Generating Future Weather Data for Building Energy Simulations: A Review of Methods, Applications and Challenges

Abstract

With an increasing awareness of climate change and its effects on the built environment, climate change adaptation is changing traditional building design practices. Future weather data are essential for building energy simulation (BES) that informs a resilient and energy-efficient building design under climate change. While general circulation models (GCMs) provide future climate predictions, their outputs often require downscaling to improve spatial and temporal resolution and further methodological processing to generate weather data suitable for building-scale analysis. This study aims to examine the methods for generating and utilizing future weather data for BES, with a particular focus on bias correction and uncertainty quantification in GCM predictions. This study summarizes the prevailing methods for bias correction of GCM outputs and the generation of representative future weather data. The characterization of GCM uncertainty and its implications for BES results are discussed. It is shown that GCM outputs can be effectively used for BES to evaluate long-term effects of climate change under various climate scenarios, and the most cost-effective approach often involves a combination of statistical downscaling and adjustment of grid cell size, which balances the need for high-resolution, site-specific weather data with the demand for computational resources. In addition, key challenges are identified, including the selection of appropriate GCMs and climate scenarios, the trade-off between computational cost and representativeness, and the need to include both extreme and typical weather conditions. Furthermore, future research prospects are proposed. Through a synthesis of current advancements in future weather data generation methods, this study contributes to the robustness of climate-responsive building design.

1. Introduction

Buildings worldwide account for a substantial share of energy consumption and greenhouse gas emissions [1,2]. Such substantial impact has made the built environment an important focus for climate change mitigation and adaptation studies [3]. Within this context, building energy simulation (BES) is considered an effective approach for predicting future building energy demand [4], mainly because of its ability to model complex physical processes and the interactions between the building envelope, systems and dynamic external conditions [5]. Furthermore, BES is capable of assessing building performance under future scenarios, enabling the evaluation of long-term building energy demand [6].

There is a growing emphasis on improving building energy efficiency and maintaining comfortable indoor thermal conditions under climate change (e.g., [7,8]). For example, an analysis of climate change adaptation measures for residential buildings in the Netherlands demonstrated that exterior solar shading and additional natural ventilation were the most effective for reducing indoor overheating [9]. Similarly, several studies employed BES to demonstrate effective strategies for improving building energy efficiency under climate change (e.g., [4,8]). Consequently, BES has become an essential approach for mitigating the adverse effects of climate change on buildings.

However, the adoption of BES to mitigate the adverse effects of climate change presents a challenge due to the inherent uncertainty in future climate predictions and the limitations of downscaling future climate predictions to a resolution suitable for building-scale analysis. Climate change is characterized not only by rising ambient air temperatures but also by altered patterns of extremes. More extreme weather events, such as the increased frequency and intensity of heatwaves, are changing environmental conditions [10]. The changes in environmental conditions can lead to an increase in building cooling energy demand [11] and a higher risk of indoor overheating [12]. Key sources of uncertainty in future climate predictions include the socioeconomic uncertainty (e.g., future emissions pathways), natural climate variability, and model structure differences [13]. Depending on the methods selected, downscaling future climate predictions may fail to capture local microclimates and site-specific extreme weather conditions [14]. The cumulative effect of these factors may affect long-term predictions of future weather conditions, thereby compromising the outcomes of BES and climate change adaptation in the built environment.

A range of solutions are found in the literature to address uncertainty in climate predictions and the resolution limitations of downscaling for BES applications. To avoid reliance on a single model, a multi-model ensemble combines predictions from different GCMs [15]. The variance across the ensemble members quantifies the uncertainty and illustrates the range of possible future climates, enabling BES to be performed under a variety of possible climate conditions (e.g., [15]). This approach directly supports risk-informed design by identifying adaptation measures that perform effectively across a range of possible future climate conditions, thereby ensuring long-term building resilience and helping designers test the robustness of building adaptation strategies. Statistical downscaling using morphing employs techniques such as shifting, stretching or their combination on monthly predictions to generate hourly time series based on historical observations [16]. Spatial interpolation addresses the scale mismatch between coarse model outputs (e.g., from GCMs) and building-scale analysis. This process estimates conditions at unsampled locations from known surrounding points (e.g., [17]), thereby creating high-resolution climate data from coarse grids. These spatially downscaled data are essential for incorporating the site-specific weather conditions that are critical to BES.

In the context of climate change adaptation and BES, the generation of future weather data is essential for assessing long-term climate change impacts, such as building indoor overheating [18] and energy demand [11]. Furthermore, the systematic errors (biases) in climate predictions and the coarse resolution of model outputs necessitate the application of bias correction and downscaling methods. Hence, generating future weather data involves a critical chain of steps to ensure that the resulting future weather data are suitable for BES under climate change (e.g., from selecting GCMs and emissions scenarios to applying downscaling and bias correction).

BES typically uses weather data from a historical typical meteorological year (TMY), assuming a relatively stationary climate condition. The TMY is a structured weather dataset, often at hourly resolution for a specific location, designed for simulating and assessing building energy performance [19]. The TMY is a synthesized composite of individual representative months selected from a longer climatic record (typically 10–30 years [20]). The selection process identifies the month from the climatic record that is most “typical” for that period, based on key meteorological parameters such as solar radiation, dry bulb temperature, dew point temperature and wind speed [21]. By assembling these 12 representative months, the TMY aims to preserve the realistic temporal sequence of real weather while representing the essential characteristics of long-term average conditions [22]. For example, an analysis of energy consumption for office buildings in five Chinese cities showed that the monthly energy consumption predicted using BES under a historical TMY closely aligned with the long-term mean [23]. Moreover, a study of an office building in Hong Kong using BES demonstrated that although energy consumption predictions using historical TMY data agreed well with the long-term mean annually, the predictions showed larger monthly deviations [24].

By using the historical TMY, these examples illustrate the dual achievement of reduced calculation time and the inclusion of representative weather conditions (e.g., [25]). However, a major limitation of using historical weather data from the TMY is its core assumption of climate stationarity. As it reflects past climate norms, it fails to capture the evolving trends and often underestimates the intensity of extremes associated with climate change. Changes in both average weather conditions and extremes due to climate change are leading to more frequent and intense extreme weather events. This has been documented in several locations worldwide, including heatwaves in China [26], Australia [27] and Europe [28].

Thus, given the inherent uncertainty in generating future weather predictions and the limitations of downscaling methods regarding spatiotemporal resolution, BES under climate change requires future weather data that represent not only changing average conditions but also extremes.

Previous studies have reviewed the effects of climate change on building energy demand and indoor thermal conditions [29,30,31,32,33]. Moreover, a number of earlier studies have summarized effective passive and active solutions for buildings under climate change and highlighted strategies such as natural ventilation and insulation [7,34,35]. In addition, the methods to create weather data for building simulation were reviewed and analyzed [36,37]. For the downscaling methods and uncertainty analysis, methods for downscaling future weather data and propagating their associated uncertainties into building energy demand were reviewed [38]. The aims, key focuses, and implications of the aforementioned reviews are summarized in

Table 1. Overview of selected studies reviewing building energy demand and indoor thermal conditions under climate change.

No.	Author (Year)	Aim	Key Focus	Implications for Building Energy Simulation (BES)	Bias Correction	Uncertainty Analysis *
1	Li et al. (2012) [32]	To review studies on the impact of climate change on building energy use in different climate zones	Climate zone differences Degree-days method, and building energy simulation technique Mitigation and adaptation measures	Evaluate mitigation effects and overheating risks Update weather data to include future climate change Interactions between external climate, building envelope and heating, ventilation and air conditioning (HVAC) systems	Not included	Input
2	Yau & Hasbi (2013) [29]	To review the effects of climate change on commercial buildings and their technical services in the tropics	Contribution of buildings to climate change Climate change impacts on HVAC systems, building energy consumption, and carbon emissions	Include future climate scenarios in BES Importance of climate variability and occupant behavior	Not included	Input
3	Barbosa et al. (2016) [35]	To investigate vulnerability factors affecting thermal comfort in residential buildings under climate change and adaptive strategies	A vulnerability framework for thermal comfort in residential buildings Strategies to increase adaptive capacity in buildings	Need to downscale climate model outputs The identified vulnerability factors for BES Inclusion of climate projections and uncertainty in building codes	Not included	Input, methodology
4	Herrera et al. (2017) [36]	To assess the methods for creating weather variables for use in building simulation	Typical and extreme weather conditions Synthetic and future weather generation Requirement list for weather files	Inclusion of microclimate and extremes Limitations of using observed data	Not included	Input, methodology
5	Yassaghi & Hoque (2019) [38]	To review methods to develop high-resolution weather files and the climate change impact on building energy performance	Building energy performance assessment Climate change adaptation strategies Uncertainty analysis	Current and future weather files Use of downscaling methods Uncertainty propagation	Not included	Input, methodology
6	Bazazzadeh et al. (2021) [31]	To study the effects of climate change on building energy consumption	Comparison of climate scenarios Effect of climate change on HVAC systems and energy demand	Increased cooling load and decreased heating load Use of the degree-days method	Not included	Input
7	Li et al. (2021) [30]	To summarize studies on the effects of climate change on building energy consumption and the methods adopted	Estimate climate change impacts on building energy consumption Spatiotemporal changes in degree days and influencing factors Future directions	Compare physical model method and degree days method Effects of urbanization on degree days	Not included	Not included
8	Campagna & Fiorito (2022) [33]	To map the impacts of climate change on building energy consumption	Change in building energy demand under future climate scenarios Perform a meta-analysis to derive effect sizes	Choice of downscaling methods High heterogeneity of data	Included	Input, methodology
9	Kutty et al. (2024) [7]	To document building energy performance and adaptation measures in the Middle East Gulf states under climate change	Geographic contexts Climate change effects on building energy use and thermal performance Mitigation and adaptation measures	Use of downscaling methods for weather data Effect of building typology and adaptation measures	Included	Input, methodology
10	Tajuddeen & Sajjadian (2024) [34]	To evaluate passive and active solutions for climate change mitigation and adaptation	Assessment of passive strategies Assessment of active strategies	BES can be used for performance evaluation Combination of BES and optimization techniques	Not included	Input
11	Ren (2025) [37]	To provide an overview of the development and use of weather files in building performance simulations	Typical and extreme weather conditions Methods for generating future weather data Methods for downscaling climate model outputs	Typical, extreme, and future weather files Inclusion of urban microclimate	Included	Input, methodology

*: Uncertainty in weather file generation, including input (e.g., general circulation models, emissions scenarios), methodology (e.g., downscaling methods).

.

There are few studies that review the methods for generating and applying future weather data specifically for BES. To the best of our knowledge, methods for uncertainty quantification and bias correction of future weather data have not been sufficiently investigated for BES applications. The novelty of this study lies in its evaluation of methods for generating future weather data and the subsequent development of a methodological framework that provides guidance on the selection and application of these methods for BES under climate change. The comparison of the previous reviews and the present study is shown in

Table 2. Comparative analysis of the previous reviews and the present study.

Aspect	Previous Reviews	The Present Study	What Is New
Focus on bias correction	Lack of detailed bias correction analysis	Specific analysis on bias correction methods, including their limitations	Integration of bias correction into a methodological framework for BES under climate change
Uncertainty analysis	Few reviews provided a structured uncertainty classification.	Explicit uncertainty classification	Uncertainty classification and propagation pathways
Methodological framework for weather file generation	Lack of a methodological framework for generating future weather files	A methodological framework integrating downscaling, bias correction and uncertainty propagation	A methodological framework for method selection in generating future weather files
Trade-off analysis of cost-effectiveness	Insufficient detail in cost-effectiveness trade-off analysis	A trade-off analysis between computational cost and representativeness	Discussion of the effect of method choice on BES

.

This study thoroughly examines the methods for generating future weather data for BES under climate change, aiming to develop a methodological framework from the review findings to provide practical guidance on the selection and application of these methods. Accordingly, previous research is evaluated and synthesized around the following three objectives to achieve the aim:

(1)

To assess the effectiveness of methods employed in the generation and application of future weather data for BES.

(2)

To provide guidance on the selection of methods for generating future weather data.

(3)

To identify knowledge gaps and outline paths for future studies.

The contributions of this study are threefold: (1) it provides a thorough analysis of the methods employed to generate future weather data for BES; (2) it develops a methodological framework to inform the decision-making on method selection for generating future weather data; (3) it highlights potential pathways to improve the generation of future weather data for BES, especially by quantifying uncertainty and balancing computational cost with representativeness.

This paper is structured according to the sequence of future weather file generation: the prediction of future climate is examined (based on GCMs and emissions scenarios, Section 2). Subsequently, the improvement of data resolution (spatiotemporal downscaling, Section 3) and the reduction of the systematic errors (bias correction, Section 4) are presented, followed by spatial scale matching (Section 5), uncertainty assessment (Section 6), and microclimate and extreme weather conditions (Section 7). Finally, the implications of these sequential choices for BES are synthesized (Section 8), and the conclusions are summarized (Section 9). The structural outline of this paper is illustrated is Figure 1.

2. Climate Predictions Based on General Circulation Models

The prediction of building future energy demand requires the input weather data that include the effect of possible future weather conditions. Moreover, the growing need for climate change impact assessment, adaptation and mitigation necessitates the development of tools and assessment frameworks. Different scenario frameworks were developed to support successive Intergovernmental Panel on Climate Change (IPCC) assessment reports. The developed scenario frameworks, their key characteristics and the corresponding IPCC assessment reports are summarized in

Table 3. Scenario frameworks developed to support Intergovernmental Panel on Climate Change (IPCC) assessment reports.

Scenario Framework	Representative Scenario	Key Characteristic	IPCC Report
Early emissions scenarios	Scenario A (Business as Usual), Scenario B, Scenario C, and Scenario D [39]	Four main emissions scenarios were developed [39].	IPCC First Assessment Report, 1990 [39]
IPCC 1992 emissions scenarios (IS92)	IS92a-f [40]	Based on different socioeconomic assumptions [40]	IPCC Second Assessment Report (SAR), 1995 [41]
IPCC Special Report on Emissions Scenarios (SRES)	Six scenario groups: A1FI, A1B, A1T, A2, B1, B2 [42]	Four scenario families describing alternative socioeconomic development pathways [42]	IPCC Third Assessment Report (TAR), 2001 [43] and IPCC Fourth Assessment Report (AR4), 2007 [44]
Representative Concentration Pathways (RCPs)	Four pathways: RCP2.6, RCP4.5, RCP6.0 and RCP8.5 [45,46]	Radiative forcing-based scenarios consistent with different socioeconomic storylines [45]	IPCC Fifth Assessment Report (AR5), 2013 [47]
Shared Socioeconomic Pathways (SSPs)	Five SSP narratives: SSP1, SSP2, SSP3, SSP4 and SSP5 [48]	A scenario matrix architecture combining SSPs with climate forcing levels (RCPs) [49]	IPCC Sixth Assessment Report (AR6), 2021 [50]

.

GCMs that participated in the Coupled Model Intercomparison Project (CMIP) are employed as the effective tools for generating the future weather data under different emissions scenarios [51]. Although GCMs are available for different time periods, they often have a coarse spatial resolution and thus are less representative of local climate conditions (e.g., [25]). In addition, the outputs from GCMs typically show a coarse temporal resolution for some weather variables, e.g., monthly or daily instead of hourly resolution, missing the required finer temporal resolution for BES on an hourly basis (e.g., [52]). The CMIP was launched in the mid-1990s as an initiative by the World Climate Research Programme Working Group on Coupled Models to standardize and compare climate model simulations from different institutions worldwide [53,54]. CMIP emerged from the proven framework of the Atmospheric Model Intercomparison Project and from the need to employ coupled models to provide more robust statements about anthropogenic climate change [55]. The historical development of different CMIP phases, their key innovations and the scenario frameworks they adopted are summarized in

Table 4. Development of the Coupled Model Intercomparison Project (CMIP) phases.

CMIP Phase	Start Year	Key Innovations	Associated IPCC Report	Scenario Framework
CMIP1	1995	Established standardized framework for comparing coupled ocean-atmosphere general circulation models [53]. Provided open data-sharing for model outputs [56].	Initiated after IPCC SAR [41]	Constant forcing: climate forcing under constant conditions [53]
CMIP2	1997	Introduced transient climate simulations (a 1% per year increase in carbon dioxide) [54]. Provided a consistent benchmark for model intercomparison [57].	IPCC TAR [43]	Perturbed: a 1% per year atmospheric carbon dioxide increase [53,54]
CMIP3	2003	Provided a foundation for the IPCC AR4 multi-model ensemble [58]. Coordinated historical and future climate simulations [58].	IPCC AR4 [44]	SRES scenarios: SRES A2, SRES A1B, SRES B1 [42]
CMIP5	2008	Generally higher-spatial-resolution models and expanded output fields [59] Coordinated historical, decadal (near term) and future experiments [59]. Earth system models including biogeochemical components [59]	IPCC AR5 [47]	RCP scenarios: RCP2.6, RCP4.5, RCP6.0, RCP8.5 [46]
CMIP6	2014	A new structure (e.g., Diagnostic, Evaluation and Characterization of Klima, historical simulations and CMIP-Endorsed Model Intercomparison Projects (MIPs)) [57] Additional forcing pathways [60] A standardized data infrastructure for model outputs [57]	IPCC AR6 [50]	ScenarioMIP: combinations of SSP and forcing pathway (e.g., SSP5–8.5) [60]

.

3. Temporal and Spatial Downscaling: Methods and Applications

High-resolution weather data are of importance for building climate change adaptation. To generate high-resolution weather files, downscaling approaches have been developed to derive high-resolution, localized data from coarse-scale climate information. Commonly used downscaling approaches include dynamical downscaling using a Regional Climate Model (RCM), statistical downscaling, and hybrid downscaling that combines the dynamical and statistical downscaling [25], each addressing the challenges of spatial refinement and temporal disaggregation at finer scales.

3.1. Dynamical Downscaling: Physically Based Refinement

Dynamical downscaling employs high-resolution RCMs driven by boundary conditions from GCMs or reanalysis data (e.g., [61,62]).

3.1.1. Dynamical Spatial Downscaling

The RCMs downscale climate simulations by running at a higher spatial resolution, capable of representing topographical feathers and urban areas. This enables RCMs to simulate the mesoscale phenomena that are often absent in GCMs (e.g., urban heat island [63]). RCMs are therefore better representative of local climate conditions. Furthermore, RCMs generate a dataset that maintains physical consistency across different weather variables [25] and provides a spatially continuous field [64]. The provision of high-resolution, physically consistent climate data is important for BES, as such data are essential for calculating building energy demand under climate change.

3.1.2. Dynamical Temporal Downscaling

By integrating physical laws over time, RCMs are capable of generating results with high temporal resolution (e.g., daily or sub-daily) for a range of weather variables, e.g., air temperature [65] and precipitation [66].

3.1.3. Limitations of Dynamical Downscaling

Dynamical downscaling is characterized by high demands for computational resources and requires considerable expertise [25], and the simulation of multiple GCM-RCM combinations for uncertainty assessment may further increase the computational cost. Consequently, the spatial coverage of dynamical downscaling is limited due to the increased demands for computational resources to perform high-resolution simulations. Moreover, the performance of an RCM is largely dependent on the quality of the driving GCM (e.g., [67]) and the selection of physical parameterization schemes (e.g., [68]).

3.2. Statistical Downscaling: Data-Driven Statistical Relationships

In contrast to physical modeling, statistical downscaling establishes statistical relationships between large-scale variables (e.g., from GCMs) and local climate variables [69], requiring relatively low demand for computational resources.

3.2.1. Statistical Spatial Downscaling

High-resolution grids of different weather variables can be generated from coarse-scale GCM predictions using point-scale and grid-based downscaling methods.

Point-scale downscaling method: statistical models are developed for specific weather stations, and then spatial interpolation methods are used to generate high-resolution weather data from these point-based estimates. Common spatial interpolation methods include

• Kriging: provides the best linear unbiased prediction (e.g., [70]).
• Inverse Distance Weighting: provides a proximity-based deterministic estimation (e.g., [71]).
• Spline interpolation: generates a smoothed surface from data points (e.g., [72]).

Grid-based downscaling method: machine learning models, particularly Convolutional Neural Networks (e.g., [73]), can perform pattern-based spatial downscaling by learning the complex, nonlinear mapping from a low-resolution grid to a high-resolution grid.

3.2.2. Statistical Temporal Downscaling

The disaggregation of temporal weather data from monthly/daily resolutions to an hourly resolution can be achieved using statistical downscaling methods, including morphing, stochastic generation, and machine learning.

Morphing: historical observed time series (e.g., hourly data) are adjusted by applying changes (e.g., in mean, variability) derived from climate model predictions [16]. The changes are applied using techniques such as Shift, Stretch, and their combination. The morphing method is widely used to generate future hourly weather data for BES (e.g., [15,74,75]).

For Shift, the predicted hourly values are calculated by applying a shift to the baseline hourly weather values for month m, using Equation (1) [16]:

$$x=x_0+\Delta x_m$$

(1)

with

• x = predicted hourly value;
• x₀ = baseline hourly value for a weather variable;
• ∆x_m = a predicted absolute change in monthly mean value.

For Stretch, a stretching is applied to the baseline hourly weather data for month m to calculate the predicted hourly values using Equation (2) [16]:

$$x={\alpha }_m{x}_0$$

(2)

with

• x = predicted hourly value;
• α_m = a predicted relative change in monthly mean value;
• x₀ = baseline hourly value for a weather variable.

The combination of Shift and Stretch for month m is defined by Equation (3) [16]:

$$x=x_0+\Delta x_m+{\alpha }_m\times (x_0-{〈x_0〉}_m)$$

(3)

with

• x = predicted hourly value;
• x₀ = baseline hourly value for a weather variable;
• ∆x_m = a predicted absolute change in monthly mean value;
• α_m = a predicted relative change in monthly mean value;
• ${〈x_0〉}_m$ = baseline monthly mean value for a weather variable.

The selection of morphing techniques depends on the specific type of weather variable, e.g., Shift for atmospheric pressure, Stretch for solar irradiance, and the combination of Shift and Stretch for dry-bulb temperature. Details on the morphing method can be found in the study of Belcher et al. [16].

Stochastic generation: statistical temporal downscaling using stochastic weather generators can generate weather data at a high temporal resolution required for BES (e.g., hourly data from Meteonorm [76]). This method develops a stochastic model that generates synthetic, site-specific time series while maintaining the essential statistical properties (e.g., mean, variance) consistent with observed weather data [77,78].

Machine learning: machine learning-based nonlinear models, such as Long Short-Term Memory (LSTM) networks, can be trained to perform temporal downscaling by learning complex, nonlinear relationships that transform coarse, large-scale input weather data into high-resolution local time series [79]. This process is based on the ability of LSTM networks to capture long-term dependencies and temporal dynamics in sequential data. Consequently, machine learning-based models have become effective tools for generating high-resolution temporal data (e.g., [80,81]).

3.2.3. Limitations of Statistical Downscaling

For observational data availability, statistical downscaling methods that rely on observational data for spatial interpolation are highly dependent on the number and spatial distribution of weather stations. Consequently, the application of spatial interpolation is constrained in regions with limited observational data. Moreover, the stationarity assumption in statistical downscaling requires that the statistical relationships derived from the historical period will remain valid under a future, changing climate (e.g., [82]). Hence, high-quality and sufficiently long observational datasets are important for statistical downscaling (e.g., [83]). Weather data generated using statistical downscaling can be constrained by historical data, thus making them less representative of potential future extreme weather conditions [25].

Statistical downscaling can introduce physical inconsistencies among downscaled weather variables [25,84]. These inconsistencies often arise from the independent processing of individual weather variables.

3.3. Hybrid Downscaling: An Integrated Approach

Hybrid downscaling combines dynamical and statistical downscaling approaches to utilize their respective strengths, e.g., the physical consistency of dynamical models and the computational efficiency of statistical methods, thereby mitigating the limitations of each approach.

The hybrid downscaling can be categorized into the following three principal frameworks:

3.3.1. Dynamical-Statistical Downscaling

The dynamical downscaling approach (e.g., an RCM) is used to downscale coarse-resolution climate model output to an intermediate resolution (e.g., 10–30 km). In the subsequent statistical downscaling, the output from the dynamical downscaling approach is employed as predictors to train a statistical model using high-resolution weather data as the predictands. This statistical model can further downscale the output from the dynamical downscaling to a higher resolution (e.g., 1–5 km). This framework captures the physical consistency of the RCM while achieving higher resolution at a fraction of the computational cost.

For example, Tran Anh and Taniguchi [85] used a hybrid downscaling approach that combined an RCM and artificial neural networks to downscale rainfall data over the Red River Delta in Vietnam. In this approach, artificial neural networks were trained on coarser RCM output to generate high-resolution data. The results demonstrated that such combination achieved accuracy comparable to that of high-resolution RCM simulations while reducing computation power by 89%.

3.3.2. Statistical-Dynamical Downscaling

A statistical method is used to correct biases in the output from the large-scale climate model (e.g., from a GCM). Subsequently, this bias-corrected output is used as the boundary conditions to drive the RCM for dynamical downscaling. This framework performs statistical bias correction on outputs from large-scale climate models to improve the driving conditions for dynamical downscaling, thereby reducing the propagation of model biases.

For example, a high-resolution climate dataset for the Midwestern United States was developed by dynamically downscaling bias-corrected Community Earth System Model output using the Weather Research and Forecasting model [86].

3.3.3. Model-Embedded Downscaling

This involves deeply embedding machine learning within the dynamical model’s framework, thereby providing physical consistency while achieving significant improvement in computational efficiency. The typical strategies include physics-informed machine learning (PIML) and parameterization replacement.

• PIML: PIML integrates physical principles and domain knowledge into models, enabling the creation of physically consistent models with enhanced capabilities such as increased data efficiency and an accelerated training process [87]. Therefore, machine learning models are designed to obey the physical principles (e.g., conservation laws) by incorporating these physical principles into the loss function or model architecture. For instance, Feng et al. [88] developed a physics-informed neural network framework for downscaling river flow simulations to the subgrid scale. By embedding the Saint-Venant equations into the learning objective, the model achieves satisfactory accuracy in downscaling river flow from a coarse grid with limited observational data assimilated, demonstrating the effectiveness of the PIML strategy for downscaling. Additional approaches to integrate physics into machine learning for weather and climate modeling are summarized by Kashinath et al. [87].
• Parameterization replacement: a trained machine learning model (e.g., artificial neural networks [89]) is used to replace specific, computationally expensive physical parameterization schemes, offering a pathway to more efficient climate simulations.

Compared to conventional statistical downscaling methods (e.g., morphing and stochastic generation), machine learning approaches, including PIML and deep learning models such as LSTM, rely less on stationary statistical assumptions and may offer potential advantages for weather file generation, particularly in capturing evolving nonlinear temporal dependencies and improving physical consistency under non-stationary climate conditions [87,90].

3.3.4. Limitations of Hybrid Downscaling

For dynamical-statistical downscaling frameworks, the quality of the final output largely depends on the driving RCM simulations, as the results from the subsequent statistical downscaling may be limited if the driving RCM poorly represents key climate processes. For statistical-dynamical downscaling frameworks, bias-correcting individual variables from a large-scale climate model can introduce physical inconsistencies with other variables, potentially leading to uncertainties when used to drive the RCM. The integration of machine learning typically requires representative datasets for model training and validation, and such integration may also increase model complexity. Moreover, some machine learning models can act as “black boxes”, often making it difficult to interpret the underlying physical processes behind their predicted results (e.g., [84,91]). In addition, the ability of machine learning models to generalize beyond their training data (e.g., to extreme weather conditions) is often limited. Therefore, the use of hybrid downscaling requires careful and thorough evaluation of the advantages and limitations of each framework in a specific context.

4. Bias Correction for Modifying Model Outputs

The outputs of GCMs and their downscaled results often contain systematic errors (biases) relative to actual observed conditions [92]. Several factors lead to these biases, such as coarse spatial resolution and simplified representations of thermodynamic and physical processes [93,94]. These biases can propagate into subsequent impact assessments, potentially resulting in misleading conclusions in BES.

Bias correction can be applied to reduce the systematic errors (e.g., [95]). Common bias correction methods include distribution-based methods and machine learning.

4.1. Distribution-Based Methods

Distribution-based bias correction methods adjust climate model outputs to achieve consistency in key statistical properties (e.g., mean, variance) with observed data from a historical reference period (e.g., [96]).

The Delta method calculates the differences between current and future climate model outputs and then applies these differences to observed data [97], focusing primarily on mean bias adjustment. Although computationally efficient, the Delta method does not address biases in variability and extremes.

Quantile mapping is considered an effective alternative for bias-correcting climate model outputs [98]. For quantile mapping, a transfer function is established based on the cumulative distribution functions of climate model simulations and observations during a historical period [99]. In contrast to the Delta method, this process adjusts the mean, as well as the variability and quantiles of the distribution. For extreme events, Rohith and Cibin [100] proposed an extremes-weighted empirical quantile mapping to improve the bias correction of extreme precipitation events from climate models by applying separate corrections to extreme and non-extreme precipitation data. Several methods based on quantile mapping have been proposed, including parametric quantile mapping [101], non-parametric quantile mapping [102], and detrended quantile mapping [99].

4.2. Machine Learning

A machine learning model is trained on historical simulations and observations and is subsequently applied to perform bias correction on climate model data (e.g., [103]), allowing it to capture non-liner relationships and complex interactions. However, data imbalance may lead to algorithmic bias toward majority classes in machine learning models. The algorithmic bias that leads to regression to the mean and underestimation of extreme events can be effectively reduced by using methods such as conditional Generative Adversarial Networks (GANs) to better capture extreme weather data [104]. Different machine learning methods, such as Random Forest and Support Vector Machine, perform differently in bias-correcting climate model data. The effectiveness of different machine learning methods for bias-correcting GCM rainfall and temperature data in Nigeria was assessed by Tanimu et al. [105]. Moreover, deep learning models, particularly convolutional neural networks [106] and recurrent neural networks [107], demonstrate the capacity to capture the spatial and temporal dependencies in climate data for bias correction. However, machine learning methods also present limitations, such as overfitting risks (e.g., [108]) and lack of model interpretability [109].

Unlike conventional distribution-based methods (e.g., quantile mapping), which typically assume a stationary transfer function between modeled and observed distributions, GAN-based methods learn nonlinear mappings from data through adversarial training [110]. GAN-based methods may better represent spatial intermittency and the characteristics of extreme events under non-stationary climate conditions (e.g., [111]).

4.3. Limitations of Bias Correction Methods

The stationarity assumption that historical bias patterns will remain constant under future climate conditions may not hold [112,113], especially for extreme events (e.g., [114]). The statistical properties of the climate conditions may change in non-stationary ways under a changing climate [113].

The physical consistency between variables in the original climate model could be disrupted if bias-correcting a variable independently [115]. Maintaining physical consistency in bias correction requires the preservation of fundamental physical principles (e.g., [116]), as well as temporal dependence and spatial coherence (e.g., [117]). The lack of physical consistency may introduce additional uncertainties, especially for applications sensitive to multivariate relationships, such as BES.

Bias correction methods reported in previous studies are compared in

Table 5. Comparison of bias correction methods reported in previous studies.

Author (Year)	Region	Methods Compared *	Variables	Metrics **	Key Findings
Kim et al. (2022) [118]	Thorverton basin	QM	Daily precipitation and flow	Ensemble spread, Gamma distribution parameters, FDCs, percentage error	Bias correction of both precipitation and flow performed best in reducing bias.
Kupilik et al. (2024) [119]	A sub-Arctic region	EQM, GP regression	Daily maximum temperature	RMSE, PSS, Mean Bias	EQM showed a higher RMSE and greater distributional mismatch than GP regression.
Zhang et al. (2024) [92]	Queensland, Australia	LS, QM, Dis_T_F	Precipitation, minimum and maximum temperature, radiation, vapor pressure, mean sea level pressure	KGE, PSS	LS and EQM are the best approaches for mean climatology.
Okirya & Du Plessis (2025) [120]	Uganda	QM, LT, DM, PR	Annual maximum rainfall	RMSE, MAE, Pbias, NSE, KS test	QM outperformed other methods (e.g., RMSE was reduced from 29.20 mm to 19.00 mm).
Song & Chung (2025) [121]	Six continents: South America, North America, Africa, Europe, Asia, Oceania	QDM, EQM, DQM	Daily precipitation	RMSE, MAE, R2, Pbias, NSE, KGE, MdAE, MSLE, EVS, JSD	EQM performed best across most metrics (RMSE: 0.30, MAE: 0.18, R2: 0.98, KGE: 0.87, NES: 0.93, EVS: 0.99).

*: QM: Quantile Mapping. EQM: Empirical Quantile Mapping. GP regression: Gaussian process regression. LS: Linear scaling. Dis_T_F: Statistical distribution-based transfer functions. LT: Linear Transformation. DM: Delta Multiplicative. PR: Polynomial Regression. QDM: Quantile Delta Mapping. DQM: Detrended Quantile Mapping. **: FDCs: Flow duration curves. RMSE: Root Mean Square Error. PSS: Perkins Skill Score. KGE: Kling–Gupta efficiency. MAE: Mean Absolute Error. Pbias: Percent bias. NSE: Nash–Sutcliffe Efficiency. KS test: Kolmogorov–Smirnov test. R²: Coefficient of Determination. MdAE: Median Absolute Error. MSLE: Mean Squared Logarithmic Error. EVS: Explained Variance Score. JSD: Jensen–Shannon divergence.

.

5. Strategies for Matching Spatial Scales

5.1. Grid Resolution Adjustment

The spatiotemporal resolution of weather data, including both grid cell size and time step, significantly affects BES results. A high spatial resolution enables the capture of local effects (e.g., urban heat islands) that influence building energy demand [122]. Extreme events may not be well captured at a coarse temporal resolution, for example, when using monthly or daily averages [123]. In addition to spatial and temporal downscaling, another important step is to adjust grid cells of different datasets to a common size to enable the comparison of weather data with notably different native spatial resolutions (e.g., [17]). Depending on the building type, microclimate conditions, and model configurations, further increases in spatial resolution may lead to diminishing improvements in BES results, while increasing computational and data-storage demands.

5.2. Station Data Compared to Gridded Data

Climate models are typically run on fixed grids at coarser resolutions, whereas observational networks provide point-scale measurements. Additional considerations are required when comparing gridded data from climate models with point observations from weather stations (e.g., for bias correction [124]), because these two data types represent different spatial representations: area averages versus point samples. Therefore, achieving spatial consistency between observational and gridded data is essential for evaluating climate change impacts. Common methodologies to address this spatial mismatch include (1) the interpolation of station data onto a grid spatially comparable to climate model output; (2) the comparison of area averages from station data with those from the corresponding climate model grids; (3) data aggregation of climate model output or station data.

Spatial interpolation of station data is widely adopted to generate gridded datasets that are consistent with the spatial resolution of climate model output. The commonly used spatial interpolation methods are summarized in Section 3.2.1. However, in complex terrain or data-scarce regions, interpolation may result in non-negligible errors, such as the smoothing of extreme values.

Instead of using spatial interpolation of station data, an alternative methodology is to compare model output with the average of station observations within the corresponding model grid cell. This methodology is, however, dependent on the distribution of available stations within the grid cell. In data-sparse regions, the spatial mismatch may be addressed using data aggregation, e.g., temporal aggregation of point measurements [124], enabling a statistically consistent comparison between gridded data and station observations.

6. Assessing Uncertainty with Multiple General Circulation Models and Emissions Scenarios

In addition to the uncertainties associated with spatiotemporal downscaling and bias correction, differences in GCM structures and socioeconomic development pathways can also introduce uncertainties in the generation of weather files for BES applications.

6.1. Multi-Model Ensemble: Different Model Structures

The use of multiple climate model predictions is essential for uncertainty assessment. GCMs vary in aspects of their model structure such as the representations of physical processes and spatial resolution, leading to a wide range of predictions for climate variables (e.g., precipitation [17]). A multi-model ensemble approach is often adopted because it provides a more comprehensive assessment of the uncertainty range compared to using a single GCM simulation (e.g., [15]). Moreover, the suitability of GCMs for specific regions can be evaluated by comparing GCM outputs with observed historical weather data to identify models that perform well, thereby informing the selection or weighting of models within the ensemble (e.g., [125]).

6.2. Emissions Scenarios: Socioeconomic Uncertainty

Differences in predicted weather conditions, such as increases in ambient air temperature, across scenario frameworks (e.g., RCPs [45,46] and SSPs [48]) are another important source of uncertainty in the weather files used for BES under climate change.

These scenario frameworks represent distinct greenhouse gas concentration pathways for climate change studies. The scenarios range from low emissions (e.g., RCP2.6/SSP1–2.6) to high emissions (e.g., RCP8.5/SSP5–8.5), representing different assumptions about socioeconomic development and policy choices. Therefore, the adoption of multiple emissions scenarios is important to study the possible range of future climate conditions that are related to different socioeconomic development pathways. Moreover, the effect of emissions scenarios differs across time periods, i.e., the effect is limited in the near-term period (e.g., 2035) but becomes noticeable in the long-term period (e.g., 2090) [15].

Given the multiple sources of uncertainty resulting from the preparation of hourly weather data for BES (e.g., from temporal and spatial downscaling, and bias correction), a probability analysis is recommended to inform decision-making instead of using a single deterministic prediction [25]. The probability analysis enables the generation of a probabilistic range of outcomes, thereby propagating and quantifying uncertainty from the input weather data to the BES outcomes. Moreover, the probability analysis provides a more robust basis for the assessment of building performance under climate change.

Earlier studies have demonstrated that uncertainty can be quantified through ensemble-based simulations, probabilistic predictions, and sampling-based uncertainty propagation methods. For example, Coronato et al. [126] reported that cooling energy demand for a building in Rosario city, Argentina ranged from 27 to 37 MJ/m² under a moderate emissions scenario and from 51 to 70 MJ/m² under a high emissions scenario, showing increased uncertainty under higher emissions scenarios. Similarly, Troup et al. [127] adopted ensemble-based BES simulations using 14 GCMs and two RCP emissions scenarios, showing an increase of 0.5–10% in annual source energy use in the 2090s depending on the location and emissions scenario. Moreover, Tian & de Wilde [128] employed probabilistic climate projections and Latin Hypercube Sampling to quantify uncertainties in future building energy performance predictions.

7. Microclimate and Extreme Weather Conditions

The generation of weather files for BES based on TMY weather data from open rural areas often neglects two essential aspects: the effects of urban environment on microclimate and the occurrence of extreme weather conditions. Climate change may enhance the intensity and frequency of extreme weather events [129]. Moreover, urban microclimate effects, such as the Urban Heat Island (UHI) effect, may amplify the impacts of climate change [130]. Therefore, it is important to include urban microclimate effects and extreme weather conditions in weather files used for BES.

Urban environment can significantly change local climate conditions through mechanisms such as the UHI effect [131] and wind flow modification [132]. Consequently, including these changes in the weather files can influence the results of BES [133]. For example, a decrease of 1% to 10% in winter heating energy consumption and an increase of 2% to 14% in cooling energy consumption due to urban microclimate effects were reported for a school building in Montreal, Canada [134]. However, TMY weather data collected from open rural areas (e.g., from airports) tend to neglect urban microclimate effects, resulting in misleading results for the simulation of building energy demand and indoor thermal conditions. Therefore, site-specific weather data generated using Computational Fluid Dynamics (CFD) (e.g., ENVI-met models [135]) or energy balance models (e.g., the Urban Weather Generator (UWG) [136]) are subsequently employed in BES to include urban microclimate effects. ENVI-met is a high-resolution CFD-based model suitable for detailed microscale analysis [137,138], while the UWG is based on an energy balance model enabling integration with BES at the mesoscale [139]. These approaches enable the integration of urban microclimate effects into BES, while often demanding high computational resources and requiring detailed geometric and material information. A comparison between a CFD-based model (i.e., ENVI-met) and an energy balance model (i.e., the UWG) is presented in

Table 6. Comparison between ENVI-met (a Computational Fluid Dynamics (CFD)-based model) and the Urban Weather Generator (UWG) (an energy balance model) (based on [137,138,139,140,141]).

Aspect	ENVI-Met (CFD Model)	The UWG (Energy Balance Model)
Core methodology	CFD	Energy balance model
Spatial scale	Microscale (e.g., grid size = 5 m)	Mesoscale (e.g., urban scale)
Temporal resolution	Second-scale time-step simulation	Hourly-resolution weather data generation
Vegetation modeling	Detailed (vegetation modeling based on physical processes)	Simplified (fraction-based vegetation representation)
Coupling with BES	Indirect (e.g., post-processing of microclimate outputs)	Direct (e.g., as a weather file modifier)
Outputs	Spatially distributed microclimate outputs	Urban canopy-level time-series weather data
Computational cost	High	Low
Typical application	Local microclimate, street canyons, green infrastructure	BES, urban heat island estimation

.

The peak building cooling load under extreme weather conditions was found to be higher than under typical weather conditions in future climate scenarios [25]. Extreme weather files are relevant for heating, ventilation and air conditioning (HVAC) design, as they inform peak cooling load, HVAC system sizing, and system resilience under extreme climate conditions [142,143]. Moreover, extreme weather events, which may affect power plants’ production capacity and electricity transmission line losses, together with the high building energy demand, can lead to high stress on the power grid (e.g., [25]). Therefore, it is important to adjust the distribution of extreme weather conditions (e.g., through bias correction) to better represent predicted future climate extreme events, such as heatwaves, when generating future weather data for BES. Future extreme weather data for BES can be generated using three approaches: (i) selecting a representative extreme year [144], (ii) constructing a synthetic year by combining monthly extreme weather conditions [145], and (iii) applying stochastic weather generators to analyze extreme statistics [146]. Commonly used indicators for extreme weather conditions are summarized in

Table 7. Indicators for extreme weather conditions.

Indicator	Definition	Example Calculation	Application Scenario	Reference
Maximum or minimum value	The maximum or minimum value over a period.	$T_{\max }^{\left(p\right)}=\underset{h\in p}{\max }\left(T_h\right)$ with $T_{\mathrm{m}\mathrm{a}\mathrm{x}}^{\left(p\right)}$ = maximum temperature in the period p [°C]; Th = air temperature at hour h in the period p [°C]; all hours within the period p (h∈p).	Peak building energy demand for heating or cooling.	[147]
Return period	The average time interval between occurrences of an event exceeding a threshold.	$F_X\left(x_{T_r}\right)=1-{\displaystyle \frac{1}{T_r}}$ with Tr = return period (years); $x_{T_r}$ = return level (e.g., air temperature, °C); Fx(⋅) = cumulative distribution function.	Characterization of the frequency of occurrence and risk of extreme events.	[146]
Heating degree days (HDD) or cooling degree days (CDD)	The cumulative sum of daily positive deviations of outdoor dry-bulb temperature above (for CDD) or below (for HDD) a base temperature.	$CDD={\displaystyle \sum _{d=1}^D\max \left(T_d-T_b,0\right)}$ with CDD = cooling degree days [°C⋅day]; Td = daily average outdoor dry-bulb temperature on day d [°C]; Tb = base temperature [°C]; D = total number of days.	Annual building energy demand for heating and cooling.	[148,149]
Frequency of exceedance	Number of hours with a variable exceeding a threshold.	$H_t=\left\{\begin{array}{c}1, \mathrm{if} X_t>X_h\\ 0, \mathrm{if} X_t\le X_h\end{array}\right.$ $N={\displaystyle \sum _{t=1}^n{H}_t}$ with Xt = hourly variable (e.g., outdoor dry-bulb temperature, °C); Xh = an hourly threshold (e.g., outdoor dry-bulb temperature, °C); Ht = hourly exceedance conditions; N = number of hours exceeding a threshold; n = Number of hours.	Assessment of cooling hours and thermal comfort.	[150]
Heatwave	A period of high temperature during which the daily ambient temperature is not below a threshold for a minimum number of consecutive days.	$T_a\ge T_{td}$ for Nm consecutive days. with Ta = daily ambient temperature (e.g., daily maximum temperature, °C); Ttd = a threshold [°C]; Nm the minimum number of consecutive days.	Evaluation of overheating risk and system stress during heatwave events.	[9,151]

.

8. Discussion

8.1. Application to Building Energy Simulation

8.1.1. Key Application Areas

The application of temporal and spatial downscaling can provide high-resolution weather data essential for analyzing extreme events. Downscaling methods improve both temporal and spatial resolution of weather data, and this high-resolution weather dataset (e.g., hourly based weather files) can be used as an input for BES to assess the effects of extreme events (e.g., heatwaves) and urban microclimate (e.g., UHI) on building energy demand and indoor thermal conditions. Moreover, climate change adaptation strategies for buildings can be developed based on the downscaled weather data (e.g., [18]). Renewable energy resource assessment requires high-resolution weather data to represent local climate conditions.

Bias correction is commonly applied to reduce biases in climate model outputs. For example, in reginal climate studies, biases in dynamically downscaled RCM outputs, including biases inherited from the driving GCMs and structural biases of the RCM [152], can be reduced using bias correction.

The use of multi-model ensembles and different emissions scenarios provides a range of possible future weather predictions for BES. This enables the design of buildings that are more robust (climate-robust buildings) to climate change and supports a broader analysis of extreme events. Different emissions scenarios are essential for the long-term assessment of building energy demand. In addition, the emissions scenarios link future climate predictions to socioeconomic factors, informing building design under different mitigation and adaptation strategies.

The integration of microclimate and extreme weather conditions into BES is important for assessing heating, ventilation and air conditioning system performance, as it enables buildings to be evaluated under extreme events while including the effect of microclimate. Furthermore, extreme events and microclimatic effects may increase peak building energy demand and affect grid impact assessments. In addition to building energy applications, weather data that include extreme weather events and microclimatic effects can be used in BES for health assessments [153] and thermal comfort evaluation. The relevance of future weather data to BES is reflected in building performance metrics: for example, an increase in dry-bulb temperature can lead to higher cooling degree days, and changes in solar radiation can affect heating energy demand. The key application areas for generating and utilizing future weather data in BES are graphically depicted in Figure 2.

8.1.2. Effect of Method Choice on Building Energy Simulation

Dynamical downscaling shows good applicability to future scenarios and provides high spatiotemporal resolution due to its physical mechanisms and high physical consistency. However, these advantages may require considerable computational demand and high input data requirements. Statistical downscaling and distribution-based methods (e.g., the Delta method, quantile mapping) are computationally efficient and have relatively low input data requirements, but they rely on the stationarity assumption and are therefore less representative of non-stationary extremes, especially under a changing climate. The hybrid approach combines physical and statistical elements to achieve comparable performance in weather file generation while reducing computational costs. Multi-model ensembles improve the representation of uncertainty in future weather predictions (e.g., extreme events), albeit with increased computational costs and data requirements. Microclimate methods are effective for high-resolution studies, and approaches focusing on extreme weather conditions are suitable for climate change impact assessments; however, their applicability is often limited to specific locations.

Therefore, although the performance may vary by method (e.g., because of methodological uncertainty), the choice of methods necessitates consideration of application requirements, study objectives, and available computational resources. The commonly used approaches and methods in weather file generation and their implications for BES are compared in

Table 8. Effect of method choice on BES *.

			Generalizability		Stationarity Assumption and Extremes		Output Data Resolution		Physical Principles		Computational Resources and Data Requirements
	Approaches and methods		Spatial transferability	Future scenario applicability	Stationarity assumption	Representation of extremes	Spatial resolution	Temporal resolution	Physical consistency	Physical mechanisms	Computational cost	Input data requirements **
Temporal and spatial downscaling	Dynamical downscaling		M	H/L	(-)	M	H/L	H/L	H/L	H/L	H/L	H/L
	Statistical downscaling		L/S	M	H/L	L/S	M	H/L	L/S	L/S	L/S	M
	Hybrid downscaling		M	H/L	M	M	H/L	H/L	M	M	M	H/L
Bias correction	Distribution-based methods	The Delta method	L/S	M	H/L	L/S	(-)	(-)	L/S	(-)	L/S	L/S
		Quantile mapping	L/S	M	H/L	M	(-)	(-)	L/S	(-)	L/S	M
	Machine learning		L/S	M	M	M	(-)	(-)	M	M	M	H/L
Strategies for matching spatial scales	Grid resolution adjustment		(-)	(-)	(-)	L/S	H/L	(-)	L/S	(-)	L/S	L/S
	Station data compared to gridded data		(-)	(-)	(-)	L/S	H/L	(-)	L/S	(-)	L/S	L/S
Assessing uncertainty	Multi-model ensemble		H/L	H/L	(-)	H/L	(-)	(-)	M	M	H/L	H/L
	Emissions scenarios		(-)	H/L	(-)	(-)	(-)	(-)	(-)	(-)	L/S	L/S
Microclimate and extreme weather conditions	Microclimate		L/S	M	(-)	M	H/L	H/L	H/L	H/L	H/L	H/L
	Extreme weather conditions		M	H/L	L/S	H/L	M	H/L	M	M	M	M

*: High/Large (H/L); Moderate (M); Low/Small (L/S); Not applicable (-). **: Input data requirements include spatial and temporal resolution, time span, and amount of data. Darker background colors indicate higher levels, while lighter colors indicate lower levels of each methodological aspect.

.

8.1.3. Methodological Framework

The methods for generating future weather files have been widely used in BES studies to assess climate change effects (e.g., [18,125]). Climate model outputs (e.g., from GCMs) can be bias-corrected against historical observations and downscaled using different downscaling approaches to generate high-resolution weather data for microclimate analysis and extreme weather assessments. For example, high-resolution weather data downscaled from GCM outputs coupled with urban climate models can be used to simulate the evolution of the UHI effect under future climate conditions (e.g., [154]). The methodological framework to generate future weather data based on the statistical downscaling approach, the dynamical downscaling approach, and the hybrid downscaling approach is shown in Figure 3.

For studies where computational cost is the primary constraint and long-term average condition is of importance, statistical downscaling (e.g., morphing) combined with distribution-based bias correction (e.g., quantile mapping) can be adopted (e.g., [155]). For BES applications requiring high physical consistency across different weather variables, dynamical downscaling is generally preferable despite its higher computational cost [25]. When the research focus is on extreme events (e.g., heatwaves), dynamical or hybrid downscaling combined with Extremes-Weighted Empirical Quantile Mapping [100] is recommended. For uncertainty-focused studies, the use of a multi-model ensemble under different emissions scenarios is suggested. The decision criteria for method selection in typical BES contexts are summarized in

Table 9. Decision criteria for method selection based on application requirements.

Application Requirement	Recommended Method *	Explanation
Spatiotemporal downscaling + low computational cost	Statistical downscaling + distribution-based bias correction	Computationally efficient
Spatiotemporal downscaling + high physical consistency	Dynamical downscaling	High physical consistency but high computational cost
Focus on extremes	Dynamical/hybrid downscaling + EW-EQM	Representation of extremes
Uncertainty quantification	Multi-model ensemble + different emissions scenarios	Climate model structural uncertainty and socioeconomic uncertainty
Mesoscale (e.g., urban scale) + low computational cost	The UWG (energy balance model)	Direct coupling with BES
Microscale + high spatial resolution (e.g., grid size = 5 m)	ENVI-met (CFD model)	Local microclimate modeling

*: EW-EQM: Extremes-Weighted Empirical Quantile Mapping [100].

.

The generation of hourly weather data for 2020 is presented, focusing on the inclusion of a multi-model ensemble with low computational cost under average conditions. Shanghai, one of the largest cities in China, was selected as a case study (31.17° N, 121.43° E). Air temperature was selected as the weather variable. The Delta method was selected to bias-correct the monthly weather data, and subsequently, the bias-corrected monthly weather data were downscaled to an hourly resolution using the morphing method. Data from the fifth generation European Centre for Medium-Range Weather Forecasts (ECMWF) reanalysis-Land (ERA5-Land) [156] and the CMIP6 GCM outputs from the Earth System Grid Federation (ESGF) [157] were used. Three CMIP6 GCMs were selected for the present study, including AWI-CM-1-1-MR [158], BCC-CSM2-MR [159], and MRI-ESM2-0 [160].

For the Delta method, the monthly mean air temperatures and the monthly mean daily maximum and minimum air temperatures obtained from the CMIP6 GCM outputs [161,162,163] and derived from ERA5-Land [156] for the baseline period 1985–1995 were used to bias-correct the monthly air temperatures from CMIP6 GCM outputs for 2020 [164,165,166] under the SSP2–4.5 scenario [167,168]. The bias-corrected projected air temperatures (X_bc,proj) were calculated using Equation (4) [97]. The CMIP6 GCM outputs obtained from ESGF at a nominal resolution of 100 km were interpolated to the 0.1° × 0.1° ERA5-Land grid using bilinear interpolation.

$$X_{bc,proj}=X_{ref,base}+\Delta X$$

(4)

with

• X_bc,proj = bias-corrected projected air temperature [°C];
• X_ref,base = reference air temperature during the baseline period [°C];
• ∆X = temperature change factor [°C].

The temperature change factor (∆X) can be calculated using Equation (5).

$$\Delta X=X_{gcm,proj}-X_{gcm,base}$$

(5)

with ∆X the temperature change factor [°C], X_gcm,proj the GCM-simulated projected air temperature [°C], and X_gcm,base the GCM-simulated air temperature during the baseline period [°C].

The bias-corrected monthly data for 2020 were subsequently employed for morphing, with hourly ERA5-Land data [156] for the period 1985–1995 adopted as the baseline weather data. For morphing, the predicted hourly air temperature (dbt) was calculated using a combination of the stretching and shifting algorithms (Equation (6) [16]).

$$dbt=db{t}_0+\Delta TEM{P}_m+\alpha db{t}_m\times (db{t}_0-{〈db{t}_0〉}_m)$$

(6)

with

• dbt = predicted hourly air temperature [°C];
• dbt₀ = baseline hourly air temperature [°C];
• ∆TEMP_m = predicted change in monthly mean air temperature [°C];
• αdbt_m = scaling factor for stretching;
• ${〈db{t}_0〉}_m$ = baseline monthly mean air temperature [°C].

The scaling factor for stretching (αdbt_m) can be calculated using Equation (7).

$$\alpha db{t}_m={\displaystyle \frac{\Delta TMA{X}_m-\Delta TMI{N}_m}{{〈db{t}_{0\max }〉}_m-{〈db{t}_{0\min }〉}_m}}$$

(7)

with

• αdbt_m = scaling factor for stretching;
• ∆TMAX_m = predicted change in monthly mean daily maximum air temperature [°C];
• ∆TMIN_m = predicted change in monthly mean daily minimum air temperature [°C];
• ${〈db{t}_{0\max }〉}_m$ = baseline monthly mean daily maximum air temperature [°C];
• ${〈db{t}_{0\min }〉}_m$ = baseline monthly mean daily minimum air temperature [°C].

The performance of the bias correction method was assessed by comparing hourly downscaled air temperature (without bias correction) and hourly bias-corrected downscaled air temperature against ERA5-Land data for 2020. The performance was evaluated using RMSE with Equation (8), MAE with Equation (9), and R² with Equation (10).

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n{(y_i-{\widehat{y}}_i)}^2}}$$

(8)

$$\mathrm{MAE}={\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n\left|y_i-{\widehat{y}}_i\right|}$$

(9)

$$R^2=1-{\displaystyle \frac{{\displaystyle \sum _{i=1}^n{(y_i-{\widehat{y}}_i)}^2}}{{\displaystyle \sum _{i=1}^n{(y_i-\overline{y})}^2}}}$$

(10)

with n the number of hours, y_i the hourly air temperature obtained from ERA5-Land, ${\widehat{y}}_i$ the morphed hourly air temperature, and $\overline{y}$ the average air temperature obtained from ERA5-Land.

The performance metrics (RMSE, MAE, R²) calculated using hourly data for the entire year of 2020 are presented in

Table 10. Comparison of downscaled (without bias correction) and bias-corrected downscaled air temperatures using RMSE, MAE, and R² for the entire year of 2020.

	RMSE [°C]		MAE [°C]		R2 [-]
GCM	Downscaled Without Bias Correction	Bias-Corrected Downscaled	Downscaled Without bias Correction	Bias-Corrected Downscaled	Downscaled Without Bias Correction	Bias-Corrected Downscaled
AWI-CM-1-1-MR	4.61	3.26	3.54	2.37	0.686	0.843
BCC-CSM2-MR	4.96	3.07	3.91	2.24	0.637	0.861
MRI-ESM2-0	3.53	3.29	2.67	2.43	0.816	0.840

, while Figure 4 shows the performance metrics calculated separately for each month of 2020.

The results summarized in Table 10 indicate that the bias correction method effectively reduces the systematic errors and improves the agreement between the downscaled GCM outputs and reference ERA5-Land air temperature data, as indicated by lower RMSE and MAE and higher R² values compared to the downscaled results without bias correction. Moreover, the degree of improvement varies across different GCMs. For example, MRI-ESM2-0 shows the smallest improvements (e.g., RMSE decreases from 3.53 °C to 3.29 °C). The largest improvements are observed for BCC-CSM2-MR (e.g., RMSE decreases from 4.96 °C to 3.07 °C).

The monthly comparisons (Figure 4) show that the bias-corrected downscaled outputs generally outperform the downscaled outputs without bias correction across all three GCMs, as evidenced by lower RMSE and MAE values and higher R² values in most months. Although negative R² values remain in some months, the overall reduction in the systematic errors (Table 10) demonstrates the effectiveness of bias correction in improving the performance of statistically downscaled air temperature data derived from CMIP6 GCM outputs.

8.2. Challenges and Future Prospects

In addition to the progress outlined above, three major challenges are identified, each pointing towards specific research directions for future work.

The selection of GCMs and climate scenarios requires a careful evaluation for BES applications due to the uncertainties arising from different GCM structures and socioeconomic development pathways. BES requires high-resolution weather files, while a GCM may perform differently depending on the geographic region and climatic variable (e.g., [125]). Moreover, the process of downscaling and bias correction may introduce further uncertainties into weather files. The adoption of multi-model ensembles and probabilistic approaches offers a robust strategy to quantify these uncertainties, rather than single deterministic predictions. A quantitative comparison of downscaling and bias correction methods could be included in future research (e.g., prediction errors for building energy demand under different climate scenarios).

Methodologies for generating future weather files involve different computational trade-offs. For example, dynamical downscaling using RCMs can simulate local weather conditions with high spatiotemporal resolution, but it often requires considerable computational resources [169]. Another approach is statistical downscaling, which is computationally efficient, as it downscales GCM outputs to local weather conditions based on historical observations. However, this approach typically increases either spatial or temporal resolution separately, and the physical consistency among different weather variables is often limited [169]. These trade-offs necessitate careful consideration of application requirements and available computational resources to balance computational cost with representativeness.

The development of future typical weather files (e.g., TMY) may not adequately represent extreme weather conditions, whereas the inclusion of both typical and extreme weather data in BES enables the design of buildings that are not only optimized for average conditions but also resilient to extreme weather events. A thorough analysis of extreme events, such as return period estimation, could be the focus of future work.

9. Conclusions

This article investigates key methods for generating future weather data for BES applications. Moreover, the methodological framework to downscale future weather data is proposed. The major challenges for BES applications are identified, and future prospects are discussed. The selection of methods necessitates careful consideration of the generalizability (e.g., spatial transferability if applied to another location) and the trade-offs between computational cost and representativeness (e.g., high-resolution weather data). Furthermore, the physical consistency may be compromised if the downscaling methods are used to process the variables individually. The following conclusions are made:

• Dynamical downscaling with RCMs enables high-resolution simulations of local climate conditions, often with high computational costs. Statistical downscaling requires relatively low computational resources, but it depends on the availability of observational data and can be constrained by historical records. The use of hybrid downscaling requires careful evaluation based on the application context, due to the respective advantages and limitations of dynamical and statistical downscaling.
• Bias correction reduces the systematic errors in GCM outputs and their downscaled results, but its dependence on the stationarity assumption may limit its applicability in BES, especially for extreme weather events.
• The selection of methods for downscaling and bias correction involves different trade-offs, requiring consideration of specific characteristics of model-variable-region combinations.
• Differences in GCM structures and socioeconomic development pathways highlight the necessity of employing a multi-model ensemble approach and probability analysis to address uncertainties in future weather predictions.
• The inclusion of microclimate data in future weather files enables the integration of local weather conditions into BES. Furthermore, the consideration of extreme weather conditions shifts the focus of building design from responding to typical weather conditions to achieving resilience to extreme weather events.

This study demonstrates that appropriate processing (e.g., downscaling and bias correction) of GCM outputs is important for the use of future weather files in BES to assess long-term climate change effects. The most cost-effective approach often combines statistical downscaling with adjustment of grid cell size, thereby balancing the need for high-resolution, site-specific weather data with the demand for computational resources. This study summarizes the key methods and strategies for generating future weather data for BES, thereby providing a guideline to support the development of high-resolution, bias-corrected future weather files while minimizing computational costs.