Multiobjective decision-making, also known as multiobjective optimization, indicates that the needs of numerous goals impact the system scheme’s selection and decision-making direction. The rapid economic development and the acceleration of urbanization have promoted the prosperity of the construction industry, but at the same time, the problem of building energy consumption is also becoming more and more serious. In order to overcome this issue, this study proposes an optimization method for the impact selection of green building energy consumption. The machine learning (ML) algorithm, random forest (RF) is used to discover and design criteria relating to the influence of green building energy consumption. Then, based on the data of building information modeling (BIM) dynamic simulation, the optimization of the RF model in multiobjective decision-making in green building energy consumption is presented. Furthermore, a comparative study of the proposed system with the existing systems and deep learning (DL) models is also conducted. Performance of the proposed system is measured in terms of accuracy and ROC curve. The proposed system achieved a training accuracy of 95% and a testing accuracy of 83% which is superior to the earlier approaches. The experimental results show that the RF algorithm can effectively determine the relationship between the influencing factors of green building energy consumption. This approach also enables policymakers to better understand the complex relationships between green building energy consumption, which may subsequently improve the acceptability of decision-making.

1. Introduction

Energy is among the most significant resources in the modern era, and it is at the heart of economic and social activities in developed and industrialized nations. Because of industrial progress and population expansion, worldwide energy demand has risen dramatically in recent years. Buildings consume a substantial quantity of energy, resulting in major harmful effects on the environment. Regarding the factors that influence green building energy consumption, in a scenario when there are two or more goals, regardless of the scope of the goals or by what standard they are measured, the green building energy consumption is likely to be increased, whenever there is a conflict of interests between the goals. In addition, the approach of multiobjective assessment and decision-making may be used to evaluate and decide on competition or the inconsistency of its dimensions. Moreover, it is capable of simultaneously addressing multiple conflicting objectives across multiple dimensions, which means that it has a larger consideration space and more considerations than the single-objective decision-making method [14], and it can maximize the utility of decisions made in this manner. It is also possible to take into consideration the interests of all parties to a certain extent. However, when we compare it to single-objective decision-making, the most significant challenge that we observe in multiobjective decision-making is that it is difficult to bring all objectives to an optimal state, different results are attained, and the most prominent one is selected for the final decision.

Green building energy consumption planning is fundamentally a standard multiobjective decision-making problem with numerous objectives. The exchange of information between decision-makers and analysts (either system analysts or decision support systems) must be frequent for decision-makers to take the initiative throughout the entire decision-making process, from defining the model’s goals and constraints to analyzing and comprehending the results. This is a procedure that requires interaction. Interactive multiobjective decision-making strategies include the interactive replacement ratio based on the notion of noninferior replacement ratio, the interactive relaxation method (IRM), the weight-space interaction method, and the interactive stepwise equilibrium method (ISTM). The gradual tolerance method of adjusting the target value (satisfaction, superiority degree) or the percentage of the target value and the weight-space interaction approach is an example of interactive multiobjective decision-making employing the idea of a noninferior replacement ratio, which allows decision-makers to get important preference information. The ISWT technique, the IRM method, and the weight-space interaction method all need a significant degree of interaction. Each interaction step in the ISTM approach necessitates the solution of a more sophisticated auxiliary plan. The dimension of the decision variables in this auxiliary planning is greater than the dimension of the decision variables in the original planning, and it is required to provide a large capacity of satisfactory target value at the same time, resulting in an unnecessary burden on both analysts and decision-makers. This approach also yields a weak noninferior solution, and its noninferiority cannot be guaranteed because of the nature of the solution. While the advantages of easy and transparent interactive processes offered by various step-by-step tolerance techniques and methods for altering reference points outweigh the disadvantages of not being able to use the information provided by noninferior replacement rates, there are certain exceptions. The solution obtained using the progressive tolerance technique is a weak noninferior solution, and its noninferiority cannot be guaranteed. This is due to the fact that the tolerance goal function is turned into constraints and is thus not included in the auxiliary scalar program that is built. Furthermore, when the objective function vector has many dimensions, the stepwise permissive technique is not always successful as it is difficult to define the permitted objective function and its broad capacity, in the case when the objective function vector has few dimensions. Determining the amount of adjustment required by the method of altering the reference point is equally difficult to establish.

Influencing factors in algorithms are used to automatically solve sequential decision problems, either by planning a given Markov Decision Process (MDP) model (for example, through dynamic programming methods) [5] or by interacting with an unknown MDP that has been identified for green building applications. Learning (for example, using temporal difference approaches) [6] is a significant problem for AI researchers and practitioners. Based on these issues, the desirability or inadvisability of behaviors, as well as their consequences, are often codified into a single scalar reward function in different studies. It is customary for an autonomous agent to interact with an MDP with the purpose of maximizing his or her predicted total sum of rewards throughout the course of the relationship. There are numerous jobs in which scalar reward functions are the most natural, such as trading in the financial markets, where financial trading agents can be paid depending on the amount of money they have made or lost over the course of the previous period. However, many tasks are more naturally described by multiple, potentially conflicting goals. For example, traffic control systems should minimize latency while simultaneously increasing throughput. Autonomous vehicles should minimize travel time while simultaneously minimizing fuel costs. There has been a great deal of research on multiobjective issues across a wide range of decision-making domains [710]. Multiobjective decision-making in continuous situations is the subject of an expanding and albeit fragmented body of literature.

A large number of academics, both domestic and international, concentrate their emphasis on only one or two influencing variables while researching the influencing factors of building energy use as discussed in the related work section. Therefore, in order to maximize efficiency, this research looks into the influence of a variety of factors on the energy consumption of green buildings at the same time, rather than separately. A large number of related studies of literature were reviewed in order to identify the influencing aspects that needed to be researched. The link between the influencing elements was then investigated by utilizing the BIM dynamic simulation data, in order to develop an integrated model, which is then presented in the shape of the proposed system.

The rest of the paper is organized as follows: Section 2 is about the related work, which highlights the work done in the literature on green building energy consumption. Section 3 describes the research design along with different influencing parameters. It further highlights differential privacy and its properties, out-of-package estimation under differential privacy, and its applications and data source. The result and analysis of the paper are given in Section 4. Section 5 finally concludes the overall saying of the paper.

2.1. Multiobjective Decision-Making

In decision-making, multiobjective decision-making refers to the process of identifying, selecting, and optimizing options for various goals. Since the reform and opening up, the theory and technique of multiobjective decision-making have made significant advances and developments, and it has now emerged as a significant topic in the fields of scientific engineering, management, and operations research, among other disciplines. It is possible to represent the decision-making process for the multiobjective decision-making issue of generic discrete variables in the following ways. First and foremost, define the main goal of this choice dilemma. Another extremely crucial phase that may need several iterations is the development and identification of all candidates. Last, identify the characteristics of the potential solutions. It is necessary to have a minimum of two of these characteristics, and each program’s character level should be represented in a suitable manner, which can be either digital or textual in nature. The decision recommendation should be carried out if the decision-maker is satisfied. Otherwise, the decision-maker should return to the first stage and proceed from there.

For discrete decision-making problems, there are a variety of strategies available in the literature, including the character limitation method and its improvement, the combined social welfare function method, the weighted average method, and the ELECTRE (Elimination and Choice Translating Algorithm) method, among others. When faced with a deterministic multiobjective decision-making problem, with a finite number of candidate solutions and discrete variables, the ELECTRE method, which can be translated as the elimination and selection transformation method or exclusion selection method, is used to solve the problem [11]. It was first presented by Benayoun and Roy, among others, and then refined and completed by Roy. After more than two decades of development, it has gained widespread recognition and acceptance. This approach has been widely used all around the world [12]. ELECTRE is a method of exclusion and selection that is fundamentally different from the other methods. This approach rejects certain noninferior choices initially so that decision-makers can make decisions on the fly, or arranges all of the candidates in a logical sequence, with the first candidate being the most sensible choice, depending on the situation. The ELECTRE technique II is used in this study, which can list all of the candidate programs in priority order, with the first program being the most reasonable choice, as the starting point. Without a doubt, this ranking reflects the preferences of those in charge of making decisions.

2.2. Green Building Approach

Carbon dioxide emissions from the atmosphere are resulting in a number of difficulties, including socioeconomic problems, global warming, and energy supply limitations. As the public’s knowledge of, and support for, green and sustainable development is growing, governments throughout the globe are paying more and more attention to green development and society [1315]. The construction business is one of the three fundamental industries with the highest energy use, along with agriculture and manufacturing. It is condemned for being the world’s greatest consumer of energy and natural resources, using more than 30% of global resources and generating greenhouse gases accounting for more than 40% of global emissions [16]. A variety of environmental concerns have resulted in increased resource consumption in the building sector, which has been more concerning to the general public, as well as the importance of sustainable and healthy development [1720]. Eventually, the construction sector began to embrace green building practices. There is no universally accepted definition of green construction. However, Global Green Building Council (GGBC) defines green buildings as structures that limit their negative impact on the climate and natural environment while also having a beneficial influence throughout their design, construction, and operation life cycles. When it comes to green buildings, they not only pay attention to the performance of energy-saving and emission reduction, such as increasing energy and water efficiency, improving indoor air quality, and reducing environmental pollution, but they are also concerned with the economic aspects of their construction as well as social sustainability. The green construction sector has played a significant part in this development. As a result, governments and experts have expressed interest and concern about its role in promoting green growth in the global economy. The use of innovative green construction technologies in the areas of energy conservation, land conservation, water conservation, and material conservation is extremely important for the long-term growth of urban ecology.

China is currently through a time of tremendous economic growth, and the rapid expansion of urban building has significantly aided the advancement of the construction sector in China. According to a report, China’s building energy consumption accounts for more than 27% of the country’s total energy consumption, and it is expected to rise by 1% every year going forward. Analysis of the elements influencing the energy consumption of green buildings is critical in order to address the energy crisis and ease the problem of building energy consumption, which is now a major concern in the United States. An enormous amount of theoretical analysis and empirical study on the influencing variables of green building energy consumption and associated concerns has been carried out in recent years by a large number of domestic and international experts. According to a scholar, new technologies for lowering building energy consumption may be broken down into three categories that are optimization of the building’s maintenance structure, complete exploitation of building energy, and intelligent control technology. According to the results of his research, different ventilation techniques have varying effects on the airtightness of urban structures. He also highlighted the relevance of airtightness in reducing the energy consumption of urban buildings. The impact of benchmark building parameters, such as the body shape coefficient, exterior walls, exterior windows, and other variables on energy consumption is investigated by a scholar using simulation values and software analysis to investigate the impact of these factors on building energy consumption. Some experts [21] advocated for reducing construction materials throughout a structure’s life cycle in order to reduce the amount of energy consumed by the structure. The direct and indirect impacts of high-albedo roofing on the energy consumption of residential buildings have been explored by some researchers in detail [22].

A wide range of stakeholders has various degrees of influence over the design and construction of environmentally friendly structures. As a result of the involvement of a varied range of stakeholders with a diverse range of backgrounds and aims in the green building development process, the actions of these stakeholders can have an impact on the achievement of green building objectives in the long run. In Albin’s opinion [23], green building is related to cooperation between businesses such as suppliers, design teams, and businesses that are engaged in sustainable development projects. According to Berardi [24], when it comes to the implementation of green buildings, the instability, and fragmentation of an ad-hoc assembly of diverse stakeholders is a barrier to new technology acceptance. The evaluation of environmental and economic benefits plays a crucial role in stakeholder decision-making in green construction projects [25], and it is an aspect that should not be overlooked. To close the energy performance gap in buildings, many scholars have proposed that an approach and methodological framework involving stakeholder engagement be developed, which would include in-depth case studies of the state-of-the-art green office building to close the energy performance gap in buildings [26].

The involvement of motorists in the development of green construction practices has also been shown to be quite beneficial. A review of the current development state of green construction in China has resulted in some academics making some recommendations, which include raising awareness among stakeholders, boosting technological research and communication, and developing standards and regulations [27]. It is critical to consider the costs and consequences of green buildings from a life-cycle perspective when analyzing their costs and consequences in order to aid in the promotion of green buildings [28]. Researchers have emphasized the importance of using well-developed green project management evaluation models to assess the costs and benefits associated with environmentally friendly building designs [29]. Furthermore, there is a need to increase knowledge and understanding of sustainable development among all stakeholders, including policymakers, owners, designers, builders, as well as members of the general public [30]. A significant amount of real-world research indicates that the level of support received from senior management has a direct impact on the adoption of green construction practices. Owing to China’s enormous population and high building density, as well as the amount of recoverable energy per square meter of building space, developing sustainable construction solutions that are suitable for a range of scenarios is crucial [31]. Incorporating environmentally friendly technologies into a building’s design and interacting with other building components are both necessary components of green construction [32]. The use of procedures such as soft landings, in which professionals are hired after the completion of building construction to ensure that the structure is actually functioning as intended by the designer, has been implemented in several nations. One advantage of this strategy is that it allows for the creation of a feedback loop for a project. Nevertheless, it also creates a number of challenges for contract administration [33, 34]. In order to better understand the elements that motivate people to embrace green building techniques, many research works have been carried out on the basis of literature reviews. While most studies are accomplished from a life cycle or stakeholder perspective, the great majority of them are done from an individual standpoint. As a result, this study investigates all of the influencing factors that are involved from the perspectives of the life cycle and stakeholder groups. This technique has the potential to improve both the comprehensiveness and the accuracy of factor collection.

3. Research Design

3.1. Random Forest Algorithm

Random forest (RF) is a supervised machine learning (ML) algorithm, which may be utilized for both classification and regression problems. It is based on ensemble learning, which is a method of integrating numerous classifiers to solve a complicated problem and increase the model’s performance. RF combines several decision trees on different subsets of a data set and averages the results to increase the data set’s predicted accuracy. Instead of depending on a single decision tree, it collects the forecasts from each tree and predicts the final output based on the majority votes of predictions. The more trees in the forest, the higher the accuracy and the lower the risk of overfitting. However, there is some trade-off between the number of trees and the given data set that is the RF accuracy increases with the number of trees but then decreases after a certain point. For instance, in our case, we take different numbers of trees, ranging from 300 to 900, and check the accuracy of each set of trees as shown in Figure 1. The accuracy increases until the number of trees reaches 700, then falls as the number of trees increases. More detail is given in Section 4, the results section.

3.2. Differential Privacy and Its Properties

Definition 1. (Differential Privacy). If an algorithm M fulfills ε-differential privacy, then it is considered to be secure.In the equation (1), Algorithm M can produce any value set, and is an arbitrary subset of all value sets that can be output by the algorithm. D and D′ are two data sets that differ by no more than one record, and ε is the budget for privacy protection.
The sequence composition and parallel composition qualities of differential privacy protection technology are two of the most essential combination properties found in differential privacy protection technology. In the privacy of the proof method as well as the allocation of the privacy budget, these two qualities are extremely vital to consider.

Property 1. (Sequence Combination): Suppose we have a database D and n random algorithms, and each of the algorithms is capable of achieving -differential privacy. Then the sequence combination of A on D is capable of achieving AA-differential privacy, where is the random algorithm.

Property 2. (Parallel compositionality): Let D be a private database, which is divided into n disjoint subset , and let A be any random algorithm that satisfies ε-differential privacy. Then the series of operations of algorithm A on satisfy ε-differential privacy.

Property 3. (Parallel compositionality): For simplicity, consider D to be a private database that has been partitioned into n distinct subsets, and A to be a randomly generated method that meets the requirements of differential privacy. In such a case, a set of operations of algorithm A meet the ε-differential privacy requirement.

3.3. Out-of-Package Estimation under Differential Privacy and Its Applications
3.3.1. Estimation of Out-of-Package in the Presence of Differential Privacy

When producing each decision tree (DT), samples must be picked randomly and with replacement, which means that around one-third of the data will not be retrieved, and these data are referred to as the decision tree’s out-of-package data. The RF method makes use of the characteristics of the data collection. Each of the DT’s functions, including its significance and classification ability, as well as the correlation calculation, is dependent on the data that came with the package. When using out-of-package data to produce judgments on a DT, the out-of-package must be approximated. This is the ratio of misclassified data to the total quantity of out-of-package data while making decisions on the DT. It has been demonstrated that the out-of-bag estimation is a fair and unbiased estimate of the generalization error of the ensemble classifier and that it may properly assess the classification ability of the decision tree when used in conjunction with other methods. This study, on the contrary, introduces the bag under differential privacy in order to safeguard the privacy of the collected out-of-bag data. It is defined as outer estimation, which is the process of including noise when computing out-of-bag estimates to meet the differential privacy criterion.

Definition 2. (Out-of-package estimation under differential privacy): Out-of-package estimate B for a DT in a RF is given by equation.where O is the size of the out-of-package data, M is the size of misclassified data, ε is the privacy budget of the DT, and N(ε) is the differential privacy noise function.
However, it can be seen that the added differential privacy noise N(ε) disturbs the real amount of misclassified data. However, according to the definition of differential privacy, only a small amount of noise can be added to achieve the goal of protecting privacy without compromising data availability. Therefore, the Laplace noise function is used in this paper in the case of the aforementioned law violation. As a result, the lower the value of B, the greater the classification accuracy of the DT will be.

3.3.2. Filtering Using a Decision Tree

In light of the fact that out-of-package estimation can be used to quickly and accurately estimate the classification ability of a DT, this paper calculates the weight of the DT in accordance with Paul’s method and then uses this weight to represent the classification ability of the DT before performing the DT screening. As a result, the weight of the tth DTis defined by equation (3).

The out-of-package estimation under the differential privacy of the DT is represented by the symbol in the formula. If the DT has more weight, it is easy to observe that the less out-of-package data are incorrectly categorized when the DT is classified, which also indicates that the classification of the DT is reduced.

(1) Filter by Feature. The out-of-package estimate is used in this research study. In addition to the DT screening, to efficiently evaluate the relevance of features in the data set, this study also refers to the approach presented by Paul to screen the features in the data set throughout the process of DT building.

(2) The Weight of the Feature. Before beginning with feature screening, we must first determine the feature’s assessment criterion, which is referred to as the feature weight in this study. Following the approach presented in the literature, first determine the weight assigned to the feature on the node, which is more clearly stated as follows.

Definition 3. (Weight of feature on the node): On a nonleaf node in the DT, the weight of feature j on this node is defined in formula (4).The Gini index is denoted by the term in formula. In accordance with its definition, it can be deduced that the smaller its value, the more effective this feature’s classification influences on the classification process. As a result, when the feature weight is higher, it indicates that the Gini index is lower and that the classification impact is better. When there are several nodes on a DT, take the average of the feature weights on the nodes and use it as the weight of the feature in the DT to more correctly quantify the feature weight. This is described as follows.

Definition 4. (Weight of feature in DT): The weight assigned to feature j in the DT is given as follows:where N is the number of nonleaf nodes in the decision tree and is the feature weight assigned to feature j on the nonleaf nodes in the DT. Nevertheless, because the classification performance of each DT varies, the weighted average of the feature weights in the DT is used to determine the relevance of a particular feature in the RF method, which is described as follows.

Definition 5. (Weight of feature in RF algorithm): In the RF algorithm, the weight of a feature for the DT’s is equal to:where denotes the weight assigned to feature j in the DT, and denotes the overall weight assigned to the DT.
(3) The Division of Features.To determine how important a feature is, the weight assigned to the feature in the RF algorithm (hereinafter referred to as the feature weight) is used in this paper. After that, the features are divided according to its importance, the nonimportant features are identified, and the DT is improved. Furthermore, throughout the division process, it is required to identify outlier characteristics, which are outlined as follows.

Definition 6. (Outlier features): For the features subset R with weight, the outlier features satisfies the following condition.where the feature weight of feature j, μ denotes the average of the weights in R, and σ denotes the standard deviation of the feature weights.
As can be seen from the definition, the weight of outlier features is small, and it is far from the average value of feature weights. As a result, in this paper, outlier features are defined as features that are less important than other features, as opposed to features that are more important than others. The characteristics can be classified according to their weight values.For example, in formulas (8) and (9), Г is the feature set with the biggest f before the weight, where f is the total number of DT features, R is the complement of Г, and A is the outlier feature set in the complement of Г. Then, in R, identify the features that are outliers and are denoted as unimportant features Г′. Once these characteristics have been classified as significant or unimportant, the classification cannot be modified.
During the RF generation process, it is necessary to update both key features Г and nonimportant features Г′ to reflect recent changes. In order to maintain the existing features in Г and Г′, the features for the remaining features R are moved into the important feature set Г. This allows the existing features in Г and Г′ to remain unchanged.where is the feature weight of feature j, and is the minimal weight in the critical feature Г. Next, all of the nonessential R features that stand out are removed and placed in the insignificant feature set Г′.
The creation of a DT in the context of movie screening. Following the division of the features, nonimportant features may be used to filter the features in the data set during the process of the DT building, which will reduce the number of features in the data set. If a node is the most significant in the DT generating process, the node is changed into a leaf node if the optimum decision feature turns out to be an insignificant feature. The whole privacy budget is used to add noise to the count value on the leaf node, which decreases the overall amount of noise and increases the classification accuracy.

3.4. Data Sources

We generated impact data for green building energy usage in different cities using BIM simulation software, which is used to create the data. The COM standard is used to develop the Navisworks 3D simulation platform, which allows for the 3D display of architectural landscape models and simulation data. The CAD entity modeling technology is used to model the three-dimensional simulation system for the large-scale urban architectural landscape, and the three-dimensional architectural simulation model is rendered in the Navisworks visualization software, resulting in the creation of a three-dimensional simulation image of the architectural landscape. The next section describes the concrete implementation scheme for the three-dimensional simulation system of the architectural landscape.

We built a green building energy consumption influencing factor and integrated model based on BIM dynamic simulation and multiobjective decision-making utilizing the RF approach employing 80% of the simulation data as the training set and 20% of the data as the test set. We used the training set for the model training and the test set is used for the model validation.

4. Results and Analysis

The main goal of this study is to analyze the effect of the RF model on multiobjective decision-making in green building energy consumption using BIM dynamic simulation data. In the first step, we choose the number of DT’s to include in the RF, as shown in Figure 2, where the horizontal axis represents the number of DT’s and the vertical axis represents the accuracy of the training set. Table 1 shows the trade-off between the number of DT’s and accuracy.

The influence of the number of DT’s on the accuracy of the model on the test set is determined, as shown in Figure 2. Using this illustration, we can observe that as the number of DT used in RF in our model grows, the accuracy of the model grows as well. On the whole, it climbs for a while before decreasing again. The highest accuracy is achieved when the number of decision trees is 700 or more in number.

Furthermore, we also utilized the other two most commonly used deep learning (DL) models such as convolutional neural network (CNN) and extreme learning machine (ELM) as well. As demonstrated in Figure 3, the accuracy of the RF is the highest, followed by the accuracy of the CNN, and the accuracy of the ELM is the lowest of the three utilized models. Our model has attained promising accuracy as compared to the other used models as shown in Table 2. The proposed green building energy consumption affecting elements as well as our integrated model based on BIM dynamic simulation and multiobjective decision-making can accurately anticipate the outcomes.

In this research study, we proposed an ensemble model (RF) that achieves 95% accuracy on the training data set and 83% accuracy on the test data set, as illustrated in Figure 4. The error rate of the proposed model is also lesser as compared to the other models as shown in Table 2.

5. Conclusion

The prediction of building energy consumption aids in the efficient operation of buildings, identification and diagnosis of faults, and the control of demand-side resources. To increase the prediction performance of building energy consumption, this work suggests a new ensemble model that incorporates an RF algorithm. The model exhibits excellent generalization ability as well as rapid parameter learning speed and is capable of accurately simulating complex nonlinear relationships. It has the potential to outperform standard approaches in terms of prediction accuracy. In addition, to examine the data from the BIM-based simulation, the proposed model is employed. The experimental results demonstrate that the predictive performance of this model is much better than that of the other base models, which indeed is a good sign. The proposed model attained a training accuracy of 95% and a testing accuracy of 83%, which is superior to the other used models, that is, CNN (82.5%) and ELM (81.6%). Despite the fact that the proposed RF ensemble model has improved in terms of accuracy in predicting building energy consumption, this model still needs further refinement. In this study, a relatively simple correlation analysis method is employed. In the future, researchers may employ a variety of feature selection strategies. It is recommended that the researchers should take into account more complex procedures. Further study can enhance the quantity of training data available and employ a huge amount of energy consumption data to train the network, which will improve the model’s generalization ability and prediction accuracy.

Data Availability

The data used to support the findings of this study are included within the article.

Conflicts of Interest

The author declares that there are no conflicts of interest.