Intelligent building construction cost optimization and prediction by integrating BIM and elman neural network

Abstract

This study aims to address the challenges of capturing design changes, supply chain fluctuations, and labor cost variations to improve the accuracy and real-time nature of intelligent building construction cost predictions. It seeks to accurately forecast and optimize project costs. The study innovatively constructs an intelligent building construction cost prediction model based on Building Information Modeling (BIM) and Elman neural networks (ENNs), denoted as the BIM-ENN model. The BIM-ENN model first introduces BIM technology to digitize and visualize information related to building structures, electromechanical systems, and pipelines. The digitized data obtained through BIM technology is then used as input data for the ENN, which optimizes the neural network parameters to predict and optimize intelligent building construction costs. Finally, the BIM-ENN model is experimentally evaluated. The results demonstrate that the prediction value of the construction cost of the intelligent building by this model closely matches the original information price, with a prediction accuracy of 95.83 %. Compared with the single ENN, the root mean squared error of the BIM-ENN model is less than 75, and the determination coefficient is above 0.95. This indicates that this model can explain more than 95 % of the construction cost prediction results, making it a feasible solution for actual prediction problems and achieving satisfactory results. The intelligent building construction cost prediction model reported here exhibits high accuracy and reliability. It can successfully forecast construction costs, providing robust decision support for the digitalization and intelligent development of construction enterprises. The practical significance of this study lies in providing the construction industry with an accurate cost management tool that helps enterprises optimize budget control and resource allocation, enhancing risk assessment and management capabilities. Moreover, the potential impact of the BIM-ENN model is its ability to elevate prediction standards within the construction industry, promote technological integration and innovation, enhance enterprise competitiveness, and drive the industry's transition towards digitalization and intelligent sustainable development.

1. Introduction

The rapid development of the global economy and the increasing competition in the construction market have made cost management a crucial component of the construction industry These scientific management practices are essential for the successful implementation of projects, budget control, and profit management [1,2]. Accurate cost budgeting is critical for ensuring projects are completed on time and within budget, as well as for the economic viability of enterprises [3]. However, due to the complexity and uncertainty inherent in construction projects—such as design changes, supply chain fluctuations, and variations in labor costs—traditional methods often rely on historical data and empirical judgment. This reliance may fail to promptly reflect the latest market changes and project progress. As a result, traditional cost forecasting methods struggle to meet the industry's growing demands for increased forecast accuracy and efficiency. Although deep learning has shown outstanding predictive performance in various fields, its application in intelligent building cost prediction is still in its infancy and has not received sufficient attention. Moreover, while some studies have utilized artificial neural networks (ANNs) for cost prediction, these models frequently struggle to adequately capture the time-series characteristics of construction costs, resulting in inadequate prediction accuracy. Therefore, there exists a pressing need to integrate advanced information technology and artificial intelligence methods to enhance the accuracy and real-time performance of construction cost predictions. This integration is essential for addressing these challenges and meeting the evolving demands of industry development.

The advent of Building Information Modeling (BIM) technology presents new opportunities for improving construction cost prediction. BIM technology is widely recognized for its ability to digitally manage essential information in construction projects. It encompasses geometric structures, materials, time, and cost data, thereby offering a robust data source for accurate cost forecasting [4,5]. Conversely, the Elman neural network (ENN) within ANNs is distinguished by its robust data processing and pattern recognition capabilities, showing significant promise in addressing complex and nonlinear problems. Equipped with internal time delay modules, the ENN can effectively capture the time-series characteristics of data, crucial for predicting time-dependent fluctuations in construction costs [6]. The ENN is distinguished by its unique network model incorporating travel delay and a local feedback function. Its topology includes a specialized middle layer known as the correlation layer, while its internal time delay module provides distinct advantages for addressing real-time price fluctuations over specific periods [7,8]. As a result, it demonstrates significant strengths in forecasting construction factor prices.

In conclusion, the integration of data visualization in BIM technology with the learning capabilities of neural networks enables more accurate construction cost predictions, thereby offering effective support for decision-making and project control in construction. The aim of this study is to develop and validate an intelligent construction cost prediction model integrating BIM and ENN, denoted as the BIM-ENN model. The BIM-ENN model seeks to improve prediction accuracy and real-time capability, addressing complex factors like design changes, supply chain fluctuations, and labor cost variations. By digitally processing and visually managing construction project data using BIM technology, and harnessing ENN's capacity to analyze time-series data, this study aims to optimize the cost prediction process comprehensively. The study contributes by proposing an innovative cost prediction method, validated through experiments for its high accuracy and reliability. Moreover, it introduces a new perspective and tool for the digital transformation and intelligent advancement of the construction industry, thereby enhancing efficiency and effectiveness in cost management and project decision-making for construction enterprises.

This study addresses several shortcomings evident in existing research on multiple fronts. Firstly, it achieves the comprehensive digital integration of critical information across the entire lifecycle of construction projects through the application of BIM, a facet often overlooked in prior studies. Secondly, the incorporation of the ENN, specifically leveraging its internal time-delay module, provides a robust tool for capturing the time-dependent variations in construction costs, a dimension insufficiently explored in current literature. Additionally, the introduction of the Particle Swarm Optimization (PSO) algorithm serves to optimize the parameters of ENN while effectively mitigating issues such as local optima and overfitting commonly encountered in traditional neural networks. This approach significantly enhances the predictive accuracy and adaptability of the intelligent building construction cost prediction model.

2. Literature review

2.1. Analysis of the application of BIM technology in the construction industry

In recent years, the application of BIM technology in the construction industry has garnered significant attention and practical application. BIM serves as a digital modeling method that integrates geometric, material, time, and cost data of construction projects. Moreover, it provides comprehensive building information, enhancing the understanding and coordination among project participants. Menassa (2021) [9] summarized the evolution of intelligent building models in construction engineering and facility management, highlighting the progression from BIM to digital twin models. Chen et al. (2022) [10] assessed the implementation benefits and challenges of BIM in enhancing the energy efficiency of intelligent buildings. Their research highlighted that integrating BIM in intelligent building projects can potentially facilitate sustainable development, energy conservation, and emissions reduction. It also provides valuable insights for decision-making in intelligent building projects. Wang et al. (2022) [11] introduced a query classification method based on transfer learning for oral dialogue in intelligent building information systems. Experimental results demonstrated the method's effectiveness in facilitating dialogue within intelligent building systems, offering a practical solution for oral interactions. In another study, Li et al. (2023) [12] utilized BIM data combined with intelligent modeling to accurately model the edge components of precast shear wall structures. Their approach significantly enhanced the precision of BIM, leading to improved coordination between architectural design and construction processes.

Through the analysis of the literature reviewed above, it becomes evident that BIM technology has demonstrated significant effectiveness in construction project management and optimization. However, there remains a notable gap in research integrating BIM with machine learning or neural networks. Specifically, studies utilizing neural networks for processing BIM data to predict construction costs have been insufficiently explored. Some research has explored the integration of BIM data with intelligent modeling, such as Li et al. (2023) in their study on modeling prefabricated shear wall edge components. Nevertheless, these efforts predominantly focus on enhancing model accuracy and facilitating design coordination rather than on cost prediction.

2.2. Review and analysis of neural network applications

The construction cost of intelligent buildings stands as a pivotal index in enhancing the competitiveness of construction enterprises. Scholars have extensively researched the predictive capabilities of deep learning across various domains. Lv et al. (2022) [13] proposed an intelligent prediction and evaluation approach based on deep learning, specifically applied in the realm of intelligent medical care to assess and forecast interactive multimedia data. Their trials underscored the exceptional accuracy and efficacy of this approach in intelligent medical care settings. Similarly, Agga et al. (2022) [14] developed an effective hybrid deep learning architecture for short-term photovoltaic power forecasting. Combining convolutional neural network (CNN) and long short-term memory network (LSTM), their experimental results highlighted the CNN-LSTM model's robust performance and accuracy in predicting short-term photovoltaic power. This model significantly supports the management and operational efficiency of solar power production systems. Alshboul et al. (2022) [15] employed the extreme gradient boosting machine learning method to develop a model for predicting the costs of green buildings. Their study demonstrated that this method achieved a prediction accuracy exceeding 90 % in forecasting green building costs, offering substantial decision support for sustainable building initiatives. In another study focused on improving wind forecasting, Pujari et al. (2023) [16] proposed a green deep learning approach based on evolutionary neural architecture search. Their experimental results highlighted the efficacy of deep learning in managing and optimizing wind energy planning. The approach significantly enhanced performance and accuracy in wind forecasting, providing valuable insights for the effective utilization of wind energy resources.

Simultaneously, scholars have conducted relevant research on the application of machine learning methods in the construction field. For instance, Almasabha et al. (2023) [17] studied the structural performance of reinforced concrete pipes buried under deep fill. They found that machine learning models can effectively predict the performance of pipes under different soil conditions, thereby enhancing the reliability of structural design. Al-Shboul et al. (2023) [18] explored the effectiveness of machine learning models in predicting soil radon emission rates. Their study demonstrated that these models can accurately forecast radon emission rates under various environmental conditions, thereby assisting in environmental monitoring and risk assessment. Shehadeh et al. (2024) [19] proposed an automated management algorithm for detecting slope displacement in construction, utilizing machine learning technology to prevent disasters effectively and improve construction safety and management efficiency. Furthermore, Alshboul et al. (2024) [20] compared the applications of LightGBM, XGBoost, and GEP models in predicting shear strength for steel fiber-reinforced concrete management. Their findings suggest that these machine learning models can significantly enhance the accuracy of shear strength prediction and management practices.

The studies highlight the substantial potential of neural networks in predicting construction costs, particularly when integrated with data from BIM technology, to enhance prediction accuracy and efficiency. However, while these findings provide valuable insights into intelligent building cost prediction, research on integrating BIM data with optimized neural network models for cost prediction remains limited. Specifically, there is a scarcity of studies utilizing neural network models optimized for time-series forecasting, such as ENNs, in this context.

2.3. Review and analysis of studies on the integration of BIM and machine learning

Several scholars have explored the integration of BIM with machine learning or neural networks. For instance, Hong et al. (2020) [21] utilized neural network methods to predict the net costs associated with BIM adoption. Their research demonstrates the effectiveness of this approach in forecasting BIM-related costs, thereby enhancing decision-making accuracy and efficiency. Alshboul et al. (2023) [22] conducted a comparative analysis of various machine learning algorithms to predict the shear strength of deep beams reinforced with steel fibers, highlighting significant improvements in prediction accuracy over traditional methods. Similarly, Almasabha et al. (2023) [23] developed a machine learning-based model for predicting the shear strength of synthetic fiber-reinforced concrete beams without stirrups, illustrating the applicability of such models in structural engineering contexts. Halalsheh et al. (2022) [24] employed the boosted decision tree algorithm to forecast the breakthrough curve of modified zeolite in water treatment applications, demonstrating high accuracy and practicality in environmental engineering. Furthermore, Abbasnejad et al. (2024) [25] utilized mathematical modeling and ANNs to evaluate the effectiveness of BIM implementation. Their findings underscore the method's efficacy in assessing the benefits of BIM adoption, advocating for its widespread use in the construction industry.

These studies illustrate the potential of machine learning technology to enhance the efficiency and prediction accuracy of BIM applications, while also offering robust theoretical and practical support for future applications of machine learning in the construction and related engineering fields. However, despite the application of integrating BIM with machine learning or neural networks to specific issues such as cost prediction or structural analysis, these studies often focus on singular domains or specific types of construction projects, lacking generality and diversity. Moreover, current research remains insufficient in achieving real-time prediction and feedback systems that integrate BIM with machine learning or neural networks.

2.4. Summary

Through analysis of the above studies, this study adheres to principles of scientific rigor and practicality in model selection. Firstly, given the complexity inherent in predicting construction costs, the ENN is chosen for its capability to handle nonlinear and dynamic system characteristics. ENN effectively captures the time-series properties of construction costs through its internal time delay module. Secondly, to enhance ENN's prediction accuracy and generalization capability, the PSO algorithm is introduced to optimize ENN parameters. PSO, known for its effectiveness as a global optimization method, aids ENN in avoiding local optima, thereby improving model search capabilities and prediction stability. Moreover, existing literature supports the combined application of ENN and PSO in intelligent building cost prediction, highlighting their potential to enhance prediction accuracy. Considering advancements in technology, data processing capabilities, prediction accuracy, and practical application value, this selected model is poised to provide robust technical support for intelligent construction cost prediction.

3. Construction cost prediction method based on BIM and ENN in the field of architecture

3.1. Digital analysis of intelligent building based on BIM technology

BIM technology serves as a digital tool in engineering design, construction, and management. It integrates building data and information models, facilitating the sharing and transfer of information throughout the entire lifecycle of project design, operation, and maintenance. Engineers and technicians leverage BIM to enhance production efficiency, reduce costs, and expedite construction timelines by effectively interpreting and responding to diverse building information [26,27]. Fig. 1 illustrates the advantages of BIM technology.

Fig. 1.

Illustrations of advantages of BIM technology.

In Fig. 1, the advantages of BIM technology primarily encompass visualization, parameterization, simulation, collaboration, and continuity [[28], [29], [30]]. These aspects highlight BIM's capability to digitally and visually represent intelligent buildings, as illustrated in Fig. 2.

Fig. 2.

Digital Representation of intelligent building using BIM technology.

In Fig. 2, the process of using BIM to digitize and visualize building information is detailed as follows: First, comprehensive building information is collected, including architectural drawings, engineering specifications, technical specifications, and related documents. These documents provide essential details about the structure, layout, materials, and equipment used in the building. The extraction and processing of BIM data are shown in Table 1. Using this information as a basis, professionals can then construct a three-dimensional model of the building through BIM software or modeling tools. During this modeling process, every component of the building, such as the structural framework, exterior features, internal mechanical and electrical equipment, and piping systems, is meticulously depicted. The goal of this step is to ensure accuracy in geometric shape, dimensions, and location, thereby accurately reflecting the physical characteristics of the building [31]. Next is the step of associating attribute information with the model, which involves assigning specific attribute data to each element in the model, such as material types, quantities, costs, and potential safety risks encountered during construction. This attribute information forms the critical data foundation for subsequent cost prediction and analysis, ensuring that the model is not only visually accurate but also highly practical at the data level [32]. Finally, through BIM technology, all building information and attribute data are integrated and visually displayed.

Table 1.

BIM data extraction and processing.

Step	Specific implementation method
Data fetch	Extract geometric information (e.g., shape and size), attribute information (e.g., material strength, density, cost), component relationships (e.g., connections between beams and columns), and spatial topology (e.g., building layouts) from the BIM model.
Data cleaning and preprocessing	Remove duplicate data, fill in missing values, handle anomalies, and normalize or standardize data to ensure consistency in scale and magnitude.
Data feature selection	Select features with high correlation and predictive ability through statistical analysis, feature importance evaluation, or expert domain knowledge.

In Table 1, data cleaning and preprocessing are crucial steps to ensure the accuracy and quality of the dataset, directly affecting the performance of data analysis and machine learning models. Firstly, identifying and removing duplicate data avoids redundancy and ensures data uniqueness. Next, imputing missing values is vital; common methods include using statistical measures (such as mean, median, or mode) or predictive models based on other data to estimate the missing values. Additionally, handling outliers is essential, involving the identification of anomalies in the data and deciding on appropriate strategies, which may include deletion, replacement, or transformation. Data normalization or standardization ensures that different features have the same importance during model training by adjusting the numerical ranges to prevent certain features from dominating due to their larger scale. Concurrently, feature engineering enhances model performance by creating new features, transforming existing ones, or selecting the most informative subset of features. Data type conversion ensures that all data in the dataset are in the correct format, while splitting the dataset into training and testing sets allows for the evaluation of the model's generalization capability on different data subsets. Finally, a consistency check of the dataset ensures that all records conform to business logic and expectations. This series of data cleaning and preprocessing steps lays a solid foundation for building an accurate and reliable intelligent construction cost prediction model.

3.2. ENN algorithm and its optimization

The fundamental structure of the ENN comprises an input layer, a hidden layer, and an output layer, with the recurrent connections in the hidden layer being its core feature. These recurrent connections enable the network to process current input data while recalling hidden states from previous time steps, thereby capturing and utilizing the temporal dependencies within the data [33]. This short-term memory function is crucial for understanding and predicting complex systems that change over time, such as construction costs.

The advantage of selecting the ENN algorithm lies in its ability to model dynamic changes in sequential data, which is particularly important for intelligent construction cost prediction. Construction costs are influenced by various factors, such as market fluctuations and changes in material prices, which often vary over time. The ENN, through its internal time-delay module, can effectively capture these time-series characteristics, thereby providing more accurate and real-time predictions. Furthermore, this capability of the ENN makes it highly effective in handling nonlinear and dynamic systems, which is particularly valuable in the construction field, given that construction projects often entail complex nonlinear relationships and dynamic changes. The topological structure and algorithm flow of ENNs are shown in Fig. 3.

Fig. 3.

Topological structure and algorithm flow of ENN (a. Topological structure; b. Algorithm flow).

In Fig. 3a, the ENN comprises an input layer, a hidden layer, and an output layer. The hidden layer's recurrent connections between neurons allow the network to retain and utilize its previous state and information. The ENN method is adept at updating and propagating data over time. At each time step, the network receives inputs for the current data and the hidden state from the previous time step, updates the hidden layer state using an activation function [34,35]. Subsequently, this updated hidden state is passed forward to the next time step and serves as input for the output layer. This process enables the network to leverage temporal sequence information from the input sequence to predict future outcomes. The specific formulation of the ENN is presented in Equation (1).

$$y\left(k\right)=g\left(w_{3}x\left(k\right)\right)$$ (1)

In Equation (1), y represents the m-dimensional output vector, and $w_3$ is the weight matrix connecting the hidden layer to the output layer. The function g(∗) is the activation function for the output neurons, typically a linear weighted combination of the hidden layer's output, often implemented using the Purelin function. The vector x(k) denotes the n-dimensional hidden layer node vector at time k, which can be expressed as in Equation (2).

$$x\left(k\right)=f\left(w_1{x}_c\left(k\right)+w_2\left(u\left(k-1\right)\right)\right)$$ (2)

Here, u is the r-dimensional input vector, $w_2$ is the weight matrix connecting the input layer to the hidden layer, and $w_1$ is the weight matrix connecting the feedback loop to the hidden layer. The function f(∗)is the activation function for the hidden layer neurons, typically the sigmoid (S) function. The n-dimensional feedback state vector xc(k) is given by Equation (3).

$$x_c\left(k\right)=x\left(k-1\right)$$ (3)

The training error function e of the ENN can be expressed as Equation (4):

$$e\approx \sum _{k=1}^m{\left(y_k-{\bar{y}}_k\right)}^2$$ (4)

In Equation (4), $y_k$ refers to the ideal output data of the ENN, ${\bar{y}}_k$ refers to the actual output of the neural network, and m refers to the number of samples trained by the neural network.

The ENN consists of an input layer, a hidden layer, and an output layer, with its distinctive feature being the recurrent connections in the hidden layer. These connections enable the network to process current input data while retaining and utilizing hidden states from previous time steps, thus capturing temporal dependencies within the data. This short-term memory function is essential for comprehending and predicting complex systems that evolve over time, such as construction costs. The ENN algorithm offers several advantages, particularly its capability to model dynamic changes in sequential data, which is crucial for intelligent construction cost prediction. Construction costs are influenced by diverse factors like market fluctuations and fluctuations in material prices, which frequently vary over time. Through its internal time-delay module, the ENN effectively captures these time-series characteristics, thereby enabling more precise and real-time predictions. Moreover, this feature makes the ENN well-suited for managing nonlinear and dynamic systems, which is highly beneficial in the construction sector where projects often involve intricate nonlinear relationships and dynamic changes.

Once the ENN achieves satisfactory performance during the training phase, it progresses to the deployment phase, where its application involves predicting new or unknown data. During deployment, input data undergoes identical preprocessing steps as in the training phase to maintain consistency in data format and scale. Following preprocessing, the processed data is inputted into the trained ENN model for forward propagation, promptly generating predictive outputs. These outputs are directly applicable for decision support or further analysis. Crucially, in the deployment phase, evaluating the stability and generalization capability of the ENN model becomes pivotal for assessing its practical application value.

While the ENN possesses a short-term memory function enabling it to simulate dynamic system changes and predict construction factor prices in real-time, practical applications reveal several challenges. These include extensive iterations, slow learning speeds, susceptibility to overfitting, and a tendency to converge towards local extremities. Therefore, this study employs the PSO algorithm, an efficient evolutionary computing technique, to adjust the training hyperparameters of the ENN. The PSO algorithm simulates the foraging behavior of birds to optimize key ENN hyperparameters, such as the learning rate and number of hidden layer nodes, thereby enhancing the network's training efficiency and predictive performance. Additionally, the application of the PSO algorithm helps mitigate potential "jitter" during the training process and guides the network away from local optima, aiming to find global or near-global optimal solutions [36,37]. Fig. 4 illustrates the integration flow of PSO with the ENN algorithm for optimizing neural network training hyperparameters.

Fig. 4.

Flow chart of optimization algorithm based on PSO and ENN.

As depicted in Fig. 4, the PSO process is an optimization algorithm that simulates the foraging behavior of birds. It initializes a group of random particles, each representing a potential solution, to explore and find the optimal solution within the defined search space. The process of enhancing the ENN using PSO unfolds as follows.

• 1)Initialization: PSO starts by randomly initializing the positions and velocities of a group of particles. Each particle corresponds to a configuration of ENN parameters, including weights and biases.
• 2)Fitness computation: ENN constructs and trains a network for each particle using the training data. The fitness of each particle is evaluated based on network performance metrics, such as error or accuracy.
• 3)Update of particle positions and velocities: Using individual best position (p_best) and global best position (g_best), the positions and velocities of each particle are updated. This update process is governed by parameters such as inertia weight (w), cognitive factor (c₁), social factor (c₂), and random numbers (r₁ and r₂). Equations (5), (6) describe the velocity and position update of particles:

$$v\left(t+1\right)=wv\left(t\right)+c_1{r}_1\left(p_{best}-position\left(t\right)\right)+c_2{r}_2\left(g_{best}-position\left(t\right)\right)$$ (5)

$$position\left(t+1\right)=position\left(t\right)+v\left(t+1\right)$$ (6)

Here, v(t+1) represents the next generation speed; position(t+1) denotes the subsequent position update; w denotes the inertia weight; c₁ and c₂ are learning factors; r₁ and r₂ represent random numbers; p_best signifies the individual optimal position of particles; g_best refers to the global optimal position of the entire particle swarm.

• 4)Repeat the above steps until stopping criteria are met, such as reaching the maximum number of iterations or a predefined fitness level. In each iteration, particles update their positions based on the current velocity and compute fitness again.
• 5)Optimization of ENN. The PSO algorithm optimizes the initial connection weights and biases of ENN, mitigating local optima issues during network training and thereby enhancing ENN performance. The optimized ENN can better capture data patterns and improve prediction accuracy.
• 6)Enhancement of ENN. Introducing the PSO algorithm allows ENN to escape local minima during training, exploring a broader parameter space to discover global optimal or near-global optimal solutions. This enhances the generalization capability of ENN and improves prediction accuracy on new data.
• 7)Selection of the optimal solution. After completing iterations, the parameters of ENN corresponding to particles with the highest fitness are selected as the final optimized solution.

The specific hyperparameter settings for this optimization algorithm are presented in Table 2.

Table 2.

Hyperparameters for optimization algorithm.

Hyperparameter name	Value
Learning Rate	0.0001
Hidden Units	128
Activation Function	ReLU
Batch Size	64
Epochs	100
Ensemble Strategy	Weighted Averaging
Number of Iterations	200
Initial Point Selection	Random

Therefore, the PSO algorithm can be optimized and tailored to specific problems and requirements in this study when applied to the construction cost prediction of intelligent buildings. Enhancements in algorithm performance and convergence speed can be achieved by adjusting parameters such as the inertia weight, learning factor, and random number values, and by implementing adaptive strategies.

3.3. Intelligent building construction cost prediction model based on BIM and ENN

To effectively predict the construction cost of intelligent buildings, this study introduces BIM technology and establishes an intelligent building model. Following data visualization and normalization preprocessing, the ENN is optimized using the PSO algorithm, thereby establishing the construction cost model for intelligent buildings. The data from the BIM intelligent building model is fed into the ENN as input, and the parameters of the neural network are optimized to predict the construction cost of intelligent buildings. The specific intelligent building construction cost prediction model based on BIM and ENN is illustrated in Fig. 5.

Fig. 5.

Framework of intelligent building construction cost prediction model based on BIM and ENN.

Fig. 5 illustrates the framework of the intelligent building construction cost prediction model based on BIM and the ENN. In this model, BIM technology is initially employed to digitize and visualize various aspects of building data, including geometric information (shape and size), attribute information (material strength, density, and cost), component relationships (beam and column connections), and spatial topology (building layout). Subsequently, the data undergoes preprocessing, which includes cleaning, removal of abnormal values, filling missing values, and other operations to ensure data quality and consistency. Next, the preprocessed BIM data serves as input for the ENN. The ENN optimizes parameters such as weights and thresholds using the PSO technique. This optimization process enhances the ENN's capability to accurately predict the construction cost of intelligent buildings. By utilizing BIM data as input, the ENN learns the relationships between architectural elements and construction costs, enabling it to forecast costs based on the input data.

When applying the PSO algorithm to optimize the ENN, consider a three-layer network as an example. The coding length S can be defined according to real coding rules as shown in Equation (7):

$$S=S_1\cdot S_2+S_2\cdot S_3+S_2\cdot S_2+S_2+S_3$$ (7)

In Equation (7), $S_1$ represents the number of neurons in the input layer, $S_2$ represents the number of neurons in the hidden layer, and $S_3$ represents the number of neurons in the output layer of the network.

Next, set the algorithm parameters. The initial population size plays a crucial role in determining both the convergence speed and the quality of the final search results. However, these factors often exhibit an inverse relationship. Therefore, it is essential to judiciously select the initial population size popu, ensuring it strikes a balance between being neither too large nor too small. The size FF of the subpopulation can be defined by Equation (8):

$$FF=popu/\left(B\_size+T\_size\right)$$ (8)

In Equation (8), $B\_size$ denotes the number of superior subgroups, and $T\_size$ represents the number of temporary subgroups.

The fitness function serves as a crucial metric in individual search evolution, characterized by its simplicity in calculation, universal applicability, positive values, and continuity when choosing this parameter. Typically, the trained network employs training data for prediction, where the mean square error (mse) of the prediction output is inverted to determine the fitness value val of the current individual, as illustrated by Equation (9):

$$val=\frac{1}{mse\left(T-A\right)}$$ (9)

In Equation (9), T represents the expected output value of the training data, and A denotes the predicted output value of the training data.

This model enables researchers to extract building information through BIM technology and predict construction costs using ENNs. The performance and accuracy of ENNs are enhanced by optimizing their parameters with the PSO algorithm. This approach provides a valuable technique for estimating construction costs within the domain of intelligent buildings and promotes the digital and intelligent advancement of construction enterprises.

3.4. Simulation experiment evaluation

To assess the performance of the intelligent building construction cost prediction model based on BIM and ENN developed in this study, research data are collected using web crawling technology from various sources. These sources included government and industry reports, academic papers, patents, publicly available corporate data, social media, news websites, and open data platforms. The data spanned from April 2019 to October 2022 in Xi'an City, focusing on the artificial costs of hot-rolled third-level seismic rebars (HRB400Eφ18-25 mm). Key components included average wages for construction workers and management personnel, equipment costs (average rental prices for mechanical and measurement equipment), and information costs. The study introduces a rolling prediction approach with a time frequency of four to achieve long-term forecasting of the model and ensure the timeliness of results. Given the substantial price fluctuations in steel bars, fluctuations at each node can vary by up to 800 CNY per ton, contributing to data variability. Prior to network training in this study, original data must undergo normalization. Equation (10) illustrates the standardization of dimensionless data, commonly achieved through the maximum-minimum method to scale the original data to the range [0, 1]:

$$x_k=\left(x_k-x_{\min }\right)/\left(x_{\max }-x_{\min }\right)$$ (10)

In Equation (10), $x_{\min }$ represents the minimum value in the data sequence ($x_{\min }=-1$), and $x_{\max }$ denotes the maximum value in the sequence ($x_{\max }=1$).

To evaluate the performance of the model developed in this study, comparative experiments were conducted with the ENN, the (back propagation) BP neural network [38], LSTM algorithm [39], and the algorithm proposed by Li et al. (2023). The evaluation focuses on several metrics including fitting effect, prediction accuracy, root mean squared error (RMSE), coefficient of determination (R²), and Area under the Curve (AUC).

Fitting Effect assesses how well the model aligns with historical data, typically gauged by the proximity between predicted values and actual values.

RMSE is a widely used metric to quantify the accuracy of model predictions. It computes the square root of the average of the squared differences between predicted values and actual values. A smaller RMSE indicates higher prediction accuracy of the model.

The determination coefficient (R²), also known as the coefficient of determination or square of the correlation coefficient, quantifies the extent to which the model elucidates the variability of the data. It ranges from 0 to 1, where a value closer to 1 signifies a stronger explanatory power of the model.

Prediction accuracy directly quantifies the percentage of agreement between the model's predicted outcomes and the actual results, reflecting the model's precision in practical applications.

The AUC metric represents the area under the Receiver Operating Characteristic (ROC) Curve, utilized for evaluating the performance of classification models. A higher AUC value indicates greater classification capability of the model.

4. Results and discussions

This chapter presents detailed simulation experimental results of the intelligent construction cost prediction model based on BIM and ENN. Through comparative analysis, the model demonstrates superior performance over existing methods in crucial metrics including prediction accuracy, RMSE, R², and AUC. These results not only validate the efficiency and reliability of the model but also underscore its practical significance and potential for application in construction cost management.

4.1. Comparison results and discussion of fitting performance

The cost of intelligent building materials is normalized, as depicted in Fig. 6.

Fig. 6.

Normalized standard results.

In Fig. 6, after normalization, the price information of hot-rolled Grade III seismic reinforcement concentrates between −1 and 1. This preprocessing reduces oscillations in material cost data during gradient updates, accelerates gradient descent, and facilitates quicker convergence of the network, thereby shortening training time.

The fitting effect of price information under different algorithms is further compared, as depicted in Fig. 7. Fig. 8 illustrates the prediction accuracy of various model algorithms for construction costs.

Fig. 7.

Fitting results of information price under different algorithms.

Fig. 8.

Prediction accuracy results of different algorithms.

In Fig. 7, Fig. 8, it is evident that across different months, the model proposed in this study achieves the closest prediction values to the actual ones, with an average prediction accuracy of 95.83 %. Following closely is the model algorithm suggested by Li et al. (2023), which achieves a prediction accuracy of 91.54 %, marking a 4.29 % difference compared to the BIM-ENN model proposed here. The sequence of model performance, from highest to lowest accuracy, is as follows: Model by Li et al. (2023) > LSTM > ENN > BP neural network. The intelligent building construction cost prediction model based on BIM and ENN developed in this study demonstrates a robust capability in predicting construction costs, achieving an overall prediction accuracy of 91.51 %. This accuracy effectively captures the fluctuations in steel bar prices over specific periods and meets the requirements for construction cost prediction in intelligent building projects.

4.2. Comparison and discussion of model optimization performance under different algorithms

Further comparison and analysis of RMSE and R² values across different algorithms are presented in Fig. 9, Fig. 10.

Fig. 9.

Comparison of RMSE results under different algorithms.

Fig. 10.

Comparison of determination coefficient results under different algorithms.

In Fig. 9, Fig. 10, upon comparing the RMSE and determination coefficients across different algorithms, it is evident that these metrics do not exhibit a consistent trend with increasing months. Notably, this study achieves the smallest RMSE, consistently below 75, and the highest determination coefficient, exceeding 0.95, indicating that over 95 % of the cost prediction outcomes are explainable. This underscores the feasibility and efficacy of predicting construction costs in intelligent buildings, meeting the requisite prediction standards. In contrast, the model proposed by Li et al. (2023) shows RMSE values exceeding 95 and determination coefficients consistently below 0.95. Specifically, the BP neural network algorithm exhibits the largest RMSE, exceeding 130, and the lowest determination coefficient, consistently below 79.51. Therefore, this study introduces the PSO algorithm within the evolutionary thinking framework to enhance the ENN model, forming a hybrid prediction model for the construction costs of intelligent buildings. This approach mitigates the tendency of single ENN models to converge towards local extreme value. Compared to other models, it demonstrates robust applicability in addressing practical prediction challenges, yielding satisfactory results particularly in predicting steel bar prices.

Further comparison of the AUC values across different algorithms to analyze their sensitivity is presented in Fig. 11.

Fig. 11.

AUC curve under different algorithms.

Fig. 11 illustrates the comparison of AUC values across different models. The BIM-ENN model reported here demonstrates superior sensitivity and specificity performance across most thresholds. Notably, it exhibits higher sensitivity at higher thresholds (e.g., sensitivity of 0.95 at threshold 0.1, with 1-specificity only 0.10) and achieves an AUC value of 0.95, indicating strong discriminative ability. In contrast, the model proposed by Li et al. (2023) shows slightly lower performance, maintaining good sensitivity and specificity at various thresholds but inferior to the BIM-ENN model (e.g., sensitivity of 0.92 at threshold 0.1, with 1-specificity of 0.15). LSTM and ENN display higher sensitivity at lower thresholds, but their sensitivity declines as the threshold increases, while specificity remains relatively high, suggesting a potential for more false positives. The BP neural network performs the least favorably, exhibiting lower sensitivity and specificity at all thresholds compared to other models (e.g., sensitivity of 0.80 at threshold 0.1, with 1-specificity of 0.25), indicating weaker discriminative ability. Overall, the BIM-ENN model excels in terms of AUC, sensitivity, and specificity compared to the comparison models, particularly demonstrating higher sensitivity and lower 1-specificity at higher thresholds, thereby enhancing practical application effectiveness.

4.3. Discussion

Through performance analysis of the BIM-ENN model, it demonstrates exceptional prediction accuracy across different months, averaging 95.83 %. This accuracy surpasses that of comparative models by 4.29 %, particularly outperforming the model algorithm proposed by Li et al. (2023). These findings underscore the BIM-ENN model's superiority in predicting construction costs for intelligent buildings, providing a dependable tool for accurate cost estimation, consistent with Baduge et al. (2022) [40]. This result not only significantly exceeds traditional methods but also highlights the model's efficiency and reliability in predicting intelligent building construction costs.

Furthermore, through comparison of RMSE and determination coefficients, this study's model consistently exhibits smaller RMSE values and larger determination coefficients across different months. This underscores the BIM-ENN model's robust performance in predicting costs, effectively capturing cost fluctuations and supporting feasibility in managing intelligent building construction costs. In contrast with other models, particularly the BP neural network, this model demonstrates superior performance in RMSE and determination coefficient, aligning with the perspectives of Alshboul et al. (2022) [41] and Pham et al. (2023) [42]. Therefore, the smaller RMSE and higher R² values of the BIM-ENN model indicate its capability to accurately capture fluctuations in cost data with minimal error, thereby explaining a significant portion of cost variations—an essential factor in optimizing and managing construction costs.

The BIM-ENN model exhibits significant advantages over traditional methods in several respects. Firstly, leveraging the digitalization and visualization capabilities of BIM technology enables the intelligent building construction cost prediction model to accurately depict real-world cost scenarios in construction projects. Secondly, integrating the ENN with the PSO algorithm enhances the model's capability to handle time-series data, effectively capturing dynamic trends in cost variations. Moreover, the model's enhanced generalization ability allows it to perform effectively across diverse datasets and adapt to varying construction projects and market conditions.

However, several limitations are acknowledged in this study. Firstly, the BIM-ENN model's performance is somewhat constrained by the quality and availability of data. While it currently delivers accurate predictions based on existing data, subtle variations in construction cost estimates may not be fully captured. Therefore, future research should prioritize optimizing data collection and preprocessing methods, and consider leveraging advanced deep learning algorithms to enhance the model's predictive precision. Secondly, while this study focuses primarily on cost estimation, the construction industry faces other significant challenges such as scheduling and quality maintenance. Future endeavors could broaden the scope to encompass a wider array of construction issues, thereby offering more comprehensive decision support and optimization solutions.

The intelligent building construction cost prediction model proposed in this study holds significant practical implications for the construction industry. By integrating BIM technology and ENN, the model provides an efficient and accurate cost prediction tool, assisting construction enterprises in making informed decisions during project planning and execution phases. Accurate cost predictions can guide enterprises to make timely procurement decisions amidst market price fluctuations, thereby reducing the risk of cost overruns. Furthermore, the model's real-time capabilities enable enterprises to swiftly adapt to design changes and supply chain fluctuations, allowing for timely adjustments in budgeting and resource allocation. This enhances project economic efficiency and market competitiveness.

From a theoretical perspective, this study introduces a novel research approach for predicting construction costs in intelligent buildings. By integrating BIM technology and ENN, this study not only validates the efficacy of these technologies in cost prediction but also demonstrates their potential to improve prediction accuracy through algorithmic optimization. Moreover, the study employs the PSO algorithm to optimize ENN parameters, offering a fresh perspective on mitigating local optima issues during neural network training. To illustrate the model's impact on decision-making, consider the following scenario: the model predicts a 5 % increase in total project costs due to seasonal rises in labor expenses. Armed with this prediction, project managers can proactively adjust human resource planning by modifying work schedules, increasing labor availability, or reallocating tasks. This data-driven decision-making approach enhances the scientific and systematic aspects of project management, thereby increasing practical value in real-world applications.

5. Conclusion

This study successfully develops an intelligent building construction cost prediction model based on BIM technology and the ENN. By introducing the PSO algorithm, the prediction model not only enhances the prediction accuracy of the ENN but also improves its generalization capability. Experimental results demonstrate that the model achieves an average prediction accuracy of 95.83 % across different months, with an RMSE of less than 75 and a determination coefficient exceeding 0.95. These metrics outperform other models documented in existing literature. These findings indicate that this model can provide the construction industry with a more precise and reliable cost management tool.

The practical significance of the BIM-ENN model in the construction industry is profound. It enables construction companies to predict construction costs more accurately, thereby optimizing budget control and resource allocation. Additionally, it enhances capabilities in enterprise risk assessment and management. Moreover, the application of this model has the potential to elevate forecasting standards within the construction sector, foster technological integration and innovation, improve enterprise competitiveness, and propel the industry towards sustainable digital and intelligent development.

This study contributes by proposing an innovative cost prediction method that demonstrates high accuracy and reliability through experimentation. It introduces a new perspective and tool for the digital transformation and intelligent upgrading of the construction industry. By integrating BIM technology with optimized ENN, this study introduces a novel and efficient cost management tool for the construction industry. This approach aims to enhance the efficiency and effectiveness of cost management and project decision-making within construction enterprises.

Despite the significant achievements in intelligent construction cost prediction, there is a need to further expand the scope and depth of the research. Future studies could extend this model to other aspects of construction management, such as scheduling, quality control, and environmental impact assessment. Moreover, with the advancement of big data and artificial intelligence technologies, future research could explore the application of these new technologies in construction management to further enhance the performance and applicability of prediction models. Continuous technological innovation and application expansion can contribute significantly to the digitalization and intelligent development of the construction industry.

Data availability

All data generated or analysed during this study are included in this published article [and its supplementary information files].

CRediT authorship contribution statement

Yanfen Zhang: Writing – review & editing, Writing – original draft, Visualization, Funding acquisition, Formal analysis, Data curation, Conceptualization. Haijun Mo: Writing – review & editing, Writing – original draft, Visualization, Funding acquisition, Formal analysis, Data curation, Conceptualization.

Declaration of competing interest

The authors declare the following financial interests/personal relationships which may be considered as potential competing interests:Yanfen Zhang reports financial support was provided by The Educational Science planning project of Guangdong Province (Higher Education Special Project). If there are other authors, they declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.