names. Finally, the Python geemap helper library was used in conjunction with the Google Earth Engine API to determine the built-up surface fraction as a proxy for urban density within a one-kilometre radius from each weather station, in line with the World Meteorological Organisation’s definition of a “Local Scale”35. The most stable release, the 2020 epoch of the Global Human Settlement Layer—Sentinel 2 (GHSL-S2) dataset was accessed via the Google Earth Engine API. Urban density categorisation was based on the Local Climate Zone (LCZ) framework by Stewart & Oke (2012), which considers the thermal admittance of materials and the height, density, and surface cover of urban features36. A Building Surface Fraction (BSF) value—the equivalent of plinth area computation in Malaysia—exceeding 40 is indicative of LCZ 1, 2, or 3, all of which refer to compactly developed LCZs36. Computational simulations Daylight simulations were conducted using a virtual Rhinoceros model and the selected weather files. The Rhinoceros 8 software is compatible with Ladybug and Honeybee plug-ins, which utilise the Radiance prediction engine to execute daylight simulations. The weather file was the sole manipulated variable, while the model, surface material settings, and sensor grid were kept constant. The simulation settings were as follows (Table 1):Table 1Model settings in Rhino and Grasshopper for computational simulations.ItemDescriptionLocationBilik Kuliah 6, 14th Floor, Menara Razak, Universiti Teknologi Malaysia, Kuala Lumpur, Malaysia (3°10′21"N 101°43′09"E)TypologyUniversity classroomOrientation0˚ North (north-facing external apertures)Clear classroom dimensions5.24 m (W) × 6.86 m (D) × 2.86 m (H)Wall finishesLight blue skim coat on partition wallFloor finishesGray carpetCeiling finishesWhite gypsum boardWall reflectance0.465Floor reflectance0.072Ceiling reflectance0.64Glass transmittance value0.3Daylight metricsASE, sDA, and UDISimulation grid density0.5 mSimulation workplane height0.85 mASE, sDA, and UDI simulated results were recorded for sensitivity analysis. Table 1Model settings in Rhino and Grasshopper for computational simulations.ItemDescriptionLocationBilik Kuliah 6, 14th Floor, Menara Razak, Universiti Teknologi Malaysia, Kuala Lumpur, Malaysia (3°10′21"N 101°43′09"E)TypologyUniversity classroomOrientation0˚ North (north-facing external apertures)Clear classroom dimensions5.24 m (W) × 6.86 m (D) × 2.86 m (H)Wall finishesLight blue skim coat on partition wallFloor finishesGray carpetCeiling finishesWhite gypsum boardWall reflectance0.465Floor reflectance0.072Ceiling reflectance0.64Glass transmittance value0.3Daylight metricsASE, sDA, and UDISimulation grid density0.5 mSimulation workplane height0.85 mASE, sDA, and UDI simulated results were recorded for sensitivity analysis. Model settings in Rhino and Grasshopper for computational simulations. ASE, sDA, and UDI simulated results were recorded for sensitivity analysis. Sensitivity analysis A two-stage sensitivity analysis exercise was employed to explore the interchangeability of weather files. The magnitude of change in sDA, ASE, and UDI between the simulated scenarios were assessed relative to the dominant PC determined via PCA. The use of a single virtual model with identical properties ensured that any observed variance in simulated results was directly caused by meteorological discrepancies among the weather files. A case study approach was used to assess the ability of PCA features to predict CBDM performance. The Petaling Jaya (PJ) weather file (WMO index: 486480) from Malaysia contains data from an urban area in the Klang Valley conurbation that includes Kuala Lumpur, Putrajaya, and Selangor. Klang Valley files were anticipated to be most frequently used on the basis of geographical proximity, as the majority of the Malaysian buildings that have green building ratings, and thus a higher probability of environmental simulations, are located in the Klang Valley. The first stage analyses intra-cluster performance of one selected country, Malaysia. The entire country is broadly categorised under the Af climate type. Malaysia has three distinct regions formed through geographical boundaries, namely East Peninsular Malaysia, West Peninsular Malaysia, and East Malaysia. A preliminary observation of weather station locations suggested an uneven distribution of stations across the country. All weather files were classified accordingly, in addition to their terrain and Köppen-Geiger climate categorisations. The sensitivity analysis was conducted using continuous regression plots to compare the predictive power of CBDM metrics between geographic proximity and the PCA-driven selection. The Great-Circle Distance via the Haversine formula was used to calculate distance, to account for the spherical geometry of the Earth. The Haversine formula is as follows:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$a= {sin}^{2}\left(\frac{\Delta lat}{2}\right)+cos\left({lat}_{1}\right)\times cos\left({lat}_{2}\right)\times {sin}^{2}\left(\frac{\Delta lon}{2}\right)$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d=2\times R\times arctan2\left(\sqrt{a},\sqrt{1}-a\right)$$\end{document}where: lat = latitude (°), lon = longitude (°), R = radius of the Earth (°) The simulation results were also assessed to explore the relationship between PCA distance (PCAd) and CBDM simulation biases. A combined Euclidean bias of all CBDMs was considered acceptable if it was within 10% of the benchmarked file’s results. The findings from Malaysia were used to develop a weather file selection protocol for tropical climates if the reliability of geographic proximity-based methods was successfully disproved. To ensure universal transferability of the proposed protocol across the global Class A climate, an out-of-sample cross-validation was constructed using dynamic reference benchmarks. A proportional stratified sampling approach was adopted based on the distribution of the four largest Köppen-Geiger Class A representatives, namely Aw, Am, Af, and As. Based on the tropical climate dataset, the ‘sample’ function of the pandas module in Python was used to select one weather file as the new benchmark. Using thresholds preliminarily identified in the Malaysian case study, the module was rerun until at least two files from the same Köppen-Geiger class meet the protocol thresholds. The weighted PCA distance and combined Euclidean CBDM bias of all candidate files in the sample set were calculated relative to the benchmark file. Intra-cluster Ordinary Least Squares (OLS) regression was used to assess the applicability of the proposed protocol. The R2 value was assessed alongside t-statistic, p value, and Coefficient of Variation of Root Mean Square Error (CVRMSE) to determine the validity of the proposed protocol. Results and discussion Weather file analysis At the time of writing, the COB master list of weather files contained 36,380 entries, of which 10,885 were located within the torrid zone. The weather files available for the torrid zone were predominantly TMYx files, which represent 83.75% of the tropical climate files available in the COB database. Files were discarded for the following reasons: 33,019 files for being either non-TMYx or redundant time-ranged TMYx files; 50 files for missing solar data; 191 files for having a mean kT ≥ 0.82; 5 files for the inability to compute kd based on the TMYx file contents. No files contained improbable kT and kd value pairs that were too high or too low. Terrain categorisation revealed that the 3115 TMYx files topographically comprised 1173 coastal, 1119 inland, and 823 highland stations. These files were then categorised by urban density. The data resulted in a right-skewed histogram, suggesting that the majority of the weather stations were located in areas with low urban density (LCZ < 0.4). 41.89% of the TMYx files were sourced from airports. Furthermore, 44.16% of these airports fell under the coastal terrain category. Coastal airport stations demonstrate a higher degree of meteorological homogeneity (SDPC1 = 1.3469) compared to inland airport stations (SDPC1 = 1.5556). 1936 files represented Class A climates, suggesting a high weather station density for this climate class. Despite representing only 19% of the Earth’s land area, Class A TMYx files comprised 62.15% of the torrid zone TMYx files 28. Conversely, Class B climates covered 30.2% of the Earth’s land area 28, but represented only 17.34% of the available torrid zone TMYx files. Aw was the dominant Class A climate type, covering 11.5% of the global land area 28. It was represented by 36.98% of the torrid zone TMYx files. Am and Af climates had 363 and 330 TMYx files, respectively, for 7.5% of the Earth’s land area. PCA was executed for the entire Class A climate (refer Table 2). Generally, Class A climates exhibited a high degree of cloud scattering, with elevation (+ 0.51), distance to coast (+ 0.52), and kd (+ 0.46) dominating PC1. PC2 differentiated inland from coastal files. By contrast, urban density was decoupled onto PC3.Table 2Principal Component (PC) features and explained variances of the overall tropical climate and four Class A Köppen-Geiger climate types, where N = number of stations, kT = clearness index, kd = diffuse index, UD = urban density, Ele = elevation, DC = distance from coast, and Lat = absolute latitude.FeatureClass A (N = 1,936)Aw(N = 1,152)Am(N = 363)Af(N = 330)As(N = 91)PC1 (%)kd, Ele, and DC (32.3%)kT, kd, and Lat (28.5%)kd and Lat (33.0%)kT, kd, and Lat (39.1%)kd, Ele, and DC (36.3%)PC2 (%)kT, kd, Ele, and DC (27.8%)Ele & DC (26.5%)kT, DC, and UD (21.9%)UD and Ele (17.8%)UD and Lat (22.7%)PC3 (%)UD (16.7%)UD (16.4%)UD and kT (15.2%)UD, Ele, and DC (16.1%)kT and UD (14.1%)PC4 (%)Lat and kT (12.3%)Lat (11.7%)Ele and Lat (13.1%)Ele and DC (13.0%)UD and Lat (13.1%)Explained Variance89.1%83.1%83.2%86.0%86.2% Table 2Principal Component (PC) features and explained variances of the overall tropical climate and four Class A Köppen-Geiger climate types, where N = number of stations, kT = clearness index, kd = diffuse index, UD = urban density, Ele = elevation, DC = distance from coast, and Lat = absolute latitude.FeatureClass A (N = 1,936)Aw(N = 1,152)Am(N = 363)Af(N = 330)As(N = 91)PC1 (%)kd, Ele, and DC (32.3%)kT, kd, and Lat (28.5%)kd and Lat (33.0%)kT, kd, and Lat (39.1%)kd, Ele, and DC (36.3%)PC2 (%)kT, kd, Ele, and DC (27.8%)Ele & DC (26.5%)kT, DC, and UD (21.9%)UD and Ele (17.8%)UD and Lat (22.7%)PC3 (%)UD (16.7%)UD (16.4%)UD and kT (15.2%)UD, Ele, and DC (16.1%)kT and UD (14.1%)PC4 (%)Lat and kT (12.3%)Lat (11.7%)Ele and Lat (13.1%)Ele and DC (13.0%)UD and Lat (13.1%)Explained Variance89.1%83.1%83.2%86.0%86.2% Principal Component (PC) features and explained variances of the overall tropical climate and four Class A Köppen-Geiger climate types, where N = number of stations, kT = clearness index, kd = diffuse index, UD = urban density, Ele = elevation, DC = distance from coast, and Lat = absolute latitude. The individual climate types could be loosely grouped into humid equatorial (Af and Am) and seasonal tropics (Aw and As). Atmospheric conditions and latitude dominate PC1. kd loading was strongly positive between + 0.561 to + 0.565, while latitude was strongly negative between − 0.507 and − 0.508. Am climates (kT = − 0.384) were cloudier than Af climates (kT = − 0.438), indicating a higher degree of cloud-dome scattering for sites closer to the equator. Aw and As climates were more discrete. In Aw, PC1 was purely atmospheric and latitudinal, with insignificant geographic terrain correlation (elevation = + 0.038 and distance to coast = − 0.057). Instead, geographical features were driven through PC2, where elevation (+ 0.661) and distance from coast (+ 0.656) explained 26.5% of the variance. In As, the inverse held true: PC1 was dominated by geographical features, while atmospheric features were mostly driven by PC3 (kT = + 0.799). The reliability of geographic labels for weather station categorisation was explored through a confusion matrix (refer Fig. 1), which revealed significant deviations between the k-means clusters and static geographic labels. The coastal geographical label exhibited the highest agreement, with consensus achieved for 82.72% of the actual coastal stations. These findings validated the viability of the 20 km Internal Boundary Layer (IBL) delineation from the coastline, as identified by Hsu 31.Fig. 1A comparison between geographically-informed and k-means assigned terrain categorisation. A comparison between geographically-informed and k-means assigned terrain categorisation. However, significant divergence was observed within land-based categories. The geographical highland label utilised a threshold of 500 m. However, the k-means algorithm, which considered atmospheric conditions in addition to geographic data, resulted in an 82.56% overlap between the geographically-labelled inland and highland sites. An analysis of the k-means cluster centroids was executed to identify more appropriate labels using meteorological characteristics. K-means clustering grouped the dataset by similarity to assess interchangeability between weather files (refer Fig. 2). k = 3 was determined to be the optimum number of clusters that explained 43.94% of the variance, with a silhouette score of 0.2746. The centroid for the coastal lowland cluster exhibited a clear kT value of 0.64, a low kd value of 0.36, a low elevation (21.9 m), a short distance to the coastline (10.9 km), and a high LCL (543.0 m). The low elevation resulted in the lowest absolute cloud altitude among the clusters (564.9 m). The proximity of these weather stations to large water bodies with limited topographical obstructions revealed a distinct micro-climate compared to weather stations further inland.Fig. 2PC1 (geographic factors) and PC2 (atmospheric factors) revealed centroids that diverged from pure geographic labels for Class A climate weather stations. PC1 (geographic factors) and PC2 (atmospheric factors) revealed centroids that diverged from pure geographic labels for Class A climate weather stations. An orographic transition zone between coastal lowland areas and highlands could be found at an average altitude of 242.4 m. The cluster centroid was located 249.9 km from the coastline, where kT decreased to 0.51, while kd increased to 0.47. Diffuse solar radiation was more prevalent, given the presence of clouds and other atmospheric particles. The LCL of the centroid was identified for a post-hoc analysis. At an average LCL height of 450.2 m, it was close to the mean elevation of the clear highland cluster (482.8 m). This potentially explains the drawbacks of relying purely on geographical thresholds for weather station categorisation. It also suggests that the transition cluster occurs close to foothills or areas with similarly higher terrain. Finally, a clear highland cluster was characterised by both high terrain elevation and LCL (482.8 m and 831.7 m, respectively). This may be attributed to the mass elevation effect, which refers to the thermodynamic effects of mountain masses [Schroeter (1908) as cited in 33]. The heat absorbed by high-altitude ground planes decreases relative humidity and increases dew-point compression, resulting in an absolute cloud altitude that is nearly double that of the transition zone centroid. Highly elevated sites experience less orographic cloud shading and narrower diffuse sky components than transition zone sites (kT = 0.63 and kd = 0.38). This exercise refined weather file categorisations by integrating geographical nomenclature with atmospheric characteristics. Sensitivity analysis An intra-cluster sensitivity analysis was carried out on Malaysia. It is designated as Af under the Köppen-Geiger climate classification system. As of 2026, Malaysia has 105 publicly-available weather files representing 28 unique locations. Upon removal of 25 custom files and 52 time-ranged TMYx files, 28 TMYx files were available for analysis. Peninsular Malaysia is represented by 18 TMYx files across nine states; Perlis, Negeri Sembilan, as well as the Federal Territories of Kuala Lumpur and Putrajaya have none. East Malaysia, located across the South China Sea, has 10 TMYx files. File metadata revealed that the elevations of the weather stations ranged between 3.4 and 1545 m above mean sea level (AMSL), with the Cameron Highlands, Pahang station being an outlier. The next-highest weather station is in Petaling Jaya, Selangor, at 47 m AMSL. 82.14% of the TMYx file locations in Malaysia were sourced from airport weather stations. Notably, the PCA-derived clusters also identified discord between geographical and meteorological categorisations in line with the k-means algorithm, despite the homogeneity among station locations. Topographically, there are 20 coastal files, seven inland files, and one highland file. Only 45% of the coastal files (Langkawi, RMAF Butterworth, Kota Bharu, Labuan, Kota Kinabalu, Kudat, Tawau, Penang International Airport, and Sandakan) indicated geographical-meteorological alignment, while the sole highland file from Cameron Highlands was meteorologically inland. All inland files were aligned. The 28 Malaysian TMYx files were used for CBDM simulations. sDA and UDI were analysed while ASE was excluded as all files except Sitiawan, Perak resulted in ASE = 0. Overall, there are four climatically inland stations and one coastal station in East Peninsular Malaysia; 10 inland and two coastal stations in West Peninsular Malaysia; as well as six inland and five coastal stations in East Malaysia, namely Sabah and Sarawak. A one-way ANOVA for sDA returned a marginally significant result (F = 2.73, p = 0.054), whereas UDI distributions were more stable (F = 4.33, p = 0.009). This may be attributed to the high internal heterogeneity between the zoned files. Furthermore, UDI filters for a wider range of illuminance (100–2000 lx) compared to sDA (300 lx), allowing it to consider a larger compliance coverage. Using the PJ file as a benchmark, PCA-driven file selections were found to be more robust than pure geographic proximity for both sDA and UDI (refer Fig. 3). Geographic proximity demonstrated negligible predictive ability (R2 = 0.0196) for sDA variance. Information on sDA bias risk is challenging to gauge from proximity, despite the larger bias magnitude illustrated in Fig. 4. On the other hand, the PCA method was nearly six times better at predicting sDA performance drift. Likewise, PCA-driven UDI explained UDI performance bias nearly 2.3 times better than geographic proximity alone, predicting up to 35.8% of the variance.Fig. 3PCA-driven file selection had stronger predictive power of CBDM performance bias compared to pure geographic proximity for Malaysia (Af climate).Fig. 4sDA and UDI relative bias, benchmarked against PJ, highlights that PCA features play a larger impact on CBDM reliability compared to geographical proximity. PCA-driven file selection had stronger predictive power of CBDM performance bias compared to pure geographic proximity for Malaysia (Af climate). sDA and UDI relative bias, benchmarked against PJ, highlights that PCA features play a larger impact on CBDM reliability compared to geographical proximity. Development of a PCA-based selection protocol The proposed “Climate Sensitivity Index” (CSI) protocol comprises a suite of tiered assessments for screening weather files based on PC features. The preliminary CSI protocol was developed using Malaysia (Af class) before being validated across the tropical climate zone. The PJ file was maintained as the benchmark, as it represented an urban site with limited equivalent weather files in Malaysia. This section explores the possibility of identifying its closest weather file proxy within the same Köppen-Geiger class. The raw data required for screening are kT, kd, elevation (in metres), distance from coast (in kilometres), urban density (calculated using the LCZ framework), and latitude. As a preparatory step, log transformations were executed on elevation and distance from coast, as these may contain large values that may skew the data. The function log(x + 1) was implemented for urban density to account for potential inputs of zero, while log(x + 2) was used for elevation to address low-lying sites located below mean sea level. Global PCA distance (PCAd) is proposed as the screening layer under the CSI protocol. Only four PCs, explaining 89.1% variance in tropical climate weather files, were used. PC1 to PC4 coefficients derived from the PCA exercise has been attached as Supplementary Information (Table B). The PCAd of all Malaysian files was calculated relative to PJ then analysed in conjunction with the relative biases of sDA and UDI. Figure 4 illustrates the PCAd of all Malaysian weather files relative to PJ. PCA identified Subang, Selangor as the closest “environmental twin” to Petaling Jaya, Selangor (PCAd = 0.32). Subang is geographically and climatically similar to PJ, although its urban density is approximately half. These findings also disprove the conventional distance-based file selection approach, as Malacca is approximately 115 km away from PJ, compared to Subang (approximately 12 km), KLIA (approximately 40 km) or Mersing (approximately 254 km). The nature of the CBDM metric appears to influence the concordance between climatic similarity and CBDM results. Generally, shorter PCAd significantly reduced the risk of CBDM simulation bias. The PCAd provided an 78% probability of simulating sDA and UDI within 10% tolerance of the PJ benchmark when PCAd < 0.7. Primary screening via PCAd < 0.7 resulted in nine candidate weather files. Besides Subang, this included Ipoh, Perak (PCAd = 0.33), KLIA, Selangor (PCAd = 0.51), Senai, Johor (PCAd = 0.55) and Malacca (PCAd = 0.56). Ipoh and Temerloh, Pahang were anomalies. Their PCAd are relatively close to PJ, but their CBDM relative biases exceeded the ± 10% tolerance. Conversely, several East Malaysian TMYx files, namely Sri Aman and Sibu, Sarawak, as well as Tawau, Sabah, achieved distant CBDM agreement despite having PCAd values exceeding 1.2. Mersing, Sarawak also stood out remains an anomaly, possibly due to the binary nature of sDA, the wide range of UDI, or a feature that was beyond the scope of this study. However, it is worth noting that both Malacca and Mersing candidate files were within 5% bias for CBDM metrics. Tightening the PCAd threshold to 0.4 in the Malaysian context yielded the Subang and Ipoh files. Next, the PCAd thresholds for the CSI protocol were verified using proportional stratified sampling across all Class A climates to explore the possibility of broadening the weather file search globally throughout its Köppen-Geiger classification zone. PCAd < 0.4 yielded random sample sets between four and 18 files, including a benchmark file. 105 TMYx files, comprising 13 randomly-identified benchmarks and 92 candidate files, were randomly selected, used for simulations, and compared (refer Fig. 5). According to the k-means clusters identified in Fig. 2, 44 files belong to the Clear Highland cluster, 41 files to the Coastal Lowland cluster, and 20 to the Orographic Transition Zone cluster. The utilisation of PCAd demonstrated significant structural fidelity to the k-means cluster framework; each randomly-generated sample set exhibited complete cluster homogeneity, containing exclusively coastal lowland, orographic transition, or clear highland files.Fig. 5Analysis of CBDM relative biases by k-means cluster and weighted PCA distance. Analysis of CBDM relative biases by k-means cluster and weighted PCA distance. The CSI protocol achieved an overall accuracy rate of 86.67% across Class A climate files with a 10% CBDM simulation bias threshold. Across the sample set, the protocol achieved 100% success for Am and As climates; 96.61% for Aw climates; and 93.75% for Af climates. The relationship between PCA distance and CBDM simulation bias was relatively stable across Class A climates. R2 ranged between 18.1 and 23.4%, proving the universality of the protocol, with a t-statistic of 5.109 and a p-value of 1.49 × 10–6. The t-statistic and p value prove the validity of the model despite the low R2, especially in consideration of the micro-climatic noise present in dynamic sky conditions. For all clusters, the mean Euclidean CBDM bias ranged between 4.65 and 5.45%, which is below the 10% acceptable tolerance level. The RMSE and mean are both small (3.39% and 5.03%, respectively), suggesting that proxy selection errors are low. CVRMSE is therefore mathematically inflated to 67.37%. However, the highly-significant p value and small RMSE within the 10% tolerance validates the viability of the proposed CSI protocol as a weather file selection methodology for Class A climates. Comparatively, the R2 based on geographic proximity to benchmark files is negligible for all clusters (R2 = 0–10%). The proposed CSI protocol procedure is provided as S1 in the supplementary file. Conclusion Based on regional and tropical climate-level analyses of weather file implications on CBDM outcomes, geographic proximity that is topographically- and meteorologically-agnostic is an inadequate metric for weather file suitability. By identifying that a four-component PCA captures up to 89.1% of the critical drivers of naturally-daylit indoor environments, this study proposes a standardised protocol for selecting “environmental twin” tropical climate files for more reliable simulation accuracy regardless of geographical boundaries. In countries like Malaysia, whose weather files cover only a limited range of meteorological and topographical conditions relative to the country’s terrain, the CSI protocol provides a wider range of potential weather files that may be a better “environmental twin” to actual site conditions. Weather file selection for reliable CBDM simulations is not necessarily constrained to country borders, as evinced by the global Class A validation. CSI-based weather file selection could overcome the limitation of geographic and meteorological dissimilarity between the actual site and weather stations. It is worth noting that this study was conducted specifically for CBDM simulations, and not for thermal or Computational Fluid Dynamics (CFD), which involve more macro-climatic features. Furthermore, we recommend that the PCA weights be updated at five- or ten-year intervals as updated earth observation layers become available, in order to keep the protocol relevant as tropical cities continue to develop. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1