From Urban Design to Energy Sustainability: How Urban Morphology Influences Photovoltaic System Performance

Abstract

In response to the pressing need for sustainable urban development amidst global population growth and increased energy demands, this study explores the impact of an urban block morphology on the efficiency of building photovoltaic (PV) systems amidst the pressing global need for sustainable urban development. Specifically, the research quantitatively evaluates how building distribution and orientation influence building energy consumption and photovoltaic power generation through a comprehensive simulation model approach, employing tools, such as LightGBM, for the enhanced predictability and optimization of urban forms. Our simulations reveal that certain urban forms significantly enhance solar energy utilization and reduce cooling energy requirements. Notably, an optimal facade orientation and building density are critical for maximizing solar potential and overall energy efficiency. This study introduces novel findings on the potential of machine learning techniques to predict and refine urban morphological impacts on solar energy efficacy, offering robust tools for urban planners and architects. We discuss how strategic urban and architectural planning can significantly contribute to sustainable energy practices, emphasizing the application of our results in diverse climatic contexts. Future research should focus on refining these simulation models for broader climatic variability and integrating more granular urban morphology data to enhance precision in energy predictions.

1. Introduction

The United Nations “World Population Prospects 2024” [1] report states that the world’s population is projected to grow from 8.2 billion in 2024 to approximately 10.3 billion by the mid-2080s. Accompanying this growth is an escalating demand for energy, driven by extensive development and leading to the large-scale exploitation of non-renewable resources. The building sector, particularly in urban areas, predominates in energy consumption. The widespread use of fossil fuels is causing a gradual depletion of global reserves and could inflict irreversible damage on the Earth’s environment and climate [2]. In response, numerous countries have devised renewable energy strategies aimed at achieving zero emissions by 2050, with solar energy identified as the greatest potential clean energy source for large-scale application [3,4]. Urban rooftops offer a substantial foundation for distributed solar installations, though much of their potential remains underutilized [5]. Implementing solar energy systems enhances urban sustainability significantly [6]. The potential of solar energy in urban blocks, especially with photovoltaic panels on rooftops, is heavily influenced by the nearby structures and the general layout of the urban area [7,8]. Nevertheless, excessive solar irradiation can elevate interior building temperatures, particularly in sunlit areas, increasing the energy required for cooling [9,10]. Therefore, investigating the interplay between neighborhood morphology and solar potential and finding a balance between the efficiency of photovoltaic systems and the energy needed for cooling is essential.

1.1. Urban Morphology and Building Energy Consumption

Building energy consumption is influenced by numerous factors, with the urban form being one of the key determinants of building energy efficiency. The urban form can be classified through several dimensions, such as the density, geometric shape, and building type, each impacting building energy consumption directly [11,12]. Urban density is commonly measured using indicators like the floor area ratio, building density, open space ratio, and average floor levels. Studies indicate that the connection between density and building energy use is not linear [13,14]. Some research has demonstrated that, beyond a certain density threshold, the building energy use intensity (EUI) decreases, forming a “U-shaped” relationship with energy consumption [15]. For instance, a simulation study in Portland indicated that the EUI is reduced once the building density surpasses a certain level [16]. Javanroodi, K. et al. [17] observed that the cooling energy intensity in both office buildings and residences tends to decline as the FAR increases. Nonetheless, other studies suggest that high density could boost cooling energy consumption due to the urban heat island effect [18].

Geometric parameters at both urban and architectural scales, such as building height, shape factor, number of floors, and street aspect ratio, play crucial roles in determining energy dynamics. Mangan et al. [19], through the evaluation of 120 urban morphology models, demonstrated that building height and the street aspect ratio considerably affect both energy consumption and cost-effectiveness. Further research highlights that the geometric features of street canyons significantly influence energy consumption in office and residential buildings. In urban canyons, the impact on office buildings can reach up to 30% and up to 19% for residential buildings [17]. Shareef’s [20] findings in the UAE suggest that orientation is the predominant factor impacting cooling loads and energy consumption in urban blocks. Roman Loeffler’s [21] key indicators for thermal gain and loss in buildings are instrumental in enhancing energy efficiency. Ensuring good compactness in block constructions aids in minimizing cooling loads and optimizing indoor lighting conditions.

Building typologies, including a courtyard, a point block, and slab configurations, exert a considerable impact on energy consumption. Research consistently shows that courtyards, in particular, demand lower energy across diverse climatic scenarios. Chua and Beng’s [22,23] further research on courtyards showed that in Thessaloniki-Greece, the minimum cooling requirements are for summer days of the courtyard.

In summary, the wide range of urban forms for a series of land use are important factors to predict building energy consumption and may have significant impacts and differ with location and the climate scenario. Further studies need to be conducted on how different urban morphology indicators can be contextualized and implemented in the wide range of global cities, which intends to help identify the most adaptive functional or construction forms that maximize energy efficiency while minimizing energy consumption.

1.2. Urban Morphology and Solar Energy Potential

The urban morphology shapes the potential of solar energy harvesting [24]. The roof and the facade of buildings actually act as a platform for the placement of solar collectors and photovoltaic systems; they also support optimized systems with the help of tailor-made urban layouts and morphological indicators. The floor area ratio, building density, geometric configurations, and block types are a few among the various important indicators of diverse urban forms that this review probes about their specific impacts on the efficacy of the utilization of solar energy.

Studies have shown that relative heights, distances between buildings, and overall layouts play a pivotal role in influencing the reception of solar radiation on building surfaces. It was further supported by the study by Tian and Xu [25] in residential areas of Wuhan, claiming that the floor area ratio, building density, average height, and interspacing are paramount morphological factors in determining solar potential. Similarly, J. Zhang et al. [26], through an analysis of 30 distinct urban block types, revealed marked differences in their effects on solar gains and energy efficiency. Vulkan et al. [27] assessed the solar potential on the rooftops and facades of high-density urban residences, examining the contributions of various building surfaces to the city’s total photovoltaic output. Additionally, Zhao’s [28] research on rooftop photovoltaic efficiency under different scenarios established the correlation between the usable rooftop area, array layout, and shading effects with photovoltaic efficiency, pinpointing the optimal photovoltaic design configurations for urban contexts.

Further research has revealed that certain urban block morphologies, particularly courtyards and mixed configurations, generally surpass other forms in terms of energy performance, attributable to their expansive layouts and optimal orientations. A case study in Nanjing highlighted that U-shaped blocks exhibit the greatest annual photovoltaic potential [29]. Moreover, a pioneering study in Singapore applied a new typological method to explore the interplay between block forms and solar energy, which demonstrated that site coverage and block configurations markedly affect solar potential [30].

Urban morphology has a direct and complex role in enhancing solar efficiency. Realizing such potential could achieve a remarkable gain in solar harvesting and photovoltaic efficiencies within buildings, which are the results of sustainable urban goals by enhancing energy efficiency and decreasing carbon emissions while incorporating such energy considerations in designs and strategies among policymakers, urban planners, and architects.

1.3. Application of Machine Learning Algorithms

Indeed, an increasing amount of research does acknowledge that urban form has a complex impact on block energy consumption. Traditional linear regression models, in fact—such as the Ordinary Least Squares (OLS) regression model—do not suffice in capturing the effects of urban form on energy consumption within urban blocks; sometimes, they even lead to inconsistent results because, in reality, the relationship between urban form and energy consumption is not always linear [31,32]. Furthermore, the study of urban morphology has evolved from a two- to three-dimensional analysis, greatly increasing the amount of data being handled, so as to better respond to the analytical and design requirements in the built environment [33].

With all these, advanced machine learning algorithms, such as LightGBM [34], XGBoost [35], Random Forest [36], Artificial Neural Networks [37], and Naïve Bayes [38] using extremely large datasets, expose complex, nonlinear relationships and complicated factor interactions [39,40]. More specifically, an increasingly growing body of research examines the dynamics between urban morphologies and city environments that now incorporate such algorithms in an effort to increase analytic accuracy. Yu et al. [41] used the Random Forest, among four machine-learning algorithms, for describing the nonlinear correlations between urban outdoor temperatures and urban forms. Similarly, Huang et al. [42] studied how urban morphology modulates the dispersion of air pollution at the high-density urban block level by means of a combination of methods: Decision Trees, Support Vector Machines, and Neural Networks, the latter being best in terms of performance. Of further note, Dan Assouline [43] has very successfully combined Support Vector Machines with Geographic Information Systems for predicting the potential for rooftop solar photovoltaic installations in Swiss municipal regions.

In response to the inherent opacity of machine learning processes, the integration of SHAP values to elucidate machine learning models has gained traction in contemporary research. Song et al. [44] introduced an enhanced interpretable machine learning model to estimate global solar radiation in China, utilizing a sophisticated XGBoost algorithm refined through particle swarm optimization and paired with GIS-based methodologies. Their research revealed a temporal decline in solar radiation and photovoltaic potential across China, prompting policy suggestions aimed at addressing regional imbalances between the photovoltaic potential and installed capacities. Additionally, Felix Wagner et al. [45] leveraged interpretable machine learning to probe the effects of urban morphology on Berlin’s traffic dynamics. Their findings underscored the pivotal role of central urban designs in curtailing commute distances and pinpointed a significant reliance on vehicles in economically disadvantaged areas, providing crucial insights for future urban development strategies.

Urban design considerations in photovoltaic capabilities have been pointed out by several studies. A better urban morphology and block design improve not only the efficiency of solar collection within buildings but also magnify the overall efficacy of photovoltaic systems. These are thus very important optimizations for reaching sustainable urban energy goals. However, the lack of research at the block scale has in part hindered an understanding of the implementation and optimization of these techniques for best results in solar energy utilization. These gaps in research into the block-scale urban morphology suggest that the potential in the urban spaces, especially in the denser contexts of a city’s form, might not be fully taken advantage of. Stated differently, a more granular review of urban morphology at the block level could unveil subtle building interplays that affect solar reception—data useful for more effective architectural design layouts and urban planning strategies.

Consequently, fulfilling sustainable urban development goals necessitates conducting more empirical research at the block scale. This research should employ high-precision geographic and environmental data to simulate and evaluate how different urban forms affect solar potential. Furthermore, the integration of advanced simulation tools and interpretable machine learning methods can deepen our understanding of the nexus between urban morphology and energy efficiency. Such insights are critical for guiding practical urban design and policy-making, with the ultimate aim of achieving greater energy efficiency and minimizing carbon footprints.

This paper is structured as follows: Section 2 presents field research on residential districts within the designated study area, analyzing shared traits to formulate an ideal model. This section also employs simulation software to estimate neighborhood energy consumption and photovoltaic outputs. Section 3 details the machine learning principles and the formulaic models utilized in this study. Section 4 examines the results from these simulations and enhances these findings through further simulations with machine learning predictive models, incorporating SHAP values for model interpretation. Finally, Section 5 and Section 6 delve into how design factors affect building energy consumption and outline prospective research avenues and their broader implications.

2. Materials

2.1. Study Area

The focus of this study is on the prototypical residential blocks of Yichang City in Hubei Province. Employing both field surveys and an analysis of satellite imagery (as illustrated in Figure 1), the research examined 17 residential blocks, comprising 10 older and 7 newly constructed blocks. The investigation highlighted a higher prevalence of older residential blocks, characterized by systematic land use and uniform elevation profiles. The observed floor area ratios ranged from 4.56 to 5.48, while building densities varied between 28.5% and 32.4%. The architectural landscape is primarily composed of low-rise and high-rise residential buildings, with the absence of skyscrapers.

Accordingly, this study conceptualized an ideal residential block with uniform dimensions, each measuring 300 m in both length and width, thereby occupying a total land area of 90,000 square meters. The designated floor area ratio (FAR) of 5.0 and a building density of 30% were derived from empirical field studies reflecting the current state of residential blocks in the Yichang area. The FAR, representing the ratio of the total building floor area to the plot area, is a critical measure of construction intensity. In this context, an FAR of 5.0 indicates that the total building floor area is five times that of the plot area, reflecting a high building density and the effective utilization of vertical space.

The building density, set at 30%, indicates that buildings occupy 30% of the total plot area, with individual building footprints of 1200 square meters. This measure of the proportion of land area covered by buildings is a key parameter for the utilization of space at ground level. By controlling the building density, the plan effectively dictates the allocation of open spaces and green areas, impacting ventilation, lighting, and overall residential comfort.

To assess the shadow impacts from adjacent structures, the configuration included 72 external buildings, each measuring 50 m in length (east-west), 30 m in width (north-south), and 30 m in height. Based on the specifications outlined in

Table 1. Neighborhood parameter setting.

Neighborhood Setting Parameters	Setpoint
Neighborhood location	Yichang City, Hubei Province
Block width	300 m
Block length	300 m
Block plot ratio	5.0
Neighborhood building density	30%
The height of the building	4 m
The floor area of the building	1200 m2

, an idealized parametric model of the residential block was illustrated in Figure 2.

2.2. Datasets

This study utilized parametric design techniques on the Grasshopper platform to manage 12 design variables within the site, including the open space area and building width, to develop a model of a residential block. Grasshopper 1.0 is a parametric design plugin for Rhinoceros 7, a widely used three-dimensional NURBS modeling software [46]. Rhinoceros 7 is celebrated for its versatile modeling capabilities and comprehensive data interfaces, and it supports a multitude of plugins (for modeling, analysis, and programming) that enhance its functionality. As computer technology becomes more widespread, more researchers are turning to digital simulation methods. Grasshopper offers significant advantages in terms of flexibility, speed, and cost-efficiency compared to traditional field data collection, particularly in complex projects requiring numerous iterations and optimizations. Through parameter adjustments, Grasshopper enables the flexible and precise control over block creation [47]. Moreover, Grasshopper supports various plugins that simulate impacts from the natural and built environments, such as lighting, energy efficiency, and photovoltaic efficiency.

This study distills design characteristics of residential blocks from field research, defining 12 design parameters (referenced in

Table 2. Block design control parameter setting.

Design Parameter Name	Actual Setting Range	Normalized Setting Range
A Plot Vacant Space Location (AVSL)	A0, A1, A2, A3	0–1
B Plot Vacant Space Location (BVSL)	B0, B1, B2, B3	0–1
C Plot Vacant Space Location (CVSL)	C0, C1, C2, C3	0–1
D Plot Vacant Space Location (DVSL)	D0, D1, D2, D3	0–1
A Plot Building Facade Width (AFW)	20 m to 60 m	0–1
B Plot Building Facade Width (BFW)	20 m to 60 m	0–1
C Plot Building Facade Width (CFW)	20 m to 60 m	0–1
D Plot Building Facade Width (DFW)	20 m to 60 m	0–1
Low-rise Building Ratio (LBR)	0 to 60%	0–1
Building Height Control Line North-South Shift (NSHS)	0 m to 300 m	0–1
Building Height Control Line East-West Shift (EWS)	0 m to 300 m	0–1
Building Height Control Line Clockwise Rotation (CRC)	0° to 360°	0–1

, Figure 3) to modulate the ideal block’s form. The site is partitioned into four primary plots—A, B, C, and D—each subdivided into four subplots, labeled A0 through A3, B0 through B3, C0 through C3, and D0 through D3. Each plot reserves three subplots for building footprints and one for open space. The “plot open space location” parameter pinpoints the exclusive open space within each plot. The “building facade width” parameter governs the east-west length of buildings within the plot, allowing control over whether the building’s longer, sunlight-facing sides are oriented north-south or east-west. Buildings are categorized by height into high-rises (≥12 m) and low-rises (<12 m), with the “low-rise building ratio” detailing the percentage of buildings under 12 m, adjustable between 0 and 60%. The “building height control line” theoretically influences the perceived height of buildings within the block—closer proximity to this line renders buildings taller, and greater distance makes them appear shorter. This line’s placement is adjusted through its north-south and east-west movements and its clockwise rotation angle. Photovoltaic panels, each measuring 15 m by 5 m with a 15° tilt, are deployed across open grounds and rooftops, with installations prohibited within a 10 m shadow radius around ground-level buildings and mandated in open spaces exceeding 200 m² within the block to optimize solar exposure.

2.3. Data Collection

EnergyPlus 8.1.2, integrated within Grasshopper 7, is an advanced energy simulation software that leverages data from representative years and days to conduct simulations, delivering highly reliable results applicable to real-world production settings [48]. This tool effectively simulates the energy demands and consumption of HVAC systems within buildings by meticulously designing circuits, branches, and nodes to guarantee system stability. EnergyPlus is capable of simulating both single-loop and multi-loop systems, quantifying energy needs for both cooling in the summer and heating in the winter and modeling the thermal behavior of buildings under diverse climatic conditions. This facilitates a comprehensive analysis of HVAC system efficiency and overall energy consumption.

Honeybee is an open-source architectural energy simulation plugin for Rhino 3D and Grasshopper, which facilitates a detailed energy analysis and optimization for building designs [49]. Capable of importing and analyzing standard meteorological files in EPW format, it conducts thorough analyses encompassing annual solar radiation, shadow distribution, and sightline evaluations. Honeybee seamlessly integrates with prominent tools, such as OpenStudio 3.8.0, EnergyPlus 9.6.0, Radiance 5.4a, and Daysim 4.0, allowing for extensive simulations that assess building energy consumption, thermal comfort levels, and the efficacy of natural lighting solutions [50].

OpenStudio, crafted by the U.S. Department of Energy, is a comprehensive open-source toolkit designed for building design, analysis, and simulation. The suite includes the OpenStudio SDK and OpenStudio Application [49] and supports a range of energy modeling engines, such as EnergyPlus, Radiance, and DOE-2. Its user-friendly graphical interface permits architects and engineers to construct and modify building models seamlessly, while also enabling effective interactions with other energy simulation tools. Moreover, OpenStudio’s robust model library helps users swiftly develop energy models, which can be integrated with Building Energy Management Systems (BEMSs) to optimize building energy management practices.

This investigation addresses a location in Yichang City, Hubei Province, characterized by a hot summer and cold winter climate regime, necessitating substantial cooling energy during summer months and comparatively less during winter. Accordingly, this study is informed by an inquiry into how block morphology and the key energy metrics of cooling energy consumption, overall electricity usage, and photovoltaic energy generation interact. The Honeybee plugin of the Ladybug Tools 1.8.0 suite within the Grasshopper parametric platform is used to analyze EC for cooling, ET total electricity consumption, and EP photovoltaic generation. To highlight the impact of design parameters on consumption, the model observes ASHRAE standards in prescribing values for consumption parameters. The simulation configurations and objective function settings are elaborated in

Table 3. Ideal model simulation setting and simulation objective function setting table for blocks.

Analog	Settings	Description
Climatic zones	4 ASHRAE Climate zone	Yichang is a typical climate region with hot summers and cold winters
Year of construction	ASHRAE 90.1 2015	The buildings of the area surveyed Basically built for 2010–2020
Architectural function	Midrise Apartment	Residential blocks
Type of building structure	Steel Framed	Most of the buildings in the survey area are high-rise residential buildings
Solar panel electricity conversion rate	21%	Combined with the cost consideration of photovoltaic solar panels, polycrystalline silicon photovoltaic solar panels are used
Study the objective function	Unit	Description
$E_{Cooling}$	$\mathrm{K}\mathrm{W}\cdot \mathrm{h}/{\mathrm{m}}^2$	The electricity consumed by cooling the neighborhood through the HAVC system in the summer
$E_{Photovoltaic}$	$\mathrm{K}\mathrm{W}\cdot \mathrm{h}/{\mathrm{m}}^2$	The block generates electricity throughout the year through a photovoltaic solar system
$E_{Total}$	$\mathrm{K}\mathrm{W}\cdot \mathrm{h}/{\mathrm{m}}^2$	The electricity consumed by the neighborhood throughout the year

.

Latin Hypercube Sampling (LHS) is an advanced stratified sampling method designed for approximating the random sampling of multivariate distributions. It is widely adopted in computer simulations and Monte Carlo integrations [51]. By ensuring that each sample point is maximally spaced across each parameter dimension, LHS comprehensively covers the parameter space, thereby enhancing sampling effectiveness and representativeness. This approach not only reduces the requisite number of samples but also diminishes experimental costs while enhancing the reliability and repeatability of experiments. For this study, the Latin Hypercube Sampling method facilitated the acquisition of 1000 sets of block design control parameters, with detailed hyperparameter configurations presented in

Table 4. Latin Hypercube Sampling settings table.

Hyperparameters	Settings	Description
Samples	1000	The number of samples to be generated
Core Algorithm	Minimizing Correlation	Algorithm used to generate initial LHS samples
Weights	1	The weight of each dimension, if you want the importance of different dimensions to be different
Resampling	Yes	Whether resampling of certain intervals is allowed to improve sample coverage or conform to a specific distribution

.

3. Methods

This study developed a comprehensible machine learning regression prediction model, referenced in Figure 4. Initially, the Latin Hypercube Sampling method secured 1000 combinations of design parameters and their corresponding target function values, which underwent extensive analysis. Pearson correlation and multicollinearity analyses were then employed to probe the influence of design parameter variations on the target function. Given the nonlinear relationships observed between certain design parameters and the target function, advanced machine learning regression algorithms—LightGBM, AdaBoost, and Random Forest—were utilized for prediction. Following a comparative performance review, the optimal model was selected, and its predictions were elucidated using the SHAP explainability framework, enhancing the interpretability of the results.

3.1. Pearson Correlation Coefficient

This study commenced with an analysis of the influence of design control parameters on objective functions, employing correlation analysis methodologies. Specifically, the Pearson correlation coefficient [52], denoted as “r”, was applied to assess the linear relationship between two variables, X and Y. This coefficient is defined as follows [52]:

$$\mathrm{r}={\displaystyle \frac{\sum _{i=1}^{\mathrm{n}}\left(X_i-\overline{X}\right)\left(Y_i-\overline{Y}\right)}{\sqrt{\sum _{i=1}^{\mathrm{n}}(X_i-\overline{X}{)}^2}\sqrt{\sum _{i=1}^{\mathrm{n}}(Y_i-\overline{Y}{)}^2}}}$$

(1)

where $\overline{X}$ and $\overline{Y}$ are the means of the samples $X_i$ and $Y_i$, respectively. The Pearson correlation coefficient “r” spans from −1 to 1. An r value of 0 indicates an absence of a linear correlation. Negative correlations are denoted by r values between −1 and 0, intensifying as r approaches −1. Positive correlations are indicated by r values between 0 and 1, strengthening as r nears 1. Perfect negative and positive linear correlations are represented by r = −1 and r = 1, respectively [52].

Nonetheless, the Pearson correlation coefficient exhibits certain limitations in delineating relationships between variables. It is confined to measuring linear correlations and is incapable of capturing nonlinear dynamics. Should the data exhibit complex nonlinear relationships, the Pearson correlation coefficient might not provide an accurate representation of these interactions. Moreover, its efficacy diminishes with high-dimensional data, where it struggles to manage intricate multivariate relationships effectively. In contrast, machine learning algorithms excel in handling high-dimensional data and in discovering latent patterns among variables. Consequently, this research has adopted machine learning predictive models to mitigate the limitations inherent to the Pearson correlation coefficient.

3.2. Multicollinearity Analysis

Multicollinearity, a ubiquitous phenomenon in multiple linear regression analyses, arises when two or more explanatory variables within a model exhibit a notable linear correlation [53]. This correlation escalates the standard error of the regression coefficients, thereby compromising the precision and reliability of statistical estimates. When multicollinearity is pronounced, it renders the estimation of model parameters unstable, impairing both the model’s interpretative capability regarding the impacts of independent variables and its predictive accuracy. In quantitative research methodologies, tools, such as the Variance Inflation Factor (VIF) and Tolerance (TOL), are routinely utilized to identify and quantify the implications of multicollinearity [54].

In this study, multicollinearity was assessed using the Tolerance (TOL) and Variance Inflation Factor (VIF). The outcomes of this analysis, as presented in

Table 5. Factor multicollinearity analysis.

Design Parameter Name	Variance Inflation Factor	Tolerance
A Plot Vacant Space Location (AVSL)	1.002172	0.997833
B Plot Vacant Space Location (BVSL)	1.001757	0.998246
C Plot Vacant Space Location (CVSL)	1.001309	0.998692
D Plot Vacant Space Location (DVSL)	1.002082	0.997922
A Plot Building Facade Width (AFW)	1.003556	0.996457
B Plot Building Facade Width (BFW)	1.004524	0.995497
C Plot Building Facade Width (CFW)	1.002182	0.997823
D Plot Building Facade Width (DFW)	1.003851	0.996164
Low-rise Building Ratio (LBR)	1.003556	0.996457
Building Height Control Line North-South Shift (NSHS)	1.001960	0.998044
Building Height Control Line East-West Shift (EWS)	1.001835	0.998168
Building Height Control Line Clockwise Rotation (CRC)	1.003508	0.996504

, revealed that the maximum VIF value was 1.004524, adhering to the criteria of VIF < 10 and TOL > 0.1. Consequently, it is verified that the design parameters selected for this research are devoid of multicollinearity.

3.3. Building a Machine Learning Prediction Model

Machine learning predictive models are formulated by analyzing the relationship between provided input data and associated outputs, specifically focusing on how variations in input influence the output. The precision of these models is directly proportional to the volume of data in the training set. It has been identified that design parameter combinations generated through random sampling do not sufficiently encompass all conceivable variable combinations, resulting in less distinctive combinations that compromise the accuracy of the predictive models. Conversely, employing a robust sampling method can enhance the representativeness of design variable combinations, facilitating the construction of more effective machine learning predictive models with fewer data points. In this study, the Latin Hypercube Sampling method [55,56,57] was selected to develop the training set.

A considerable amount of research categorizes algorithms, such as Support Vector Machines (SVM) [58], LightGBM [59], Random Forest [60], AdaBoost [61], and XGBoost [62], among high-performance machine learning predictive modeling tools. In this study, AdaBoost, LightGBM, and Random Forest algorithms were implemented to create predictive models for estimating building photovoltaic power output and energy consumption, alongside models for classifying urban block morphology. The development of these machine learning models was facilitated using the scikit-learn library in Python 3.2.

3.3.1. Light Gradient Boosting Machine

LightGBM (Light Gradient Boosting Machine) is a distributed gradient boosting framework intended to be much faster and more efficient when handling large-scale data. It fits multiple decision trees at separate iterations by building each tree to reduce residuals, step by step, in order to improve them by a small local amount towards the right decision. Unlike other gradient boosting methods, LightGBM uses histogram-based techniques for binning continuous feature values and a leaf-wise method of tree growth. This novel approach does not only foster the identification of split points quickly because continuous features are discretized, but it also optimally minimizes the losses by considering the splits at leaves with maximum gains. In addition, LightGBM copes with high-dimensional sparse datasets and provides multithreading and distributed computing that speeds up the processes of training. Its built-in regularizations make the overfitting possibility less likely. This coefficient is defined as follows:

Initial model [59]:

$$F_0\left(x\right)=\arg \underset{\gamma }{min}\sum _{i=1}^{n}L\left(y_i,\gamma \right)$$

(2)

where L denotes the loss function, which typically utilizes metrics, such as Mean Squared Error (MSE) or cross-entropy loss.

Round M iteration [59]:

$$F_m\left(x\right)=F_{m-1}\left(x\right)+\eta h_m\left(x\right)$$

(3)

where $\eta$ is the learning rate and $h_m\left(x\right)$ is the m weak learner (decision tree).

Steps to update [59]:

$$h_m\left(x\right)=\arg \underset{h}{min}\sum _{i=1}^{n}L\left(y_i,F_{m-1}\left(x_i\right)+h\left(x_i\right)\right)$$

(4)

This formula illustrates that, within each iterative cycle, a new decision tree is constructed by fitting the residuals from the current model.

3.3.2. Adaptive Boosting

AdaBoost (Adaptive Boosting) is a complex ensemble machine learning algorithm that improves the accuracy of classifiers. It builds a strong classifier by combining many weak classifiers. In AdaBoost, for the given dataset, each weak classifier does its prediction, and the sample weights within the algorithm are altered according to the performance of that classifier. This update is specially designed to pay more attention to the samples that have previously been misclassified so as to improve the chances of proper classification in subsequent rounds. The formula is given by the following:

Weight Initialization [61]:

$$D_1\left(i\right)={\displaystyle \frac{1}{N}}$$

(5)

In this formulation, $N$ signifies the total count of samples present in the training dataset, and $D_1\left(i\right)$ specifies the weight of the i-th data point in the first iteration.

Training of Weak Learners:

In each iteration t, a weak learner h_t is carefully selected to minimize the weighted error, considering the distribution of weights D_t at that stage. The weighted error is calculated as follows [61]:

$${ϵ}_t={\displaystyle \frac{{\sum }_{i=1}^N{D}_t\left(i\right)·1(y_i\ne h_t\left(x_i\right))}{{\sum }_{i=1}^N{D}_t\left(i\right)}}$$

(6)

In this formula, 1 (condition) is the indicator function, which assigns a value of 1 if the condition holds true and 0 otherwise.

Weight Calculation for Weak Learners [61]:

$${\alpha }_t={\displaystyle \frac{1}{2}}\log {\displaystyle \frac{1-{ϵ}_t}{{ϵ}_t}}$$

(7)

Here, ${\alpha }_t$ signifies the weight of the t-th weak learner, delineating its role in the final classifier.

Weight Update for Data Points [61]:

$$D_{t+1}\left(i\right)={\displaystyle \frac{D_t\left(i\right)\cdot \exp \left(-{\alpha }_t\cdot y_i\cdot h_t\left(x_i\right)\right)}{Z_t}}$$

(8)

$Z_t$ is employed as a normalization constant to ensure that $D_t\left(i\right)$ is a proper probability distribution [61]:

$$Z_t=\sum _{i=1}^N{D}_t\left(i\right)\cdot \exp \left(-{\alpha }_t\cdot y_i\cdot h_t\left(x_i\right)\right)$$

(9)

Construction of the Final Strong Classifier [61]:

$$H\left(x\right)=\mathrm{s}\mathrm{i}\mathrm{g}\mathrm{n}\left(\sum _{t=1}^T{\alpha }_t{h}_t\left(x\right)\right)$$

(10)

$H\left(x\right)$ represents the ultimate classifier, a cumulative weighted sum of all $T$ weak classifiers, with the final classification outcome determined through the sign function.

3.3.3. Random Forest

The fundamental building blocks of a Random Forest are decision trees, also known as Classification and Regression Trees (CART) [63]. This methodology employs a binary recursive partitioning technique, extensively utilizing binary trees to split the current dataset into two distinct subsets under specified splitting rules. Each non-leaf node in the generated decision tree branches into two, and this bifurcation continues on the subsets until they culminate in leaf nodes, signifying that no further division is feasible. The intrinsic randomness of Random Forest [64] is evident during the sampling phase, where sample selection is performed with replacement, leading to diverse trees each learning distinct segments of the overall dataset’s features. The final predictive results are then aggregated through a voting process for classification tasks, or by averaging for regression tasks, ensuring a comprehensive and robust outcome.

This research will generate 1000 design parameter combinations and their corresponding output values utilizing Latin Hypercube Sampling and building performance simulation methodologies. Distinct predictive models will be constructed using LightGBM, AdaBoost, and Random Forest regression algorithms. Detailed configurations for the machine learning regression hyperparameters are outlined in

Table 6. LightGBM regression hyperparameter setup.

Parameter Name	Parameter Value
Data Split Ratio	0.8
Data Shuffling	Yes
Cross-validation	Yes, 3-fold
Base Learner	gbdt
Number of Base Learners	100
Number of Leaves	31, 41, 51
Learning Rate	0.01, 0.05, 0.1
Number of Trees (Estimators)	50, 100, 200
Normalization of Features	Yes
Normalization of Targets	Yes
Metric for Optimization	Neg mean squared error
Random State	42
Verbose Output	Level 2
L1 Regularization Term	0
L2 Regularization Term	1

,

Table 7. AdaBoost regression hyperparameter setup.

Parameter Name	Parameter Value
Data Split Ratio	0.2
Data Shuffling	Yes
Cross-validation	Yes, 3-fold
Base Learner	Decision Tree Regressor
Number of Base Learners	50, 100, 200
Base Learner’s Maximum Depth	3, 4, 5, 6
Feature Normalization	Yes
Target Normalization	Yes
Random State	42
Verbose Output	Level 2

and

Table 8. Random Forest regression hyperparameter setup table.

Parameter Name	Parameter Value
Data Split Ratio	0.7
Data Shuffling	Yes
Cross-validation	No
Base Learner	Random Forest
Number of Base Learners	100
Maximum Tree Depth	10
Minimum Samples per Leaf	10
Sample Feature Sampling Rate	1
Tree Feature Sampling Rate	1
Node Split Threshold	0
Minimum Weight of Samples in Leaf	0

, with any parameters not specified in these tables defaulting to standard settings. This approach ensures a comprehensive assessment of the predictive capabilities across different algorithmic frameworks.

3.4. Establishment of SHAP Interpretability Analysis Framework

Although machine learning models are renowned for their superior predictive capabilities in complex settings, their operations are often perceived as “black boxes,” complicating efforts to understand their internal decision-making mechanisms. To address this challenge and enhance the interpretability of machine learning predictive models, this study incorporated the SHAP (Shapley Additive explanations) interpretability analysis framework. The SHAP methodology employed herein is a widely acclaimed framework for machine learning explanations and has been successfully applied in practical engineering contexts [65]. The computational formula for Shapley values is detailed below:

$${\varphi }_i=\sum _{S\subseteq F\setminus \left\{i\right\}}{\displaystyle \frac{\left|S\right|(\left|F\right|-\left|S\right|-1)!}{\left|F\right|!}}\left[f_{s\cup \left\{i\right\}}\left(x_{s\cup \left\{i\right\}}\right)-f_s\left(x_s\right)\right]$$

(11)

In the aforementioned expression, φ¡ signifies the Shapley value attributed to the i-th feature, $F$ encompasses all features, and $S$ denotes a subset of all features excluding the i-th feature. The calculation of the Shapley value quantifies the variation in the predicted outcomes of models trained with and without the inclusion of the i-th feature. This is achieved by traversing all conceivable feature subsets that exclude feature i, computing the differential in predictions elicited by the inclusion of feature i, and summing these differentials using a weighted approach to ascertain the Shapley value of the feature. Consequently, the Shapley value of a feature essentially captures the magnitude by which the inclusion of a specific feature shifts the predicted value of a sample compared to the average predicted value.

4. Results

4.1. Data Processing

In this investigation, 1000 sampling iterations were performed, yielding 953 valid objective function results after 170 h, with a success rate of 95.3%, as depicted in Figure 5. An analysis of these 953 sets of design parameters and their corresponding objective function values demonstrated significant sensitivity to changes in design parameters. Specifically, the total annual $E_{Cooling}$ values for the entire block varied significantly, ranging from a maximum of 68,921 kWh to a minimum of 34,322 kWh, a variation of 50.20%. Furthermore, the total annual $E_{Photovoltaic}$ values for the block experienced substantial fluctuation, from a high of 3,027,100 kWh per year to a low of 1,009,200 kWh per year, marking a 66.67% change. Similarly, the total annual $E_{Total}$ values for the block varied from a peak of 2,438,900 kWh per year to a minimum of 787,360.278 kWh per year, a variability of 67.72%. These outcomes underscore that alterations in design parameters can significantly influence the energy performance of block buildings, highlighting the necessity for designers to pay close attention to the substantial variability in E values resulting from identical changes in parameters.

In light of the considerable magnitudes of photovoltaic power generation and total electricity consumption simulated in this research, coupled with the disparate units used for design parameters, the study implemented z-score standardization to preprocess the 953 datasets. This method was selected to ensure uniform data scaling and to enhance the performance of the analytical algorithms. The z-score normalization formula is articulated as follows:

$$z={\displaystyle \frac{\left(X-\mu \right)}{\sigma }}$$

(12)

where $X$ represents the individual data point, $\mu$ is the mean of the dataset, and $\sigma$ is the standard deviation of the dataset. This formula adjusts each data point relative to the dataset’s mean, scaled by the standard deviation, thus facilitating comparisons and integrations across varied scales and units in the dataset.

All datasets slated for machine learning regression predictions in the later stages of this investigation have been subjected to standardized preprocessing.

4.2. Overview of the Simulation Results

The current research evaluated the Pearson correlations between 953 design parameter sets and their respective objective function values, depicted in a heatmap in Figure 6.

The analysis reveals that although most design parameters are weakly correlated with the objective functions, a select few show robust correlations. Specifically, for block photovoltaic generation, LBR is significantly negatively correlated with EP (correlation coefficient −0.44). DFW, CFW, BFW, and AFW demonstrate strong positive correlations with EP (correlation coefficients of 0.23, 0.18, 0.17, and 0.15, respectively), suggesting their pronounced impact on EP. Conversely, other parameters exhibit weak correlations, highlighting their minimal influence on EP. For building cooling energy consumption, LBR’s negative correlation with EC is markedly significant (correlation coefficient −0.95), while other parameters show negligible correlations, indicating their limited influence on EC. With regards to total block electricity consumption, DFW, BFW, CFW, and AFW show fairly strong negative correlations with ET (−0.26, −0.22, −0.21, and −0.20 correlation coefficients, respectively), whereas LBR is rather positively correlated with ET (correlation coefficient 0.24).

In general, the Pearson correlation coefficients of total electricity consumption (ET), cooling energy consumption (EC), and photovoltaic power generation (EP) with most design parameters are quite low, indicating the lack of a significant linear relationship between them. This has led to the development of a wide range of nonlinear regression predictive models, such as Random Forest, AdaBoost, and LightGBM, which deal with looking deeply into the relationship between the inter-variables of the design parameters and the objective functions. These models become very useful in revealing complex, non-linear relationships that would not easily be detected in a linear analysis.

4.3. Machine Learning Regression Predicts Model Results

This paper has conducted experiments on 953 sets of design parameters for objective functions with regression models using LightGBM, AdaBoost, and Random Forest. Results show that based on the whole set of tested functions, LightGBM is consistently superior to the AdaBoost and Random Forest regression model. See Figure 7.

Specifically, the LightGBM regression model excelled in cooling energy consumption, achieving an R² score of 0.918 within the prediction dataset, which underscores its exceptional performance. In terms of photovoltaic power generation, the LightGBM model registered an R² of 0.868, demonstrating its high effectiveness. Furthermore, in the context of total building energy consumption, the LightGBM model attained an R² of 0.885, affirming its robust performance.

In light of the markedly enhanced performance of regression predictive models constructed using the LightGBM algorithm over alternative approaches, this study will persist in employing the LightGBM algorithm for a further examination of design parameters. The performance of the training and test sets under different modes of the LightGBM machine learning model is shown in from Figure A1, Figure A2 and Figure A3.

This investigation reveals that the low-rise building ratio and building facade width are the predominant influencers on block energy consumption. Specifically, the low-rise building ratio and B Plot building facade width are pivotal feature parameters within the cooling energy consumption regression model, contributing feature proportions of 21.35% and 10.63%, respectively. Furthermore, the same parameters for the D Plot in the photovoltaic power generation model account for 16.58% and 11.77%. Additionally, for total building energy consumption, the D Plot building facade width and A Plot building facade width stand out as the essential feature parameters, with respective contributions of 13.71% and 12.24%, as detailed in Figure 8.

4.4. Shap Interpretability Analysis Framework Results

The LightGBM model constructed in this research proficiently captures the interrelationships among features, as evidenced by exemplary performance metrics [34]. Employing the training dataset, this study calculates the marginal contributions of each feature to the predicted outcomes using their SHAP values, thereby delineating the influence of each feature on the dependent variable.

To rigorously analyze the effects of individual features on the dependent variable, this study introduces Beeswarm plots as the analytical instrument. Each point on the Beeswarm plot symbolizes a unique instance, with color variations representing the spectrum of feature values—with blue indicating lower and red indicating higher values [66]. The x-axis of the plot depicts the SHAP values associated with the features, where the magnitude of these values signifies the degree of impact a feature exerts on the outcome. Positive SHAP values denote a beneficial influence on building energy consumption, whereas negative values indicate adverse effects.

Figure 9 delineates the impact levels of various design factors on building photovoltaic (PV) power generation. The findings underscore that the low-rise building ratio (LBR) exerts the most substantial impact on building PV output, evidenced by an average SHAP value of 0.4388, which corresponds to a negative effect. Notably, the building facade width also significantly influences PV generation; this effect is most significant in Plot D, with the highest recorded average SHAP value of 0.2275, whereas Plot A exhibits the smallest impact, with an average SHAP value of 0.2043. Across all four plots, building facade widths contribute positively to PV power generation. The influence of other design factors on building PV power generation remains comparatively minimal, with average SHAP values not exceeding 0.05, underscoring their limited impact on photovoltaic outputs.

Figure 10 elucidates the specific impacts of various design factors on building cooling energy consumption. The study reveals that the majority of design factors negatively influence cooling energy consumption, with the low-rise building ratio (LBR) demonstrating the most substantial impact, evidenced by an average SHAP value of 0.8312. Echoing its effects on photovoltaic power generation, the building facade width markedly influences cooling energy consumption, negatively impacting consumption across all plots. Plot B experiences the most considerable impact, with an average SHAP value of 0.0571, while Plot D shows the least, noted at 0.0520. Furthermore, other design factors that ease cooling energy have quite average SHAP values and stay below 0.050. This means that they hardly contribute to more consumption of energy in terms of cooling.

Figure 11 portrays a clear distinction of how different design factors affect the overall building’s energy consumption. The low-rise building ratio (LBR) and the width of the building façade for various plots are realized as major influencers on the overall energy consumption in block buildings. In particular, the LBR exhibits the most favorable impact in relation to total energy use, with an average SHAP value recorded at 0.3090. Demonstrating a trend similar to that of photovoltaic power generation, the building facade width exerts a great effect on total energy use, the negative influence of which was recorded across all plots. The most significant effect is observed in Plot D, at an average SHAP value of 0.2730, whereas in Plot A, it is observed at only 0.250. For the other design factors, the average SHAP values are very small and even below 0.050, which shows that they have a very minor impact on the total energy demand.

The present study utilized the analyses based on the average SHAP value to develop an insight into the individual influences of the design variables, namely the low-rise building ratio (LBR) and building facade width for different plots, on photovoltaic power output, the cooling energy demand, and total energy in buildings. The results further clarify the way in which these design variables impact energy usage. More precisely, the LBR decreases the photovoltaic output but contributes to a decrease in the total energy consumed. On the other hand, the building facade width consistently favors the photovoltaic performance for each plot and even suggests that with increased facades, power production could be enhanced. However, increasing the width of the facade had a negative impact on the need for cooling energy and total energy consumption, leading to the balancing act that architects have to perform—weighing the benefits of increased efficiency for power generation against the negatives of increased energy usage.

5. Discussion

In the preceding analysis, average SHAP values were utilized to investigate the impact of specific design factors on building photovoltaic power generation, cooling energy requirements, and total energy consumption. The evaluation identified five critical design factors—LBR, AFW, BFW, CFW, and DFW—that influence these energy dimensions differently. The ideal model’s best and worst solutions for building photovoltaic power generation, cooling energy consumption, and total energy consumption are shown in Figure A4, Figure A5 and Figure A6. In pursuit of energy conservation and the efficient harnessing of photovoltaic resources, it becomes imperative to strategically balance these design factors. Such a balance aims to optimize photovoltaic efficiency while concurrently minimizing both cooling and overall energy consumption in buildings. Section 5 will further elucidate the mechanisms through which these design factors modulate building energy dynamics.

5.1. Low-Rise Building Ratio

In the modeling phase, as the plot ratio and total building area were constrained, modifications to the low-rise building ratio (LBR) affected only the average height of the buildings. A decrease in the LBR leads to an increase in the average building height and diminishes the height disparities among buildings. Conversely, an increase in the LBR results in a larger number of buildings shorter than 12 m, compelling the inclusion of several taller structures to adhere to the established plot ratio and total area limits. The augmentation of the LBR corresponds to an increase in the height of these taller buildings, underscoring the direct impact of LBR adjustments on urban architectural profiles.

In this study, we extensively investigated the influence of the low-rise building ratio (LBR) on photovoltaic power output in buildings. The results suggest that reducing the LBR and therefore increasing the number of high-rises is likely to increase the rooftop’s exposure to sunlight by reducing the mutual shading effect between buildings. This plane far extends the effective sun hours for PV panels, thereby leading to a commensurate increase in output from solar power generation. On the other hand, an increase in LBR, which implies more low-rise buildings, will decrease shadows on individual buildings but may also decrease aggregate sunlight exposure on rooftop photovoltaic installations because of lower building heights. This reduction in sunlight exposure will substantially decrease the overall efficiency in power generation, hence the need to carefully consider building heights when optimizing photovoltaic systems in urban designs.

The investigation on the impact of the LBR on building cooling energy consumption indicates that a decreased LBR causes an increment in the cooling load. This increase is predominantly due to a reduced natural ventilation flow among closely spaced high-rise buildings and an increased level of solar radiation exposure to the facades of the buildings, which further enhance the internal temperature. Moreover, clustering taller buildings can heighten the urban heat island impact, thereby further straining cooling systems. Increasing the LBR, on the contrary, encourages better ventilation and reduces the direct contact of large building surfaces with high solar radiation, hence significantly reducing the associated energy demands for cooling. This end further necessitates good architectural planning in urban spaces to strike a good balance between the building density and height and energy efficiency targets.

Based on an analysis of the relationship between total energy use of the building and its LBR, one can generally observe that as the LBR decreases, total use increases. This trend is because, in many instances, elevated energy production from the photovoltaic systems, associated with high-rise buildings, is normally offset through an increase in heating energy use. On the contrary, a higher LBR results in improved energy performance of the building. Although this can lead to a slight decrease in the photovoltaic energy yield, the strong decrease in cooling demand leads to a considerable overall energy saving. These results elucidate that the distribution of building heights in an urban and architectural design must be considered as a crucial step in ensuring effective energy efficiency. This approach is vital toward the development of sustainable urban environments in meeting energy demands but also contributing to the greater goals of energy conservation and sustainability.

5.2. Building Facade Width

The modeling process, the part over which control of the building footprint was placed, stated that because of the changes in the width of the facade involving different plots, it will give the primary light-receiving surfaces. When the facade is 20 m in width, then the length of the building will have a length of 20 m towards the east-west and towards the north-south at 60 m. As this includes a larger width of the facade, then the north-south is the primary means of the light-receiving surface. When the facade is 60 m in width, then the principal light-receiving surface becomes the east-west orientation surface.

In our study, we investigated the effect of changing the width of facades on energy performance in buildings. Changing the facade width from 20 m to 60 m shifts the main light-receiving surface from the north-south to the east-west direction and greatly changes the energy profile of the building. The photovoltaic effectiveness of north-south facades is much better than in other facades, since they can obtain the sunlight for a longer period and more uniformly in the day. This fact is also important in cities like YiChang, where both efficiency and total photovoltaic output could be improved. The east-west facades have high solar radiation at the time of sunrise and sunset, but due to the elevation angle of radiation, power generation efficiency is low, which is reflected by the great importance of the facade orientation in building energy performance.

In the context of energy consumption for cooling, it becomes relevant because buildings having major surfaces that receive light in the east-west direction and receive great exposure to sunlight, especially in the summer months. This results in a raised indoor temperature compared with the surface. As a result, the demand for cooling is much more in the building. In contrast, north-south facing surfaces have long exposure to sunlight during the day but the intensity would be relatively low. The milder exposure hardly raises the same temperature inside a building, reducing the use of cooling energy. These findings demonstrate that the orientation of the primary light-receiving surfaces of a building carries an enormous effect on cooling requirements and therefore should be well thought out during the architectural design stages. Such considerations are crucial for increased energy efficiency and the control of operational costs within building environments.

The orientation of a building’s primary light-receiving surfaces has an important bearing on its energy profile in the context of total energy consumption. Our study showed that the mean energy use for buildings with north-south facade orientations is lower, with smaller cooling demands and higher photovoltaic energy generation efficiencies. Conversely, for an east-west orientation, the overall energy usage is generally higher because of more cooling loads, even though the photovoltaic power outputs are likely to be higher at some part of the day. On that note, in-depth considerations of the facade width and the orientation of the main light-receiving surfaces seem to be an important step for the maximization of energy gains. This strategic consideration is very important because it helps to minimize operational costs and advance sustainable energy practices in managing buildings.

6. Conclusions

In a word, the present study analytically determines the critical effects of such important architectural design variables as the low-rise building ratio (LBR) and facade width on building energy performance in regions with severe summer heat and winter cold. It shows that these parameters play an indispensable role in managing photovoltaic power generation, cooling energy consumption, and overall energy use. Particularly, in the areas of difficult climatic conditions, many elements to improve neighborhood energy performance can be implemented well via strategic design optimizations. For example, a low LBR may enhance the efficiency of the collection of sunlight by photovoltaic installations, so it increases power generation in summer but at the same time increases the need for cooling due to increased heat gain indoors. Conversely, a high LBR would lower the cooling loads at the expense of a slightly reduced efficiency of solar power.

Moreover, the change in the widths of facades deeply impacts both electricity generation from sunlight and the cooling energy demand, with the key switch in orientation of the main surfaces that receive light—from east-west to north-south—defining the dynamics of energy in summer. North-facing and south-facing facades are particularly helpful in extreme seasonal temperature regions as they allow greater uniformity of sunlight through the course of the day so that the amplitude of temperature fluctuations is kept in check to a much lower degree, which reduces cooling loads and thus conserves energy. These findings underscore the necessity of intentional architectural planning that optimizes energy use and the need for more granular studies to enhance building design strategies appropriate for various environmental contexts.

These are very important findings in guiding the future of architectural design and urban planning in extreme seasonal patterns where hot summers and cold winters are a major feature of regions. Under such circumstances, it is important to properly configure design parameters optimizing energy performance. Future studies need to investigate the best values of these design variables that may apply over different geographic and climatic states, and how these can contribute effectively to improved indoor environmental quality and occupant comfort may be realized. This important effort will help guide the building industry toward sustainable, energy-efficient practices that complement architectural development with mainstream environmental sustainability and energy-conservation objectives.

This study leverages computer simulations and machine learning regression models to probe the effects of architectural design factors on the photovoltaic capabilities, cooling energy demands, and total energy consumption of buildings. However, the research is subject to certain limitations. Due to technical and logistical constraints, direct photovoltaic data collection through field measurements was unachievable, precluding direct validation of the simulation outcomes. While the simulation tools deployed have undergone extensive validation and generate results closely mirroring real-world scenarios, future research will aim to corroborate and calibrate these findings with actual field data, thus improving the models’ accuracy and applicability. Moreover, constrained by temporal and computational resources, the analysis focused exclusively on the cooling energy consumption during Yichang’s summer season, neglecting winter heating demands. Although the impact of cooling is more pronounced than heating in this region, addressing winter energy consumption remains essential. Accordingly, future studies will incorporate simulations of winter heating energy demands, facilitating a comprehensive assessment of design factors based on energy utilization and enabling the formulation of more expansive design strategies to enhance annual energy efficiency.