Optimizing rural building design with an intelligent framework integrating BES ANN and MCDM

Abstract

This study addresses the multi-objective trade-offs among energy consumption, thermal comfort, and construction cost in rural buildings by proposing a performance optimization framework that integrates Building Energy Simulation (BES), Artificial Neural Networks (ANN), and Multi-Criteria Decision-Making (MCDM). The method combines DesignBuilder modeling with JePlus batch simulations, incorporates the Morris method for key parameter sensitivity analysis, and utilizes MATLAB to construct an ANN-based prediction model. The TOPSIS approach is then used to select the optimal design solution. This framework significantly improves prediction accuracy and optimization efficiency under high-dimensional design spaces, overcoming the limitations of conventional platforms in convergence speed and computational complexity. A case study of a typical rural house in Chuzhou, Anhui Province, demonstrates that the optimized model reduces total energy consumption by 61.64% and discomfort hours by 32.04%, with an additional cost of ¥73,519.6, achieving a well-balanced improvement in overall performance. The study contributes a novel BES–ANN–MCDM framework, offering a replicable pathway and theoretical foundation for performance-driven, energy-efficient rural building design.

Introduction

As the level of urban construction in China increases and the focus of development shifts to villages and towns, rural areas have become a prominent research topic in recent years. According to 2020 statistics, the rural housing construction area in China covers 23.3 billion square meters, accounting for 42% of the country’s total residential construction area. This large stock signifies significant potential for rural development¹. Currently, energy and environmental issues are of widespread concern, with the building and construction sector responsible for approximately 30% of global energy consumption and 27% of global carbon emissions². As urban and rural areas in China develop in tandem, villagers’ demands for better living environments continue to rise. Rural residences differ significantly from urban ones in terms of construction, with most being self-built and self-financed based on personal preferences or traditional practices³, often lacking consideration for energy-efficient building design. Recognizing this, “actively promoting energy efficiency in rural residences” became a major task for the Chinese government in 2017⁴.

As economic standards and quality of life improve, residents increasingly demand higher levels of indoor comfort, which in turn leads to higher building energy consumption⁵. In recent years, residents’ economic status has also influenced decision-making in construction programs, with the incremental cost of building rural residences being a key concern. This has often resulted in highly subjective and limited construction approaches⁶. Therefore, a comprehensive assessment of the trade-offs between energy consumption, indoor thermal comfort, and cost-effectiveness is both necessary and urgent. Retrofitting building envelopes emerges as one of the most effective strategies for enhancing energy efficiency and improving indoor thermal comfort in building components^7,8. In summary, retrofitting the building envelope of rural residences is essential for reducing energy consumption and discomfort during heating seasons while minimizing incremental costs.

Energy retrofitting of rural buildings

Given the large number of existing rural residences, energy-efficient retrofitting is often more cost-effective and environmentally beneficial than complete reconstruction⁹. Envelope retrofitting provides a viable solution for reducing energy consumption and improving the indoor climate. Several recent studies have focused on identifying the most suitable retrofitting strategies for existing rural residences across different climatic conditions¹⁰. Design variables for envelope retrofitting include insulation of exterior walls, roofs, and floors, the thermal performance of exterior windows, and the type of vertical and horizontal shading¹¹. These design variables have typically been optimized under deterministic conditions, often treating parameters with inherent uncertainties—such as indoor set-point temperature¹², occupancy¹³, window-opening behavior¹⁴, and equipment usage¹⁵—as fixed values. Alexandre found that better building envelope designs significantly reduce energy consumption and improve comfort, emphasizing the importance of climate adaptation strategies, such as passive design, in mitigating the adverse effects of rising temperatures¹⁶.

In terms of research methodology, Cao used the orthogonal test method and entropy weight method to address the multi-objective optimization problem of building envelopes. Their approach assigned different entropy weights to energy efficiency, incremental cost, return on investment, carbon emission reduction, and discomfort hours to compute the final updating strategy¹⁷. Wang utilized the energy simulation program OpenStudio to first conduct energy analyses and then calculate the payback period by comparing the incremental cost economics. This analysis served as the basis for comprehensive optimization¹⁸. Pan and Mei aimed to balance energy efficiency and incremental cost, considering the heating conditions of rural residential buildings in cold regions during the heating period. They established a low-energy design framework through comparisons between a demonstration house and a reference house, incorporating factors such as spatial layout, building type, enclosure system, and heating system¹⁹.

These studies indicate that while single-objective optimizations, such as energy-saving renovations, can be effectively addressed using energy simulation software and mathematical methods, multi-objective optimization involving numerous variable factors requires more complex solutions. This often entails establishing multiple samples for comparison or using entropy weighting and mathematical models to select the best option. However, such approaches face several challenges: (1) Only fixed values of variables can be adjusted, resulting in a limited number of sample cases. (2) The coupling relationships between optimized objectives are not always considered. (3) After creating a detailed model, repetitive parameter adjustments for different sample cases can be time-consuming and labor-intensive.

Multi-objective optimization in rural buildings

In building design or retrofitting, the optimization of design schemes significantly influences the green performance of the building. The design parameters (independent variables) and building performance (dependent variables) are typically high-dimensional²⁰. Traditional optimization methods, which rely on manual calculations and basic software simulations, often struggle to address the complexity posed by numerous independent variables and the computational effort required. This makes it challenging to effectively compare optimization objectives²⁰. With advancements in science and technology, multi-objective optimization methods have become increasingly prevalent in building design. These approaches typically involve performance simulation integrated with algorithmic optimization to enhance building performance across multiple objectives, usually two or three. One such approach leverages the interaction between mathematical-based software (e.g., MATLAB or Python) and energy simulation software (e.g., EnergyPlus, TRNSYS, or ESP-R). Cross-platform integration often requires co-simulation interfaces, such as Jeplus or DesignBuilder, to link the various tools seamlessly.

As these interfaces are primarily connected to energy simulation software, optimization efforts frequently focus on energy savings, carbon emissions, comfort, and cost. For instance, Shao et al. employed the NSGA-II algorithm in combination with EnergyPlus simulation software, the MOBO optimization engine, and the weighted summation method to develop a multi-parameter, multi-objective optimization process for rural residences in the Hanzhong region²¹. Wang proposed a novel multi-objective optimization approach that integrates the NSGA-II algorithm with the entropy-weighted TOPSIS decision-making method to determine the optimal non-ergodic design parameters for rural tourism buildings²⁰. Kang introduced a data-driven integrated method based on PSO, SVM, and NSGA-III to evaluate and optimize objectives such as carbon dioxide emissions, incremental cost, and time discomfort. This approach facilitated the optimization of passive design parameters²².

Another promising approach utilizes the parametric design platform Rhinoceros, along with integrated plug-ins (e.g., Grasshopper, Octopus, Ladybug, and Bee) for geometric modeling, performance simulation, evaluation, and optimization. In addition to energy simulations, these plug-ins enable other analyses such as lighting and wind environment assessments. This method is widely adopted by researchers to optimize building shapes or component sizes. For example, Sun used Rhino and Grasshopper to construct models and conduct performance simulations, analyzing the impact of different building shapes in the Jinan area. Their study aimed to identify building types with low energy consumption and incremental costs, thereby improving early-stage design decision-making²³. Li et al. proposed a multi-objective optimization framework for community morphology based on the Rhino-Grasshopper platform. Their research examined the influence of morphological parameters on performance objectives and suggested energy-efficient design strategies for new rural communities²⁴.

All the intelligent optimization methods mentioned above are genetic algorithms that rely on automatic search solutions, with NSGA-II and NSGA-III being the most widely used. For example, Guo compared the performance of NSGA-II, IDBEA, MSOP-II, and NSGA-III in optimizing four dimensions: energy consumption, effective insolation, comfort time ratio, and incremental cost of the building envelope²⁵. Kang applied NSGA-III to multi-objective optimization problems, integrating predictive models such as Random Forest, Back Propagation Neural Network, and Support Vector Machine. These predictions were then combined with NSGA-III to enhance both prediction accuracy and optimization performance²². Unlike NSGA-II, NSGA-III is specifically designed for high-dimensional objective optimization problems. It employs reference points to maintain a uniform distribution of solutions, making it more suitable for complex multi-objective scenarios. The genetic algorithm generates a set of solutions along the Pareto front, demonstrating superior overall performance compared to other algorithms.

Previous studies often determined the optimal solution from the Pareto front by calculating the Euclidean distance between the Pareto frontier set and the ideal point. However, these approaches typically did not consider the varying importance of different objectives. Scholars such as Luo addressed this limitation by incorporating the weights of objectives, although they assumed equal importance for all objectives²⁶. In practice, the weights of multiple objectives are often unequal. To address this disparity, Mostafazadeh employed the TOPSIS method, a powerful tool for solving real-world multi-attribute or multi-criteria decision-making problems. TOPSIS seeks to find an equilibrium point in a multi-objective optimization problem while incorporating the weights of each objective²⁷. In this study, the entropy-weighted TOPSIS method is used to derive the optimal solution from the Pareto front, ensuring a balanced consideration of varying objective weights.

Research gaps and study focus

Despite increasing efforts in applying multi-objective optimization to building performance, current frameworks still face notable challenges, particularly in the context of rural residential retrofits. Most existing studies adopt generalized optimization approaches that do not fully account for the unique constraints of rural settings—such as construction informality, diverse occupant behaviors, and limited economic investment²⁸. Furthermore, high-dimensional parameter interactions, inconsistent objective weighting, and low adaptability to practical rural scenarios continue to hinder the effectiveness and applicability of many optimization methods. These limitations underscore the need for a more robust and context-aware framework²⁹.

In the context of existing optimization frameworks, it is also important to place this research within the broader landscape of multi-objective optimization for building performance. In recent years, significant contributions have been made by both domestic and international scholars in the field of multi-objective optimization for building performance. However, several shortcomings remain:

• (1) Traditional sensitivity analysis methods often struggle with the complexity of high-dimensional parameters. This study proposes the BES–ANN–MCDM optimization framework and incorporates the Morris sensitivity analysis method, which enhances the existing research frameworks by effectively identifying influential factors and improving optimization accuracy and robustness.

• (2) Traditional multi-objective optimization frameworks often rely on genetic algorithms such as NSGA-II or GA. While effective in solving multi-objective problems, these algorithms struggle with high-dimensional design spaces. NSGA-III, an advanced algorithm designed for such high-dimensional optimizations, requires further investigation, particularly in the context of rural building design. Its application to energy-efficient rural housing and building performance optimization remains largely unexplored and needs more empirical testing.

• (3) A key contribution of this study is the application of the proposed optimization framework to rural housing. The study considers 14 relevant parameters, such as wall and roof insulation thickness, U-value of windows, window-to-wall ratio, shading depth, and air changes per hour. By integrating these parameters, this research validates the BES-ANN-MCDM framework’s practical applicability in rural settings. The study demonstrates the framework’s potential for improving energy-efficient design in rural buildings, contributing to more sustainable rural architecture.

Research methodology

This study proposes an integrated optimization framework—BES–ANN–MCDM—that combines building energy simulation, artificial neural network modeling, and multi-criteria decision-making analysis. By harnessing the complementary strengths of various software platforms and algorithms, the framework enables fast, accurate, and reliable optimization results without compromising simulation precision. The methodology consists of six main steps: (1) constructing the baseline building model; (2) calibrating the model using field-measured data; (3) performing sensitivity analysis via the Morris method to identify key design variables; (4) conducting batch simulations using jEPlus to generate a comprehensive dataset; (5) training ANN models in MATLAB to predict performance indicators; and (6) applying MCDM methods to rank Pareto-optimal solutions and determine the most suitable design alternative.The detailed workflow of this methodology is illustrated in Fig. 1.

Fig. 1

Framework of BES-ANN-MCDM.

For building energy simulation, DesignBuilder was employed as the primary modeling and visualization platform, interfaced with the EnergyPlus simulation engine to perform dynamic analysis of thermal performance, energy consumption, and comfort indicators. DesignBuilder offers a graphical interface and precise parameter control, significantly improving modeling accuracy and efficiency. In addition, jEPlus was used as a parametric simulation manager to automate batch EnergyPlus simulations across multiple design configurations. This integrated toolchain not only accelerates simulation but also ensures a comprehensive dataset for ANN model development. The combined use of these tools balances accuracy, scalability, and flexibility, effectively reducing the simulation cycle and supporting the optimization of complex building designs. The software integration and simulation workflow are illustrated in Fig. 2.

Fig. 2

Framework of software & tools.

In the proposed workflow, DesignBuilder is used for geometry and parameter definition, and the model is manually exported in EnergyPlus IDF format. jEPlus acts as a batch controller by modifying the IDF and associated CSV files to generate parameter combinations. Simulation outputs are manually imported into MATLAB in CSV format for optimization and analysis. This process uses structured data flow without the need for format conversion—IDF and CSV serve as common interfaces between software tools.

BES

Building simulation software

Among the various tools available for building energy simulation, the combination of DesignBuilder and jEPlus offers distinct advantages. DesignBuilder, as a graphical modeling platform based on the EnergyPlus engine, allows for efficient and intuitive input of building geometry, construction details, occupancy schedules, and HVAC systems, thereby reducing modeling complexity and improving simulation accuracy³⁰. Meanwhile, jEPlus provides robust parametric control and enables automated batch execution of EnergyPlus simulations, making it an essential tool for conducting multi-variable sensitivity analysis and generating large-scale datasets³¹. The synergy of these two tools not only enhances modeling efficiency and data consistency, but also significantly reduces simulation time, thereby improving the feasibility and adaptability of building energy simulation in multi-objective optimization workflows.

Sensitivity analysis

In this study, the Morris Sensitivity Analysis (Morris SA) method was employed to identify key input variables that influence output performance indicators, including energy consumption, thermal comfort, and construction cost. The reason for choosing Morris SA is its ability to efficiently rank and identify the most influential parameters in complex, high-dimensional models. Originally proposed by Morris, this method is widely recognized for its computational efficiency and has been extensively used for the preliminary screening of high-dimensional input spaces, especially when computational resources are limited^32,33. Compared with other global sensitivity analysis techniques, Morris SA is particularly suitable for early-stage variable ranking in complex simulation environments, such as building energy models, where the number of input variables is large, and computational efficiency and interpretability are critical considerations³⁴. The method’s ability to handle large input spaces with relatively low computational cost makes it an ideal choice for optimizing model parameters in the early stages of design.

The Morris method evaluates sensitivity using two main statistical metrics: Mu* (the absolute mean) and Sigma (the standard deviation). Mu* measures the overall strength of an input variable’s influence on the output; a higher Mu* indicates greater overall importance. Sigma captures the uncertainty or interaction effects between variables; a larger Sigma suggests nonlinear relationships or significant interactions with other inputs. Thus, Mu* serves to identify dominant influencing factors, while Sigma helps flag variables requiring further investigation due to potential complexity.The formulas for these indices are as follows:

$$Mu*:\,\,\,\,\,\,\,\,\,\,\,\,\,\mu _{i}^{*} = \frac{1}{r}\mathop \sum \limits_{{j = 1}}^{r} \left| {EE_{i}^{{\left( j \right)}} } \right|$$

(1)

Where EE_i^(j) represents the elementary effect of input variable X_i in the j-th trajectory, and r is the number of trajectories.

$${\text{Sigma}}:\,\,\,\,\,\,\,\sigma _{i} = \sqrt {\frac{1}{{r - 1}}\mathop \sum \limits_{{j = 1}}^{r} \left( {EE_{i}^{{\left( j \right)}} - EE_{i} } \right)^{2} }$$

(2)

where EE_i is the mean of the elementary effects for input variable X_i.

In this study, the Morris tool integrated within jEPlus was used to perform the sensitivity analysis. The results were visualized using a Mu*–Sigma scatter plot to facilitate the identification of input parameters with both high impact and stability, supporting the subsequent modeling and optimization process.

ANN and multi-objective optimization

Artificial Neural Networks are powerful data-driven tools used to develop highly accurate models³⁵ and have been widely applied in the field of building energy efficiency. Despite the theoretical challenges in explaining the internal mechanisms of neural networks, they have demonstrated exceptional performance in predicting and optimizing complex systems³⁶. In this study, we employ an artificial neural network model to address the multi-objective optimization problem, ensuring that each objective remains independent throughout the optimization process. In current research, Surrogate models can overcome the building performance simulation complexity and replace time-consuming simulation engines within multi-objective optimization studies³⁷.

The basic architecture of the neural network model is shown in Fig. 3. The network consists of an input layer, multiple hidden layers, and an output layer, with information transmitted between layers through weighted connections. Each node receives a weighted input from the previous layer, applies a nonlinear activation function, and passes the result to the nodes in the next layer.The ultimate goal is to achieve the optimal solution for multi-objective building design through optimization algorithms. While NSGA-II has been widely used to address Multi-Objective Optimization problems with two objectives in building performance, few studies have explored NSGA-III, which has shown superior performance in solving problems with three or more objectives²². Therefore, this paper employs NSGA-III to solve the multi-objective optimization problem.

Fig. 3

Flowchart of neural network and algorithm optimization.

In this study, MATLAB was selected for both the modeling and optimization stages instead of relying on integrated visual platforms such as Grasshopper or jE + EA. This choice was driven by the need for greater computational efficiency and modeling accuracy. Compared to visual programming platforms, MATLAB provides superior flexibility, stability, and control in numerical operations, algorithm implementation, and result processing, significantly reducing the total computation time and improving workflow efficiency. When handling large datasets and iterative optimization tasks, MATLAB enables faster execution and more consistent output, along with advanced error evaluation and parameter tuning capabilities to ensure accuracy and convergence quality. Therefore, this study adopts a multi-tool strategy, using jEPlus for batch simulation and MATLAB for post-processing, modeling, and optimization, effectively enhancing the efficiency and robustness of the overall workflow. As shown in Table 1, the DesignBuilder + jEPlus + MATLAB workflow demonstrates superior performance in simulation accuracy, parameter control, and optimization efficiency compared to integrated platforms like Grasshopper or jE + EA. This makes it especially well-suited for high-precision, data-intensive, and multi-objective research scenarios in building performance optimization.

While the MATLAB-based workflow offers high computational precision and flexibility, it is important to acknowledge the advantages of integrated platforms. These platforms provide user-friendly graphical interfaces, particularly beneficial for architects and designers, and support seamless 3D modeling. Moreover, integrated environments allow streamlined processes where modeling, simulation, and optimization can be executed within a single platform, reducing tool-switching complexity and enhancing early-stage design exploration.

Table 1 Comparison of frame patterns.

MCDM

Multi-Criteria Decision Making is widely applied in complex optimization problems such as building energy performance evaluation, where trade-offs among conflicting objectives must be made. MCDM provides a systematic approach to rank and select optimal solutions based on multiple performance criteria and their respective weights. In this study, MCDM is used to evaluate and rank the Pareto-optimal solutions derived from the optimization process. Considering both robustness and practical applicability, the TOPSIS is adopted as the core method. Initially proposed by C.L. Hwang and K. Yoon, the TOPSIS method ranks alternatives based on the shortest distance to the ideal solution and the farthest distance from the negative ideal solution³⁸. The decision-making results are highly sensitive to the weights and normalization methods used. The entropy weighting method provides an objective way to determine the weights and is commonly applied in TOPSIS-based approaches³⁹. Moreover, according to previous studies, Min-Max Normalization is the most frequently used normalization method in entropy-weighted TOPSIS calculations. Therefore, this method is adopted in the present study. The detailed procedure for implementing the entropy-based TOPSIS method is outlined below.

• (1) A judgment matrix was constructed. The matrix is normalized using the MMN method, taking into account the differences in units and orders of magnitude of the attributes. The structure of the normalized matrix can be expressed as follows.

  $$\begin{array}{*{20}c} {\begin{array}{*{20}c} {Z_{1}^{'} Z_{2}^{'} \cdots Z_{m}^{'} Z^{\prime} = \begin{array}{*{20}c} {A_{1} } \\ {A_{2} } \\ { \vdots A_{n} } \\ \end{array} \left[ {\begin{array}{*{20}c} {Z_{{11}}^{'} } & {Z_{{12}}^{'} } & \cdots & {Z_{{1m}}^{'} } \\ {Z_{{21}}^{'} } & {Z_{{22}}^{'} } & \cdots & {Z_{{2m}}^{'} } \\ \vdots & \vdots & \ddots & \vdots \\ {Z_{{n1}}^{'} } & {Z_{{n2}}^{'} } & \cdots & {Z_{{nm}}^{'} } \\ \end{array} } \right],} \\ \end{array} } \\ \end{array}$$

  (3)

  Where A₁, A₂, …, An denote the alternatives; Z’₁, Z’₂, …, Z’_m represent the evaluation indexes of the alternatives; Z ‘_ij denotes the performance value of the j^th index in the i^th alternative after normalization; i = 1, 2, …, n; and j = 1, 2, …, m.

• (2) Where E₁, E₂, …, En is the list of design measures in Table 6; n is the total number of measures; n = 14; S₁, S₂, …, S_m represent the targets in the SA; m = 3 is the total number of targets (EC, TNC and ESEC); and S_ij is the SRC value from the i^th design variable to the m^th target.

  $$\:\begin{array}{c}{\text{E}}_{\text{i}}=-\frac{1}{\text{l}\text{n}\text{n}}\sum\:_{\text{i}=1}^{\text{n}}\:{\text{p}}_{\text{i}\text{j}}\text{l}\text{n}{\text{p}}_{\text{i}\text{j}},\end{array}$$

  (4)

  Where n is the number of programs. P _i·j are the contribution of different indices which can be calculated according to the following equation:

  $$\:\begin{array}{c}{\text{p}}_{\text{i}\cdot\:\text{j}}=\frac{{\text{Z}}_{\text{i}\text{j}}^{{\prime\:}}}{\sum\:_{\text{i}=1}^{\text{n}}\:{\text{Z}}_{\text{i}\text{j}}^{{\prime\:}}},\end{array}$$

  (5)

  The attribute weights (w_i) can then be calculated using the following equation:

  $$\:\begin{array}{c}{\text{w}}_{\text{i}}=\frac{1-{\text{E}}_{\text{i}}}{\sum\:_{\text{i}=1}^{\text{n}}\:\left(1-{\text{E}}_{\text{i}}\right)},\end{array}$$

  (6)

  Finally, the quantized values can be put into the above decision matrix:

  $$\:\begin{array}{c}{\text{x}}_{\text{i}\text{j}}={\text{w}}_{\text{j}}{\text{z}}_{\text{i}\text{j}},\end{array}$$

  (7)

  Where w_j is the weight of the j^th index.

• (3) Positive ideal solutions (x+) and negative ideal solutions (x) are determined using a weighted normalized decision matrix.

  $$\:\begin{array}{c}{\text{x}}^{+}=\left\{{\text{x}}_{1}^{+},{\text{x}}_{2}^{+},\dots\:,{\text{x}}_{\text{m}}^{+}\right\}=\left\{\text{m}\text{a}{\text{x}}_{\text{i}\text{j}}\right\},\end{array}$$

  (8)
  $$\:\begin{array}{c}{\text{x}}^{-}=\left\{{\text{x}}_{1}^{-},{\text{x}}_{2}^{-},\dots\:,{\text{x}}_{\text{m}}^{-}\right\}=\left\{\text{m}\text{i}{\text{n}}_{\text{i}\text{j}}\right\},\end{array}$$

  (9)

• (4) Euclidean distance from each alternative to the positive ideal solution

  $$\:\begin{array}{c}{\text{D}}_{\text{i}}^{+}=\sqrt{\sum\:_{\text{j}=1}^{\text{m}}\:{\left({\text{x}}_{\text{i}\text{j}}-{\text{x}}_{\text{j}}^{+}\right)}^{2}}\end{array},$$

  (10)
  $$\:\begin{array}{c}{\text{D}}_{\text{i}}^{-}=\sqrt{\sum\:_{\text{j}=1}^{\text{m}}\:{\left({\text{x}}_{\text{i}\text{j}}-{\text{x}}_{\text{j}}^{-}\right)}^{2}}\end{array},$$

  (11)

• (5) Calculate the relative proximity (Ci) of each alternative and rank the order of preference.

  $$\:\begin{array}{c}{\text{C}}_{\text{i}}=\frac{{\text{D}}_{\text{i}}^{-}}{{\text{D}}_{\text{i}}^{-}+{\text{D}}_{\text{i}}^{+}},\end{array}$$

  (12)

  Where 0 ≤ Ci ≤ 1. The relative proximity (Ci) value indicates the relative strength of each alternative. A higher Ci value suggests a better alternative, while a lower Ci value indicates a less favorable one. Alternatives are ranked in descending order of their Ci values, with the alternative having the highest Ci value being the best solution.

Establishing the benchmark model and objectives

Model information

Site survey

In recent years, new types of rural areas have been rapidly developing in China, particularly in the southeastern regions. Unlike the coastal cities of southeastern China, Anhui Province possesses a considerable amount of rural resources and is significantly influenced by the economic radiation of the southeastern economic belt. This has led to rapid economic development and numerous rural renewal and reconstruction projects. As a result, farmhouse renovation in Anhui Province offers an opportunity to explore new development directions for rural areas in southeastern China, making it a valuable subject for research. The case study is located in Chuzhou City, a prefecture-level city in Anhui Province, situated between latitudes 32°27′27″N to 32°57′36″N and longitudes 118°39′19″E to 119°13′23″E. Additionally, rural development in Chuzhou City is thriving, with abundant resources available for farmhouse renovation. Therefore, this study focuses on existing rural residences in Chuzhou City, Anhui Province, and includes a survey of data from 53 rural households, as shown in Table 2.

Table 2 Rural residence survey statistics.

Through a detailed survey of local rural residences, covering characteristics such as architectural style, layout, structure, materials, and the year of construction, it was found that existing rural houses exhibit a high degree of similarity in architectural form. The survey results indicate that single-story houses are gradually being phased out in rural areas, particularly in newly constructed homes, where two-story houses have become the dominant building type, accounting for approximately 94.3% of the total. Therefore, this study focuses on two-story houses as the research subject. Based on functional requirements, spatial layout, housing scale, and the living habits of rural residents, and in combination with field survey data, a typical case was selected to construct a benchmark model for rural residences in the region. The floor plan of this model is shown in Fig. 4. The measured indoor and outdoor air temperatures, as well as the indoor and outdoor relative humidity, are presented in Fig. 5. The summer measurements were conducted from July 16 to July 21, 2024, and the winter measurements were taken from February 8 to February 12, 2025.

Fig. 4

Plan view of the benchmark model.

Fig. 5

Benchmark farmhouse temperature and humidity.

Weather files

Local climate conditions must be considered in the design process for optimal thermal performance. This study further demonstrates that only through climate-responsive and tailored design strategies can both thermal comfort and energy efficiency be effectively achieved^40,41. In the dynamic simulation, the weather data used comes from the China Standard Weather Data (CSWD) available on the EnergyPlus website. However, due to the lack of meteorological data for Chuzhou in the CSWD database, this study used meteorological data from the neighboring city of Bengbu as a substitute. Bengbu and Chuzhou are located in similar climatic regions, and they exhibit high similarity in climate indicators such as temperature, humidity, and precipitation. Comparative analysis of the relevant climatic characteristics shows that the meteorological data from Bengbu can accurately represent the climate conditions of Chuzhou. Therefore, the meteorological data from Bengbu is considered a reasonable substitute for Chuzhou’s climate data. The specific weather data is shown in Fig. 6.

Fig. 6

Psychrometric chart based on the weather file.

Benchmark model

The benchmark model is created using DesignBuilder, a 3D modeling and simulation software based on EnergyPlus. The specific model is shown in Fig. 7. In the simulation, the dimensions of doors and windows are set according to the actual architectural drawings. As the building facades of rural houses vary, the façade details are adjusted accordingly, and the maintenance structure is defined based on research findings during the modeling process. The materials and calculation methods used in the model comply with established standards. The base values for each parameter in the baseline model are listed in Table 3. Additionally, parameter values and typical daily operation schedules for equipment, lighting, and personnel activities in the corresponding thermal zones are determined according to local standards, as shown in Table 4. As shown in Table 5, this schedule is based on the default residential settings in DesignBuilder, which are primarily derived from ASHRAE and CIBSE standards, and has been further adjusted to reflect the characteristics of the studied rural buildings.

Fig. 7

Benchmark model diagram.

Table 3 Baseline model information sheet.

Table 4 Internal load and temperature settings for major zones.

Table 5 Occupancy schedule for activities.

Input parameters

The parameters related to building characteristics, equipment systems, and personnel activity are set according to relevant local standards. The threshold and variation ranges for each parameter are listed in Table 4. During the batch simulation phase, Jeplus simulation software, which integrates with EnergyPlus and DesignBuilder, is used to conduct large-scale simulations by varying a range of parameters. The simulation input parameters are detailed in Table 6.The design variables for envelope retrofitting include the insulation of exterior walls, roofs, and floors, the thermal performance of exterior windows, and the type of vertical and horizontal shading¹¹. In this study, the retrofit simulation modules are divided into four categories: exterior walls and roof, exterior windows, and shading. Optimizing heat transfer through the building envelope is crucial for designing energy-efficient buildings⁴², so the thermal insulation performance of the walls is a key parameter considered in the study.

Building energy loss through windows is estimated to account for 20–40% of the total building energy loss⁴³. Therefore, window characteristics, such as window area and U-value, are also considered in the simulation. Additionally, applying appropriate shading strategies can significantly reduce direct solar heat gain through windows, optimizing both cooling and artificial lighting loads⁴⁴. As a result, the study also takes into account the depth and angle of shading panels.The impact of passive building strategies, such as building orientation, on the simulated environment is also explored. Heating, Ventilation, and Air Conditioning (HVAC) systems are among the most significant contributors to global energy consumption⁴⁵. Therefore, the batch simulation in Jeplus is coupled with HVAC parameters, including temperature settings and the number of air changes per hour. A field survey revealed that the farmhouses primarily use split-type air conditioners for cooling and heating, with most of them having an energy efficiency rating of 3. The minimum allowable energy efficiency ratios for cooling and heating are 3.2 and 2.2, respectively.

Table 6 Jeplus input parameters.

Model validation and confirmation

To quantify the accuracy of the building energy simulation results in comparison to the measured data, the evaluation method provided by ASHRAE 14-2002 was used. This standard employs two dimensionless error indices—Mean Bias Error (MBE) and Coefficient of Variation of the Root Mean Square Error (CV(RMSE))—to assess the deviation between the simulation results and the actual measured values. The calculation formulas for these indices are as follows:

$$\:\begin{array}{c}\text{M}\text{B}\text{E}=\frac{\sum\:_{\text{i}=1}^{\text{N}}\:\left({\text{M}}_{\text{i}}-{S}_{\text{i}}\right)}{\sum\:_{\text{i}=1}^{\text{N}}\:{\text{M}}_{\text{i}}}\end{array}$$

(13)

$$\:\begin{array}{c}\begin{array}{c}\&\text{C}\text{V}\left(\text{R}\text{M}\text{S}\text{E}\right)\hspace{0.25em}\hspace{0.25em}\hspace{0.25em}\hspace{0.25em}\&\&=\frac{\sqrt{{{\Sigma\:}}_{\text{i}=1}^{\text{N}}\left[\frac{{\left({\text{M}}_{\text{i}}-{S}_{\text{i}}\right)}^{2}}{{\text{N}}_{\text{i}}}\right]}}{\frac{1}{\text{N}}\sum\:_{\text{i}=1}^{\text{N}}\:{\text{M}}_{\text{i}}}\end{array}\end{array}$$

(14)

Where M and S represent the i^th measured and simulated data, respectively, and N is the total number of measurement or simulation data points. The closer the two evaluation indices are to 0, the higher the consistency between the simulation results and the measured data, indicating that the building model and parameter settings closely reflect the real conditions. Conversely, larger values suggest greater discrepancies. According to the standard, if the Mean Bias Error is less than 10% and the Coefficient of Variation of the Root Mean Square Error is less than 30%, the simulation results are considered acceptable. This means the deviation from the actual data is within the permissible error range, and the simulation results align well with the measured data. In this case, the parameter settings in the building model accurately reflect the real-world situation, and the model is deemed sufficiently accurate.

Objectives and assessment indicators

Energy consumption (EC)

Usually, the annual energy consumption of a building includes heating and cooling energy consumption, lighting energy consumption, and other electrical equipment energy consumption. Considering that the energy used for indoor lighting can indirectly reflect the adequacy of indoor natural lighting, the energy consumption in this study includes cooling, heating and lighting energy consumption. Therefore, the energy consumption can be calculated from the function. Where EC is the annual energy consumption per unit area of the building in kWh/m²; $\:{\text{e}}_{\text{h}}$ is the heating energy consumption generated during the heating period (kWh); $\:{\text{e}}_{\text{c}}$ is the cooling energy consumption generated during the cooling period (kWh); $\:{\text{e}}_{\text{l}}$is the total energy consumption generated by the artificial lighting for the whole year (kWh); and A is the total building area (m2).

$$\:\begin{array}{c}\text{E}\text{C}=\frac{{\text{e}}_{\text{h}}+{\text{e}}_{\text{c}}+{\text{e}}_{\text{l}}}{\text{A}}\end{array}$$

(15)

Time not comfortable (TNC)

Thermal comfort is a key metric in assessing both building energy efficiency and indoor environmental quality. To accurately evaluate the impact of the thermal environment on occupants, this paper utilizes EnergyPlus simulation software in combination with the ASHRAE 55-2004 standard. The ASHRAE 55 standard defines thermal comfort conditions, using the Predicted Mean Vote (PMV) and Predicted Percentage of Dissatisfaction (PPD) models to assess comfort levels.

According to the ASHRAE 55-2004 thermal comfort model, the PMV model calculates a thermal comfort score based on factors such as temperature, humidity, wind speed, metabolic rate, and clothing insulation. A PMV value close to 0 indicates that the environment is considered neutral and comfortable for the human body. On the other hand, PPD reflects the percentage of people dissatisfied with the thermal conditions, with a lower PPD indicating greater comfort for a larger portion of the population. Equation (1) is one of the core equations in the comfort evaluation model of ASHRAE 55, used to calculate the PMV value:

$$\:\begin{array}{c}\text{P}\text{M}\text{V}=\text{f}\left(\text{T},\text{R}\text{H},\text{v},\text{M},\text{C}\text{l}\text{o}\right)\end{array}$$

(16)

Where T is the air temperature, RH is the relative humidity, v is the air flow rate, M is the metabolic rate, and Clo is the clothing insulation value. The thermal comfort score derived from this formula is used to evaluate the thermal comfort of occupants under the current environmental conditions.

To further quantify the degree of thermal comfort, this paper introduces the indicator ‘TNC’ (Time Not Comfortable) to assess the duration of discomfort. This metric is analyzed alongside the simulation data to explore the effectiveness of various design solutions in improving the indoor thermal environment. For this study, the PMV range for comfortable conditions is defined as between − 0.5 and + 0.5, which is consistent with the ASHRAE 55-2004 standard. PMV values outside this range indicate discomfort, with values further from zero representing more extreme discomfort.

Energy-saving envelope cost (ESEC)

Incremental cost of energy consumption as an indicator of the economics of optimizing energy efficiency in farm buildings. This metric specifically calculates the additional costs associated with components related to energy-saving technologies, such as insulation, exterior windows, equipment systems, and airtightness⁴⁶. In the optimization process, we focus on the incremental costs associated with insulation and windows, as these are the primary controllable parameters. This combined cost is referred to as the Energy Saving Envelope Incremental Cost (ESEC). The ESEC is calculated using the following function:

$$\:\begin{array}{c}\text{E}\text{S}\text{E}\text{C}={\text{C}}_{\text{I}}+{\text{C}}_{\text{W}\text{D}}\end{array}$$

(17)

$$\:\begin{array}{c}{\text{C}}_{\text{I}}={\text{P}}_{\text{I}}\left({\text{T}}_{\text{W}}\sum\:_{\text{i}}^{\text{O}}\:{S}_{\text{E}\text{i}}\left(1-\text{W}\text{W}{\text{R}}_{\text{i}}\right)+{\text{T}}_{\text{R}}{S}_{\text{R}}\right)\end{array}$$

(18)

$$\:\begin{array}{c}{\text{C}}_{\text{W}\text{D}}={\text{P}}_{\text{W}\text{D}}\left(\sum\:_{\text{i}}^{\text{O}}\:{S}_{\text{E}\text{i}}\text{W}\text{W}{\text{R}}_{\text{i}}\right)\end{array}$$

(19)

In the formula, ESEC represents the incremental cost of the building’s energy-efficient envelope (CNY); CI is the incremental cost of building insulation (CNY); CWD is the incremental cost of constructing windows (CNY); PI is the unit price of insulation (CNY/m³); TW is the thickness of the exterior wall insulation layer (m); i denotes the four orientations (S, N, E, W); S_Ei.

is the enclosure area of the i^th orientation (m²), including both window and wall areas; WWR_i is the window-to-wall ratio of the i^th orientation; TR is the thickness of the roof insulation layer (m), denoted by variable X5; SR is the surface area of the roof (m²); and PWD is the unit price of the window (CNY/m²).

Results and analysis

Sensitivity analysis results

In this study, the Morris sensitivity analysis method is applied to assess the influence of uncertain input variables on the optimization objectives. One of the objectives, total energy consumption, is divided into heating, cooling, and lighting demands. Since lighting energy shows a perfect linear regression fit (R² = 1.0), its analysis is omitted. Sensitivity analyses for heating and cooling demands are shown in Figs. 8 and 9, while the analysis of uncomfortable time is presented in Fig. 10. The sensitivity analysis of incremental envelope cost is not detailed separately due to its dependency on unit prices and material areas.

The Morris analysis provides two core indicators: Mu* and Sigma, which reflect the global influence and variability of each input variable, respectively. As shown in the Morris plot, XPS thickness and heating setpoint exhibit the highest Mu* values, indicating they are the dominant factors affecting heating energy demand. A thicker insulation layer (XPS) significantly reduces heating demand, while a higher heating setpoint and window U-value both lead to increased energy consumption. The solar heat gain coefficient shows a strong negative correlation with heating energy, suggesting that maximizing solar gain helps reduce heating loads. Room air change rate and indoor airflow rate also display moderate impacts, mainly by increasing heating demand when values are higher. While window-to-wall ratio and sunshade angle contribute to heat loss control, their impact remains comparatively limited.Overall, variables with high Mu* and Sigma values—including XPS thickness, heating setpoint, and window U-value—should be prioritized for optimization. These findings highlight the need to enhance insulation performance, fine-tune heating temperature control, and make full use of passive solar heat gain in rural building design.

Fig. 8

Regression and Morris plots under heating objectives.

Under the objective of cooling energy consumption, the regression results show that the cooling setpoint has a significant and negative influence on cooling demand, indicating that lowering the setpoint increases energy use. The solar heat gain coefficient (SHGC) exhibits a strong positive effect, suggesting that higher SHGC values lead to greater cooling loads. The XPS insulation thickness also plays a critical role, with a large and negative coefficient, meaning that increasing insulation thickness effectively reduces cooling energy consumption.In addition, shading depth and window-to-wall ratio (WWR) significantly influence cooling performance. Deeper shading and higher WWR can reduce solar heat gain and cooling demand. While window U-value and ventilation rate also impact cooling loads, their effects are relatively modest.

The results of the Morris sensitivity analysis further validate these findings. The Mu* value of the cooling setpoint is the highest among all variables, emphasizing its importance as a key control parameter. Both SHGC and XPS thickness also show high Mu* values, indicating substantial and consistent impacts. Moreover, high Sigma values for XPS thickness and shading suggest variability or interaction effects, which should be considered during design optimization. In summary, the most influential factors affecting cooling energy consumption include the cooling setpoint, solar heat gain coefficient, XPS insulation thickness, and window-to-wall ratio. Effective cooling strategies should prioritize increasing insulation levels, optimizing solar shading, fine-tuning window configurations, and reasonably adjusting the cooling setpoint.

Fig. 9

Regression and Morris plots under the cooling objective.

For the objective of discomfort time, the standardized coefficient of the window-to-wall ratio is the highest and positively significant, indicating that increasing the window area tends to increase indoor discomfort. Larger window surfaces can intensify indoor temperature fluctuations and solar heat gain, thereby prolonging discomfort periods. The solar heat gain coefficient (SHGC) also shows a strong positive effect, further confirming that excessive solar radiation aggravates thermal discomfort. In contrast, the coefficient of XPS insulation thickness is negative, suggesting that thicker insulation improves thermal stability and reduces discomfort time.Both the depth of the external shading and the shading angle also exhibit negative effects on discomfort time, indicating that enhanced shading strategies can effectively reduce indoor overheating. Roof insulation thickness and window U-values similarly contribute to improved comfort, though their impact is relatively moderate.

The results of the Morris sensitivity analysis align with the regression findings. The Mu* values of the window-to-wall ratio and SHGC are the highest among all variables, highlighting their dominant role in discomfort generation and the need for careful control. Meanwhile, shading depth and shading angle show high Mu* but low Sigma values, indicating they exert a strong yet predictable influence, making them ideal targets for passive design optimization. In summary, the key factors influencing discomfort time include the window-to-wall ratio, solar heat gain coefficient, XPS insulation thickness, and shading design. To reduce discomfort time, design strategies should focus on increasing insulation thickness, optimizing window layouts, refining shading schemes, and limiting excessive solar gain.

Fig. 10

Regression and Morris plots under thermal comfort objectives.

Multi-objective optimization results

Optimization was performed using the neural network fitting toolbox in MATLAB, combined with the NSGA-III genetic algorithm for multi-objective screening and optimization. The objective of the multi-objective optimization (MOO) is to minimize energy consumption (EC), discomfort time, and incremental construction cost for each region. As described in Sect. 4.1, the influence of each input parameter was first evaluated using Morris sensitivity analysis. Variables with low Mu* values were identified as insignificant and excluded from subsequent jEPlus batch simulations.Due to the extremely large number of possible input combinations (2,009,306,250 cases), full simulation was not feasible. To address this, the Latin Hypercube Sampling (LHS) method was adopted to extract 10,000 representative samples. LHS is a stratified sampling technique that ensures the entire parameter space is efficiently and evenly explored, improving the coverage and representativeness of training data. These samples were then used for neural network training in MATLAB.The model fitting results are shown in Fig. 11. The R-squared values of the neural network models for EC, discomfort time, and incremental cost were 0.98093, 0.94191, and 0.99999, respectively. These results indicate excellent predictive performance. The trained models accurately capture the relationships between input variables and the three objectives. The convergence of mean squared error (MSE) curves during training also demonstrates strong generalization and model reliability.

Fig. 11

Neural network validation diagram.

A multi-objective genetic algorithm with 200 generations and a population size of 100 was employed to calculate the Pareto solution set. The number of Pareto frontier solutions generated was 100, and the results are presented in Fig. 12. The optimal values for EC, TNC, and ESEC reached 15.189(kWh/m²), 3570.863(h), and 1733.601(CNY), respectively. For comparison, the pre-retrofit values for EC and discomfort time were 58.25(kWh/m²), and 5664.5(h), respectively. From the Pareto-optimal solution set, the following observations were made: When EC is minimized at 15.189(kWh/m²), discomfort time is reduced to 4639.841(h), achieving a 73.92% reduction in energy consumption and an 18.09% reduction in discomfort time. The incremental cost in this case is 18,882.029(CNY). When discomfort time is minimized at 3570.863(h), EC increases to 17.752kWh/m²), and the incremental cost rises to 106,923.313(CNY). Energy savings are 69.52%, and discomfort time is reduced by 36.96%. When incremental cost is minimized at 1733.601(CNY), EC is 17.110(kWh/m²), and discomfort time is 4362.399(h). Energy savings are 70.63%, and discomfort time is reduced by 22.99%. These results highlight the trade-offs between energy consumption, discomfort time, and incremental cost, providing valuable insights for optimizing building retrofitting strategies.

Fig. 12

Genetic algorithm for objective optimization graph.

The parallel coordinate plot of the 100 individuals in the Pareto solution set, generated from the multi-objective optimization, is shown in Fig. 13, illustrating the distribution characteristics of the Pareto front. Regarding the input variables, most of them are relatively concentrated. For instance, the value of XPS thickness primarily ranges from 0 to 0.5, reflecting the significant impact of insulation material thickness on the optimization results. The Cooling Sp value ranges from 25 to 28, indicating that the variation in the cooling setpoint temperature is relatively small, which may be related to the thermal comfort objective. Additionally, the value of Orientation ranges from 329 to 360, indicating that the orientation design is mostly concentrated around the south-facing direction. In terms of the distribution of objective variables, energy consumption mainly falls within a lower range, concentrated between 15 and 25(kWh/m²). The number of discomfort hours is primarily distributed between 3000 and 5000, meeting the objective of achieving relatively low thermal comfort. The cost distribution spans from 15,000 to 130,000(CNY), with color coding showing that solutions with higher costs typically correspond to lower energy consumption and comfort levels. This further validates the trade-off characteristics between objectives in the multi-objective optimization.

Fig. 13

Parallel coordinate plot of the solution set.

MCDM results

The results in Sect. 4.2 show that the optimal solutions for farmhouse optimization are evenly distributed across the Pareto front. However, this distribution does not suggest that one solution is superior to the others. To further refine the decision-making process, the entropy-based TOPSIS method is applied to select the optimal solution from the Pareto optimal set. The relative proximity is determined by calculating the distances of each solution from the positive ideal solution (DPD_PDP) and the negative ideal solution (DND_NDN). A higher proximity value indicates a solution closer to the positive ideal, signifying better performance.

Using the calculation method outlined in Sect. 2.3, the entropy weights for EC, TNC, and ESEC are derived as 0.282, 0.247, and 0.471, respectively. This suggests that incremental cost carries a higher weight, while energy consumption and thermal comfort have relatively lower weights. The relative proximity of the 100 Pareto scenarios was calculated, and the results, shown in Fig. 14, indicate that scenario 54, with a relative proximity of 0.573, represents the optimal solution based on the integrated decision-making approach.

Fig. 14

Relative proximity of pareto solution sets.

In order to further illustrate the rationality of the comprehensive optimal solution, the schemes under different optimal objective states were selected from the set of Pareto feasible solutions for analysis and comparison. The optimal solutions for individual objective EC, uncomfortable time, ESEC, and the optimal TOPSIS solution are summarized in Table 7.

Table 7 Optimization of different scenarios.

The results in Table 7 highlight the varying characteristics and applicability of different scenarios in terms of EC, TNC, ESEC. The base scenario, serving as the reference point, shows the highest values for both energy consumption and discomfort time, illustrating the inefficiencies of the unoptimized building design in terms of energy performance and indoor comfort.

In contrast, the TOPSIS optimal solution strikes a favorable balance between all three objectives, reducing energy consumption to 22.32(kWh/m²), discomfort time to 3849.7(h), and incremental construction cost to 73,519.6(CNY). This solution is ideal for contexts where a comprehensive consideration of all objectives is necessary and offers the most cost-effective option. On the other hand, the energy-optimal solution excels in energy control, with an EC value of 15.19(kWh/m²), and features a lower incremental cost of construction (18,882.0). However, it results in a longer discomfort time (4,639.8), making it suitable for projects with budget constraints that prioritize high energy efficiency, but can tolerate longer periods of discomfort.

Fig. 15

Comparative analysis of energy consumption.

Figure 15 shows the fluctuations in the Base Model’s external infiltration heat loss, lighting energy consumption, equipment energy consumption, and zone sensible heating/cooling loads. Notably, during extreme weather conditions, energy consumption fluctuates significantly, indicating a lack of precise energy efficiency control. In contrast, the Optimized Model exhibits a marked reduction in energy consumption fluctuations, especially in reducing summer cooling demand and stabilizing winter heating demand, showcasing the substantial improvement in energy efficiency brought about by the optimized design and the reduction in the impact of external environmental changes on energy consumption fluctuations. Figure 16 further highlights the variations in electricity consumption and lighting energy consumption between the Base and Optimized Models. It is evident that the Optimized Model significantly reduces energy consumption for lighting and air conditioning, especially during winter and summer, with much smaller fluctuations in energy consumption. This indicates that the optimized design not only reduces building energy consumption but also enhances the overall system energy efficiency.

Fig. 16

Comparative analysis of heat balance.

Further analysis reveals that the discomfort time-optimal solution significantly reduces discomfort time to 3570.9(h) by increasing the incremental cost to 106,923.3(CNY). This solution offers the best thermal comfort performance, making it ideal for scenarios with very high comfort requirements. On the other hand, the incremental cost-optimal solution achieves baseline energy efficiency and comfort optimization with the lowest construction cost (17,333.6). This scenario features an energy consumption of 17.11(kWh/m²) and an uncomfortable time of 4362.4(h), making it a cost-effective option for projects with budget constraints. Overall, the TOPSIS optimal solution provides a well-balanced compromise across all objectives, making it suitable for situations requiring multi-objective optimization. In contrast, the energy consumption-optimal and discomfort time-optimal solutions cater to energy-saving and comfort priorities, respectively. The incremental cost-optimal solution presents a practical optimization approach for cost-sensitive projects.

Conclusion and outlook

This study establishes a BES–ANN–MCDM framework tailored for the multi-objective performance optimization of rural residential buildings. By integrating DesignBuilder modeling, JePlus batch simulation, Morris sensitivity analysis, artificial neural network prediction, and TOPSIS-based multi-criteria decision-making, the framework forms a scalable and efficient optimization process. Compared with traditional integrated platforms, the proposed framework significantly reduces computational time while maintaining high prediction accuracy and enhancing overall optimization efficiency.

This study establishes a three-objective optimization framework that integrates BES, ANN, and entropy-weighted TOPSIS to address energy consumption, thermal comfort, and economic performance simultaneously, thereby enhancing both prediction accuracy and decision-making robustness. In contrast to most existing studies that rely on fixed parameter settings or small-scale case assumptions, this research incorporates the Morris sensitivity analysis method to identify key variables in high-dimensional design spaces, enabling more targeted and efficient optimization. Furthermore, the study focuses on real-world rural residences—particularly those in underrepresented regions—thus improving the practical applicability and generalizability of the proposed approach.

While this study demonstrates the effectiveness of the BES–ANN–MCDM framework, several limitations warrant further investigation. First, the proposed framework still has significant room for methodological improvement. Enhancing the surrogate modeling with more advanced machine learning algorithms (e.g., XGBoost, LightGBM), exploring alternative optimization strategies, and integrating other decision-making tools such as fuzzy AHP or grey relational analysis could further strengthen the robustness and adaptability of the framework.Second, the current parametric modeling strategy faces challenges in handling complex interactions among high-dimensional variables. As the number of objectives and design variables increases, the computational burden of batch simulations grows rapidly, potentially limiting scalability. Future research could explore more efficient sampling methods or dimensionality reduction techniques to address this issue. Additionally, knowledge graph-based optimization approaches, as demonstrated in recent studies, offer promising potential for improving optimization efficiency by systematically organizing and leveraging domain knowledge. Incorporating such methods into future frameworks could further enhance scalability and decision-making robustness, particularly in complex high-dimensional scenarios⁴⁷. Lastly, rural building studies still face several inherent limitations. Many existing studies—including this one—focus on representative cases within a limited region, often neglecting the diversity of rural residential types, construction practices, and socioeconomic conditions across broader contexts. Future work should aim to expand the applicability of optimization frameworks by incorporating a wider range of rural building types and regional contexts.