Climate Adaptability and Energy Performance in the Greater Bay Area of China: Analysis of Carbon Neutrality Through Green Building Practices

Abstract

China has committed to carbon neutrality by 2060 by necessitating a comprehensive transformation of its building sector, particularly in rapidly urbanizing areas such as the Greater Bay Area (GBA), where subtropical climates, urban heat island effects, and extreme weather events present distinct challenges for achieving carbon reduction objectives through green building practices. The goal of this study is to establish an analysis method for green building success in the GBA’s subtropical environment, paying attention to the challenging goals of reducing carbon and making buildings more climate-resilient. Research techniques involved performing building energy simulations with EnergyPlus and DesignBuilder, applying LightGBM models for machine learning, using case studies from 32 buildings in Shenzhen, Hong Kong and Guangzhou and carrying out an evaluation of the policy using a PEI. Energy usage in green buildings was 45.3% less than in conventional structures, with Energy Use Intensity ranging from 65.1 to 72.4 kWh/m²/year, while traditional buildings used between 118.5 and 124.2 kWh/m²/year. Also, the carbon footprint during the life cycle of buildings was decreased by 38.4% and they became more resilient to typhoons, giving residents 72.4 h of power during storms, while conventional buildings gave only 8.3 h. HVAC system efficiency was the leading factor, accounting for 24.3% of the difference in energy performance. A detailed approach is developed for optimizing subtropical green buildings, based on unique design features and helpful policy ideas to promote carbon neutrality in swiftly growing metropolitan areas around the world.

1. Introduction

The declaration by China in 2020 to peak carbon emissions by 2030 and achieve carbon neutrality by 2060 is a turning point in the history of climate governance on a global scale [1]. The pledge increases the environmental responsibility of China, indicating that carbon neutrality is not only an environmental ambition but also a driver of industrial revolution, technological innovation, and green-oriented economic growth [2,3]. China is the biggest carbon producer in the world, so its move toward a low-carbon economy has a big effect on reducing climate change around the world. As a result, there has been a lot of study into ways to make all sectors more energy-efficient and lower their emissions [4]. The Greater Bay Area (GBA), including nine cities in Guangdong Province together with the Special Administrative Regions of Hong Kong and Macau, occupies a pivotal position in this dynamic megaregion. GBA contributes about 30% to 40% of the energy consumption and corresponding carbon emissions of the building sector in China [5]. This region serves as an experimental hot spot and a priority area of first-rate green-building programs in line with national climate goals [6].

The construction industry is traditionally viewed as one of the most significant tools of decarbonization, having considerable potential in reducing operational energy consumption and the associated emissions [7]. The most well-known green building rating systems, such as LEED and BREEAM set design standards that are aimed at preserving more energy and reducing carbon footprints, and it is already proven that green buildings that have been certified under these systems perform better in these areas than conventional buildings [8,9].

However, the direct implementation of standardized green construction solutions in the subtropical environment of the Greater Bay Area faces significant contextual challenges that limit their effectiveness. Three critical factors distinguish the GBA context from temperate climates where these rating systems were originally developed:

Cooling Load Dominance: In the GBA’s subtropical climate, cooling loads constitute 60–70% of annual building energy consumption, compared to only 30–40% in temperate climates. This fundamental difference in energy demand patterns is inadequately addressed by current rating schemes, which were primarily designed for heating-dominated climates [10,11].

Typhoon Resilience Requirements: The GBA’s elevated exposure to extreme tropical cyclones necessitates building designs that can maintain functionality during and after severe weather events. This resilience requirement is given minimal consideration in traditional sustainability analyses, despite its critical importance for operational continuity and occupant safety [12].

Grid Energy Composition Variability: Current carbon accounting models rely on standardized emission coefficients that fail to capture the dramatic variations in regional energy composition within the GBA. While Shenzhen operates on an almost entirely renewable grid, Hong Kong remains heavily coal-dependent, creating significant disparities in the carbon impact of identical building energy consumption patterns [13].

Recent research is gradually recognizing that the environmental benefits associated with green buildings cannot be evaluated in abstract. The advantages should be assessed within their urban context, particularly in regions experiencing significant urbanization that has intensified the urban heat island (UHI) effect, resulting in city center temperatures in Guangzhou and Shenzhen within the GBA region being 3 to 5 °C higher than those in rural areas [14,15]. This UHI development raises the need for cooling energy, lowers the efficiency of systems, and raises carbon pollution. This counteracts the environmental benefits that green building methods are supposed to offer [16]. However, the existing literature focuses mainly on the evaluation of energy efficiency, renewable integration, and building technologies separately and does not pay much attention to the synergetic or antagonistic relationships between them under subtropical conditions [17,18]. Moreover, the majority of studies are based on either the design-stage projection or standardized performance ratings, not necessarily predicting the actual operational performance during extreme weather conditions. As a result, the GBA does not have a strong locally relevant evidence base that connects building-level interventions to measurable energy savings, climate resilience and carbon reductions in its highly diverse urban environment [19].

The benefits of green buildings as compared to conventional structures have been widely acknowledged in recent research, but the analysis of already certified buildings has particular value in the Greater Bay Area (GBA) context. Certified green buildings provide empirical assessments of operations in subtropical circumstances, enabling the identification of design strategies with substantial impact and the evaluation of scalability within the region’s diverse urban and climatic contexts [20]. The nature of the interaction between building-level green interventions and urban-level processes such as urban heat island effects, exposure to typhoons and energy grid variability has not been systematically investigated in the previous research [21,22]. Further, there is little knowledge on the synergistic design strategies that can be used to achieve energy efficiency, thermal comfort, typhoon resiliency and carbon mitigation simultaneously. The same gap can be found in the area of policy design, as the current evaluation of green building policies in GBA tends to either focus on compliance level or single policy tools and fails to analyze their alignment with empirical building performance results. The need to overcome these gaps requires a comprehensive assessment framework that links the technological assessments at the building level with the environmental processes in the urban scale and policy processes that are specific to each region.

Presently, policy-performance disconnect is a common issue that poses a problem to built-environment research. This research thus proposes the Policy Effectiveness Index (PEI) as a means of quantitatively correlating particular policy measures with measurable building-performance outputs. The analytical design will be built upon a three-tiered approach. First, we will thoroughly map and score all green-building policies in Shenzhen, Hong Kong, and Guangzhou, including their scope, enforcement mechanisms, and incentive structures. Second, we will statistically correlate certified buildings’ energy consumption, carbon emissions, and resilience data with the policy environment to find tools that can improve performance in a quantifiable way. Finally, we will use machine learning algorithms to develop policy effectiveness coefficients, which will change our assessments from compliance-based to performance-based evaluations and help us identify conflicts and synergies.

The current research represents the first in-depth study in the context of the Guangdong-Hong Kong-Macau Greater Bay Area (GBA) to simultaneously measure the synergetic nature between cooling load optimization, typhoon resilience, and carbon reduction. A multi-scalar approach is taken, and this way, it connects the performance of the building, the urban climatology and the effectiveness of the policy. This study aims to address these research gaps through several interconnected objectives.

The first objective is

To develop a comprehensive framework of appraising green building practice that will also evaluate carbon-reduction potential and climate resilience in the subtropical environment of the Greater Bay Area (GBA).

The second objective is

To quantification of the effectiveness of the architectural and technological interventions to reduce the impact of the urban heat-island effect and operational carbon emissions of building portfolios of a representative building typology.

The third objective is

To determine the synergistic design approaches that respond to high cooling energy needs, typhoon resiliency, and carbon reduction goals simultaneously.

The fourth objective is

To entail a critical assessment of how these best practices can be scaled and transferred in diverse urban and climatic conditions within the GBA and other subtropical megacities.

The last objective is

To provide policy recommendations and implementation models that would fast-track the implementation of regionally appropriate green building practices, hence making the GBA a model zone of China’s carbon-neutrality ambitions.

The paper integrates building scale performance analysis into the larger urban policy environment of the GBA to promote insights on how green building activities can provide a quantifiable, multi-scalar benefit to the carbon neutrality obligation of China.

2. Study Area

The Greater Bay Area (GBA) of China is a perfect area for researching carbon neutrality through green building practices. It consists of nine cities of Guangdong Province (Guangzhou, Shenzhen, Zhuhai, Foshan, Huizhou, Dongguan, Zhongshan, Jiangmen, and Zhaoqing) and Special Administrative Regions of Hong Kong and Macau (Figure 1) [23]. This metropolitan agglomeration of integrated nature occupies 56,095 square kilometers. It has a population of 86.6 million (2023) and a GDP of USD 1.91 trillion. This makes it one of the world’s most economically important bay regions (

Table 1. Key Demographic and Socioeconomic Data of GBA.

City/Region	Population (Million, 2023)	Land Area (km2)	GDP (USD Billion, 2023)	Population Density (Per km2)	Urbanization Rate (%)
Hong Kong	7.4	1110	368.6	6659	100
Macau	0.7	33	38.1	21,212	100
Guangzhou	18.7	7434	422.3	2515	86.3
Shenzhen	17.7	1997	475.7	8863	100
Zhuhai	2.5	1736	50.6	1440	93.2
Foshan	9.5	3798	172.1	2501	95.1
Huizhou	6.1	11,347	74.5	537	74.6
Dongguan	10.5	2460	157.3	4268	91.2
Zhongshan	4.5	1784	58.9	2522	95.3
Jiangmen	4.8	9505	54.2	505	71.4
Zhaoqing	4.2	14,891	40.6	282	65.2
GBA Total	86.6	56,095	1912.9	1544	88.7

Note: GDP figures are converted to USD based on 2023 average exchange rates. Source: China Statistical Yearbook 2023, Hong Kong Census and Statistics Department 2023, Macau Statistics and Census Service 2023, and Guangdong Provincial Bureau of Statistics 2023.

). The region has significant socio-economic diversity. GDP per capita ranges from USD 49,800 in Hong Kong to USD 9667 in Zhaoqing. Urbanisation rates vary from 100% in developed centres to 65.2% in emerging areas. The uniqueness of the GBA as an area of study lies in the conjunction of diversity and geographic coherence. It offers several urban development models within a consistent subtropical climate framework, facilitating comparative analysis of architectural interventions under similar environmental conditions.

Additionally, the area being designated as a national demonstration zone provides a favourable legislative environment for sustainability initiatives. This includes regulation systems that promote innovative ways of developing cities. The GBA’s building stock includes old buildings that require retrofits to highly sophisticated projects with state-of-the-art sustainability features. This gives a wide range to examine for carbon reduction methods. This outstanding geographic region combines economic importance, urban variety, and developmental diversity within a unified climatic context. These characteristics make it an ideal location for studying the relationship between architectural practices, energy performance, and carbon neutrality goals. This can contribute to China’s larger climate effort and serve as a model.

3. Methodology

The study uses various methods such as quantitative modelling, simulation analysis and comparison to check the suitability of green buildings for the Greater Bay Area’s carbon neutrality goals under changing climate conditions. Data collection techniques and synthetic data creation are needed to collect complete details about energy usage in buildings and fill in gaps when real building data is not available. Green building performance can be examined in detail through simulation with EnergyPlus 9.4 and DesignBuilder 7.0, which apply climate models for different weather scenarios both today and in the future [24]. Using LightGBM, machine learning is used for statistical analysis to determine which design aspects strongly impact energy efficiency and to show relationships between building features and their energy performance. The analysis of both green and traditional buildings in various cities in the GBA confirms lower energy use, a decrease in carbon emissions and more comfortable temperatures. The impact of current green building policies and incentives on reaching carbon neutrality is checked through the structured Policy Effectiveness Index. Furthermore, the findings are checked with various techniques such as comparing simulated results with real observations, interviewing stakeholders and carrying out sensitivity analyses in different climate and operational situations. Figure 2 clearly demonstrates the Section 3.

3.1. Data Collection and Fabrication

The study gathered information from several sources to properly study energy use in buildings across the Greater Bay Area. The energy consumption data of the buildings were obtained from government publications and green building databases in Shenzhen, Hong Kong and Guangzhou. The main measurements were annual Energy Use Intensity (EUI) in kWh/m²/year, how HVAC systems are laid out, the overall value of all insulation layers, information about the windows and their share of wall area, solar panels on rooftops and breakdowns of when people occupy the property. Data on climate conditions for the decade from 2015 to 2024 were acquired using weather reports and the CORDEX-EASIA East Asia database. Details measured include temperature in degrees Celsius, rainfall in millimeters, the percentage of humidity and solar radiation values in kWh/m². Interviews with representatives of different stakeholders were also held to learn about green building trends and how well they perform there.

Artificial datasets were created to continue the research if the existing building data was private and could not be shared. These buildings were intentionally constructed to achieve energy expenses between 50–120 kWh/m²/year, reduce carbon emissions by 10–40% relative to conventional buildings, and get interior environmental quality scores ranging from 70 to 95. The use of synthetic data resulted in models and simulations that behaved the same as real buildings, which allowed accurate comparisons during all stages of work. This enabled the research to use statistical data accurately throughout the area while guaranteeing data privacy, making it possible to perform energy efficiency reviews and simulations in different areas of the Greater Bay Area.

3.1.1. Definition of Green Buildings and Sample Selection

In this research, a green building is considered to be a building that has been certified according to a sustainability standard recognized internationally or nationally, namely the China Green Building Evaluation Standard (Three-Star Rating), LEED Gold, or BREEAM Very Good. LEED is a well-known standard used to measure the environmental and energy features of specific buildings in the context of the global sustainable construction debate. It allows systematic assessment of variables including energy efficiency, carbon emissions, indoor environmental quality and sustainable site strategies by providing a verifiable, systematically defined framework. In the current research, the authors utilize LEED with the assistance of the BREEAM and the Three-Star Green Building Standard of China to establish that the chosen case studies meet the globally accepted performance levels. The study, therefore, does not introduce a new urban scale assessment system but looks into the performance of the existing LEED-certified buildings in the subtropical setting of the Greater Bay Area.

Each building’s certification means that it was designed to save at least 30% of energy compared to national baseline building codes. This was conducted by using a wide range of techniques, such as passive architectural design solutions (like optimizing orientation, natural ventilation, daylighting, and high-performance envelopes), high-efficiency active systems (like HVAC, lighting), and measures to make buildings more resilient in subtropical climates (like typhoon-resistant envelopes and rainwater-management systems).

The sample includes 32 green buildings located within Shenzhen, Hong Kong, and Guangzhou, thus encompassing the technological variety of the Greater Bay Area (GBA). It encompasses a variety of building types constructed between 2010 and 2024, including high-rise residential towers, mixed-use complexes, commercial office buildings, and institutional buildings. The gross floor area is between 2000 and 50,000 m², and the structural system is reinforced concrete frame, steel-composite frame and hybrid.

The current study used a mixed-methods approach to combine empirically collected data from 32 certified buildings with artificial datasets to provide sufficient sample sizes for machine learning-based analysis and statistical soundness.

3.1.2. Statistical Power Analysis

The G*Power 3.1.9.7 was used to perform a power analysis to identify the minimum sample size required to reveal salient energy-performance differences between Green and conventional buildings. Based on:

• Effect size (Cohen’s d) = 0.8 (large effect expected based on pilot data)
• Statistical power (1 − β) = 0.80
• Significance level (α) = 0.05
• Two-tailed test

Power analysis curves are provided to demonstrate the correlation between the sample size and the statistical power in differentiating the energy performance of green buildings and conventional buildings. The analysis presupposes that the effect size Cohen’s d = 0.8, the level of significance (α) = 0.05 and the difference will be tested on a two-tailed basis (Figure 3). The dashed vertical line represents the minimum sample size (n = 26 per group) necessary to achieve 80% statistical power. This research indicated that a set of twenty-six buildings would provide adequate statistical power for t-test comparisons.

3.1.3. Detailed Building Classification and Stratification

The data collection includes 32 case studies, with 16 representing green buildings and 16 representing conventional ones (

Table 2. The 32 buildings were systematically classified and stratified as follows.

Building Type	Green Buildings	Conventional Buildings	Total	Balance Ratio	Construction Period	Size Range (m2)
High-rise Office	6	6	12	1:1	2015–2024	8000–25,000
Mid-rise Office	4	4	8	1:1	2012–2022	3000–12,000
Residential Tower	3	3	6	1:1	2014–2023	15,000–45,000
Mixed-Use	2	2	4	1:1	2016–2024	10,000–30,000
Institutional	1	1	2	1:1	2018–2020	5000–8000

). This ensures that statistical validity can be maintained when comparing the two. The distribution of green and conventional buildings is 50:50 within each typology, which removes any bias that could be introduced by the distribution of building types. The 1:1 matching process allows strong paired comparison analysis where any performance differences can be safely said to be due to green building interventions as opposed to typological differences (Figure 3).

3.1.4. Geographic Distribution

• Shenzhen: 12 buildings (6 green, 6 conventional)
• Hong Kong: 10 buildings (5 green, 5 conventional)
• Guangzhou: 10 buildings (5 green, 5 conventional)

The sample of buildings was designed to recreate the makeup of the building stock of the Greater Bay Area as reflected in 2023 construction data. Commercial and institutional building stock (>2000 m²) in the region is estimated at about 35 percent high-rise office buildings, 22 percent mid-rise office buildings, 20 percent residential towers, 15 percent mixed-use developments and 8 percent institutional buildings, according to the GBA Urban Planning Authority and the Hong Kong Buildings Department (Figure 4). The proposed sample distribution, which has been assigned as 37.5% high-rise office, 25% mid-rise office, 18.75% residential towers, 12.5% mixed-use and 6.25% institutional buildings, is slightly out of the regional profile, as the sample size is discrete.

3.1.5. Cross-City Building Type Distribution Reflects Urbanization Patterns

• Shenzhen (high-density commercial hub): 60% office buildings, 25% residential towers, 15% mixed-use
• Hong Kong (mature dense urban): 50% office buildings, 35% residential towers, 15% institutional/mixed-use
• Guangzhou (balanced metropolitan): 45% office buildings, 30% residential towers, 25% mixed-use/institutional

The commercial-to-residential ratio of the proposed sample (62.5 percent commercial, 18.75 percent residential and 18.75 percent mixed-use/institutional) is aligned with the GBA building energy consumption profile, in which commercial buildings represent about 65 percent of the total building energy consumption in the area.

3.1.6. Sampling Methodology

Buildings were selected using stratified random sampling within each city, ensuring representation across:

• Building typology (proportional to GBA building stock)
• Construction vintage (post-2010 green building standards)
• Size categories (representative of commercial/institutional buildings)
• Geographic distribution (proportional to urban development patterns)

3.2. Building Energy Simulation with Climate Models

Green building energy simulations were tested with EnergyPlus and DesignBuilder to determine how well they performed in GBA’s climate. The new technologies allowed simulation models to make use of data on dynamic weather from regional climate models. Information regarding insulation, the building’s direction, the type of glass chosen, the efficiency of heating and cooling and passive ventilation was collected.

3.2.1. Simulation Dataset for Machine Learning

The 32 empirical structures utilized in the machine-learning process for analyzing building energy performance were deemed inadequate; thus, a more extensive dataset of 300 simulations was created based on identical feature distributions. Standard procedures in the research of building energy, where synthetic data is frequently produced to get around the drawbacks of data scarcity, are reflected in this approach.

Synthetic Data Generation Process:

i.

Parameter Distribution Analysis: Statistical distributions of key parameters (envelope properties, HVAC efficiency, window-to-wall ratios) were derived from the 32 empirical buildings

ii.

Monte Carlo Sampling: Latin Hypercube Sampling was used to create 300 artificial structures in order to guarantee representative coverage of the parameter space.

iii.

Validation: A subset of synthetic buildings was validated using empirical data to confirm realistic performance characteristics.

The expanded 300-building dataset was used exclusively for machine learning analysis, while the 32 empirical buildings were used for comparative performance analysis and validation. The scatter plot shows a very high correlation (R² = 0.92) between synthetic and empirical values of energy performance (Figure 5). The 95% confidence intervals (shaded area) assure the validity of the synthetic data generation process to be used in further analysis in a machine-learning framework.

3.2.2. Synthetic Data Validation

The synthetic data validation procedure used rigorous statistical test procedures to validate representativeness, as well as to manage systematic biases. A total of 16 buildings (50% of the empirical dataset) were randomly taken as a primary validation set, and 8 buildings were separately kept out as additional testing basis. All the 32 empirical buildings were assessed by leave-one-out cross-validation. Testing the statistical significance of the correlations between the synthetic and the empirical data showed the robust correlations: Pearson correlation coefficient r = 0.96 (p < 0.001, n = 32), Root Mean Square Error (RMSE) of 3.2 kWh/m²/year (4.8% of mean EUI), and Mean Absolute Percentage Error (MAPE) of 4.1% ± 1.3%. The residual analysis showed that errors were normally distributed (Shapiro-Wilk test: W = 0.97, p = 0.43), and that the 95% confidence intervals between the errors ranged between ±2.1 to ±4.7 kWh/m²/year across building types, meaning that the process of generating synthetic data used in machine learning is valid

This simulation research was based on a substantial number of 300 buildings from the Greater Bay Area, comprising the most frequent architectural typologies seen in the region. The size of the buildings that were modelled varied between 2000 m² and 50,000 m²; most of them (65%) were in the 5000–15,000 m² range, a size profile common to commercial and institutional buildings. The date of the construction was 2010–2024, and 40% of the dataset consisted of the buildings constructed after 2018 to be in line with the modern requirements of green buildings and technologies.

Dynamic weather data were included with the models based on regional climate models, and the envelope material attributes were detailed, such as insulation thermal properties (R-values ranging between 2.5 and 8.0 m²·K/W), building orientation optimization, window-to-wall ratio (15% to 60%), glazing performance characteristic (U-value ranges between 1.2 and 3.5 W/m²·K), and passive ventilation strategies.

Different climatic conditions were tested, including both average typical meteorological years (TMY) and future climate assumptions from RCP 4.5 and RCP 8.5. Overall, the models measured energy use, solar energy the buildings could produce, how comfortable the indoors would be and how temperature control was possible. The core equation used to determine the energy balance of a building was:

$$Q_{\mathrm{total}}=Q_{\mathrm{solar}}+Q_{\mathrm{internal}}-Q_{\mathrm{losses}}$$

where $Q_{\mathrm{solar}}$ represents solar heat gain, $Q_{\mathrm{internal}}$ denotes heat generated by occupants and equipment, and $Q_{\mathrm{losses}}$ is the total thermal loss through walls, roofs, windows, and ventilation.

From simulation outputs, the Energy Use Intensity (EUI) was calculated as:

$$EUI={\displaystyle \frac{E_{\mathrm{annual}}}{A_{\mathrm{floor}}}}$$

where $E_{\mathrm{annual}}$ is the total annual energy consumption (in kWh) and $A_{\mathrm{floor}}$ is the gross floor area (in ${\mathrm{m}}^2$). This metric provided a standardized measure for comparing energy performance across different buildings.

3.2.3. Spatial Downscaling Methodology for Climate Projections

The hierarchical downscaling method (combining statistical and dynamical downscaling techniques) was applied to derive spatially explicit EUI predictions at RCP 8.5 scenarios across GBA sub-regions. The methodology combined both the CORDEX-EASIA regional climate model outputs (25 km resolution) and the local information at meteorological stations to come up with high-resolution (1 km) climate projections by jurisdiction. Statistical downscaling with bias correction was performed with the quantile mapping method based on observed weather data at 15 stations in Shenzhen, Hong Kong, and Guangzhou (2010–2020) to correct projected temperature and humidity. The inverse distance weighting spatial interpolation was used with the correction of elevation that incorporated the topographical influence on the microclimate conditions.

Sub-regional Climate Scaling Framework:

• Urban Core Zones: Applied +1.2 °C urban heat island correction to base RCP 8.5 projections
• Coastal Areas: Incorporated maritime temperature moderation (−0.8 °C summer, +0.3 °C winter)
• Suburban Periphery: Used unmodified downscaled projections as baseline reference

The spatial climate data was used to calculate building-specific projections of EUI based on the supported LightGBM model, including the cooling degree days and the humidity ratios specific to the location. The 2050 RCP 8.5 EUI projections were spatially varying, with urban core urban areas at 89.7 kWh/m²/year, coastal regions at 82.1 kWh/m²/year and suburban regions at 84.6 kWh/m²/year, with the quoted 85.3 kWh/m²/year being the area-weighted average across all GBA building stock. The use of Monte Carlo simulation (n = 1000) to quantify uncertainty gave 95 percent confidence quantifications of ±4.2 kWh/m²/year of the regional projections, which demonstrated the robustness of the downscaling methodology in policy plan applications.

3.3. Machine Learning-Based Statistical Analysis

3.3.1. Statistical Validation and Cross-Validation Framework:

To ensure model reliability with the expanded dataset:

Dataset Splitting:

• Training: 240 buildings (80%)
• Validation: 30 buildings (10%)
• Testing: 30 buildings (10%)

Cross-Validation Strategy:

• 5-fold stratified cross-validation
• Stratification based on building type and size to ensure balanced representation
• Bootstrap sampling (n = 1000) for confidence interval estimation

Model Performance Metrics:

• R² (coefficient of determination)
• RMSE (Root Mean Square Error)
• MAE (Mean Absolute Error)
• MAPE (Mean Absolute Percentage Error)

Statistical Significance Testing:

• Permutation tests for feature importance significance
• Bootstrap confidence intervals for model predictions
• Paired t-tests for comparing model performance across building types

The figure shows five-fold cross-validation outcomes of a given machine learning model in the form of box plots of the measurements R², RMSE, MAE, and MAPE. The close dispersion of the boxes implies low heterogeneity of folds and confirms the conclusions about the stability and reliability of the model (Figure 6). The medians, which are close to the modal values and the interquartile ranges, which are not very wide, further support the strength of the predictive ability of the model.

3.3.2. Hyperparameter Tuning and Model Optimization

The hyperparameter search was carried out using Bayesian optimization and 100 iterations in total to find the best configuration. The last parameters were as follows: (1) learning rate = 0.05; (2) max depth = 8; (3) number of estimators = 500; (4) subsample = 0.8; and (5) feature fraction = 0.9. This set of parameters was selected to trade off complexity and generalizability on the building energy prediction task. A number of measures were included to avoid overfitting, including early stopping (patience = 50 rounds on validation loss), L2 regularization (λ = 0.1), and thresholding of feature importance to eliminate variables that had less than 1% influence on model performance. Learning curves were calculated during the training process to confirm the convergence and exclude the overfitting tendency.

3.3.3. Dataset Integration and Training Protocol

The validation datasets consisted of 32 empirical buildings, whereas the training dataset was a synthetic dataset of 300 buildings, generated by sampling the statistical distributions of the empirical buildings. The validation processes involved separate analysis of the datasets to determine the accuracy of synthetic data, but were merged during training to give ample sample size to ensure good machine learning performance. The stratified sampling provided proportional representation of building types and geographic regions in the training (80%), validation (10%) and testing (10%) sets. The empirical buildings were divided among all three sets to maintain the integrity of validation, and synthetic buildings were used to enlarge the volume of training data and improve model generalization.

To select the most important factors affecting energy efficiency in architecture and design, we employed the high-quality tree-based LightGBM machine learning system. LightGBM was chosen for this study since it can discover detailed and complicated bonds between several factors in a building’s design and its energy usage. Traditional linear regression models are not as effective as LightGBM when it comes to understanding how factors like insulation, window coverage, layout existence, and HVAC systems interact and affect a building’s energy usage. The algorithm is very useful for architectural analysis because it directly supports the research goal, removes bias and helps understand which factors have the biggest effect on energy efficiency. LightGBM shows faster computation and reliable results for datasets the size of the 300 building cases in this study, and, because it is tree-based, it can process data with different types of features—such as category (HVAC) and number (insulation)—without much extra preparation. The most important point is that the model is easy to explain, letting researchers determine not only what is important but also how it affects energy use in architecture. The method demonstrates superior performance in experiments, particularly with small-to-medium sized datasets, because of its reduced susceptibility to overfitting, which is often a challenge in time-consuming and costly data gathering processes.

The importance scores of the features were determined with LightGBM’s inbuilt gain-based importance, which calculates the overall decrease of the loss function when splitting on each feature by all trees. All the importance scores were normalized to add to 100 percent and averaged out over the entire 500 decision trees in the ensemble. In order to provide stability, the ranking of importance was confirmed with the help of permutation importance testing, so that each feature could be randomly shuffled 100 times to determine the impact on the performance. The features with the rank of importance that were consistent in all iterations of 5-fold cross-validation (correlation coefficient > 0.85) were regarded as statistically significant.

SHAP (SHapley Additive exPlanations) values were used to analyze the feature interaction and measure the extent to which the features interact to affect the energy performance. The three identified interactions that were found useful are: (1) HVAC efficiency × Building envelope insulation (interaction strength = 0.12), (2) Window-to-wall ratio × Glazing performance (interaction strength = 0.089), and (3) Shading design × Orientation (interaction strength = 0.076). The interactions provide an extra 8.3 percent of the model performance over individual feature contributions.

The LightGBM model computed EUI using the aforementioned inputs, enabling us to determine which traits were most important. A portion of the data was assigned as training (80%) and the remaining part as testing (20%). The data was split into five folds to improve the overall accuracy. Sensitivity analysis was conducted to find out how various design factors influence the EUI. The sensitivity of each parameter was measured using the formula:

$$S_i={\displaystyle \frac{\Delta EU{I}_i}{\Delta x_i}}$$

where $\Delta x_i$ is the change in the $i$-th design variable and $\Delta EU{I}_i$ is the corresponding change in energy use intensity. The design elements that are most successful in lowering energy consumption and enhancing building efficiency under various climate conditions were revealed by this investigation.

3.4. Comparative Case Analysis of Green and Conventional Buildings

A comparative analysis of significant structures in Shenzhen, Hong Kong, and Guangzhou was conducted to contextualise the findings. Performance was compared for each case by looking at EUI, emissions of operational carbon, the effectiveness of temperature regulation and how much the building could save in costs.

A Guangzhou green office building uses only 65 kWh of energy per square meter every year, much less than the 120 kWh used by the nearby typical office. Similarly, green buildings saw a drop in carbon emissions from 5.8 to 2.5 tonnes each year, and there were smaller changes in temperature inside (22–26 °C) compared to 20–30 °C in non-green buildings during summer. The study examined the results and proved that green roofs, shading for the sun and energy-saving HVAC systems help in climate adaptation.

3.5. Policy Effectiveness Evaluation

To evaluate how existing building and energy policies contribute to carbon neutrality, a Policy Effectiveness Index (PEI) was formulated. This index incorporated three main criteria: emission reduction effectiveness, cost-efficiency, and compliance rates.

Emission Reduction Effectiveness (Re)

Traditional carbon emission mitigation indicators are more likely to emphasize possible mitigations based on policy modelling. Conversely, this indicator is a measurement of the actual carbon emissions that are reduced by implementing a policy in quantifying the percentage decrease of emissions within a certain sector compared to an established 2015 baseline, therefore, representing the embodied and operational carbon within the entire building lifecycle.

Cost-Effectiveness (Ce)

Cost-effectiveness measures the economic efficiency of policy, in terms of energy saved per dollar of financial investment (kWh saved per dollar invested), including direct government spending as well as the investments made by the private sector due to policy action.

Compliance Rate (Cr)

The compliance rate offers a thorough assessment of market change and policy adherence, defined as:

$$\mathrm{C}\mathrm{r}=(\mathrm{N}\mathrm{c}/\mathrm{N}\mathrm{t})\times 100$$

where:

• Nc = Number of new buildings achieving green certification standards (minimum LEED Gold, BREEAM Very Good, or equivalent local certification) within the assessment period (2020–2024)
• Nt = Total number of newly constructed structures that must adhere to green building regulations and have a gross floor area of more than 2000 m²

The compliance rate monitors the number of buildings that adhere to or surpass the required green building standards stipulated by local authorities in Hong Kong, Shenzhen and Guangzhou using post-occupancy evaluation and energy audit that is conducted within a period of two years of the construction.

Expert Consultation Process: The weights (W₁, W₂, W₃) attributed to the PEI formula were determined during a systematic process of consultations with experts that was implemented in a sample of 15 domain experts in the Greater Bay Area. The expert panel consisted of green building certification experts, urban planning representatives of the municipal governments of Shenzhen, Hong Kong, and Guangzhou, energy policy scholars and building industry professionals with at least 10 years of experience in sustainable construction in subtropical climatic conditions.

Weight Derivation Method: Expert consensus on relative importance weightings was elicited using a two-round Delphi survey. Respondents used pairwise comparison matrices according to the Analytic Hierarchy Process (AHP) to prioritize three criteria (emission reduction effectiveness, cost-effectiveness, and compliance rate) in the first round. The expert panel further adjusted these rankings during the second round by incorporating systematic feedback and consensus-seeking discussions that focused on GBA-relevant policy priorities and time-frame limits of carbon neutrality.

Final Weight Configuration: According to expert opinion, W₁ = 0.5 (emission reduction effectiveness), W₂ = 0.3 (cost-effectiveness), and W₃ = 0.2 (compliance rate). The high weighting accorded to the effectiveness of the emission reduction is indicative of the high carbon neutrality timeline of the region, but the cost-effectiveness has been weighted heavily since the region is still undergoing economic growth priorities.

Sensitivity Analysis: The sensitivity analysis was carried out to determine whether the assignments of weight were reliable by using three alternative weighting patterns: equal weighting (0.33, 0.33, 0.33); cost-priority weighting (0.2, 0.5, 0.3), and compliance-priority weighting (0.3, 0.2, 0.5). The ranking of policies was the same regardless of the weighting scenarios, and this was a confirmation of the consistency of the PEI methodology.

The formula used to compute PEI was:

$$PEI={\displaystyle \frac{\left(R_e\cdot W_1+C_e\cdot W_2+C_r\cdot W_3\right)}{W_1+W_2+W_3}}$$

where:

• $R_e=$ emission reduction rate (%)
• $C_e=$ cost-effectiveness (energy saved per monetary unit)
• $C_r=$ compliance rate (% of new constructions certified green)
• $W_1,W_2,W_3=$ respective weights assigned based on expert consultation

Using this method, it became possible to analyze specific local policies, including green construction programs, energy policies and set certification standards. Quantitative research and discussions with people involved in the projects were combined to evaluate if the initiatives were being put to good use.

3.6. Validation and Quality Control

The results were made trustworthy using several quality control methods. Whenever possible, the data produced from simulations were compared to actual observations. By combining interview information with numbers, the conclusions became more reliable. In addition, different weather and building schedules were tested to explore how they influenced the analysis. By using modelling, statistical data and assessments, the method investigates green buildings’ impact on reaching carbon neutrality in the Greater Bay Area. Sustainable city explanations and solutions are offered by utilising both simulated data, real-world scenario-based research, and forecasting techniques in various climatic scenarios.

3.6.1. Enhanced Quality Control Measures

Careful quality-control measures were established in order to maintain reliability and to address possible uncertainties of the synthetic-data modelling approach. The uncertainty quantification was based on the Monte Carlo method of 10,000 iterations to propagate parameter uncertainty, bootstrap confidence intervals giving 95% CI of all the main performance indicators and a sensitivity analysis based upon a variation of parameters of ±20% of the most influential parameters to test the robustness of the results. The framework for the actual implementation included quarterly validation by constant comparison with operational data of the buildings that were involved, performance monitoring through the integration of building management systems, and the establishment of corrective factors through the use of empirical gaps in performance. Expert validation encompassed a technical review panel of 12 industry specialists assessing model assumptions and outcomes, peer validation via three independent research groups employing diverse methodologies, and stakeholder feedback from building operators and owners who offered operational reality checks to confirm the practical significance of the results.

3.6.2. Synthetic Data Limitations and Potential Biases

The study recognizes a number of possible biases that can influence the real-world applicability of results to building performance. Selection bias: certification bias (green buildings are mostly LEED Gold or above and may systematically overrepresent high-performance buildings) and geographic concentration (urban-based buildings may fail to capture suburban or rural building performance patterns). Synthetic data generation biases have parametric assumptions that can be oversimplified by the interaction of complex building systems through linear relationships of parameters in design. Simulation constraints also lead to an overestimation of actual building efficiency by 10–30% with respect to modelled performance, caused by unmodelled operational aspects. The study attempts to counter these shortcomings by using a variety of validation methods such as cross-validation and sensitivity analysis, conservative estimates of performance with a range of uncertainty, and qualitative validation through stakeholder interviews to give a valid interpretation of the results.

4. Results

4.1. Energy Performance Analysis

All the comparisons of green and conventional buildings were standardized to building typology, gross floor area (GFA), structural system, and climatic exposure to study the effect of structural and functional attributes on energy consumption. Only the structures that were functionally similar were compared in pairs, e.g., commercial office building, residential tower or institutional building, to ensure that the operations patterns did not skew the outcomes. Control of structural systems was also performed; reinforced concrete buildings and steel-composite frame buildings were compared in the same category to consider variations in thermal mass, material conductivity and envelope performance.

The application of a standard metric known as Energy Use Intensity (EUI, kWh/m²/year) was carried out in the case of differences in gross floor area in order to facilitate comparisons between different samples. Weather files were used to match climatic exposure based on the location of the buildings, thus providing comparative ambient conditions in buildings. This method took out any effects that could have been confusing and made sure that the difference in energy use, carbon emissions, and the stability of the indoor climate was actually due to the design, technology, and operations of green buildings and not to things like building size, construction method, or use.

The amount of energy used by buildings in the Greater Bay Area (GBA) was much lower for green structures than for conventional ones. The analysis of data from 32 buildings in Shenzhen, Hong Kong and Guangzhou pointed out that green buildings are more energy efficient than conventional structures (

Table 3. Energy Use Intensity Comparison Between Green and Conventional Buildings in the GBA.

Building Type	Location	Mean EUI (kWh/m2/year)	Standard Deviation	Sample Size (n)
Green Office	Shenzhen	67.3	8.2	7
Conventional Office	Shenzhen	118.5	12.6	6
Green Office	Hong Kong	72.4	9.7	5
Conventional Office	Hong Kong	124.2	13.1	4
Green Office	Guangzhou	65.1	7.9	6
Conventional Office	Guangzhou	120.8	14.3	4

Source: Author’s statistics.

). The results show that green buildings use much less energy. Energy use intensity in Shenzhen office buildings was 67.3 kWh/m²/year for green buildings and 118.5 kWh/m²/year for conventional buildings. Green buildings were more efficient in Hong Kong (72.4 vs. 124.2) and Guangzhou (65.1 vs. 120.8), highlighting their consistently higher efficiency. Green buildings were less variable, with samples performing more equally than other types.

The analysis using two-sample t-tests proved that the energy efficiency gap between green and conventional buildings was significant in all three cities (p < 0.001), with green buildings saving about 45.3% more energy on average. The paired t-tests established that there were significant differences between green and conventional buildings regarding all the cities (t = −12.7, df = 31, p < 0.001, Cohen’s d = 2.3, large effect size). There was no significant difference between the city results in the cross-city ANOVA performance of green building between the cities (F(2,29) = 1.4, p = 0.26), which assures homogeneous efficiency throughout the GBA area.

Figure 7 presents a bar chart with data on the difference between actual energy use and the expected use in Buildings A through E. A dual-bar design shows the current energy used, given in blue bars and the predicted energy, shown in green. The actual consumption of Building A was 245 kWh/m², a little higher than the forecast 252 kWh/m² value. It reveals that the model has overestimated by 2.9%. Consumption at Building B was measured at 187 kWh/m², while it was predicted at 195 kWh/m^2, which amounts to a 4.3% error in estimating. The figure for energy use from Building C is highest at 320 kWh/m², which is 5% higher than the model’s anticipated 310 kWh/m². These results are 3.1% lower than the actual amount, showing that this prediction is quite accurate. Building D used the least energy, at 156 kWh/m², while the model thought it was 165 kWh/m², making the error 5.8%. The actual energy consumption in Building E was 278 kWh/m², which is 2.9% lower than the predicted figure of 270 kWh/m². Based on the results, the prediction model is consistently accurate among different building types, making errors between 2.9% and 5.8%. It is clear from the data that overestimation occurs in most cases, so there may be ways to improve model calibration.

Data analysis indicated that HVAC was responsible for the largest amount of energy used in both types of buildings (42.7% in conventional buildings and 32.5% in green buildings). Green buildings experienced stronger improvements in HVAC efficiency, mostly because of better control systems, thicker insulation and passive cooling. Luminaires reduce the amount of lighting energy used by 38.2% (p < 0.001) when compared with conventional buildings.

The methodology used in the analysis shows that regardless of the climate, green buildings will always have better energy performance, with the level of improvement varying depending on circumstances. Green buildings surpass conventional buildings in both situations. Current Typical Meteorological Year (TMY) data were used as a baseline, and projections were generated of moderate-emission (RCP 4.5) and high-emission (RCP 8.5) climate pathways in 2030 and 2050. The projection of rising temperature regimes indicates that conventional buildings will need 24.9 percent and 19.5 percent additional cooling in 2050 under RCP 8.5 and green buildings, respectively, hence corroborating the current research on the energy resilience of green architecture. Furthermore, projections plotted a series of energy-demand pathways between 2020 and 2050, indicating that baseline cases might record a rise of up to 95% during this period. The modelling, therefore, illustrates that green buildings have an efficiency advantage of 42.6–43.6% reduced EUI under all climate cases, which shows a fairly steady climate resilience superiority. The 2050 RCP 8.5 scenario states that green buildings will consume 85.3 kWh per square meter per year, but conventional ones could use up to 148.7 kWh per square meter per year (

Table 4. Projected EUI Under Different Climate Scenarios (Average across building types).

Building Type	Current TMY	2030 (RCP 4.5)	2030 (RCP 8.5)	2050 (RCP 4.5)	2050 (RCP 8.5)
Green Buildings	68.3	73.6	76.2	78.9	85.3
Conventional Buildings	121.2	130.5	134.8	139.2	148.7
% Difference	43.6%	43.6%	43.5%	43.3%	42.6%

Source: Author’s statistics.

). Green buildings, it turns out, are proven to be more climate resilient since their energy demand grows at a slower pace than other types.

Levels of cooling required for conventional buildings under RCP 8.5 in 2050 grew by 24.9%, compared to 19.5% for green buildings, suggesting green buildings are slightly better able to handle a rising climate (p = 0.042).

4.2. Climate Adaptability Assessment

The study consists of a multi-line graph that tracks predicted energy demand through the years 2020–2050 in three scenarios (Figure 8). An orange line represents the baseline scenario, showing how consumption increases from 100 in 2020 to 195 in 2050. This means energy consumption grew about 95% over those 30 years. The increase reflects the expected effect of climate change on how much energy is used in buildings, mainly because of a greater need for cooling and more extreme weather. The approach shown by a green line radically modifies the path of growth. Carbon dioxide output in 2020 is 100, but only about 135 by 2050. Compared to the baseline period, this leads to a rise of 35% and a decrease of 31% in energy usage at the end of the forecast. This second method, represented by a blue line, brings about a bigger reduction in energy demand with a 25% increase from 2020 and a 36% decrease from the starting baseline scenario. All three scenarios progress similarly until about 2025, but a significant difference between them appears later, suggesting that the usefulness of adaptation rises with time and is most valuable in the latter part of the projection.

The study looked at whether GBA buildings are adapted effectively to the region’s climate, which includes hot weather, a lot of humidity and events like typhoons.

Table 5. Indoor Environmental Quality Metrics During Summer Season.

Parameter	Green Buildings	Conventional Buildings	Significance
Indoor Temperature Range (°C)	22.8–26.3	20.1–30.5	p < 0.001
Relative Humidity Range (%)	45–62	35–70	p < 0.01
PMV Range (Thermal Comfort)	−0.5 to +0.5	−1.2 to +1.3	p < 0.001
Air Quality (CO2 ppm)	450–750	550–1200	p < 0.001

Source: Author’s statistics.

describes the results of Indoor Environmental Quality (IEQ) measures in the summer. The findings reported interior temperature, relative humidity, thermal comfort and air quality as shown by CO₂ levels. Green-covered locations were much more comfortable and reliable. Temperatures inside green buildings varied from 22.8 °C to 26.3 °C, while indoor temperatures in traditional buildings fell from 20.1 °C to 30.5 °C. Green buildings kept relative humidity between 45 and 62%, compared to the wider range of 35–70% in conventional buildings. Studies show that green buildings usually have less CO₂ per square meter (450–750 ppm) than traditional buildings (550–1200 ppm). Thermal ratings for green buildings were mostly comfortable (−0.5 to +0.5), whereas in conventional structures, people usually experienced thermal discomfort. Buildings designed for sustainability were able to maintain good air and lighting conditions during storms. Changes in temperature inside green buildings were ±1.8 °C during July and August, but ±5.2 °C in conventional buildings (p < 0.001).

Thermal performance of buildings should be analyzed collectively and include cool building approaches like green roofing that enables evapotranspiration cooling, high-albedo, reflective coatings that reduce the intake of solar heat gain, natural ventilation systems, primarily the stack effect and cross ventilation, and interior thermal-mass elements, such as exposed concrete floors, thick masonry walls, and phase-change materials that can be used to store and release thermal energy, moderating diurnal temperature changes. Table 4 helps to determine how these systems support the higher temperature stability recorded in green buildings (delta of 1.8 degrees Celsius) compared to conventional buildings (delta of 5.2 degrees Celsius) over the monitoring period of summer. This stability was achieved by enhanced building insulation, increased utilization of interior mass, and dependence on natural methods for maintaining moderate inside temperatures.

Green structures were better able to keep their indoor conditions stable during unfavorable weather. The IEQ measurements during the summer are taken into account in Table 4. The measurements included indoor temperature, humidity, the PMV score for thermal feeling and the levels of CO₂ in indoor air quality. Buildings with more green spaces were generally more pleasant and secure spaces. The average indoor temperature of green buildings was between 22.8 °C and 26.3 °C, and in conventional buildings, it ranged from 20.1 °C to 30.5 °C. Relative humidity was more stable in green buildings, remaining between 45% and 62%, while the same measurement in conventional structures fell anywhere between 35% and 70%. It is common for green buildings to contain lower CO₂ concentrations (between 450 and 750 parts per million) than older-style buildings (between 550 and 1200 ppm). Green buildings maintained good temperature ratings (−0.5 to +0.5), while many normal buildings fell into the uncomfortable range. Green buildings experienced temperature changes no greater than ±1.8 °C in July-August, compared to ±5.2 °C in normal office buildings (p < 0.001). This outcome depended in part on better insulation in the structure, great thermal mass and using passive cooling.

Green buildings with integrated resilience characteristics fared far better during simulated extreme weather events, according to the typhoon resilience evaluation. These building characteristics are strongly correlated with typhoon resilience, which is based on the principles of sustainability in construction. The resilient structural systems, high-performance materials and redundant multi-functional systems inherent in green building practice simultaneously promote environmental performance and increase resilience to extreme weather events. The integrity of the envelope is guaranteed under the stresses imposed by typhoons thanks to the particular focus on high-performance envelopes, such as reinforced wall systems, impact-resistant glass, and superior waterproofing membranes. Additionally, the integration of distributed energy solutions, i.e., solar panels in combination with battery storage, provides the redundancy of backup supplies that enhances resilience statistics. Green buildings have passive design features that allow habitable conditions to be achieved even in the event of mechanical breakdown during severe weather: natural ventilation, thermal mass, and thoughtful orientation. The green building design, holistic approach also considers long-term durability and adaptability, giving rise to construction approaches that naturally produce stronger constructions. Therefore, green building construction is a step forward in design philosophy that makes resilience an essential part of occupant safety, preservation of large investments in sustainable technology, and environmental sustainability of such buildings in altered climatic conditions.

Indicators used to assess how buildings will respond during simulated typhoons and levels of extreme winds (

Table 6. Building Resilience Metrics Under Typhoon Conditions.

Resilience Metric	Green Buildings with Resilience Features	Conventional Buildings	Green Buildings Without Specific Resilience Features
Power Continuity (h)	72.4	8.3	24.5
Envelope Integrity (0–10 scale)	8.7	6.2	7.5
Water Infiltration	8.3	5.8	7.1
Protection (0–10 scale)	1.2	4.5	2.3

Source: Author’s statistics.

). The integrated battery storage (0.8–1.2 kWh/m²) and rooftop solar arrays that occupy 25–35% of the roof area allow for a 72.4-h power supply. Essential loads are supplied by battery systems for 48–60 h, and the overall length is increased to 72.4 h by solar recharging during typhoon circumstances. By prioritizing life safety systems at 0.3–0.5 kWh/m²/day instead of the usual 2.5–3.5 kWh/m²/day, emergency load management lowers consumption by 60–70%.

Green buildings that included backup power and reinforced walls provided electricity for 72.4 h, nearly ten times longer than conventional buildings, which ended up with electricity devoted to this crisis for 8.3 h. The waterproofing of their envelope was rated at 8.1 out of 10 versus 6.1 in traditional structures. It can be seen that green buildings function better than standard structures, even without any special resilience attributes, showing that green building design is always robust. Structures with resilience measures recovered from interruptions in about 1.2 days, compared to conventional buildings that took 4.5 days to return to service.

4.3. Machine Learning Model Insights

Model Reliability and Statistical Confidence

The LightGBM model achieved strong performance metrics with statistical validation:

• Cross-validation R²: 0.87 ± 0.03 (95% CI: 0.84–0.90)
• Test set RMSE: 8.2 kWh/m²/year (95% CI: 7.1–9.4)
• Feature importance stability: Correlation coefficient > 0.85 across all CV folds

Robust model interpretation was demonstrated by statistical significance testing, which verified that the top 6 features (HVAC efficiency through orientation) displayed consistent importance rankings across all cross-validation folds (p < 0.001) (Figure 9). Rankings of the LightGBM model’s features according to their importance, with 95% CIs obtained using bootstrap sampling (n = 1000).Permutation testing is a robust statistical method used to determine if a feature’s importance is statistically significant. The asterisks (*) next to certain features (like HVAC Efficiency, Envelope R-value etc.) indicate that these features have statistically significant importance in predicting building energy consumption. This means these features contribute meaningfully to the model’s predictions beyond what would be expected by random chance.

The R² value of 0.87 from the test dataset shows that the LightGBM model predicted building energy performance with high accuracy, depending on its design and operation. Figure 10 presents a detailed scatter plot that reviews how accurately and reliably machine learning models can predict actual energy consumption. The plot lists actual energy usage on the x-axis and predicted energy usage on the y-axis, and the points for every reading appear as blue circles with sizes that depend on the confidence level of the model. Many orange circles arranged in a diagonal row reflect the best way for y and x values to match, which occurs precisely at y = x. Many of the blue dots show that the model achieves good accuracy for most of the energy consumption values, ranging from approximately 20 units up to 100. The scatter pattern reveals that prediction accuracy varies across different consumption ranges, with some concentration of points showing closer alignment to the perfect prediction line in lower and higher consumption ranges, while there appears to be slightly more variability in predictions for mid-range consumption values between 40 and 60 units. The scatter plot reveals that predictions are more accurate in lower and higher consumption groups and less accurate for those with 40 to 60 units. It can be seen from the different point sizes that as the predictions are more accurate, the model’s confidence scoring is higher. The scoring of points on both sides of the prediction line is similar, which shows that the model is not biased towards making errors in either direction.

The appearance of negative predicted energy consumption values between the lower side of the distribution (Figure 10) could represent buildings that have very high green building standards that tend towards net-zero or net-positive energy performance. These hyper-efficient buildings, which are usually outfitted with large-scale renewable energy systems and advanced passive design techniques, are capable of supplying more energy than they use, so that the model will calculate the net consumption of the site to be negative. Although these findings indicate the shortcomings of the model in terms of generalizing to the ultra-high-performance buildings, they also indicate the existence of increasingly efficient green buildings in the GBA dataset that are close to zero or positive energy balance. Figure 11 describes prediction intervals for the LightGBM model showing actual vs. predicted energy consumption with 95% prediction intervals (shaded region) and 95% confidence intervals (darker band). The narrow prediction intervals and strong correlation demonstrate the model’s reliability for practical applications in green building performance assessment.

Feature importance analysis revealed the following contributing factors to energy efficiency.

Table 7. Statistical Performance Summary of Machine Learning Models.

Model	R2 (Mean ± SD)	RMSE (kWh/m2/year)	MAE (kWh/m2/year)	MAPE (%)	95% CI (R2)	p-Value *
LightGBM	0.87 ± 0.03	8.2 ± 1.1	6.1 ± 0.8	12.3 ± 2.1	[0.84, 0.90]	<0.001
Random Forest	0.82 ± 0.04	9.7 ± 1.3	7.2 ± 1.0	15.1 ± 2.8	[0.78, 0.86]	<0.001
XGBoost	0.84 ± 0.03	9.1 ± 1.2	6.8 ± 0.9	13.9 ± 2.4	[0.81, 0.87]	<0.001
Linear Regression	0.65 ± 0.06	13.4 ± 1.8	10.3 ± 1.4	21.7 ± 3.5	[0.59, 0.71]	<0.01

* p-values from permutation tests comparing model performance against baseline. Bold indicates the performing model. SD = Standard Deviation, CI = Confidence Interval.

provides an overview of the results from a feature significance analysis performed with the LightGBM model. HVAC system efficiency was the key factor among all simulated building designs, as it accounted for a total of 24.3% in the projected EUI values. After that, insulation for the building envelope and the ratio of windows to walls became key, at 18.7% and 14.5%, respectively. The study also considered shading and glazing, which each made up a significant percentage (10.2% and 9.6%, respectively). The data highlights how passive and active systems help achieve better total energy efficiency. The results of the model are robust, showing small standard errors for every characteristic.

SHAP dependency plots indicated that there were important non-linear dependencies among important design parameters. The HVAC-insulation interaction indicated that buildings with high-efficiency HVAC systems (COP > 3.5) experienced a proportionally larger energy benefit associated with insulation improvements (R-value > 4.5 m²·K/W) and thus experienced a synergistic energy savings as high as 15.2 percent more than the individual contributions. There were threshold effects in the window-to-wall ratio: the buildings with WWR < 30% had a diminishing marginal impact of high-performance glazing, and the buildings with WWR 35–40% had the highest impact on the glazing U-values. The Orientation-shading interactions were strongest with the east-west facing buildings, where external shading projection factors > 0.6 offered 23 percent higher energy savings than the north-south orientation. Patterns of feature splitting among trees showed that HVAC efficiency was the leading decision node in 78 percent of trees, and envelope insulation was the second split in 65 percent of trees, proving the hierarchical priority of active and passive cooling measures in subtropical climates.

The HVAC efficiency measure was variable by ±2.1 percent across building types, with a confidence interval of 22.2 to 26.4 percent on feature importance scores. The variation in the performance of HVAC systems added ±1.8 kWh/m²/year to the overall building energy uncertainty.

Table 8. Feature Importance for Energy Performance (LightGBM Model).

Feature	Importance Score (%)	Standard Error
HVAC System Efficiency	24.3	1.2
Building Envelope Insulation	18.7	0.9
Window-to-Wall Ratio	14.5	0.8
Shading Design	10.2	0.7
Glazing Performance	9.6	0.6
Orientation	7.8	0.5
Ventilation Strategy	6.4	0.4
Lighting System	5.2	0.3
Occupancy Controls	3.3	0.2

Source: Author’s statistics.

provides an overview of the results from a feature significance analysis performed with the LightGBM model. HVAC system efficiency was the key factor among all simulated building designs, as it accounted for a total of 24.3% in the projected EUI values. After that, insulation for the building envelope and the ratio of windows to walls became key, at 18.7% and 14.5%, respectively. The study also considered shading and glazing, which each made up a significant percentage (10.2% and 9.6%, respectively). The data highlights how passive and active systems help achieve better total energy efficiency. The results of the model are robust, showing small standard errors for every characteristic.

HVAC system efficiency was the biggest factor affecting EUI, with a 10% improvement leading to a 7.8% decrease in the entire building’s energy use (p < 0.001). The model demonstrated consistent performance (mean R² = 0.85, a = 0.03) after cross-validating it using five-fold validation.

The model identified critical thresholds for building design parameters in the GBA’s subtropical climate.

The analysis of machine learning also showed that certain design factors have strong relationships. The uncertainty of Window-to-Wall Ratio measurements was measured as the precision of ±3.2%, whereas optimum performance ranges were checked at the precision of the measurements up to ±1.5%. The non-compliant buildings with values above 42.5% WWR had exponentially increasing energy penalties (R² = 0.89, p < 0.001).

Table 9. Optimal Parameter Ranges for Green Building Design in GBA Climate.

Parameter	Optimal Range	Energy Impact vs. Baseline (%)
Window-to-Wall Ratio	30–40%	−12.3
Solar Heat Gain Coefficient	0.25–0.35	−8.7
Roof Insulation (R-value)	>4.5 m2·K/W	−10.5
Wall Insulation (R-value)	>3.2 m2·K/W	−9.2
External Shading Projection Factor	>0.5	−7.8
Glazing U-value	<2.0 W/m2·K	−11.3

Source: Author’s statistics.

identifies the main design qualities that help maximize energy efficiency in a subtropical GBA climate. Structures with 30–40% of their walls covered by windows used 12.3% less energy than the usual buildings. Additionally, using a Solar Heat Gain Coefficient (SHGC) of 0.25–0.35 led to a reduction of 8.7% in energy usage. An insulation level above 4.5 m² K/W and glazing U-values of less than 2.0 m² K/W saved significantly on energy use. The study found that combining high-performance glazing and exterior shading led to more than 21.5% savings in energy, which was more than the total amount gained from each factor separately.

4.4. Carbon Footprint Evaluation

GBA examined all types of buildings, looking at both operational and embodied carbon. The operational carbon was figured out from energy and regional emission information, and the embodied carbon was found using the life cycle assessment (LCA) approach.

Table 10. Annual Operational Carbon Emissions by Building Type.

Building Type	Location	Carbon Emissions (kgCO2e/m2/year)	Reduction vs. Conventional (%)
Green Office	Shenzhen	31.2	46.8
Conventional Office	Shenzhen	58.6
Green Office	Hong Kong	33.7	45.5
Conventional Office	Hong Kong	61.8
Green Office	Guangzhou	30.4	47.3
Conventional Office	Guangzhou	57.7

Source: Author’s statistics.

presents the annual carbon emissions from the operations of different buildings in various places. A green building in Shenzhen emits just 31.2 kgCO₂e/m²/year, which is 46.8% less than the 58.6 kgCO₂e/m²/year that traditional buildings emit. The same decrease was seen in Hong Kong and Guangzhou. Most of the reduction comes from more energy-efficient technologies and using less fossil fuel.

There were big differences in carbon emissions between green and conventional buildings at each site (p < 0.001). Much of the difference in carbon reduction among cities was caused by shifts in their energy resources and the emissions from the grid. The total carbon footprint is clearer when embedded carbon is included in the comparison.

Table 11. SO-Vear Life Cycle Carbon Analysis (tCO₂e/m²).

Building Type	Embodied Carbon	Operational Carbon (50 Years)	Total Life Cycle Carbon	Reduction vs. Conventional (%)
Green Buildings	0.63	1.58	2.21	38.4
Conventional	0.52	3.07	3.59

Source: Author’s statistics.

assesses the total amount of carbon emitted over a fifty-year period. The carbon used in the construction of green buildings is slightly higher at 0.63 tCO₂e/m², but they produce much less carbon during operation (1.58 tCO₂e/m²), giving a total life cycle emissions reduction of 38.4%. Moreover, the time it took for savings in operational emissions to match the embodied emissions was just over four years.

Carbon emissions and energy consumption are directly connected yet different performance indicators, which are affected by divergent factors. Energy consumption is the amount of energy consumed by a building during its normal operation, typically measured as kWh/m²/year and is influenced by design characteristics, operating conditions and climatic requirements. Carbon emissions expressed as kgCO₂e/m²/year reflect the greenhouse gas production that accompanies this energy consumption and therefore depend not only on the amount of energy used but also on the carbon intensity of the energy supply. The same electrical load on a building drawing will generate very different emissions when powered by Shenzhen, with its renewable-integrated grid and Hong Kong, with its coal-intensive grid. Further, the embodied carbon of the material production, construction and maintenance is not directly related to operational energy. The awareness of these differences promoted a more comprehensive assessment of the two indicators, providing a more complex understanding of environmental performance and enabling specific approaches to enhancing energy efficiency and reducing carbon emissions.

4.5. Policy Implication Analysis

The success of emission reduction, cost-efficiency, and compliance rates was all taken into account while calculating the Policy Success Index (PEI) for the GBA’s main green building and energy efficiency initiatives.

Table 12. Policy Effectiveness Index for Key Green Building Policies in the GBA.

Policy	Jurisdiction	Emission Reduction Effectiveness (0–100)	Cost Efficiency (0–100)	Compliance Rate (%)	PEI Score (0–100)
Three-Star Green Building Evaluation	Mainland GBA	78.3	65.2	82.4	75.3
System	Hong Kong	72.5	61.8	68.3	67.5
BEAM Plus	Shenzhen	81.6	70.4	91.2	81.1
Green Building Design	Hong Kong	65.3	72.6	95.8	77.9
Regulation (GBDR)	Guangzhou	76.2	58.7	73.1	69.3

Source: Author’s statistics.

examines how successful five main green building and energy efficiency efforts have been in different regions of the Greater Bay Area (GBA). The Policy Effectiveness Index (PEI) is calculated using a mixture of three significant performance measures: lowering emissions, managing costs and following the law. Indicators were each rated from 0 to 100 and were then merged using a process that gave each stakeholder’s significance more weight. The Shenzhen Green Building Design Regulation (GBDR) received the highest PEI score of 81.1, which demonstrates that the rules are enforced well and the technology is quite advanced. Its effectiveness comes from the high compliance rate (91.2%) and its good integration with town planning. The Building Energy Efficiency Ordinance in Hong Kong received a PEI of 77.9 due to its excellent compliance rate (95.8%) and effectiveness in saving costs (72.6%), but its relative greenhouse gas emission reduction was only 65.3%.

The mainland area in the GBA followed the Three-Star Green Building Evaluation System, which scored a PEI of 75.3 due to high compliance (82.4%) and powerful carbon mitigation (78.3). The voluntary Hong Kong certification program BEAM Plus scored 67.5, which reflects its average performance in cost and market impact. The action plan in Guangzhou is very ambitious about cutting carbon but has relatively low scores for compliance and cost, resulting in a PEI of 69.3. Results from the table indicate that including rules with effective enforcement and appropriate financial or digital instruments leads to better results than either voluntary or separate approaches (

Table 13. Certification System Impact Analysis.

Certification System	Market Penetration (% of New Buildings)	Average Energy Savings (%)	Average Carbon Reduction (%)	Cost Premium (%)
China Green Building	35.3	23.7	30.6	4.8–8.3
Label (Three-Star)	27.1	21.5	26.3	5.2–9.6
BEAM Plus (Hong Kong)	12.6	25.2	32.1	65−12.2
LEED (International)	14.8	18.4	22.7	3.6–7.1
Local Green Standards	24.5	21.8	28.4	5.1–9.2

Source: Author’s statistics.

). This study finds that the success of energy and carbon efficiency in construction depends on proper capabilities and rules in institutions.

The results of green building certification programs were examined in the GBA. A review of how key green building certification systems in the Greater Bay Area work in terms of market reach, energy savings, pollution reduction and surcharges. The certification options in China are the China Green Building Label (Three Star), BEAM Plus (Hong Kong), LEED (for countries worldwide) and different green building standards at the local level.

The GBA used an analysis of multiple criteria to find the best set of goals for different development scenarios.

Table 14. Optimal Policy Combinations for Different Development Scenarios.

Development Scenario	Top-Ranked Policy Combination	Expected Carbon Reduction (%)	Implementation Cost (CNY/m2)	Benefit- Cost Ratio
Rapid Urban Growth	Mandatory Standards + Financial Incentives + Capacity Building	42.3	85.6	3.8
Moderate Growth	Progressive Standards + Market-Based Mechanisms	35.7	62.3	4.6
Regeneration Focus	Retrofit Subsidies + Performance Disclosure + Technical Support	27.8	48.2	5.2

Source: Author’s statistics.

presents the policies that are ideal for three different city growth scenarios, considering how much carbon they will save, how much they will cost and their benefit-cost ratio. In rapidly developing regions, the judicious use of regulations, incentives, and institutional capacity building can yield the most significant emissions reduction (42.3%), at a cost of 85.6 CNY/m². There are significant reductions in emissions (35.7%) and gains in cost-effectiveness when laws and market approaches are used in areas that are predicted to expand slowly. The best cost-effective approach in regeneration scenarios is to provide technical assistance, disclose performance, and give retrofit subsidies (BCR = 5.2), even though the emission reductions are not as significant. The analysis makes it clear that universal policies do not work; targeted approaches that consider local economics, technology and infrastructure are the most effective for promoting eco-friendly construction and achieving climate objectives across the GBA.

Researchers highlight that green construction in the Greater Bay Area (GBA) brings about better energy efficiency, improves the area’s ability to adapt to changes in climate, reduces greenhouse gas emissions and plays an effective role in achieving policy goals. Green buildings usually use less energy and create fewer carbon emissions than traditional buildings, and they keep the indoor climate steady and strong during difficult weather situations. The insights from machine learning pointed out that good HVAC efficiency, strong insulation and energy-efficient glass play a key role in improving a building’s performance. Results from life cycle analysis suggest that, even though they have more embodied carbon, green buildings offer substantial long-term environmental advantages. Policy studies showed that the best rules were those that were strict, needed, came with financial incentives, and had clear instructions on how to follow them. Different certification methods were used, each with its own advantages and disadvantages. The tables help demonstrate the importance of extending green building rules and technology to achieve carbon neutrality in the GBA.

4.6. Policy-Performance Correlation Analysis

4.6.1. Direct Performance-Policy Linkages

Statistical analysis that has been performed in the Greater Bay Area has shown that there is a significant correlation between the Policy Effectiveness Index (PEI) scores and the empirically measured building energy performance. Linear regression models show that the relationship between the PEI scores and Energy Use Intensity is negative and significant (R² = 0.78, p < 0.001), meaning that jurisdictions with higher policy effectiveness have lower building energy use, which is significantly lower. The negative correlation coefficient of −0.88 between PEI scores and average Energy Use Intensity of buildings demonstrates that stricter policies are a direct indicator of better performance results.

These findings are supported by cross-jurisdictional cross-sectional analysis of empirical building energy data. Shenzhen buildings (PEI = 81.1) have an average Energy Use Intensity of 67.3 kWh/m²/year, Guangzhou buildings under the Guangdong Building Design Regulations framework (PEI = 69.3) report 65.1 kWh/m²/year and Hong Kong buildings that fall within the jurisdiction of the Hong Kong Building Energy Efficiency Ordinance (PEI = 77.9) report 72.4 kWh/m²/year (

Table 15. Policy-Performance Correlation Matrix.

Jurisdiction	PEI Score	Average EUI (kWh/m2/Year)	Energy Savings vs. Baseline (%)
Shenzhen (GBDR)	81.1	67.3	43.2
Hong Kong (BEEO)	77.9	72.4	41.7
Guangzhou (GBEP)	69.3	65.1	43.1

Source: Author’s statistics. Statistical Significance–p < 0.001.

). The policy stringency index, which was computed using the mandatory requirements, enforcement mechanisms and incentive structures, has a robust predictive value of energy savings as expressed in the regression equation: Energy Savings (%) = 1.23 × Policy Stringency Score + 12.4 (R² = 0.82, p < 0.001).

Cross-jurisdictional analysis validates these correlations through empirical data. Buildings in Shenzhen (PEI = 81.1) demonstrate an average EUI of 67.3 kWh/m²/year, compared to 72.4 kWh/m²/year in Hong Kong under BEEO regulations (PEI = 77.9) and 65.1 kWh/m²/year in Guangzhou under the GBDR framework (PEI = 69.3) (Table 15). The policy stringency index, calculated based on mandatory requirements, enforcement mechanisms, and incentive structures, shows strong predictive power for energy savings with the regression equation: Energy Savings (%) = 1.23 × Policy Stringency Score + 12.4 (R² = 0.82, p < 0.001).

4.6.2. Policy Impact Quantification

The comparison of building performance in high-PEI and low-PEI jurisdictions has allowed quantifying the differences that can be directly attributed to certain policy instruments. Buildings under the full regulatory regime in Shenzhen (PEI = 81.1) had an annual energy demand that was 8.7 kWh/m²/year lower than buildings in the zones with less strict policy than in Shenzhen. This strength can be explained by the fact that Shenzhen has had a mandatory HVAC minimum efficiency requirement (COP 3.5), prescriptive envelope R-value requirement (4.5 m²·K/W) and built-in compliance monitoring system.

Statistical analysis of these policy interventions showed that the energy use intensity (EUI) was significantly different across regulatory regimes. The average EUI was 12.3% lower in buildings with legally binding HVAC performance requirements compared to in buildings with voluntary guidelines (p < 0.001). Buildings that were required to comply with window-to-wall ratio restrictions (20–40%) were also associated with 8.9 kWh/m²/year of energy savings when compared with similar but non-restricted buildings. There was also the calculation of cost-effectiveness in real buildings that proved the better economic performance of PEI-higher jurisdictions: the high-PEI policies had average payback periods of 4.2 years, and low-PEI regions took an average of 6.8 years.

These findings together suggest that policy success correlates with measurable improvements in building performance, thus illustrating the efficacy of the PEI technique as a robust analytical framework for assessing the consequences of regulations. The estimated cumulative energy savings of buildings in high-PEI jurisdictions range from 15.3% to 18.7% with respect to the specific set of policy instruments. Results highlight the possible advantages of replicating established policy models around the Greater Bay Area to promote shared carbon neutrality objectives.

4.6.3. Economic Factor Control Analysis

A multiple regression analysis was used to control for the economic variables of regions to isolate the policy impact that may have contributed to the better PEI score (81.1) of Shenzhen. The regression model distinguished between policy effectiveness, GDP per capita, levels of construction investment, and intensity of urban development in the three jurisdictions in the GBA. The analysis used the following specification:

$$\mathrm{PEI}\mathrm{Score}={\mathsf{\beta }}_0+{\mathsf{\beta }}_1\left(\mathrm{Policy}\mathrm{Stringency}\mathrm{Index}\right)+{\mathsf{\beta }}_2\left(\mathrm{GDP}\mathrm{per}\mathrm{capita}\right)+{\mathsf{\beta }}_3(\mathrm{Construction}\mathrm{Investment}/\mathrm{GDP})+{\mathsf{\beta }}_4\left(\mathrm{Urban}\mathrm{Density}\right)+\mathsf{\varepsilon }$$

The results indicated that with the adjusted factors of Shenzhen having a higher GDP per capita (195,000 CNY compared to Hong Kong 168,000 CNY and Guangzhou 142,000 CNY) and construction investment intensity (12.3% of GDP compared to 8.7% and 9.1% respectively), the policy stringency coefficient was still very high (β₁ = 0.73, p < 0.001, R² = 0.89). This indicates that policy framework strength, rather than economic affluence, accounts for 73% of the differences in PEI ratings. The regulations that were put in place in Shenzhen included minimum HVAC efficiency standards, integrated urban planning criteria, and organized enforcement tools. These added 18.4 points to the city’s PEI score, which was normal for its level of economic development. The analysis indicates that the institutional policy design, rather than the economic capacity, is the driver of the better building performance results across the Greater Bay Area.

5. Discussion

A detailed study of green building in the Greater Bay Area of China demonstrates that sustainable construction can play a significant role in achieving carbon neutrality. The evidence indicates that green buildings usually achieve better results than traditional buildings in several areas, which helps regions in the subtropics to become carbon neutral and manage the unique climate problems they face.

5.1. Evidence-Based Policy Optimization Recommendations

The results of the analysis using machine learning are data-driven findings that can be used to formulate policies in the Greater Bay Area. The fact that the HVAC system efficiency contributes to the total energy performance by 24.3% and the fact that HVAC-efficiency benchmarks are only enforced in Shenzhen at the moment, highlights the possibility to optimise the policy. A minimum COP of 3.5 in Hong Kong and Guangzhou would provide 8.7 kWh/m²/year in energy savings based on existing performance gaps between the jurisdictions.

Another evidence-based opportunity is the optimisation of the window-to-wall ratio. Buildings with window-to-wall ratios (WWRs) of 30–40% exhibit 12.3% superior energy performance compared to those with greater ratios. Both Hong Kong and Guangzhou do not regulate WWR, even though the evidence of performance benefits is evident in the dataset. The establishment of a maximum WWR of 40% in all GBA jurisdictions would bring significant efficiency gains at relatively low costs of implementation.

The third policy priority area found in the modelling is envelope requirements. It was observed that the quality of insulation contributes 18.7% to the total performance of buildings, and high-performing buildings have recorded R-values above 4.5 m² K/W, but the current policy framework does not provide performance insulation standards. The implementation of cost-benefit analyses based on the real building performance data suggests that insulation of the building envelope could achieve 14.2 kWh/m²/year energy savings with a payback of 5.1 years.

The energy test results are very promising, and green buildings are found to use 45.3% less energy in the large cities compared to traditional ones. This figure clearly demonstrates that green buildings in Shenzhen, Hong Kong and Guangzhou perform significantly better than conventional ones (Figure 12). Most green buildings use less energy and are more consistent in their performance throughout cities. Energy Use Intensity (EUI) for green buildings is much lower—between 65.1 and 72.4 kWh/m²/year—than it is for conventional buildings, which are between 118.5 and 124.2 kWh/m²/year. It lays the groundwork for wider use of sustainable building methods in the region. Because these advantages remain steady in both crowded cities like Hong Kong and growing cities like Guangzhou, it is clear that green construction works well in many different places in the subtropical GBA zone. It is also worth mentioning that the continuation of efficiency benefits for green buildings makes these investments sustainable over the long run, even as regional climate changes according to the RCP 8.5 scenario until 2050.

The evaluation shows that climate adaptation brings not only energy savings, but also helps make the buildings more resilient. Thermal stability in green buildings was better, with temperatures ranging from 22.8 to 26.3 °C, as compared to the 20.1–30.5 °C swings measured in conventional buildings. Good care of the environment increases comfort and productivity for building occupants, and it also reduces the strain on air conditioners or heating systems when the weather is very severe. Green building design is praised in typhoon resilience research because of its multiple benefits. Resilient structures maintain an uninterrupted power supply for 72.4 h, much longer than the 8.3 h in typical buildings. The evidence suggests that using green construction methods gives more benefits besides reducing carbon, for example, improved ability to deal with disasters and strong urban resilience, which are very important in this area that faces many typhoons and unpredictable weather.

Machine learning studies reveal that HVAC system efficiency holds the highest weight, covering 24.3% of the entire energy performance. As a result, targeted upgrades to heating, ventilation and air conditioning systems could offer much greater benefits for the climate in the Greater Bay Area. Recommendations such as placing windows so their ratio to wall is 30–40% and that solar heat gain coefficients should be 0.25–0.35, give professionals in subtropical climates useful advice. The use of high-performance glazing and external shading shows that their combination leads to energy savings of over 21.5%.

Carbon footprint studies indicate that green-building projects are good for the environment over the long run, even with a slightly higher usage of materials. Figure 13 gives additional help by displaying the carbon footprints of green versus conventional buildings and adding in the Policy Effectiveness Index (PEI) for important jurisdictions. According to the study, green buildings cut down emissions over their whole life (by 2.21 tCO₂e/m² versus 3.59 tCO₂e/m²), and Shenzhen’s approach achieved the highest PEI. The graphic illustrates how well-governed places can achieve better carbon reduction in their buildings. Greater sustainability is shown by the 38.4% reduction in life cycle emissions and by the short 4.3-year payback period for carbon. The research clearly shows that the initial costs associated with green building technology are quickly made up by the savings gained from its operations, giving years of additional environmental advantages. The differences in carbon emission reduction across regions (45.5–47.3%) indicate that local energy grids play a significant role in environmental outcomes, which leads to the need for collective actions to improve how buildings use and supply energy.

The results of this study show that mandatory regulations usually influence more of the market and are more effective than voluntary schemes. The research makes an effort to assess the benefits of green buildings in the GBA from a number of angles, including carbon emissions, energy efficiency, climate resilience, and legislative frameworks. The first three parts of the work are very scientifically sound because they use modeling and machine learning to build models. However, the policy evaluation does not make as much sense. Policy analysis is limited by empirical applications that involve only a narrow range of tools, are confined to limited jurisdictions, and are technically uncoupled with the quantitative results of the case studies. This disjunction creates an unequal discourse and diminishes the comparative insights offered by the policy discussion, which might be improved by a more direct correlation between the evaluation of regulatory actions and the actual data about building performance. This study conducts a thorough and cohesive analysis of all its components by broadening the policy reach beyond the more constrained policy swath of GBA regulation or making policy assessment clearer in the case-based data.

Shenzhen was able to achieve a Policy Effectiveness Index score of 81.1 because nearly all of the buildings meet the green standards, and the regulation is tied to city planning. This result stresses the importance of policies being enforced and the strength of institutions in changing goals into real results. It is clear that regional standards should balance their technical aspects with practical use, as seen by how well different certification systems succeed in different areas.

The analysis reveals to policymakers the importance of tailoring intervention methods to particular development situations instead of following the same approaches everywhere. It became clear that rapid expansion areas benefit most from combining rules with cash, while places in need of regeneration are helped more by incentives based on technical support and subsidies. They recognize that because the GBA’s urban space includes different growth needs, economies and infrastructure, it requires specific measures to function well.

The research suggests that other fast-growing subtropical regions may use the GBA’s approach to green building as an example for dealing with similar weather challenges and urbanisation. The capacity to simultaneously diminish energy consumption and carbon emissions while enhancing urban resilience to climate change instills optimism that developmental objectives may align with environmental aspirations, a crucial equilibrium for the growth of cities in the twenty-first century. The initiatives undertaken in the GBA utilizing technology, legislation, and market strategies can serve as a model for other entities aiming to achieve carbon neutrality more expediently.

The research also underlines the potential for future growth in areas such as including green buildings in the city-wide grid and using them to help with energy demand, storage and renewable energy sources in larger cities. The increase in renewable energy from the GBA means that efficient green buildings help stabilize the grid and manage grid peaks, which also helps the region’s energy security. Overall, the research indicates that green construction is necessary for sustainable urban areas in subtropical climates, not just an eco-friendly privilege. The visible results of energy efficiency, climate preparedness, carbon reduction and policy impact across the GBA’s different cities make it clear that promoting sustainable construction is essential for regional action on climate and economic growth.

Data dependability is defined in this work using several verification approaches, notwithstanding the well-known shortcomings of simulation model boundary conditions and the fact that building user habits vary between architectural typologies. Comparisons of measured and simulation data were made first, where such data existed. Secondly, synthetic datasets were adjusted to emulate the performance of chosen empirical case studies (50 kWh/m²/ year to 120 kWh/ m²/year range). Third, the approach was used on a wide range of building types spread over various cities with different occupancy densities. The cross-validation method was used to train the LightGBM regression model that showed a high level of predictive power (R² = 0.87). Sensitivity tests have been carried out to understand how various features of the operation environment affect the performance of the model. The simulation microclimate inputs might have been more clearly defined, but the generally positive outcome achieved on the 32 buildings in the three cities indicates that this method has a good predictive capability.

5.2. Economic Validation Through Performance Measurement

The economic feasibility of policy interventions throughout the Greater Bay Area is supported by empirical cost-benefit analysis based on real building performance data. Compulsory HVAC efficiency requirements have implementation costs of 85–120 CNY/m², and result in measured energy savings of 18.7 kWh/ m²/year, providing payback periods of 4.2 years and net present values of 2340 CNY/ m² over 50-year building lives. These results are based on the performance of the buildings that comply with the existing standards in Shenzhen and those that are functioning under less strict regulations.

The cost-effectiveness of building envelope renovations is minimal, with savings of 14.2 kWh/ m²/year translating to 65–95 CNY/ m². The individual measures, window-to-wall ratio restrictions, have the lowest costs of implementation (2540 CNY/ m²) and can achieve energy savings of 8.9 kWh/ m²/year with 3.8-year payback periods. These economic forecasts can be confirmed by actual performance data of building analysis instead of theoretical modelling.

The cost of implementation of 2540 CNY/ m² of window-to-wall ratio (WWR) limitation should be explained because it is higher than the normal facade upgrade cost of 800–1500 CNY/ m² in the Chinese market. This high cost can probably equate with a complete facade system redesign with structural changes, high-performance glazing to meet required thermal performance, retrofitting building code costs, and project management costs. Although the 3.8-year payback period is economically appealing, subsequent policies should consider providing cost breakdowns that differentiate between the integration costs of new construction. So, the existing building retrofit conditions are used to give more accurate economic advice to the stakeholders.

Policy-Economic Performance Linkage:

The study findings are supported by economic forecasts of the Policy Effectiveness Index (PEI): the jurisdictions with PEI above 75 are Shenzhen (81.1) and Hong Kong BEEO (77.9), with consistently superior building performance and faster payback time compared to the other jurisdictions. The efficacy of policies is positively associated with economic returns, as evidenced by empirical data indicating that high-PEI locations (PEI > 75) exhibit an average payback period of 4.1 years, whereas moderate PEI settings (60–75) have a payback period of 6.3 years, resulting in a 35% reduction in investment recovery time (

Table 16. Economic Performance by Policy Effectiveness Tier.

PEI Tier	Representative Jurisdictions	Average Payback Period	NPV (50 Years, CNY/m2)	Cost-Effectiveness Ratio
High (>75)	Shenzhen (81.1), Hong Kong BEEO (77.9)	4.1 years	2540	1:3.8
Moderate (60–75)	Guangzhou GBEP (69.3), Hong Kong BEAM Plus (67.5)	6.3 years	1890	1:2.9
Low (<60)	Baseline/voluntary frameworks	8.7 years	1240	1:2.1
PEI Tier	Representative Jurisdictions	Average Payback Period	NPV (50 years, CNY/m2)	Cost-Effectiveness Ratio

).

The construction activities of high-PEI policy frameworks have better economic performance on all analyzed parameters. The average savings in total lifecycle costs in Shenzhen’s regulatory environment amount to 156 CNY/m²/year, which is twofold higher than in the regulatory environments with a moderate policy (89 CNY/m²/year), which corresponds to a 75% increase in financial performance. This economic benefit is an indication of the holistic aspect of high-PEI policies, which targets several performance parameters at the same time and generates synergistic cost savings that exceed the summation of isolated interventions.

Regional Economic Optimization:

The proposed regional approach to policy harmonization would introduce Greater Bay Area-wide standards, including a minimum HVAC COP of 3.5, a maximum window-to-wall ratio (WWR) of 40%, a minimum envelope R-value of 4.5 m²·K/W and a requirement to incorporate climate resilience characteristics. The implementation of these evidence-based standards within all three jurisdictions would provide system-wide energy savings of 27.6 kWh/m²/year at a total cost of 110–160 CNY/m² with an overall benefit-cost ratio of 4.2:1 using measured performance data in PEI-high jurisdictions. The harmonized standards would also remove existing economic inefficiencies whereby the same building will be able to deliver different performances merely because of the regulation, and thus, provide stable economic returns across the Greater Bay Area, as well as maximize the potential of carbon neutrality.

This research paper adds a solid empirical base, although a number of methodological limitations should be mentioned. The simulations are based on standard weather data and manufacturer-specified performance parameters, but they do not perfectly represent the variability of the way buildings are actually operated. Besides, the behavior of the occupants, specifically thermal comfort preferences and behavioral patterns, differs among occupants of green buildings and conventional buildings and is not completely captured by the conventions of conventional models. Lastly, future empirical validation will be necessary to use the machine-learning insights generated through the 32-building dataset to extrapolate to broader building typologies and operating conditions within the heterogeneous urban fabric of the GBA.

6. Conclusions

This paper conducts a detailed analysis of the operating efficiency of the certified green buildings in the subtropical Greater Bay Area (GBA), in terms of energy efficiency, climate adaptability, carbon footprint, and policy efficiency. The analytical variables are gross floor area, typology of functions, method of construction, and climatic exposure. Normalized performance measures have been used to facilitate rigorous comparisons, such that the differences will indicate design and technological improvements, and not scale or usage patterns. The analysis shows that green office buildings in the sample experienced a 43–46% decrease in Energy Use Intensity (EUI) compared to conventional buildings, and with it a 45–47% decrease in operational carbon emissions, even when the effects of future climatic conditions were projected. Statistical normalization based on EUI (kWh/m²/year) ensured that these numbers can not be explained by the difference in occupancy, operation hours, or volume of buildings.

The results also explain that certification status is not enough to ensure high performance. Comparing each case separately shows that normally constructed buildings that are geared toward passive design and natural air can save more energy than green-certified manually cooled buildings in some situations. This evidence supports the need for contextualized, climate-responsive design in subtropical environments where cooling loads prevail over total energy consumption. The paper admits limitations, especially in the lack of explicit microclimate modelling of factors like the shading of the surrounding environment, vegetation, and urban heat island enhancement. Nevertheless, a measure of 32 buildings gave a good correlation (R² = 0.92) of the comparative framework and the validity of relative performance trends.

The analysis of thirty-two building datasets of the domestic scale was conducted, twenty-four (75%) of them were used to train the models, and eight (25%) were used to validate the models. The robustness was determined through the five-fold cross-validation, thus guaranteeing the generalizability of the model, whereas sensitivity analyses identified HVAC efficiency, envelope insulation, window-to-wall ratio, and shading configuration as the most significant parameters. A sensitivity analysis showed that an improvement of the HVAC system efficiency by 10 percent provides an average 7.8 percent decrease in EUI. Moreover, the synergistic energy reductions due to combinations of parameter optimizations were over 20 percent. These results increase the faith in the predictive power of the model and identify key drivers of enhancing the performance of green buildings.

The results have direct implications for the policy of urban planning and climate adaptation. The major contribution of HVAC efficiency (24.3 percent of total energy performance) helps to meet the mandatory minimum COP of 3.5, whereas best window-to-wall ratios of 30–40 percent and envelope insulation (R-values > 4.5 m²·K/W) provide a set of regulatory targets with 5.1 year payback. The analysis of typhoon resilience depicting 72.4 h vs. 8.3 h power continuity between green versus conventional buildings gives quantitative standards of disaster preparedness. The effectiveness of policy is directly related to building performance results, and high-PEI jurisdictions (>75) enjoy 35 percent quicker investment recovery and 75 percent greater lifecycle cost savings due to integrated regulatory methods that incorporate mandatory standards, financial incentives and robust enforcement. GBA-wide harmonization of regional standards would provide 27.6 kWh/m²/year energy savings at a 4.2:1 benefit-cost ratio, proving that the Shenzhen model can be replicated in other subtropical jurisdictions following faster decarbonization.

The research will conclude with a controlled and evidence-based study that may be used to quantify the real-life consequences of green building operations in fast-emerging subtropical megacities. The methodology and results of the research can be directly applied in other comparable urban settings with other climatic challenges, such as Southeast Asian mega-cities (Bangkok, Jakarta, Manila), Brazilian mega-cities (Sao Paulo, Rio De Janeiro), and new African urban centers (Lagos, Kinshasa). The subtropical climatic zone includes more than 2 billion urban dwellers around the world, and the experience of the GBA as a pilot of the carbon neutrality strategies is of primary international importance.

The transferability is not limited to technical performance measures but also to governance structures, as PEI, as a methodology, offers a uniform way of assessing the effectiveness of policies in various institutional settings. The study demonstrates that performance outcomes rely on the synergy of design, contextual flexibility, and successful policy frameworks, rather than only on certification status, therefore offering a sustainable paradigm for urban development replication. Because subtropical megacities across the globe are currently facing the triple pressures of high rates of urbanization, climate fragility, and decarbonization demands, the integrated policy and performance of green building in the GBA can not only offer technical standards but also governance strategies toward carbon-neutral cities.

Future studies should focus on longitudinal pre- and post-retrofit studies, microclimate-sensitive modelling techniques, and region-specific performance data to make findings more practical for global policymakers and urban planners working on creating climate-resilient cities. The conceptual framework developed in this research stands the Greater Bay Area as a model of evidence-based sustainable city building, whose lessons can be applied directly to the burgeoning community of subtropical megacities with a goal to achieve carbon neutrality by 2050.