Research on a Carbon Emission Prediction Model for the Construction Phase of Underground Space Engineering Based on Typical Resource Carbon Consumption and Its Application

Abstract

The trend of global warming remains severe. As one of the major sources of carbon emissions, the construction industry still requires large-scale and effective transformations. This study, grounded in Life Cycle Assessment (LCA) theory, carbon emission factor calculation methods, the Monte Carlo method, and feedforward neural network algorithms, develops a carbon emission prediction model based on the carbon emissions generated by typical resource consumption. The model is established in the context of typical carbon emission patterns observed during the construction phase of subway stations in China. Furthermore, the feasibility of the proposed model is validated through its application to specific engineering projects. The results demonstrate that (1) the newly developed carbon emission model can accurately predict the carbon emissions associated with the construction phase of subway stations in China; (2) actual carbon emission calculations in construction projects require the integration of data from multiple sources to ensure comprehensive coverage and avoid omissions; and (3) during the construction phase of subway stations, the use of concrete and steel constitutes significant sources of carbon emissions.

1. Introduction

The quantification of carbon emissions in construction engineering has undergone a considerable period of development. In terms of international research, as early as 2005, Japanese scholars conducted a study calculating the carbon emissions during the construction phase of a concrete structure [1]. A joint study by Harbin Institute of Technology and The Hong Kong Polytechnic University measured the carbon emissions during the construction phase of a building project in Hong Kong [2]. In 2014, a study from The Hong Kong University of Science and Technology reviewed carbon footprint data related to the production of cement and concrete materials in China and Asia, using real cases to calculate and compare carbon emissions in the Hong Kong region [3]. Subsequent studies included the calculation and analysis of the carbon footprint of materials required for highway construction in India [4]. To investigate the carbon emissions associated with traditional concrete slabs and newly developed hollow slab structures during the construction phase in South Korea, Paik and Na [5] retrieved relevant data from two national carbon footprint databases, thus completing a quantitative analysis. A study in Malaysia focused on the carbon emissions during the construction phase of a new prefabricated building structure [6]. Additionally, to examine the carbon emissions during the construction phase of an innovative reclaimed land structure in Hong Kong, Chen et al. [7] conducted a detailed study emphasizing construction processes and structural aspects.

In mainland China, relevant research has developed rapidly, evolving from initial direct carbon emission calculations to studies incorporating prediction as either a supplementary tool or a primary focus. As early as 2014, Luo et al. [8] quantified carbon emissions during the construction phase of an office building by analyzing both the construction process and material consumption, followed by regression analysis linking the results to various engineering parameters and identifying certain patterns. A study focusing on a cut-and-cover subway station project in Beijing calculated carbon emissions while also examining the relationship between emissions and engineering parameters such as station depth [9]. Pi [10] utilized multiple carbon emission factor sources to calculate and analyze emissions for a shield tunnel project in southern China. Liu Jialin [11] employed multi-source data (including industrial statistics and academic studies) to quantify emissions for general building construction in China, further exploring the relationship between carbon emissions and building area. Similarly, Xiang [12] used diverse emission factor sources to calculate the carbon emissions of a framed building project in Fujian. Li and Chen [13] used IPCC guidelines, regional statistical data, and academic research to measure emissions during the construction phase of a residential project in southern China. Huang [14] applied the national standard Calculation Standard for Carbon Emission of Buildings (GB/T 51366-2019) [15] Standard for Building Carbon Emission Calculation and industry data to analyze carbon emissions during the construction of shield tunnels and subway stations in southern China. Guo Yalin and Guo Chun [16] used emission factors sourced from IPCC guidelines and academic studies to analyze emissions during railway tunnel construction. Chen et al. [17] employed the same national standard and academic emission factor data to investigate carbon emissions in the construction phase of a subway shield tunnel in southern China. Zhang [18] utilized the CLCD (China Life Cycle Database), academic research, and local guidelines to study the carbon emissions associated with new buildings constructed within subway protection zones. Chen et al. [19] applied GB/T 51366-2019, Chinese power grid statistics, and academic carbon emission factor data combined with neural network models and other prediction algorithms to predict emissions from subway construction projects in southwest China. Xu et al. [20] used the aforementioned standard and Chinese governmental guidelines to derive emission factors for resources such as water, electricity, and fossil fuels, assessing emissions from the production to construction phases of precast concrete slabs in northwest China. Song [21] adopted Chinese standards and a wide range of academic studies and selected international carbon footprint data to comprehensively explore emissions during the materialization phase of construction. Recent research has further applied comprehensive machine learning theories, incorporating structural, mechanical, and geometric parameters of construction projects to develop various prediction models for residential building construction emissions, ultimately identifying the most suitable approach for predicting emissions during the materialization phase [22].

These findings suggest the following:

• Research on carbon emission quantification in construction projects is still predominantly based on direct calculation methods, with predictive models gradually emerging over time;
• Predictive analyses in carbon emission quantification often rely on extensive engineering data requiring rigorous data collection and processing, which involves a significant workload and poses challenges to developing generalized and efficient calculation methodologies.

To address the aforementioned challenges, this study proposes a novel approach based on predicting carbon emissions through the emissions of typical resources. First, the boundaries for carbon emission quantification during the construction phase of engineering projects are established. Subsequently, two commonly used data analysis and prediction methods are introduced. Drawing on typical studies of carbon emissions during the construction phase of subway stations in China, general bills of quantities are analyzed as references for underground construction carbon emission calculations. A detailed examination of typical resource consumption is then conducted to identify which resources can serve as predictors for overall emissions. Accordingly, two neural network prediction models are developed to estimate total carbon emissions during the construction phase based on emissions from key resources. The feasibility of this mathematical prediction approach is validated through application to specific engineering cases. The technical roadmap of this study is illustrated in Figure 1.

2. Methodologies and Theories

2.1. Carbon Emission Factor Method

Prediction should be conducted based on actual values. Based on the comprehensive review of existing research on carbon emissions in construction projects, this study adopts the carbon emission factor method to perform the actual measurement of carbon emissions during the construction phase of building projects.

2.1.1. Calculation Formulae

Generally, to conduct calculations effectively, the vast majority of studies on carbon emissions in construction engineering use the carbon emission factor method. This method determines the amount of greenhouse gases contained per unit of various resources (typically measured in kilogram equivalents of carbon dioxide, denoted as kgeCO₂). Unless otherwise specified, this unit is used throughout this paper, with other greenhouse gases (considered based on specific projects or case studies) to calculate the total carbon emissions of a construction project. The calculation method is as follows:

$$CE=\sum _{i=1}^{m}E{F}_m\times Q_m$$

(1)

where CE is the whole carbon emissions for a building project, EF_m is the carbon emission factor for all utilized resources in the construction project, and Q_m is the amount (s) of all possible resources.

Carbon emission factors represent an environmental impact assessment indicator for a material, to some extent reflecting the carbon footprint potential of different materials in different regions [23]. For construction projects, the use of various resources ultimately represents the approximate scale of the project, which can directly and effectively reflect the actual project.

2.1.2. Analysis of Major Sources of Global Carbon Emission Factors

To achieve a more precise quantification of carbon emissions in underground engineering, this study prioritizes data coverage as the primary consideration. A comprehensive retrieval of relevant carbon emission measurement data was conducted on a global scale. Given that urban rail transit systems represent an important and systematic example of underground construction projects—encompassing both tunnel and underground structural characteristics—related studies from regions with such systems are regarded as representative and authoritative within the field of underground construction. Accordingly, this study specifically collected relevant data from cities and regions worldwide (including China) that possess typical urban rail transit systems.

The Beijing and Shanghai subway systems are globally renowned. China’s Calculation Standard for Carbon Emission of Buildings (GB/T 51366-2019) [15] provides detailed research on the carbon footprint of building materials, fossil fuels, and machinery, and is an important guiding document for building engineering carbon emissions. The China Product Life Cycle Greenhouse Gas Emission Coefficient Database (CPCD) [23] is a free and open source, covering carbon footprint data for single resources and some commodities, and is suitable for engineering and academic research.

Available carbon emission databases in Asia include South Korea’s EPD (Environmental Product Declaration) database, which covers carbon emission data for some metal products, and carbon emission data for building materials and chemical materials provided by KEITI (Korea Environmental Industry & Technology Institute). The Seoul subway also has a certain global reputation, so this paper cites the above data sources.

The Paris Metro is a global model for underground railway systems. The Ecoinvent database, maintained by a Swiss institution, is a world-leading life cycle assessment database containing over 18,000 datasets covering agriculture, energy, manufacturing, transportation, and other fields and records detailed carbon emission data from resource input to production completion. Ecoinvent is characterized by its completeness and transparency and is widely used in the carbon footprint assessment of building materials and ancillary materials, with a wide geographical coverage.

The New York subway is globally renowned, and Canada’s Toronto and Vancouver also have well-developed subway systems. The Ecoinvent database contains some North American carbon emission data, and the official methods of the US EPA (Environmental Protection Agency) provide carbon emission metering methods for fossil fuels and chemical products. The USLCI (US Life Cycle Inventory) database, similar to Ecoinvent, also contains carbon emission data related to building engineering, making North American data one of the research focuses.

The Sydney railway system has a certain reputation. Australia’s carbon emission data resources include the EPiC (Environmental Impact of Construction) database, which was jointly developed by the University of Melbourne and the government, focusing on carbon emission factor lists for resources required for building engineering; in addition, the Australian government provides a simple fuel carbon footprint calculation tool for preliminary estimates of gasoline and diesel consumption carbon emissions. The AULCI (Australian Life Cycle Inventory) database, although the data are older, has detailed metering of Australia’s electricity carbon footprint.

In addition to the data from the above regions, this paper also collected carbon emission factor values used in academic research.

2.2. The Scale of Carbon Emission Measures

The study of carbon emissions in construction engineering generally requires an analysis of the research boundary. In the field of construction engineering, Life Cycle Assessment (LCA) typically divides a building project into four stages: production of building materials, construction of the building structure, building operation, and the end-of-life stage [24]. The construction process in construction engineering primarily involves material production and construction. Based on this, according to the LCA theory, this paper sets the research boundary for carbon emissions during the construction stage of construction engineering as the combined effect of building material production and building structure construction.

2.3. Theories of Data Analysis and Prediction

2.3.1. Monte Carlo Method

The Monte Carlo Data Augmentation method is a technique that employs the principles of Monte Carlo simulation to generate additional data samples. Fundamentally, the Monte Carlo method approximates complex distributions through random sampling, and by repeatedly conducting trials, it gradually converges toward the characteristics of the true distribution. When applied to data augmentation, this approach involves generating multiple random samples based on existing data or known probabilistic models to expand the original dataset. Each simulated sample can be viewed as an individual trial drawn from the underlying data distribution, and the accumulation of such samples helps capture the diversity and uncertainty inherent in the data.

In the field of machine learning, Monte Carlo data augmentation is commonly used to mitigate the issue of insufficient training data. Typically, this is achieved by introducing random noise or variations to the data (such as randomly rotating images), thereby enriching the training set and improving the model’s generalization capability. In statistical modeling, this method is also employed for simulation experiments and uncertainty analysis. By repeatedly generating potential scenario data and evaluating the model’s responses to these scenarios, researchers can gain a comprehensive understanding of the model’s robustness and predictive intervals.

Overall, the Monte Carlo data augmentation method combines theoretical rigor with practical flexibility, making it a powerful tool for building reliable machine learning models and conducting statistical inference.

2.3.2. Artificial Neural Network

Artificial Neural Networks (ANNs) are computational models inspired by biological neural systems that are designed to emulate the information-processing mechanisms of the human brain. The basic structure of an ANN consists of a number of interconnected nodes, referred to as “neurons”, which are typically organized into an input layer, one or more hidden layers, and an output layer. Each neuron is connected to others through weighted links and processes the received input signals by applying weights, followed by transmitting the processed signal through a nonlinear activation function to the next layer. Through successive transmission and transformation across layers, neural networks are capable of modeling complex patterns and nonlinear relationships within input data.

During the training process, ANNs iteratively adjust the connection weights based on a set of existing data to minimize the error between the model’s predicted outputs and the actual results. This process is commonly facilitated by the backpropagation algorithm, enabling the network to possess adaptive learning capabilities. Owing to their highly flexible structure, ANNs have been widely applied in various fields, including classification, regression, image recognition, and natural language processing, and they have demonstrated superior performance in handling large-scale, nonlinear data problems. In recent years, advancements in computational power and the proliferation of big data have further cemented ANNs as one of the core technologies in modern machine learning.

In the field of carbon emission research within civil engineering, the application of artificial neural networks has already been explored in several studies [21,25,26,27], encompassing various forms of ANN architectures. Based on the specific conditions of this study, a feedforward neural network is employed for the construction of the carbon emission prediction model. The underlying theory is elaborated as follows.

The fundamental computational formulas of artificial neural networks consist of several components. Specifically, the formula for calculating the parameters of a neuron is as follows:

$${\mathit{Z}}^{\left(l\right)}={\mathit{W}}^{\left(l\right)}{\mathit{A}}^{(l-1)}+{\mathit{b}}^{\left(l\right)}$$

(2)

$${\mathit{A}}^{\left(l\right)}=f\left({\mathit{Z}}^{\left(l\right)}\right)$$

(3)

where Z^(l) represents the weighted input of the lth layer; W^(l) denotes the weight matrix of the lth layer, with dimensions l × (l − 1); A^(l−1) is the output of the (l − 1)^th layer; b^(l) is the bias vector, which is typically initialized to zero and updated iteratively during training until it converges to a fixed vector.

The iteration of the output vector A is determined by the activation function and the decision parameters within the predictive model. Commonly used activation functions fall into three categories, namely

$$f\left(x\right)={\displaystyle \frac{1}{1+e^{-x}}}$$

(4)

$$f\left(x\right)={\displaystyle \frac{e^x-e^{-x}}{{e^x+e}^{-x}}}$$

(5)

$$f\left(x\right)=\max (0,x)$$

(6)

where x represents the predictive parameter, with its initial value determined by the initial input data, and Formula (6) is called the “ReLU” function. The initial values of the weight matrix are randomly generated, following either a uniform distribution with parameters in the range [−(6/(n⁽^l−1⁾ + n⁽^l⁾))^1/2, (6/(n⁽^l−1⁾ + n⁽^l⁾))^1/2] or a normal distribution with a standard deviation of (2/(n⁽^l−1⁾ + n⁽^l⁾))^1/2, where n⁽^l−1⁾ and n⁽^l⁾ denote the number of neurons in the (l − 1)^th (previous) layer and the lth (current) layer, respectively. Specifically, the first layer (input layer) contains 2 neurons, the final layer (output layer) contains 1 neuron, and the number of neurons in the hidden layers can be set as required based on the practical needs of the model.

The model typically employs Mean Squared Error (MSE) as the loss function to evaluate the convergence of the model and to complete the final computation. The specific formulation is as follows:

$$L={\displaystyle \frac{1}{2}}{‖{\mathit{A}}^{\left(L\right)}-\mathit{Y}‖}^2$$

(7)

where A⁽ᴸ⁾ represents the predicted value from the output layer; and Y denotes the actual target value (the total value from the initial input data).

The error term δ reflects the impact of the current layer’s output on the overall loss function, and is defined as follows:

$${\mathit{\delta }}^{\left(L\right)}=\left({\mathit{A}}^{\left(L\right)}-\mathit{Y}\right)\circ f\prime \left({\mathit{Z}}^{\left(L\right)}\right)$$

(8)

The recursive formula for calculating the error term in the hidden (intermediate) layers is given by

$${\mathit{\delta }}^{\left(l\right)}={\left({\mathit{W}}^{(l+1)}\right)}^{\top }{\delta }^{(l+1)}\circ f\prime \left({\mathit{Z}}^{\left(l\right)}\right)$$

(9)

where W^(l+1) is the weight matrix of the (l + 1)^th layer; δ^(l+1) is the error term of the (l + 1)^th layer; f′(Z^(l)) represents the derivative of the activation function with respect to the weighted input of the current layer; and ∘ denotes element-wise (Hadamard) multiplication.

The objective of neural network computation is to obtain the weights that minimize the loss function to zero. Therefore, weight updates are required in each iteration. Utilizing the error term δ, the gradients of the loss function with respect to the weights and biases are calculated as follows:

$${\displaystyle \frac{\mathsf{\partial }L}{\mathsf{\partial }{\mathit{b}}^{\left(l\right)}}}={\delta }^{\left(l\right)}$$

(10)

The parameter updates are performed using the gradient descent method, expressed as

$${\mathit{W}}^{(l+1)}={\mathit{W}}^{\left(l\right)}-\eta {\displaystyle \frac{\mathsf{\partial }L}{\mathsf{\partial }{\mathit{W}}^{\left(l\right)}}}$$

(11)

$${\mathit{b}}^{(l+1)}={\mathit{b}}^{\left(l\right)}-\eta {\displaystyle \frac{\mathsf{\partial }L}{\mathsf{\partial }{\mathit{b}}^{\left(l\right)}}}$$

(12)

where η denotes the learning rate, a fixed value that controls the step size of each parameter update.

Based on the above computational formulas, the basic calculation procedure of the feedforward neural network is as follows:

• According to the specific conditions, arbitrarily select one of Formulas (4), (5), or (6) (the choice made after the first iteration remains fixed for all subsequent computations). Using Formulas (2), (3), and (7), compute the initial value of L; if L = 0, the process terminates; otherwise, proceed to the next step.
• Apply Formulas (7)–(12) to update the elements of the weight matrix and return to Step 1.

3. Carbon Emission Prediction Model for the Construction Phase of Underground Buildings

3.1. Analysis of Typical Resource Consumption in Subway Station Construction Projects

In order to improve the accuracy of carbon emission prediction in construction projects, this study selects representative academic research results with actual calculated carbon emission values as references for prediction. Additionally, the general bill of quantities for underground construction projects is examined to guide the practical calculation of carbon emissions. Ultimately, a generalized bill of quantities for carbon emissions in underground engineering is proposed, as shown in

Table 1. General BOQs of underground building projects.

Resources	Type
Labor	Fuels and other necessary materials
Gasoline
Diesel
Electricity
Water
Other materials besides above
All possible steel materials	Building materials
Cement
Concrete of various grades
Various waterproof materials
Other materials besides above

.

The construction of this bill of quantities follows the principle of standardizing the use of construction machinery during the building phase by converting their consumption uniformly into diesel fuel and electricity usage. Simultaneously, the consumption of all other resources is categorized under “construction materials”. The specific application of this bill requires case-specific analysis and adjustments based on individual engineering projects. By utilizing this bill of quantities, the data collection process for carbon emission quantification in underground construction projects can be significantly simplified.

3.2. Typical Research Results on Carbon Emissions During the Construction Phase of Subway Stations

Studies [9,14,18,28,29] were identified through literature retrieval and screening as representative cases of research on carbon emissions during the construction phase of subway stations in China. Upon in-depth analysis, it was found that these studies considered different resources or activities in their carbon emission assessments; however, all of them explicitly reported the carbon emissions generated from the use of concrete and various types of steel during the construction phase. Therefore, it can be concluded that concrete and steel are the two most significant contributors to carbon emissions in the construction of subway stations.

3.3. Prediction Model of Carbon Emissions

Unless otherwise specified, the carbon emission prediction models developed in this study were implemented using Python 3.12 programming.

3.3.1. Unified Parameters of the Neural Network Model

According to empirical practice, the number of training samples used in a neural network should typically be 3 to 10 times the number of network parameters. Taking an example with 3 input features, hidden layers containing 128, 64, and 32 neurons, respectively, and 1 output feature, the total number of network parameters for this model is calculated as (3 × 128) + 128 + (128 × 64) + 64 + (64 × 32) + 32 + (32 × 1) + 1 = 10,881. Due to the limited availability of reference data, this study first extracts the carbon emission values—total emissions, emissions from all concrete, and emissions from all steel—from nine engineering cases reported in the aforementioned typical studies. Using the Monte Carlo method, each input feature is multiplied by a randomly selected coefficient and added with noise. Specifically, the random coefficient is drawn uniformly from the range [1.05, 1.25], and the noise follows a normal distribution N∼(0, 5 × 10⁶), both sampled randomly. This approach expands the original parameter set to 5000 samples.

By employing Python’s sklearn library, the StandardScaler and MLPRegressor modules are used to specify the basic parameters of the feedforward neural network, resulting in an initial, general, and unified neural network configuration, as shown in

Table 2. Initial parameters of NN.

Initial Parameters	Settings or Values
Size of hidden layers	(64, 32)
Learning rate	0.01
Maximum iteration(s)	10,000
Random seeds	42

.

The random seed value is fixed to ensure the uniqueness of the neural network parameters in each iteration. The input samples of the prediction model before data augmentation are shown in

Table 3. Unified parameters of the neural network model (Unit: kgeCO₂).

Resources	Total Carbon Emissions	Carbon Emissions of All Kinds of Concrete	Carbon Emissions of All Kinds of Steel
[9]	37,725,056	16,012,694	19,870,294
[14]	87,740,150	28,116,490	47,294,040
[18]	142,132,773	47,015,446	5,779,341
[18]	159,232,249	57,432,784	3,125,524
[18]	94,663,375	30,309,643	1,682,688
[18]	103,718,735	38,186,038	3,901,956
[18]	108,626,121	37,443,919	807,542
[18]	144,223,427	55,687,509	71,241,942
[28]	53,288,800	11,828,440	19,797,440

.

3.3.2. Parameter Tuning and Final Parameters

Due to the scarcity of available reference data, this study adopts the results from a study [29] as the baseline values for parameter tuning, ensuring that the final neural network model’s predictions align with actual conditions. The baseline parameters are presented in

Table 4. Reference data of carbon emissions (Unit: kgeCO₂).

Total Carbon Emissions	Carbon Emissions of All Kinds of Concrete	Carbon Emissions of All Kinds of Steel
47,718,610	17,511,797.77	23,126,232.72

.

After multiple adjustments, the final parameters of the neural network-based carbon emission prediction model for the construction phase of underground engineering are summarized in

Table 5. Final parameters of the feedforward neural network carbon emission prediction model.

Parameters	Settings or Values
Size of hidden layers	(64, 32, 16)
Activation function	ReLU
Approach of learning	Adaptive
Initial learning rate	0.0003
Maximum iteration(s)	100,000
Parameter of L2 regularization	0.003
Whether to pre-stop	True (Yes)
Minimum iteration(s) as the model does not change	3000
Random seeds	42

.

Adjusting the learning method of the neural network to “adaptive” and setting a base learning rate can reduce unnecessary computation time and enable the neural network to identify patterns that closely match the prediction data. To prevent the loss function from converging too rapidly or unreasonably, the L2 regularization technique is introduced. This approach amplifies the value of the loss function by incorporating the weight values during each iteration, thereby enhancing the stability of the model, avoiding excessive descent, and reducing the risk of overfitting. Additionally, since neural networks often exhibit rapid convergence during actual computation, an early stopping mechanism is implemented. A minimum threshold is set for halting computations when model parameters no longer change significantly, which helps to lower computational load and save processing time.

4. Case Study

4.1. Engineering Case and Calculation Parameters

This study conducts a carbon emission analysis based on the construction of a new subway station in South China. The new station is located on the northern side of a major east–west arterial road in a first-tier city, parallel to an existing subway station. It features a single-column, double-span, reinforced concrete, rectangular underground structure with four levels. The main construction phase of the station is expected to span 734 days. Figure 2 shows the actual engineering and geological conditions of this new subway station in a map.

Based on the carbon emission prediction model established in Chapter 3 and combined with the actual data of the project, the bill of quantities for the construction phase of the subway station is summarized, and the carbon emission factors for various resources are specified. The final statistical results are presented in

Table 6. BOQs and corresponding carbon emission factors for the construction phase of the new subway station.

Resources	Classifications	Unit	Quantities	Unit of Carbon Emission Factors
Labor	fuel and other necessary materials	Labor Day	462,978.80	kgeCO2/Labor Day
Gasoline	fuel and other necessary materials	kg	157,167.10	kgeCO2/kg
Diesel	fuel and other necessary materials	kg	1,200,850.70	kgeCO2/kg
Electricity	fuel and other necessary materials	kWh	6,006,359.40	kgeCO2/kWh
Water	fuel and other necessary materials	m3	61,408.90	kgeCO2/m3
Acetylene	fuel and other necessary materials	kg	1097.90	kgeCO2/kg
Rebar	construction materials	t	11,589.70	kgeCO2/t
Hot-rolled thick steel plate	construction materials	t	3984.40	kgeCO2/t
Cement	construction materials	t	3649.24	kgeCO2/t
C20 Concrete	construction materials	m3	4894.80	kgeCO2/m3
C30 Concrete	construction materials	m3	4750.30	kgeCO2/m3
C35 Concrete	construction materials	m3	12,672.90	kgeCO2/m3
C35 Waterproof Concrete	construction materials	m3	47,331.80	kgeCO2/m3
C50 Concrete	construction materials	m3	457.30	kgeCO2/m3
Galvanized steel plate	construction materials	t	38.80	kgeCO2/t
Polystyrene (3 cm thick)	construction materials	m2	1890.70	kgeCO2/m2
PVC (POLYVINYL CHLORIDE) board (1.5 cm thick)	construction materials	m2	2911.50	kgeCO2/m2
Back-attached water-stop tape	construction materials	m	5636.30	kgeCO2/m
Non-tar polyurethane	construction materials	kg	1882.60	kgeCO2/kg
Modified bitumen waterproof roll	construction materials	kg	286.20	kgeCO2/kg
Acrylic spray membrane	construction materials	kg	80,478.70	kgeCO2/kg

.

4.2. Total Carbon Emissions of the Construction Project

Based on the previously collected data, priority is given to Chinese values, with necessary supplementation using data from other regions that best align with the bill of quantities. The actual total carbon emissions are then calculated accordingly. For certain resources, unit conversion and straightforward derivation of carbon emission factors are required. The specific details are provided below.

Although the carbon emission factors for gasoline and diesel are available, their units differ from those specified earlier and require unit conversion. Specifically, in the Standard for Building Carbon Emission Calculation (GB/T 51366-2019), the emission factors for gasoline and diesel are expressed in units of tCO₂/TJ, i.e., carbon footprint per unit of energy content. Referring to the quantified lower heating values of gasoline and diesel provided in the General Principles for Calculation of Comprehensive Energy Consumption (GB/T 2589-2020) [30], expressed in kJ/kg, the conversion from kgCO₂/J to kgCO₂/kg is performed accordingly.

Based on the chemical formula of calcium carbide (CaC₂) and carbon dioxide (CO₂), the derivation method for the carbon emission factor of acetylene is as follows:

$$E{F}_{Ethyne}=E{F}_{Carbide}\times 11/8$$

(13)

where EF_Ethyne represents the carbon emission factor of acetylene in units of kgCO₂/kg; and EF_Carbide represents the carbon emission factor of calcium carbide, also in units of kgCO₂/kg.

According to the CPCD, the carbon emission factor of calcium carbide is 5.022 kgCO₂/kg. Using this method, the derived carbon emission factor for acetylene is 6.97 kgCO₂/kg.

Based on the above results, the calculation is performed using Equation (1). The actual calculated value of carbon emissions during the construction phase of the subway station project is shown in

Table 7. Actual calculation results of carbon emissions of the construction phase.

Resources	Unit	Quantities	Unit of Carbon Emission Factors	Value of Carbon Emission Factor	Resource of Data	Carbon Emissions (kgeCO2)
Labor	Labor Day	462,978.80	kgeCO2/Labor Day	0.46	Zhu [28]	212,970.25
Gasoline	kg	157,167.10	kgeCO2/kg	2.93	GB/T 51366-2019 [23] GB/T 2589-2020 [30]	460,271.84
Diesel	kg	1,200,850.70	kgeCO2/kg	3.10	GB/T 51366-2019 [23]; GB/T 2589-2020 [30]	3,722,584.27
Electricity	kWh	6,006,359.40	kgeCO2/kWh	0.57	CPCD [22]	3,425,426.77
Water	m3	61,408.90	kgeCO2/m3	0.17	GB/T 51366-2019 [23]	10,316.70
Acetylene	kg	1097.90	kgeCO2/kg	6.97	CPCD	7652.36
Rebar	t	11,589.70	kgeCO2/t	2340.00	CPCD	27,119,898.00
Hot-rolled thick steel plate	t	3984.40	kgeCO2/t	2550.00	CPCD	10,160,220.00
Cement	t	3649.24	kgeCO2/t	817.00	CPCD	2,981,429.08
C20 Concrete	m3	4894.80	kgeCO2/m3	230.00	Song [20]	1,125,804.00
C30 Concrete	m3	4750.30	kgeCO2/m3	295.00	GB/T 51366-2019	1,401,338.50
C35 Concrete	m3	12,672.90	kgeCO2/m3	317.50	Song [20]	4,023,645.75
C35 Waterproof Concrete	m3	47,331.80	kgeCO2/m3	293.09	Wang [31]	13,872,477.26
C50 Concrete	m3	457.30	kgeCO2/m3	385.00	GB/T 51366-2019 [23]	176,060.50
Galvanized steel plate	t	38.80	kgeCO2/t	2600.00	CPCD	100,880.00
Polystyrene (3 cm thick)	m2	1890.70	kgeCO2/m2	5.40	EPiC	10,209.78
PVC (POLYVINYL CHLORIDE) board (1.5 cm thick)	m2	2911.50	kgeCO2/m2	1.08	USLCI	3142.16
Back-attached water stop	m	5636.30	kgeCO2/m	3.08	Wang [31]	17,489.44
Non-tar polyurethane	kg	1882.60	kgeCO2/kg	7.70	EPiC	14,496.02
Modified bitumen waterproof roll	kg	286.20	kgeCO2/kg	0.20	EPiC	1225.79
Acrylic spray membrane	kg	80,478.70	kgeCO2/kg	0.43	EPiC	34,605.84

.

The final total carbon emissions are approximately 6.88 × 10⁷ kgCO₂.

4.3. Applications and Validations of the Carbon Emission Prediction Model

Based on the actual carbon emission calculation results from the previous section, the total carbon emissions from concrete and steel consumption are calculated separately, and the carbon emissions resulting from electricity consumption are also extracted. These three carbon emission values are then input into the prediction model proposed in this study. The prediction results are shown in

Table 8. Comparison of actual calculations and estimations (unit: kgeCO₂).

Total Calculated Carbon Emissions	Parameters of Prediction Model	Estimation of Total Carbon Emissions	Relative Bias (100%)
68,882,144.31	concrete and steel	68,009,389.82	1.27

.

Although the predicted value is slightly lower than the actual calculated value, the relative error remains within 5%. Therefore, it can be concluded that the carbon emission prediction in this study is valid.

5. Conclusions

In this study, a carbon emission prediction model for the construction phase of underground buildings is developed, utilizing the Monte Carlo method and fundamental feedforward neural network theory. The model is based on typical carbon emission values from subway station construction studies as key predictive parameters, constructing a “quantity-to-quantity” prediction approach. Drawing on the current research landscape, relevant carbon emission databases applicable to the actual measurement of construction project emissions were reviewed. Using data from these sources, the model was applied to a newly constructed subway station project in southern China to calculate the actual carbon emissions during its construction phase, thereby validating the effectiveness of the proposed prediction model. The results indicate that the newly proposed model can effectively estimate carbon emissions. Additionally, the study yields the following findings:

• The actual calculation of carbon emissions in construction projects still requires support from multiple data sources to achieve comprehensive coverage of project activities.
• During the construction phase of subway stations, the use of concrete and steel predominantly determines the carbon emission levels.

The limitations of this study include:

• Due to constraints in the availability of relevant reference data, the number of factors available for carbon emission prediction is limited.
• Owing to the incomplete coverage of carbon emission databases, certain non-China data sources were used in the actual carbon emission calculations. This may introduce discrepancies in the actual emission values, subsequently affecting the accuracy of the prediction results.

The results of this study suggest that to enhance the precision of carbon emission quantification in construction projects, it is essential to continuously establish and improve dedicated carbon emission databases tailored to the specific conditions of different types of construction projects. Furthermore, the findings highlight the significant contribution of typical construction materials to carbon emissions during the building phase. This insight can provide policymakers with a more intuitive understanding, thereby promoting the development of innovative construction materials and reducing the negative impact of civil engineering projects on climate change and environmental sustainability at its source.