Analysis of Spatio-Temporal Characteristics and Trend Forecast of Building Industry VOCs Emissions in China

Abstract

Emissions of volatile organic compounds (VOCs) from the building industry in China are increasing. Predicting future trends in China’s building industry VOCs will help the implementation of China’s construction VOCs emission reduction policy. The aim of this study is to combine Granger causality analysis, Ridge regression, GM(1,N), and categorical boosting (CatBoost) methods for the analysis of factors influencing and trend prediction of VOCs emissions from building industry in China. Firstly, the spatial and temporal characteristics of building industry VOCs emissions in China were calculated, based on building industry VOCs emissions data from 2006 to 2020 for each province and city in China. Secondly, Granger causality tests and STIRPAT models were used to examine the influencing factors of construction VOCs, combined with ridge regression to estimate the elasticity coefficients of the influencing factors and feature screening, which were used as input features for prediction. Finally, a combination of the improved GM(1,N) model and the CatBoost model was used to predict and compare the results with those of the GM(1,N) model, the support vector regression model (SVR), the random forest (RF), and the CatBoost model. The results show that the combined model with improved GM(1,N) and CatBoost has better prediction accuracy than the other models. China’s building industry VOCs emissions are increasing year on year and are not expected to reach their peak by 2030. The size of the population, the number of people employed in the building industry, and the area of housing floor space under construction are important influencing factors that affect VOCs emissions from building industry in China. Based on the predicted results for the different scenario settings, building industry VOCs emissions are lower in the short term for the baseline scenario and in the long term for the high-speed scenario.

1. Introduction

VOCs are very important trace components of the atmospheric troposphere and play an extremely important role in atmospheric chemistry [1]. VOCs have a significant impact on secondary organic pollution formation, the oxidative capacity of the atmosphere, and human health [2,3]. As a pillar of China’s national economy, the problem of high energy consumption accompanied by high pollution is also very serious in the construction industry [4]. Currently, China ranks first in the world in terms of VOCs emissions, with an annual production of over 20 million tonnes. It is estimated that among the anthropogenic sources of VOCs emissions, the paint industry accounts for about 12% of VOCs emissions and is a key industry sector of concern for energy efficiency and emission reduction in the construction industry [5]. VOCs emissions from the building industry are mainly concentrated in building decoration, painting and wood processing. Architectural coatings, as the mainstay of China’s decorative materials, have a strong market demand and the total production has ranked among the top in the world, having become the world’s largest consumer of coatings [6]. VOCs generated during the use of architectural coatings are a major source of urban VOCs and are a key area of current research in the reduction of VOCs in the building industry [7]. Emissions of volatile organic compounds from building materials are one of the main sources of indoor air pollution, negatively affecting comfort, health and productivity [8,9,10]. Building renovation is one of the most important sources of anthropogenic VOC emissions in China [11]. Decoration materials, furnishings, and household appliances in residential buildings release large amounts of VOCs, such as formaldehyde and toluene [12,13].

In order to better manage the impacts caused by air pollutants, more and more machine learning and deep learning methods are applied to air pollutant concentration prediction, such as long and short-term memory networks (LSTM) [14,15], multilayer perceptrons (MLP) [16,17], multiple linear regression [18,19], Light gradient boosting machine (LightGBM) [20,21], and extreme gradient boosting (XGBoost) [22]. These models offer better performance in predicting atmospheric pollutants than traditional statistical models, capturing the non-linear characteristics of the data. The Catboost algorithm has been successfully applied to NO₂ concentration prediction, benzene concentration prediction [23]. The Catboost algorithm applied to air pollution is less studied. Ding et al. used the CatBoost method of wavelet decomposition to estimate satellite high resolution PM_2.5 in Beijing, Tianjin, and Hebei [24]. Zhang et al. studied the formation of haze in cities around Beijing based on MF-DCCA and boosting algorithms [25]. Shahria et al. evaluated the performance of two hybrid models (ARIMA-ANN and ARIMA-SVM) and two tree-based computational models (decision tree and CatBoost) for predicting daily PM_2.5 concentrations in three air pollution hotspots using data from the Bangladesh study [26].

Grey models have better forecasting performance, especially for time series with sparse data, incomplete information, and uncertainty. The advantage of grey forecasting is that only a small sample size is required to fully resolve the uncertainty in the original data [27]. Guo et al. Predicting air quality in 18 cities in Henan Province using a compound cumulative grey model [28]. Zhou et al. proposed a new seasonal grey model for predicting PM_2.5 concentrations in the Yangtze River Delta region [29]. Hamzacebi et al. propose a grey prediction model to predict energy-related CO₂ emissions in Turkey [30]. Pai et al. used a seven-class first-order grey model to predict hourly PM₁₀ and PM_2.5 concentrations in Banqiao, Taiwan [31]. Yang introduced an advanced optimization algorithm and cycle prediction theory into the grey prediction model and established an effective multi-step PM₁₀ prediction model to achieve a reasonable prediction of PM₁₀ [32]. Shi et al. proposed a novel time-lag GM (1,N) model to predict the Beijing haze [33].

There are few studies on VOCs emission prediction and they are mainly focused on indoor VOCs concentration prediction studies. Zhang et al. proposed a deep learning domain LSTM model for predicting the concentration of volatile organic compounds in different indoor environments [34]. Simayi et al. addressed a comprehensive air quality control policy for 2013 to 2019 and projected future emissions for 2030 under two different scenarios [35]. Bachtiar et al. implemented the prediction of specific VOCs and BVOCs by using hybrid multilayer perceptrons (MLPs) and ANNs [36].

It has been found that there is little research on the influence and prediction of VOCs emissions from buildings, mainly focusing on indoor VOCs, and that many studies have stopped at examining the factors influencing VOCs emissions, without further examining the important factors as variables for predicting VOCs emissions. Existing air pollutant concentration prediction mainly focuses on PM_2.5, PM₁₀, and other air pollutants, and there are fewer studies related to VOCs prediction, and the studies on VOCs do not study the whole spatial and temporal changes and trend prediction. Current VOCs predictions are mainly based on long-term monitoring data, with less research on the prediction of small samples, while predictions are mainly made from the perspective of a single model, without considering the advantages of combining machine learning and other prediction models. This paper analyses the spatial and temporal characteristics of building industry VOCs emissions in China using ArcGIS based on building industry VOCs emissions data from 2006–2020 in Chinese provinces and cities. The STIRPAT model was constructed to screen the important factors influencing the emission of VOCs from building industry and used as input features for the prediction model. An improved combined prediction model of GM (1,N) and CatBoost was constructed to predict the trend of building industry VOCs under different contextual settings.

Compared to previous studies, this study has three main strengths.

(1)

Unlike previous studies that only considered a small number of influencing factors, we explored the influencing factors of building industry VOCs emissions in three dimensions: demographic, economic and technological, using Granger causality analysis and the STIRPAT model. In order to eliminate multicollinearity between factors, the influence of these factors on building industry VOCs emissions was estimated by means of ridge regression, which may provide some basis for developing policies on building industry VOCs emissions.

(2)

Analysis of the spatial and temporal variation of building industry VOCs emissions based on ArcGIS reveals the pattern of temporal and spatial variation in the Chinese region. The spatial pattern of building industry VOCs emissions can further demonstrate the economic and environmental interactions between cities, which provides a new reference for the coordinated development of building industry VOCs emissions in China’s provinces and cities.

(3)

A combined model based on the improved GM (1,N) and CatBoost predicted future building industry VOCs emissions in various provinces and cities was able to effectively solve the small sample, non-linear, and multi-factor data. Firstly, we set up different scenarios based on the actual situation in China and government policies, and then made projections of building industry VOCs emissions under different scenarios. Secondly, we used important factors as input features to the model to improve the accuracy of the predictions. Finally, based on the projections, some recommendations for policies to reduce VOCs emissions from building industry are presented.

2. Data and Method

2.1. Data Sources

In this paper, in order to calculate and forecast VOCs emissions from construction in China, VOCs emissions from the construction industry in each province and city were measured by the area of completed buildings in the China Construction Industry Statistical Yearbook (2006–2020). Economic indicators for the construction industry by province are obtained through the China Construction Industry Statistical Yearbook (2006–2020) years. We have selected 31 provinces, cities and autonomous regions in China for our study. Industrial data, economic data, population data related to VOCs emissions from building industry are obtained from the China Statistical Yearbook, statistical yearbooks of provinces, municipalities and autonomous regions, and the China Urban Statistical Yearbook. The data interval covered in this paper is from 2006 to 2020, with 2020 as the base year. The forecasts in this paper use data on the characteristics of the Chinese construction industry in general and on the development and operation activities of real estate development enterprises.

2.2. Emission Factor Method

Based on a comparison of the advantages and disadvantages of various measurement methods and their feasibility, the emission factor method was chosen to measure VOCs emissions from the construction industry. The definition of VOCs in the Chinese architectural coatings industry is consistent with the current EU definition, which defines VOCs in terms of boiling point, and refers to any organic compound with an initial boiling point below or equal to 250 °C at a standard pressure of 101.3 kPa. They mainly include aliphatic and aromatic alkanes, olefins, oxygenated hydrocarbons, halogenated hydrocarbons, and nitrogenous sulphur compounds, such as toluene, ethylbenzene, xylene, propylbenzene, allylbenzene, ethylene glycol, propylene glycol, cyclohexanone, ethyl acetate, dimethyl carbonate, and dimethylethanolamine. As the source of VOCs in the industry is mainly from the building and exterior decoration sectors, where paints account for a large proportion of the total, the data in this paper are calculated by calculating the total amount of paint used in the construction industry, both outside and inside the home, as the total amount of VOCs emitted from the building industry. National emissions of VOCs from architectural coatings are calculated using the emission factor method [37], which calculates the emissions of pollutants based on emission factors and corresponding activity levels, as shown in Equation (1).

$$E=A\times EF\times (1-ER/100)$$

(1)

where $E$ is VOCs emissions from construction (t), $A$ is the amount of paint consumed per unit of output (kg/m²), $EF$ is the emission factor (g/m²), and $ER$ is the abatement rate.

As architectural coatings have to be painted in open spaces, the VOCs produced in painting are basically unorganized emissions at present, and at the current level of emission control, it can be considered that there are no end-of-pipe control measures for VOCs. The calculation of VOCs emissions from architectural coatings can therefore be simplified to Equation (2).

$$E={\displaystyle \sum (P\times A\times EF)}$$

(2)

where $P$ is the area of the building coating (m²).

According to Mou and Wei’s study, the ratio of completed housing area, external wall area and internal wall area in China is about 1:2.5:0.7 every year, the paint spraying rate for internal and external walls is 250–300 g/m³ and 400–500 g/m³ [38,39]. Therefore, for the sake of uniformity, it is assumed that the standard for spraying exterior and interior walls in each province and city is 450 g/m³ and 250 g/m³. All interior wall paints are water-based, the ratio of water-based to solvent-based exterior wall paints is 18.5:81.5. By estimating, the total annual consumption of water-based and solvent-based paints in each province and city can be derived, and based on the usage multiplied by the emission factor of the paint, the total annual VOCs emission from the construction industry can be derived. Based on the calculated emission totals, the next data analysis and projections can be carried out. The emission factors mainly refer to the relevant standards in the Ministry of Environmental Protection’s “Guidelines for the Preparation of Atmospheric VOC Source Emission Inventories” [40], as shown in

Table 1. Emission factors for VOCs in the construction industry.

Types of Architectural Coatings	Water-Based Coatings for Building Facades	Solvent Coatings for Building Facades	Architectural Interior Coatings
Emission factors	120 g/kg	450 g/kg	120 g/kg

below.

2.3. CatBoost

The CatBoost algorithm is able to handle categorical features better than GBDT [41]. CatBoost improves Greedy TBS by adding a priori distribution terms, reducing the impact of noise and low frequency data on the data distribution. CatBoost uses a more efficient strategy to reduce overfitting and makes efficient use of data information by using the entire dataset for training.

Let $S=\left\{\left.(X_1,Y_1),(X_2,Y_2),\cdots ,(X_n,Y_n)\right\}\right.$ be the observation data set, where $X_i=\left(x_{i1},x_{i2},x_{im}\right)$ is an $m$-dimensional vector containing numerical and categorical features, and $Y_i$ is the marker value.

First, the CatBoost algorithm binarizes all numerical features: floating-point features, statistical information and one-hot coding are binarized using an oblivious tree as the base predictor.

Secondly, the categorical features are transformed into numerical features:

(1)

Randomly arrange the observations to generate multiple random sequences.

(2)

Given a sequence, replace the category with the average marker value from the training data set.

$$X_{ik}=\frac{{\displaystyle \sum _{j=1}^n[X_{jk}=X_{ik}]\times Y_j}}{{\displaystyle \sum _{j=1}^n[X_{jk}=X_{ik}]}}$$

(3)

where, if $X_{jk}=X_{ik}$, then $[X_{jk}=X_{ik}]=1$, otherwise 0, the same category value is placed before the given value in the ranking.

(3)

Let $\theta =({\theta }_1,{\theta }_2,\cdots {\theta }_n)$, convert the subtype eigenvalues to numerical values:

$$X_{{\theta }_P,k}=\frac{{\displaystyle \sum _{j=1}^{p-1}[X_{{\theta }_j,k}=X_{{\theta }_p,k}]Y_{\sigma j}+a\times p}}{{\displaystyle \sum _{j=1}^{p-1}[X_{{\theta }_j,k}=X_{{\theta }_p,k}]+a}}$$

(4)

Here, the a priori value $p$ is added with the parameter $a(a>0)$,and the a priori weights help to reduce the noise in the low frequency category.

Finally, when processing the combination of features, the CatBoost algorithm performs the combination with a greedy strategy:

(1)

The first split of the tree does not make any combinations.

(2)

In the second split, all combinations and categorical features that already exist in the current tree are combined with all categorical features in the dataset, and the combined values are instantly converted to numbers.

(3)

All splits selected in the tree are considered as classifications with two values and used in combination to generate a combination of numerical and categorical features.

2.4. Improving the Grey Model

2.4.1. Multi-Factor GM (1,N) Grey Prediction Model

Grey systems are “small sample” and “information-poor” uncertain systems where some information is known and some is not, and they are used as a mathematical method for solving systems with incomplete information [42]. The grey prediction model transforms the original data into a regular new series by weakening the randomness of the original data, and the predicted values are obtained by solving a linear differential equation of the first order and reducing it by a cumulative series. The multi-factor GM (1,N) grey prediction model was constructed as follows:

(1)

Construction of the original series of the grey prediction model. Assume that the original sequence is $[x_i^{(0)}(1),x_i^{(0)}(2),x_i^{(0)}(3),x_i^{(0)}(4),\cdots ,x_i^{(0)}(n)]$, where $i=1,2,3,\cdots ,n,x_i^{(0)}$ denotes the original sequence of variables $i$, and each variable contains $n$ elements.

(2)

The cumulative sequence is calculated. Let the 1-AGO sequence of $X_i^{(1)}(i=1,2,3,\cdots ,n)$ be $X_i^{(1)}$, where:

$$x_i^{(1)}(k)={\displaystyle \sum _{k=1}^n{x}_i^{(0)}(k)(i=1,2,3,\cdots ,n)}$$

(5)

(3)

Generate a sequence of immediate neighborhood means. Generate a sequence of immediate neighborhood means of $X_i^{(1)}$, $Z_i^{(1)}$, where:

$$Z_i^{(1)}(k)=\lambda X_i^{(1)}(k)+(1-\lambda )X_i^{(1)}(k-1)i=1,2,3,\cdots ,n,k=2,3,\cdots ,n$$

(6)

(4)

Calculation of grey equation parameters. According to grey theory, the GM (1,N) model can be constructed as:

$$x_1^{(0)}(k)+a{Z}_1^{(0)}(k)={\displaystyle \sum _{i=2}^n{b}_i{x}_i^{(0)}(k)}$$

(7)

where $a$ is the development factor, $b_i$ is the driver factor, and $b_i{x}_i^{(0)}(k)$ is the driver term.

$$\mathrm{B}=\left[\begin{array}{l}-Z^{(1)}(2)X_2^{(1)}(2)\cdots X_N^{(1)}(2)\\ -Z^{(1)}(3)X_2^{(1)}(3)\cdots X_N^{(1)}(3)\\ ⋮⋮\ddots ⋮\\ -Z^{(1)}(n)X_2^{(1)}(n)\cdots X_N^{(1)}(n)\end{array}\right],\mathrm{Y}=\left[\begin{array}{l}-X_1^{(0)}(2)\\ -X_1^{(0)}(3)\\ ⋮\\ -X_1^{(0)}(n)\end{array}\right],\beta =\left[\begin{array}{c}a\\ b_1\\ b_2\\ ⋮\\ b_N\end{array}\right]$$

(8)

The least squares parameter estimate gives $\beta ={\left({\mathbf{B}}^T\mathbf{B}\right)}^{-1}{\mathbf{B}}^T\mathbf{Y}$.

(5)

Calculate the predicted cumulative series. When the magnitude of change in $X_i^{(1)}(i=1,2,3,\cdots ,n)$ is small, the GM (1,N) approximation time corresponding equation is:

$${\widehat{X}}_1^{(1)}(k+1)=\left[{\widehat{X}}_1^{(1)}(k+1)-{\widehat{X}}_1^{(1)}(k)\right]$$

(9)

(6)

Calculation and correction of residuals. The difference between the predicted value and the actual value is made, and the resulting ${\epsilon }^0(k)$ is called the residual. Assume that the residual series ${\epsilon }^0=\left[{\epsilon }_1,{\epsilon }_2,\cdots ,{\epsilon }_n\right]$ of $X_1^{(0)}$, the GM(1,1) model is constructed for the residual series to obtain $\mathbf{P}={\left[{\alpha }_{\epsilon },{\beta }_{\epsilon }\right]}^T$, and the simulated value of ${\widehat{\epsilon }}^0(k+1)$ can be obtained from Equation (10).

$${\widehat{\epsilon }}^0(k)=X_1^{(1)}(k)$$

(10)

2.4.2. Markov GM (1,N) Grey Prediction Model

The GM (1,N) model predicts data trends by accumulating the original series to generate new, more regular series, and mining the underlying patterns of data change [43]. Multi-factor GM (1,N) grey forecasting models are typically used to forecast with small amounts of data and little data volatility.

Specifically, the Markov GM (1,N) grey forecasting model was constructed as follows:

(1)

State interval division. A scatter plot is made based on the relative errors predicted by the multi-factor GM (1,N) grey prediction model, with each interval in the scatter plot corresponding to a state, denoted as state $E_i$.

(2)

Transfer probability calculation. Let the number of transfers from state $E_i$ to state $E_j$ be $n_{ij}$ and the number of intervals be $N_i$. Then, the transfer probability can be expressed as:

$$P_{ij}=\left\{\begin{array}{l}\frac{n_{ij}}{N_i},N_i>0\\ 0,N_i=0\end{array}\right.$$

(11)

This results in a transfer matrix of:

$$\mathbf{P}=\left[\begin{array}{l}{P}_{11}{P}_{11}\cdots P_{1n}\\ P_{21}{P}_{21}\cdots P_{2n}\\ ⋮⋮\ddots ⋮\\ P_{n1}{P}_{n2}\cdots P_{nn}\end{array}\right]$$

(12)

where ${\displaystyle \sum _{t=1}^n{P}_{nt}=1,t\in \left[1,n\right]},P_{ij}\in \left[0,1\right],i,j=1,2,\cdots ,n$.

(3)

Calculation of forecast values.

Suppose a state is in state $E_i$ at a certain moment, and if row $k$ in the probability matrix satisfies $matx{P}_{ij}=P_{kl}$, there is a high probability that the next moment will switch from state $E_i$ to state $E_j$, at which point the prediction interval is $\left[E_{1i},E_{2i}\right]$, taking its middle value as the prediction value for the next moment.

$$\widehat{y}(n+1)=\frac{1}{2}\left(E_{1i},E_{2i}\right)$$

(13)

2.5. Construction of Granger’s Causality Model

In order to linearize the data trends and to eliminate to some extent the heteroskedasticity of the time series, the variables were logged separately in the empirical analysis [44].

(1)

First, the data are tested for smoothness. If the data are non-smooth, it means that the series contains a single product component and the data need to be differenced. The Phillips–Perron test is used in this paper and is modelled as follows:

$$\Delta y_t=\mu +p{y}_{t-1}+{\epsilon }_t$$

(14)

where $\mu$ and $p$ are parameters, ${\epsilon }_t$ is the white noise series, $y$ is the time series to be tested and the Newey–West procedure is used to adjust the standard errors.

(2)

Examine the cointegration relationship between variables. A cointegration test indicates whether there is a stable equilibrium relationship between variables. The Johanson maximum likelihood method used in this paper is a VAR system for testing the cointegration relationship between multiple variables using great likelihood estimation.

(3)

Based on the above model, Granger causality can be tested. The lags were determined from the results of LR, AIC, SC, FPE, and HQ tests, and the causal relationships between the variables were tested.

$$Y_t={\alpha }_0+{\displaystyle \sum _{i=1}^p{\alpha }_i{Y}_{t-i}}+{\displaystyle \sum _{j=1}^q{\beta }_j{X}_{t-j}}+{\xi }_t$$

(15)

The F statistic is used for Granger causality analysis and the null hypothesis for the F test is that X is not the cause of the change in Y. If the calculated value of the F statistic is greater than the critical value of F, the original hypothesis is rejected, indicating that X is the Granger cause of Y.

2.6. Construction of the STIRPAT Model and Interpretation of Indicators

The STIRPAT model retains the multiplicative structure of the “I = PAT” model, using three factors, i.e.—economic, technological, and demographic—as determinants of environmental change [45]. When using the data for econometric analysis, the model will be in logarithmic form to reduce heteroskedasticity, as follows:

$$\ln I=\ln a+b(\ln p)+c(\ln A)+d(\ln T)+\ln e$$

(16)

In order to provide a comprehensive picture of the impact of building industry VOCs emissions, population data, construction industry data, and real estate data were introduced into the model. The expressions are:

$$\begin{array}{l}\ln I=\ln a+b(\ln P_1)+c(\ln P_2)+d(\ln A_1)+e(\ln A_2)\\ +f(\ln I_1)+g(\ln I_2)+bs(\ln F)+ds(\ln G)+\ln e\end{array}$$

(17)

The variables in this equation are defined as: $I$ represents the environmental impact, expressed as total VOCs emissions; $P_1$ represents the total population size, expressed as total population; $P_2$ for construction employment; $A_1$ and $A_2$ represent the construction area and completed area of housing buildings for construction enterprises respectively. $I_1$ and $I_2$ represent the area of housing construction and completion of real estate development enterprises. $F$ denotes house construction cost. $G$ represents sales of commercial business premises. $a$ is the model coefficient, $b$, $c$, $d$, $e$, $f$, $g$, $bs$, and $ds$ are the respective variable indices, and $e$ is the error.

3. Results and Discussion

3.1. Spatial and Temporal Characteristics of VOCs Emissions from Building Industry in China

According to the data in Figure 1, it can be broadly seen that China’s VOCs emissions show a trend of high in the east and low in the west. This is due to the economic backwardness and smaller population in the west. Jiangsu and Zhejiang are two of China’s more rapidly developing economic provinces, and their VOCs emissions are also two of the highest. Tibet and Qinghai are two of the more economically backward provinces in China, and they also have two of the lowest VOCs emissions. In addition to the level of economic development, the climate and industrial structure are also important factors in the movement of people. Compared to the harsh climate in the west, the climate in the south-eastern and central regions is relatively suitable, and the industrial structure is mainly tertiary, which imposes less load on the environment and is the main reason for the movement of people.

The real estate sector is a competitive market, and it is predicated on the assumption that supply and demand population movements will inevitably lead to increased demand, which in turn will contribute to a boom in the local real estate business, leading to higher emissions in this part of the country than in other provinces. In addition, as the national economic level continues to rise, people’s pursuit of housing is also increasing, and the area of housing per capita is also rising year by year, which is another major factor contributing to the rise in VOCs emissions from the construction industry. VOCs emissions increased relatively smoothly from 2006 to 2009, increased at a faster rate from 2010 to 2014, and by and large showed a relatively stable state after 2014, which shows that China’s emission reduction policies are beginning to bear fruit. Among them, Jiangsu Province has the largest VOCs emissions from the construction industry, followed by Zhejiang Province and Shandong Province. Tibet has the lowest emissions. Emissions of VOCs from the construction industry in each province are basically positively correlated with the degree of development of its construction industry, with some provinces that are relatively less developed, such as Tibet, Qinghai, Hainan, and Ningxia, where emissions are relatively low. According to the geographical map above, it can be seen that emissions in China show a concentration of high emission areas. Moreover, the high emission areas are mainly concentrated in the east and central part of the country. This is due to the fact that there are more coastal provinces in the east, and the economic development of coastal provinces is generally faster than that of inland areas. The central region is located inland and is a transportation hub. The development of transportation will also drive the rapid development of the surrounding economy, and the population settled in the central region is also rising year by year, leading to a high demand for real estate. High emission areas are Jiangsu and Zhejiang. Medium to high emission areas are Shandong, Hubei, Sichuan, Henan, Hunan, and Guangdong. Medium to low emission areas are Anhui, Fujian, Liaoning, Chongqing, Jiangxi, Hebei, Beijing, and Shanghai. The geographic locations where the construction industry is well developed are the east and centre and, to a lesser extent, Liaoning, which covers the north-east.

Through the spatial and temporal distribution of VOCs emissions in 31 provinces in China from 2006 to 2020, it was found that China showed a pattern of significant growth in VOCs emissions between 2006 and 2014, but the rise after 2014 was relatively small and even showed a downward trend. Jiangsu and Zhejiang, two of China’s more economically developed provinces, are currently the two highest emitters, about twice as much as the second tier Shandong and Hubei, with the western region being the area where low emissions are clustered. China’s construction industry VOCs emissions show the phenomenon of regional aggregation, with high aggregation areas concentrated in the eastern coast and some central areas. China is beginning to see results in reducing emissions of VOCs from building industry, with a trend of slow growth and slight decline in building industry VOCs emissions. By analyzing the spatial characteristics of VOCs emissions in 31 provinces in China, construction VOCs emissions are mainly concentrated in the more economically developed regions, with relatively few in the northwest and northern regions where economic development is more backward.

3.2. Analysis of Factors Influencing the Emission of VOCs from Building Industry Based on Granger and STIRPAT Models

The main measures used in this paper are Granger causality and ridge regression. Granger causality is the inclusion of the lagged value of X in the regression of other variables to test whether it significantly improves the prediction of Y. If it does, then X is considered to be the Granger cause of Y. As there is often no direct correlation between some indicators in the socio-economy and building industry VOCs emissions, most of the existing studies have assumed a link between these factors and building industry VOCs emissions. The use of Granger causality, in terms of probability or distribution functions, can be used to help determine which demographic factors are statistically responsible for the STIRPAT model is a widely used and very well-established model for evaluating environmental stress, reflecting the impact of socio-economic factors on environmental stress. However, Granger causality does not give a numerical value for the influence of each factor, so the influence of each factor needs to be measured again using the STIRPAT model. This paper uses VOCs as a measure of environmental pressure and analyses the impact of human drivers, such as population size, construction development, and the development and operation activities of property development enterprises on environmental pressure based on a time-series data perspective.

3.2.1. ADF Test

If the two time series are non-stationary and satisfy the same order of single integer, a cointegration test between the two given series is required before the Granger causality test can be performed after the cointegration relationship between the series exists. Granger causality tests are sometimes sensitive to the choice of lag length, and different lags may give completely different test results.

Table 2. ADF test results (***, **, * represent 1%, 5%, 10% level of significance respectively).

Variables	t-Statistic	p-Value	Threshold Values
			1%	5%	10%
VOCs	−2.252	0.188	−4.069	−3.127	−2.702
Population	−2.03	0.273	−4.223	−3.189	−2.73
Commercial business premises	−5.552	0.000 ***	−4.223	−3.189	−2.73
Area of commercial business premises for sale of commercial properties	−1.96	0.305	−4.138	−3.155	−2.714
Other aspects of commercial property sales	2.551	0.999	−4.473	−3.29	−2.772
Sale of office space for commercial properties	2.461	0.999	−4.473	−3.29	−2.772
Residential area of commercial property sales	0.075	0.964	−4.473	−3.29	−2.772
Net value of owned construction machinery and equipment at end of year	−4.368	0.000 ***	−4.473	−3.29	−2.772
Sales area of commercial properties	0.105	0.966	−4.473	−3.29	−2.772
Technical equipment rate	−2.017	0.279	−4.069	−3.127	−2.702
Power Equipment Rate	−2.427	0.134	−4.012	−3.104	−2.691
Total power of own construction machinery and equipment	−1.882	0.341	−4.012	−3.104	−2.691
Total number of construction machinery and equipment owned	−2.558	0.102	−4.069	−3.127	−2.702
Residential area	−0.667	0.855	−4.473	−3.29	−2.772
Employment in construction	−1.426	0.570	−4.012	−3.104	−2.691
Other aspects of the construction industry	−0.578	0.876	−4.473	−3.29	−2.772
Total number of residential completions	−5.919	0.000 ***	−4.473	−3.29	−2.772
Office Building	−4.1	0.001 ***	−4.332	−3.233	−2.749
House construction cost	−1.581	0.493	−4.332	−3.233	−2.749
New housing construction	−0.807	0.817	−4.473	−3.29	−2.772
Gross construction output	−3.061	0.030 **	−4.473	−3.29	−2.772
House floor area construction area	−1.227	0.662	−4.012	−3.104	−2.691
Value of housing completions	−1.648	0.458	−4.012	−3.104	−2.691
House floor area completed	−2.278	0.179	−4.069	−3.127	−2.702
Housing floor area completion rate	−3.057	0.030 **	−4.012	−3.104	−2.691
Area of completed houses	−2.52	0.111	−4.473	−3.29	−2.772
House construction area	−2.764	0.064 *	−4.332	−3.233	−2.749

shows the results of the ADF test, including variables, t-test results, and AIC values, for testing whether the time series is smooth. If p < 0.05, it does not show significance at the level, the original hypothesis cannot be rejected, and the series is a non-stationary time series. If p ≥ 0.05, it shows significance and the original hypothesis is rejected, i.e., the series is a smooth time series.

3.2.2. Granger Causality Test

The results of the Granger causality test were used to obtain whether the causal relationship between the two variables was unidirectional or bidirectional. A lag period test based on the VAR model was first performed and a lag period of 2 was determined for all tested indicators and the results are shown in

Table 3. Results of Granger’s causality test (***, **, * represent 1%, 5%, 10% level of significance respectively).

Matching Samples		F-Statistic	p-Value
Population	VOCs	8.365	0.015 **
Commercial business premises		25.348	0.000 ***
Area of commercial business premises for sale of commercial properties		1.291	0.280
Other aspects of commercial property sales		3.558	0.086 *
Sale of office space for commercial properties		3.776	0.078 *
Residential area of commercial property sales		1.057	0.326
Net value of owned construction machinery and equipment at end of year		3.011	0.111
Sales area of commercial properties		1.328	0.274
Technical equipment rate		0.86	0.374
Power Equipment Rate		1.759	0.212
Total power of own construction machinery and equipment		0.533	0.481
Total number of construction machinery and equipment owned		1.229	0.291
Residential area		1.086	0.320
Employment in construction		10.718	0.007 ***
Other aspects of the construction industry		0.08	0.783
Total number of residential completions		3.356	0.094 *
Office Building		3.367	0.094 *
House construction cost		5.11	0.045 **
New housing construction		3.39	0.093 *
Gross construction output		0	0.993
House floor area construction area		5.729	0.036 **
Value of housing completions		0.835	0.381
House floor area completed		35.579	0.000 ***
Housing floor area completion rate		0.003	0.960
Area of completed houses		17.347	0.002 ***
House construction area		5.922	0.033 **

. Analysis of the significance of the F-statistic, if significant (p < 0.05), indicates that the original hypothesis (one set of time series is not the cause of another set of time series) is rejected and the left-hand side variables can cause changes in the right-hand side variables with Granger causality, and vice versa, there is no Granger causality. The Granger test only verifies whether there is a statistical causal relationship between the factors and can be used as a basis for determining whether there is a statistical causal relationship between the two parties, but it does not indicate a logical causal relationship.

3.2.3. Ridge Regression

We used the factors with a significance level of 5% or less in the Greeninger analysis for the ridge regression model construction. The results of the ridge regression are shown in

Table 4. Ridge regression results (***, ** represent 1%, 5% level of significance respectively).

K = 0.113	Non-Standardized Coefficients		Standardisation Factor	t	p	R2	Adjustment R2	F
	B	Standard Error	Beta
Constants	−761391.53	183139.107	-	−4.157	0.014 **	0.998	0.993	223.437 (0.000 ***)
Population	6.278	1.414	0.089	4.441	0.011 **
House floor area construction area	0.064	0.006	0.123	11.154	0.000 ***
House floor area completed	0.301	0.021	0.152	14.642	0.000 ***
House construction area	0.094	0.007	0.115	13.249	0.000 ***
Area of completed houses	1.493	0.253	0.146	5.891	0.004 ***
House construction cost	36.221	6.114	0.112	5.924	0.004 ***
Commercial business premises	3.643	0.804	0.12	4.529	0.011 **
Employment in construction	35.191	6.75	0.153	5.213	0.006 ***
Dependent variable: VOCs

. With the regression model significance p value of 0.000 *** for population, housing floor area construction area and completed area, housing construction area and completed area of real estate development enterprises, housing cost, commercial business premises, and employment in the construction industry showed significance, rejecting the original hypothesis and indicating that there is a regression relationship. Meanwhile, the model’s goodness of fit R² is 0.998, which is a relatively good performance of the model, so the model basically meets the requirements. The K value is chosen as the minimum value at which the standardized regression coefficients of the individual independent variables tend to stabilize. In general, the smaller the K value, the smaller the deviation. K = 0.125 determined by the variance expansion factor method. To determine whether the model is significant (p < 0.01 or 0.05) the F values are analyzed, which, if significant, indicate that there is a regression relationship between. The p value represents the significance of X. It is used to explore whether the relationship between the effect of X on Y shows significance (p value less than 0.05)

Formula for the model: VOCs = −761391.53 + 6.278 × population + 0.064 × house construction area + 0.301 × house construction area completed + 0.094 × house construction area of real estate development enterprises + 1.493 × house construction area of real estate development enterprises completed + 36.221 × house construction cost + 3.643 × commercial business premises + 35.191 × number of people employed in the construction industry.

3.3. Predicted Results of Building Industry VOCs Emissions in Different Scenarios

3.3.1. Influencing Factor Setting

Based on the results of the analysis of the factors influencing VOCs emissions, eight factors from the ridge regression were selected as input features for this paper, and the input features were fed into a combined model of an improved grey model and a genetic algorithm to optimize the CatBoost algorithm.

Three scenarios are set up in this paper, namely the baseline scenario, the medium rate scenario and the high rate scenario. The baseline scenario is set according to the planning proposed by government departments, taking into account the VOCs emission situation of each province in China from 2006 to 2019, and the existing changes of each influencing factor.

In terms of the size of the resident population, according to the results of the 7th National Census, the population of China will be 1.402 billion in 2020. According to the National Population Development Plan (2016–2030), the total population will peak around 2030, reaching a total national population of around 1.45 billion by 2030. The average population growth rate in China from 2006–2019 is 0.49%. The average growth rate of gross construction product from 2006 to 2019 was 13.64%. In terms of VOCs emissions, considering that the calculation of VOCs emission intensity is influenced by the total construction output and based on the historical VOCs emission intensity from 2006 to 2019, the population size, gross construction product, and construction VOCs emission intensity for the three scenarios proposed in this paper are shown in

Table 5. Setting the number of permanent residents in China, 2020–2030.

Year	Resident Population (Billion People)	Gross Regional Product (Ten Thousand Yuan)	VOCs Emissions (t)
2020	14.02	72995.7	811981.5
2021	14.09	82952.31348	807231.4
2022	14.16	94267.00904	802509.1
2023	14.23	107125.0291	797814.4
2024	14.30	121736.883	793147.2
2025	14.37	138341.7939	788507.3
2026	14.44	157211.6146	783894.5
2027	14.51	178655.2788	779308.7
2028	14.58	203023.8588	774749.8
2029	14.65	230716.3132	770217.5
2030	14.72	262186.0183	765711.7

.

During the “13th Five-Year Plan” period, the national construction industry’s added value will grow at an average annual rate of 5.1%, with an average growth rate of around 7.8% in the gross construction product from 2016 to 2020 and around 11.0% in 2021. In 2021, the construction area of housing buildings in the national construction industry will be 15.75 billion square meters, an increase of 5.4% year-on-year. In 2016–2020, the national construction area of housing buildings in the national construction industry will grow at an average rate of around 3.82%. The medium and high growth rates have been revised upwards by 0.5% and 1% respectively. The average reduction in housing construction completions in the national construction industry from 2016 to 2020 is 1.76%, with a 0.5% and 1% downward adjustment for medium and high growth respectively. In 2016–2020, with a real estate housing floor area and housing construction area growth rate of 4.75%, medium and high-speed development were revised upwards by 0.5% and 1% respectively. In 2016–2020, with real estate housing floor area housing completions down 1.76%, medium and high-speed development were down 0.5% and 1% respectively. The growth rate of housing construction was 4.43%, a downward revision of 0.5% and 1% for medium and high growth respectively. In 2016–2020, real estate commercial business premises p were down 4.34%, with medium and high growth down 0.5% and 1% respectively. Construction employment declines by an average of 4.88% between 2016 and 2020, with a downward revision of 0.5% and 1% for medium and high growth respectively. The influencing factors in different contexts are set out in

Table 6. Rules for setting influencing factors in different contexts.

Category	Baseline Scenario	Medium Speed Scenario	High Speed Scenario
Construction Business	National construction value added to increase by 5.1% per year from 2020–2030. The construction area of housing buildings in the national construction industry will increase by 3.82% per year from 2020 to 2030. National construction housing construction completion area decreases by 1.76% per year from 2020–2030. Construction employment decreases by 4.88% per annum from 2020–2030.	National construction value added to increase by 7.8% per year from 2020–2030. The construction area of housing buildings in the national construction industry will increase by 4.32% per year from 2020 to 2030. National construction housing construction completion area decreases by 2.26% per year from 2020–2030. Construction employment decreases by 5.38% per annum from 2020–2030.	National construction value added to increase by 5.1% per year from 2020–2030. The construction area of housing buildings in the national construction industry will increase by 4.82% per year from 2020 to 2030. National construction housing construction completion area decreases by 2.76% per year from 2020–2030. Construction employment decreases by 5.88% per annum from 2020–2030.
Real estate	The housing construction area of real estate development enterprises will grow at an annual rate of 4.75% from 2020 to 2030. The area of housing completed by property development enterprises fell at a rate of 1.76%. Housing cost growth rate of 4.43%. Real estate commercial business premises decreases at a rate of 4.34% from 2020–2030.	The housing construction area of real estate development enterprises will grow at an annual rate of 5.25% from 2020 to 2030. The area of housing completed by property development enterprises fell at a rate of 2.26%. Housing cost growth rate of 4.93%. Real estate commercial business premises decreases at a rate of 4.84% from 2020–2030.	The housing construction area of real estate development enterprises will grow at an annual rate of 5.75% from 2020 to 2030. The area of housing completed by property development enterprises fell at a rate of 2.76%. Housing cost growth rate of 5.43%. Real estate commercial business premises decreases at a rate of 5.34% from 2020–2030.

.

3.3.2. Predicted Results

The improved GM(1,N) model, support vector regression model, random forest, Gradient Boosting Decision Tree (GBDT), and the combined GM(1,N) and CatBoost models were used to forecast China’s construction emissions from 2014 to 2019, and the forecast results are shown in Figure 2. The SVR model performs the worst, the random forest and GM(1,N) models underestimate all true values by a large margin, the CatBoost model is a poor estimator of smaller true values, and CatBoost-GM(1,N) performs better overall and is closer to the true values. The evaluation metrics are shown in Figure 3. We calculated the mean error of the models, as shown in Figure 3. The mean MAPE value for the RF model from 2014 to 2019 was 3.71, CatBoost-GM(1,N) mean MAPE value was 0.89, CatBoost mean MAPE value was 2.83, SVR mean MAPE value was 7.57, and GM(1,N) mean MAPE value was 8.01. The CatBoost-GM(1,N) combined model has the smallest value of MAPE. The prediction results and evaluation metrics show that the combined model of improved GM(1,N) and CatBoost predicts results closer the true value. The value of MAPE for the combined model was the smallest. From the above results, it is clear that the combined model based on the improved GM(1,N) and CatBoost algorithms works best.

In this paper, the above set influencing factors are brought into the constructed combined prediction model and combined with historical data to predict the building industry VOCs emissions in different scenarios from 2020 to 2030, and the prediction results are shown in Figure 4. Emissions of VOCs from buildings in 2021 are lowest in the baseline scenario and second lowest in the medium rate scenario. Building industry VOCs emissions in 2025 are lowest in the high speed scenario and second lowest in the medium speed scenario. However, from 2026 onwards, building industry VOCs emissions are lowest in the high speed scenario, second highest in the baseline scenario and highest in the medium speed scenario. Therefore, in the short term, the baseline scenario has the lowest building industry VOCs emissions of the three scenarios, but in the long term, the medium rate scenario has the lowest building industry VOCs emissions. China’s building industry VOCs emissions are on a year-on-year upward trend in all three different scenarios, suggesting that China’s building industry VOCs emissions do not reach a peak until 2030.

In the long run, building industry VOCs emissions are lower and perform better in the high-speed scenario, so this paper considers the prediction of building industry VOCs emissions for each province and city in China in the high-speed scenario, as shown in Figure 5. The forecast trend of construction VOCs by Chinese provinces and cities from 2020–2030 shows that: ① Jiangsu, Shandong and Hubei are still the three provinces with the highest construction VOCs emissions in the next ten years, and the gap between their construction VOCs emissions and those of other provinces is still gradually widening. The three provinces of Jiangsu, Shandong, and Hubei are not expected to reach peak emissions of construction VOCs by 2030 and are still in growth. ② The emissions of construction VOCs from other provinces are divided into two echelons, with the first echelon emitting significantly more construction VOCs than the second echelon. The provinces and cities in the first echelon include Zhejiang, Sichuan, Hunan, Henan, Fujian, Jiangxi, Guangdong, and Anhui, while those in the second echelon include Chongqing, Hebei, Beijing, Shanghai, Guangxi, Yunnan, and Shaanxi. Provinces and cities in the third echelon include Jilin, Liaoning, Xinjiang, Heilongjiang, Inner Mongolia, Shaanxi, Gansu, Tianjin, Guizhou, Ningxia, Tibet, Qinghai, and Hainan. ③ In the first echelon, Zhejiang has the highest and decreasing construction VOCs emissions year on year. Sichuan and Hunan also have increasing construction VOCs emissions year on year and are expected to overtake Zhejiang by 2030. Emissions of VOCs from construction in Henan and Guangdong, although increasing year on year, have slowed down significantly and stabilized. ④ In the second echelon, Chongqing has the highest and year-on-year increase in construction VOCs emissions, but the growth is slowing. Guangxi and Shaanxi have the fastest growing building industry VOCs emissions, with these two provinces expected to outpace a large number of cities in the second echelon and second only to Chongqing in building industry VOCs by 2027. Hebei and Beijing’s emissions of building industry VOCs will be comparable to Guangxi’s in 2020 and will be less than Guangxi’s building industry VOCs by 2030. ⑤ The third echelon in which Hebei is growing at an extremely slow rate. Emissions of construction VOCs in both Gansu and Xinjiang show a year-on-year increase, but the increase is not significant. Tianjin and Ningxia’s carbon emissions have been decreasing year on year, with a relatively steady reduction.

4. Conclusions and Policy Implications

This paper begins with the estimation of building industry VOCs emission data from 2006 to 2019 for each province and city in China. The STIRPAT model for building industry VOCs emission data was then constructed with the help of building industry VOCs emission data from three dimensions: demographic, economic, and technological, and the influencing factors were screened according to the elasticity coefficients, and the screened influencing factors were used as the influencing factors for prediction. Finally, different carbon emission scenarios were constructed based on relevant policies and available data, and a combination of the improved GM(1,N) model and the Catboost model was used to forecast the carbon emissions of Chinese provinces and cities over the next ten years. At the same time, the temporal and spatial characteristics of VOCs emissions from carbon emitting buildings were analyzed for each province and city in China.

The main findings of the study are as follows:

(1)

Emissions of VOCs from buildings are increasing year on year and are not expected to peak by 2030. Jiangsu and Shandong are the two provinces with the highest emissions of construction VOCs in China and are maintaining a high growth rate, which has a profound impact on the total emissions of construction VOCs in China, and this pattern is expected to remain unchanged until 2030. Other provinces and cities emit far less construction VOCs than the two provinces of Jiangsu and Shandong, and it is expected that by 2030 Sichuan, Henan, Hunan, and Zhejiang will be the four provinces with the largest construction VOCs emissions after Jiangsu and Shandong.

(2)

The magnitude of the absolute values of the elasticity coefficients of the factors influencing total construction VOCs emissions in China were ranked, with housing cost > construction employment > population > commercial business premises > housing completion area > housing floor area completion area > housing floor area construction area > housing construction area. A larger elasticity coefficient indicates a more significant effect on emissions.

(3)

There are significant differences in the predicted results of building industry VOCs emissions under different scenario settings. In the short term (2020–2024), the baseline scenario has the lowest building industry VOCs emissions of the three scenarios. In the long term (2025–2030), building industry VOCs emissions are lower in the baseline and high-speed scenarios, with the lowest building industry VOCs emissions in the high-speed scenario in particular. It can be seen that the industrial and demographic structure has a clear impact on building industry VOCs emissions.

China’s current policies to reduce emissions of VOCs are initially effective, but have had little effect. Therefore, it is necessary for each province and municipality to tailor their emission reduction policies to their own development status and geographical location. Meanwhile, we must control the key drivers of VOCs emissions, such as the cost of housing and the number of people employed in the construction industry. By controlling key indicators, we can effectively control the incremental emissions of VOCs and ultimately achieve the goal of reducing emissions. In terms of building materials, we can choose cleaner and more environmentally friendly energy sources, increase research and development on environmentally friendly materials, and expand the range of architectural paints, which will also curb the emission of VOCs from the construction industry.