Multi-Province Collaborative Carbon Emission Forecasting and Scenario Analysis Based on the Spatio-Temporal Attention Mechanism—Empowering the Green and Low-Carbon Transition of the Transportation Sector Through Technological Innovation

Abstract

As one of the primary contributors to carbon emissions in China, the transportation sector plays a pivotal role in achieving green and low-carbon development. Considering the spatio-temporal dependency characteristics of transportation carbon emissions driven by economic interactions and population mobility among provinces, this study proposes a predictive framework for transportation carbon emissions based on a spatio-temporal attention mechanism from the perspective of multi-province spatio-temporal synergy. First, the study conducts transportation carbon emission accounting by considering both transportation fuel consumption and electricity usage, followed by feature selection using an enhanced STIRPAT model. Second, it integrates the spatio-temporal attention mechanism with graph convolutional neural networks to construct a multi-province transportation carbon emission collaborative prediction model. Comparative experiments highlight the superior performance of deep learning methods and spatio-temporal correlation modeling in multi-province transportation carbon emission collaborative prediction. Finally, three future development scenarios are designed to analyze the evolution paths of transportation carbon emissions. The results indicate that technological innovation can significantly improve the efficiency of transportation emission reduction. Moreover, given that the eastern region and the central and western regions are at distinct stages of development, it is essential to develop differentiated emission reduction strategies tailored to local conditions to facilitate a green and low-carbon transformation in the transportation sector.

1. Introduction

With the ongoing intensification of the global greenhouse effect and the increasing frequency of extreme climate events, the sustainable development of human society is encountering significant challenges. In this context, the 2 °C temperature control target established by the Paris Agreement in 2016 has become a widely recognized consensus for global climate governance. The contracting parties have committed to collaboratively constructing a carbon-neutral future. As the world’s largest developing country, China, guided by its ecological civilization strategy, has pledged to achieve carbon peaking before 2030 and carbon neutrality before 2060. It is systematically advancing the green transformation of all industries and the profound adjustment of the energy structure. Notably, in 2021, carbon emissions from China’s transportation sector accounted for 9.1% of total emissions, reaching approximately 970 million tons [1]. This makes the transportation sector one of the fastest-growing contributors to carbon emissions in China. Against the backdrop of building a “mobile China,” it faces the profound contradiction between surging demand and stringent emission reduction constraints. Under these circumstances, researching a collaborative prediction model for cross-regional carbon emissions, revealing differences in emission reduction efficiency among various policy tools through scenario analysis, and exploring scientifically feasible pathways to reduce transportation carbon emissions are of critical importance for China to achieve its emission reduction targets.

Accurate prediction and scenario analysis of carbon emissions in the transportation sector are essential for achieving the “dual carbon” goals. Existing studies primarily utilize methods such as the environmental kuznets curve (EKC) [2,3,4], gray prediction [5,6], system dynamics modeling [7,8], STIRPAT model [9,10], and deep learning techniques [11,12] to focus on predicting transportation carbon emissions at either the national or single regional scale. For example, one study developed a transportation carbon emission prediction model using deep neural networks and integrated eight scenarios to analyze the future trends of China’s transportation carbon emissions, revealing that systematic technological innovation is critical for promoting low-carbon and sustainable transportation development in China [12]. Another study performed a scenario analysis of county-level transportation carbon emissions from a full life cycle perspective, highlighting the substantial emission reduction potential of energy-saving scenarios [13]. Notably, regional transportation carbon emissions in China exhibit spatial heterogeneity: prior research has demonstrated that China’s transportation carbon emission intensity and spatial correlation network display significant imbalance [14], while urban transportation emission reduction effects vary considerably across different cities [15]. Furthermore, transportation carbon emissions among provinces or regions exhibit certain spatio-temporal dependencies. Studies indicate that the spatial dependence of carbon emissions is influenced not only by geographically adjacent areas but also by regions with bilateral economic ties [2]. However, most existing studies treat analysis units as independent entities, with limited exploration into the dynamic evolution mechanisms of multi-province transportation carbon emissions from a multi-province spatio-temporal coordination perspective. Therefore, to better capture and model the complex spatio-temporal characteristics of multi-province transportation carbon emissions, this study proposes a prediction model based on the coupling of spatio-temporal features, aiming to achieve precise predictions of multi-province transportation carbon emissions and provide technical support for identifying effective pathways to reduce transportation-related carbon emissions.

As an extension of the attention mechanism in spatio-temporal data analysis, the spatio-temporal attention mechanism encompasses two dimensions: spatial attention and temporal attention. Temporal attention is capable of identifying variations in influence across different time points and characterizing the nonlinear features of event evolution. Spatial attention effectively captures inter-regional associations between geographical entities or network nodes, uncovering the underlying interaction patterns among different regions. Compared to traditional models, the spatio-temporal attention mechanism exhibits significant advantages in addressing various complex spatio-temporal problems, attributed to its adaptive feature extraction capability, multi-scale spatio-temporal information fusion framework, and interpretable attention weight allocation mechanism [16,17,18].To address the prevalent cross-regional spatio-temporal dependency characteristics in multi-province traffic carbon emission prediction, this study proposes directly modeling the temporal and spatial dimension features of multi-province transportation carbon emission. Specifically, the graph convolutional neural network (GCN) will be employed to extract multi-scale spatial features from the multi-province traffic road network graph [19]. These extracted features will then be integrated into the time-aware attention mechanism and the spatial-enhanced attention mechanism proposed in this study, enabling collaborative modeling of spatio-temporal dependencies within the multi-province system. This approach provides a novel technical pathway for achieving precise and coordinated prediction of multi-province transportation carbon emission.

It is important to recognize that addressing spatio-temporal dependency issues is not exclusive to this study. In fact, models that can simultaneously capture temporal autocorrelation and spatial interaction effects have emerged as essential analytical tools in the study of complex systems characterized by interregional connectivity. Research incorporating spatio-temporal features has demonstrated significant success across various domains, including traffic flow forecasting [20,21,22], wind speed prediction [23,24,25], and network traffic estimation [26,27,28]. Accordingly, this study innovatively extends the spatio-temporal modeling framework to the specific context of collaborative traffic carbon emission prediction, thereby expanding its application scope.

In summary, this study builds upon existing research by integrating a spatio-temporal attention mechanism and scenario analysis methods to investigate collaborative prediction and path optimization of transportation carbon emission in multi-province systems. This provides a scientific foundation for formulating cross-scale, collaborative low-carbon transportation development strategies. The specific research contents are as follows: First, during the process of traffic carbon emission accounting, given the rapid adoption of new energy vehicles and the increasing electricity consumption in the transportation sector, carbon emissions from electricity consumption are incorporated into the traffic carbon emission calculation framework. Second, based on the extended STIRPAT model, feature selection is performed, identifying ten key factors as the primary influences on transportation carbon emission. Third, considering the complex spatio-temporal dependency characteristics of transportation carbon emission in multi-province systems, this study develops a multi-province collaborative prediction model for transportation carbon emissions based on a spatio-temporal attention mechanism (MC-STAPM) [29,30]. Finally, three typical development pathways—the base scenario, technological innovation scenario, and regional economic differentiation scenario—are designed to forecast future transportation carbon emission in each province under different scenarios and to propose multi-scale development recommendations for achieving China’s traffic carbon emission reduction targets.

The structure of this paper is organized as follows: Section 2 details the data sources, the methodology for calculating transportation carbon emissions, and the feature selection approach. Section 3 elaborates on the architecture and construction methodology of the MC-STAPM model. Section 4 provides an analysis of the prediction results for transportation carbon emissions across multiple provinces, along with the outcomes of scenario analyses. Section 5 summarizes the conclusions drawn from this study.

2. Transportation Carbon Emission Measurement and Feature Selection

2.1. Data Sources

According to the data published in the “China Energy Statistical Yearbook” (2001–2022), China’s transportation sector consumes as many as 21 different types of energy. Among these, ten types of energy—namely raw coal, washed and refined coal, coke, gasoline, kerosene, diesel, fuel oil, liquefied petroleum gas, natural gas, and liquefied natural gas—account for the majority of total energy consumption within the transportation industry. The carbon emission prediction process outlined in this article incorporates both transportation carbon emission measurement data and provincial feature data. In calculating the annual carbon emissions of China’s transportation sector, key parameters such as the average lower calorific value of various energy sources, carbon content per unit of calorific value, carbon oxidation rates, and terminal consumption volumes are required. These parameters are sourced from authoritative references: the average lower calorific value and carbon content per unit calorific value are derived from the “General Principles for Calculation of Comprehensive Energy Consumption” (GB/T2589-2020 [31]) and the “Guidelines for the Compilation of Provincial Greenhouse Gas Inventories,” respectively. The carbon emission factors for electricity in each province (municipality or autonomous region) are obtained from the “Average Emission Factors of China’s Regional and Provincial Power Grids”. The data on the number of in-use electric vehicles are sourced from the “China Automotive Industry Yearbook”. Terminal energy consumption data for the transportation sector in each province (municipality or autonomous region) are extracted from the regional energy balance tables in the “China Energy Statistical Yearbook” corresponding to the research year. Additionally, statistical data on influencing factors of transportation carbon emissions, used to construct the predictive model, are sourced from the “China Statistical Yearbook” of the relevant year. Specific parameter details are presented in

Table 1. Correlation coefficients for various energy sources.

Energy	The Average Lower Calorific Value/kJ·kg−1	Carbon Content per Unit Calorific Value/Tg·TJ−1	Carbon Oxidation Rate/%	Carbon Emission Factor
Raw coal	20,934	26.37	94	1.903
Washed and refined coal	26,377	25.4	93	2.285
Coke	28,470	29.42	93	2.856
Gasoline	43,124	18.9	98	2.929
Kerosene	43,124	20.6	98	3.192
Diesel	43,124	19.6	98	3.037
Fuel oil	41,868	20.2	98	3.039
Liquefied petroleum gas	50,242	20.1	98	3.629
Natural gas	32,238	17.2	98	1.992
Liquefied natural gas	51,498	15.32	99	2.864

and

Table 2. Carbon dioxide emission factors of power grids in each province.

Province	Electricity Carbon Emission Factors	Province	Electricity Carbon Emission Factors
Anhui	0.785	Heilongjiang	0.773
Beijing	0.709	Hubei	0.350
Fujian	0.494	Hunan	0.514
Gansu	0.534	Jilin	0.714
Guangdong	0.531	Jiangsu	0.716
Guangxi	0.474	Jiangxi	0.662
Guizhou	0.500	Liaoning	0.811
Hainan	0.577	Inner Mongolia	0.883
Hebei	0.952	Qinghai	0.203
Henan	0.795	Shandong	0.854
Shanghai	0.632	Zhejiang	0.601
Sichuan	0.189	Chongqing	0.519
Tianjin	0.855	Shanxi	0.828
Xinjiang	0.731	Shaanxi	0.762
Yunnan	0.240	Ningxia	0.772

. Any missing data points have been linearly interpolated to ensure a complete dataset. It should be noted that the missing values occurred exclusively during the early period from 2001 to 2004, and during this time, the data exhibited smooth variation without any obvious nonlinear trends. Therefore, the use of linear interpolation is appropriate in this context. The total number of interpolated data points accounts for only 3.9% of the entire sample and is unlikely to influence the overall results.

In view of the availability and reliability of data, this study focuses on the time span from 2001 to 2022. Due to data limitations in Tibet, Hong Kong, Macao, and Taiwan regions of China, these areas are excluded from the analysis. Furthermore, official statistical sources do not provide separate data for the transportation sector; instead, it is aggregated with the warehousing and postal industries. Given that the warehousing and postal sectors account for a relatively small proportion of the combined statistics, the aggregated data is adopted as the primary basis for estimating carbon emissions in the transportation industry.

2.2. Transportation Carbon Emission Measurement Model

With the increasing adoption of new energy vehicles and electric vehicles, transportation carbon emissions are no longer solely attributed to traditional fuel combustion but are also significantly influenced by the growing electricity consumption. Consequently, this paper proposes a methodology for calculating transportation carbon emissions that incorporates two key components: emissions resulting from fuel combustion and those arising from electricity consumption.

The carbon emissions resulting from fuel consumption are estimated using the methodology outlined in the 2006 IPCC Guidelines for National Greenhouse Gas Inventories, which pertains to greenhouse gas emissions in the transportation sector. Specifically, this involves calculating the carbon dioxide (CO₂) emissions from transportation activities across 30 provinces (including municipalities and autonomous regions) of China based on their respective fuel consumption data. The calculation formula is presented as follows:

$$T_p={\displaystyle \sum _{i=1}^n{E}_i}\times E{F}_i$$

(1)

where $T_p$ denotes the total CO₂ emissions resulting from fuel combustion in the transportation sector of a given province $p$ (unit: ten thousand tons), $E_i$ denotes the consumption of the i-th type of fossil fuel (unit: ten thousand tons of standard coal), encompassing various energy sources such as coal, oil, and natural gas. $E{F}_i$ denotes the CO₂ emission factor for the i-th type of fossil fuel, where the subscript $i$ indicates the specific energy type, and $n=10$ corresponds to the total number of transportation-related energy categories considered in this study. The calculation formulas for the CO₂ emission factors of different energy types are presented as follows:

$$E{F}_i=NC{V}_i\times {10}^{-9}\times C{C}_i\times O{F}_i\times {10}^3\times {\displaystyle \frac{44}{12}}$$

(2)

where $NC{V}_i$ denotes the average lower calorific value of the i-th type of fossil fuel, $C{C}_i$ denotes the carbon content per unit calorific value for the i-th type of fossil fuel, $O{F}_i$ denotes the carbon oxidation rate of the i-th type of fossil fuel, and 44/12 corresponds to the conversion factor between the atomic weight of carbon (12) and the molecular weight of carbon dioxide (44).

The maturation of the electric vehicle (EV) industry chain, coupled with policy incentives such as the “New Energy Vehicle Industry Development Plan (2021–2035),” has catalyzed a fundamental transformation in the energy consumption structure of the transportation sector. By the end of 2022, the number of public charging stations in China had grown to 1.797 million (compared to only 58,000 in 2015), establishing a comprehensive charging network that spans urban and rural areas. The widespread adoption of this infrastructure has significantly reshaped the spatiotemporal distribution features of electricity demand for transportation. Specifically, the average daily charging volume for private vehicles increased from 7.5 kWh in 2015 to 15.2 kWh in 2022, while heavy-duty truck battery swap stations now provide over 300 kWh per charge. In urban agglomeration commuting scenarios, nighttime off-peak electricity usage accounts for 63% of total demand, whereas intercity travel exhibits pronounced peaks in highway charging during holidays (e.g., a year-on-year increase of 162% in single-day charging volume during the 2023 Spring Festival). To accurately capture these dynamic trends and conduct a comprehensive assessment of carbon emissions, this study integrates indirect emissions from electricity consumption for transportation into the accounting framework and develops a model for calculating indirect carbon emissions from electricity use. It is important to note that, in line with the international academic conventions outlined in the IPCC’s “2006 Guidelines for National Greenhouse Gas Inventories,” the calculation of carbon emissions from the transportation sector includes only those emissions generated by mobile equipment such as vehicles, aircraft, and ships. Accordingly, in this study, the electricity consumption attributed to transportation encompasses only the indirect carbon emissions associated with electric vehicle usage and excludes the electricity consumption of stationary infrastructure, such as airports and subway stations. The formula for calculating CO₂ emissions generated by electricity consumption is presented as follows:

$$T_p^E=A{D}_p\cdot E{F}_p^E$$

(3)

where $T_p^E$ denotes carbon emissions from transportation power consumption in the province $p$, $E{F}_p^E$ denotes carbon emission factors of the provincial power grid in the province $p$, and $A{D}_p$ denotes electricity consumption of the transportation industry consumption in province $p$. Specifically, the traffic electricity consumption $A{D}_p$ refers to the electricity consumed by electric vehicles for transportation purposes, and its calculation formula is as follows:

$$A{D}_p=NE{V}_p\cdot VKT\cdot EC$$

(4)

where $NE{V}_p$ denotes the number of in-use electric vehicles consumption in province $p$, $VKT$ denotes average vehicle mileage, and $EC$ denotes average electricity consumption per unit distance (kWh/km).

2.3. Features Screening of Transportation Carbon Emissions Based on the STIRPAT Model

Through a systematic review of the relevant domestic and international literature in recent years, 13 key features were preliminarily identified as influencing factors for transportation carbon emissions. These features encompass three major dimensions: demographic factors, economic factors, and technological factors. Given the significance of transportation electricity consumption as an important clean energy source, the transportation electricity consumption proportion feature was additionally incorporated into this study. The initial set of features includes the following 13 items: population size, urbanization level, gross regional domestic product, growth value of transportation industry, motor vehicle ownership, public transport per capita, cargo turnover, passenger turnover, industrial structure, road length, transportation energy consumption, transportation energy intensity, and transportation electricity consumption proportion. By analyzing the correlation between each feature and carbon emissions across different provinces, it was observed that the correlation between the same feature and carbon emissions varies under different urban conditions. This variation may be attributed to the distinct development contexts of individual provinces. This section outlines the feature selection process for transportation carbon emissions across multiple provinces.

To address the limitations of the traditional impact-population, affluence, and technology (IPAT) model in variable selection, this study employs an extended version of the stochastic impacts by regression on population, affluence, and technology (STIRPAT) model, as proposed in the literature [32], for driving factor analysis. To further strengthen the model’s explanatory capacity, this study extends the system of influencing factors based on the original STIRPAT framework. The mathematical expression of the extended STIRPAT model is as follows:

$$I=\alpha P^{{\beta }_1}U{L}^{{\beta }_2}GRD{P}^{{\beta }_3}GV{T}^{{\beta }_4}MV{O}^{{\beta }_5}PTP{T}^{{\beta }_6}C{T}^{{\beta }_7}P{T}^{{\beta }_8}I{S}^{{\beta }_9}R{L}^{{\beta }_{10}}TE{C}^{{\beta }_{11}}TEI{T}^{{\beta }_{12}}TEC{P}^{{\beta }_{13}}\epsilon$$

(5)

The mathematical formulation of the extended log-linearization model is as follows:

$$\ln I=\ln \alpha +{\displaystyle \sum _{k=1}^{13}{\beta }_k}\ln X_k+\ln \epsilon$$

(6)

where $X_k$ denotes 13 initial influencing factors, ${\beta }_k$ denotes the elastic coefficient of each factor to transportation carbon emission. In the construction of the STIRPAT model, potential multicollinearity among the initial feature variables may arise. The variance inflation factor (VIF) serves as a critical metric for assessing multicollinearity among the feature variables within the model.

$${\mathrm{VIF}}_k={\displaystyle \frac{1}{1-R_k^2}}$$

(7)

where $R_k^2$ denotes the determination coefficient for the k-th variable with respect to the remaining variables. The VIF serves as a key metric for quantifying the linear correlation among independent variables. Generally, a VIF value above a certain threshold is considered to indicate the presence of significant multicollinearity.

This study utilized the stepwise regression method of stepwise regression for feature screening. Stepwise regression is a systematic approach that enhances the statistical significance of the model by iteratively adding or removing feature variables until no further substantial improvement in model performance is observed. Considering the unique features of each province, the stepwise regression method of forward selection was applied individually to each province to identify significant variables influencing transportation carbon emissions. Among the criteria, variable inclusion is based on statistical significance, measured by the p-value. A candidate variable is added to the model if its p-value is below 0.05. Variable exclusion is determined by the VIF: any variable that causes the VIF of an existing independent variable to exceed the threshold is excluded, even if it is statistically significant. The threshold follows the widely accepted rule in econometrics that a VIF above 10 indicates severe multicollinearity, which may compromise model stability.

The variable that occurred most frequently across provinces was selected as the key feature for collaborative cross-province transportation carbon emission prediction.

Table 3. The selection frequency of each candidate indicator across all provinces.

Features	Frequency
Transportation energy consumption	29
Industrial structure	11
Passenger turnover	10
Transportation electricity consumption proportion	9
Urbanization level	6
Motor vehicle ownership	5
Road length	5
Public transport per capita	4
Population size	4
Cargo turnover	4
Transportation energy intensity	3
Gross regional domestic product	2
Growth value of transportation industry	2

presents the frequency of selection for each alternative indicator across the provinces.

Ultimately, ten variables were identified: population size, urbanization level, motor vehicle ownership, public transport per capita, cargo turnover, passenger turnover, industrial structure, road length, transportation energy consumption, and the proportion of transportation electricity consumption, as presented in

Table 4. Feature classification and variable specification.

Dimension	Features	Description	Symbol	References
Demographic Factors	Population size	Year-end permanent resident population.	P	[1,33,34,35,36,37]
	Urbanization level	The proportion of the urban population relative to the year-end permanent resident population.	UL	[1,33,34,38,39]
Wealth Factors	Motor vehicle ownership	The aggregate count of different types of civilian motor vehicles.	MVO	[1,33,34,35,37,38]
	Public transport per capita	Public transport vehicle ownership per 10,000 residents.	PTPC	[37]
	Cargo turnover	The total weight of goods carried by different modes of transportation.	CT	[1,34,35]
	Passenger turnover	The total number of passengers carried by different modes of transportation.	PT	[1,33,34,35]
	Industrial structure	The share of the gross value added by the tertiary sector in the economy.	IS	[1]
	Road length	Road length.	RL	[1]
Technology Level	Transportation energy consumption	Total energy consumption in the transportation sector.	TEC	[1,33,36]
	Transportation electricity consumption proportion	The share of electricity consumption in total transportation energy consumption.	TECP

.

3. Multi-Province Collaborative Prediction Model for Transportation Carbon Emissions Based on Spatio-Temporal Attention Mechanism

3.1. Model Architecture

Accurate prediction of transportation carbon emissions is crucial for formulating effective emission reduction policies and optimizing transportation management. However, transportation carbon emission data exhibit high spatio-temporal dependencies, and traditional forecasting methods often struggle to capture their complex nonlinear relationships. Moreover, current research typically models either the temporal or spatial dimension in isolation, failing to consider the interplay between space and time. To address this, this paper proposes a multi-province collaborative transportation carbon emission prediction model based on a spatio-temporal attention mechanism (MC-STAPM).

As illustrated in Figure 1, the model is composed of an input module, a GCN module, a sliding window-based spatio-temporal perception attention module, and an output module. The input module initially cleanses the multi-province transportation carbon emission data with multi-dimensional features over n years, transforming it into a nested vector format suitable for model processing, thereby laying the groundwork for subsequent computations. Within the GCN module, the transportation network structure topology of 30 provinces nationwide is introduced. To ensure greater stability in feature propagation and information extraction within the graph structure, the GCN convolves the topology with the module’s input to extract spatial features of multi-province transportation carbon emissions. The sliding window-based spatio-temporal perception attention module first employs a sliding window mechanism to ensure rich interaction between adjacent windows. Simultaneously, to guarantee that the modeling of temporal correlations takes into account the impact of time intervals, a strided long-term temporal attention mechanism is designed. Furthermore, we propose a spatial enhancement attention mechanism, defining and appropriately partitioning various spatial connection relationships. Finally, we introduce multiple connection relationship impact factors to dynamically regulate the weights of different connections, thereby fully considering the various connection relationships among the 30 provinces. Lastly, we stack multiple sliding window-based spatio-temporal perception attention modules to capture more spatio-temporal correlations, integrating them to achieve the final prediction layer and enhancing the accuracy of transportation carbon emission prediction tasks.

3.2. Construction of MC-STAPM

The variation in transportation carbon emissions invariably occurs across both temporal and spatial dimensions. Temporally, carbon emission levels exhibit significant dynamic fluctuations. Factors such as population, the number of civilian vehicles, and urbanization levels can influence transportation carbon emissions over time. Spatially, multiple provinces influence each other; for instance, highways and railways connecting provinces form an interconnected transportation network, where increased transportation flow directly leads to the cross-regional spread of carbon emissions. Moreover, economically developed provinces attract resources and populations from surrounding areas, creating regional economic circles. The rise in transportation demand can elevate overall carbon emissions. Regional integration policies, such as the coordinated development of the Beijing-Tianjin-Hebei region, facilitate intercity transportation planning coordination, necessitating cross-province collaboration in transportation carbon emission management. Therefore, the multi-province collaborative transportation carbon emission prediction model based on the spatio-temporal attention mechanism is primarily divided into two parts. Specifically, it begins with preprocessing the data to integrate data features across both temporal and spatial dimensions. It then models the spatio-temporal aspects, capturing the impacts of time, space, and multidimensional features on carbon emissions.

3.2.1. GCN Module

The multi-province road network structure topology map can represent the connection relationships between different provinces. Introducing this topology map allows for a clearer capture of the patterns of cross-province spatial transmission and diffusion of carbon emissions. Generally, the multi-province road network structure topology map is represented by a graph connectivity matrix $A\in R^{(N\times N)}$. To more intuitively reflect the adjacency relationships among multiple provinces, we use the actual physical straight-line distance between two provinces to represent the connectivity relationship between provinces $v_i$ and province $v_j$. Based on this, each element in the graph connectivity matrix can be expressed as follows:

$$A_{ij}=\left\{\begin{array}{rr}Di{s}_{ij},\hfill & \hfill \mathrm{Cities} v_i \mathrm{and} v_j \mathrm{connected}\\ 0,\hfill & \hfill \mathrm{Otherwise}\end{array}\right.$$

(8)

where $A_{ij}$ represents the connectivity between province $A_{ij}$ and province $v_j$, and $Di{s}_{ij}$ denotes the physical straight-line distance between province $A_{ij}$ and province $v_j$ with the specific numerical values obtained from the official channels of Amap.

In our paper, a graph convolutional network (GCN) model is employed to process the graph adjacency matrix. Specifically, for the historical transportation carbon emission data X and the graph adjacency matrix A, the output after performing the graph convolution operation can be expressed as follows.

$$\widehat{X}=\sigma ({\widehat{D}}^{(-1/2)}\widehat{A}{\widehat{D}}^{(-1/2)})XW$$

(9)

where $\sigma \left(\cdot \right)$ represents the nonlinear activation function Relu. $\widehat{A}=A+I$ denotes the graph adjacency matrix after considering self-connections, where $I$ is the N-th order identity matrix, and $\widehat{D}$ represents the degree matrix calculated based on $\widehat{A}$. ${\widehat{D}}^{(-1/2)}\widehat{A}{\widehat{D}}^{(-1/2)}$ signifies the symmetric normalization applied to $\widehat{A}$, which aids in enhancing the performance of the graph convolutional neural network.

Subsequently, we integrate the features across both temporal and spatial dimensions, with the specific operation represented as follows:

$$\tilde{X}=F\left(\widehat{X}\right)$$

(10)

where $\widehat{X}\in R^{(T\times N\times C)}$ represents the output of the graph convolution, $F\left(\cdot \right)$ denotes the flattening operation performed across the spatial and feature dimensions, $\tilde{X}\in R^{(T\times \tilde{d})}$ signifies the output after feature fusion, $\tilde{d}=N\times C$ denotes the size of the dimension after flattening.

By utilizing the aforementioned method, the historical data of transportation carbon emissions can be integrated with the multi-province transportation network structure topology map information, allowing the features of transportation carbon emission propagation and diffusion patterns to be fused across both temporal and spatial dimensions. To enable the subsequent stacked different modules to capture spatio-temporal features at various scales, we proceed to construct multi-scale linear layers, with the calculation formula as follows:

$${\tilde{X}}_l={\tilde{X}}_{l-1}{W}_l+b_l$$

(11)

where ${\tilde{X}}_l\in R^{N\times \left(T/2^l\right)}$ represents the spatio-temporal feature parameters output by the l-th layer.

3.2.2. Sliding Window-Based Spatio-Temporal Perception Attention Module

To accurately predict future transportation carbon emissions, we have designed a sliding window-based spatio-temporal perception attention module. This module primarily consists of three parts: a sliding window mechanism, a strided long-term temporal attention mechanism, and a spatial enhancement attention mechanism. The sliding window mechanism is used to segment the data, enabling the model to understand local spatio-temporal correlations at different time scales. The strided long-term temporal attention mechanism is employed to capture the temporal correlations of various influencing factors. Compared to traditional attention mechanisms, it ensures that the attention weights are closely related to the time steps. The spatial enhancement attention mechanism takes the output of the strided long-term temporal attention mechanism as its input and introduces spatial correlation coefficients to directly capture implicit spatial correlations through the attention mechanism in the spatial dimension, thereby indirectly capturing local spatio-temporal correlations.

(1)

Sliding Window Mechanism

We set the size of the sliding window to be larger than the sliding step, which allows the model to share partial information between adjacent windows. The window size determines the current specific time scale, and at this scale, the feature parameter space generated by the GCN module is integrated into the attention mechanism. This sub-module uses the sliding window mechanism to segment the input data from the time dimension into windows, turning it into multiple local inputs. The specific partitioning operation is represented as follows:

$$\widehat{H}=ShiftWindow\left(H_l,U_l,S_l\right)$$

(12)

where $H_l$ represents the input of the l-th module among the stacked multiple local spatio-temporal perception modules based on the sliding window. When $l=0$, $H_0=\tilde{X}{W}_u+b_u$, where $W_u\in R^{(C\times d_c)}$, and $b_u\in R^{d_c}$. $ShiftWindow\left(\cdot \right)$ denotes the data captured by the sliding window mechanism for a specific window size, representing the local data at the current time scale, $U_l$ represents the window size, $S_l$ represents the step size, and $\widehat{H}\in R^{U_l\times N\times d_c}$.

(2)

Strided Long-term Temporal Attention Mechanism

In typical time series data, such as carbon emissions, the attention mechanism calculates attention weights to represent the importance of different time steps in the sequence. However, it treats each time step equally during computation, which fails to fully consider the impact of the time intervals between different steps. Therefore, we have designed a strided long-term temporal attention mechanism that regulates the influence of time intervals by introducing a time decay coefficient and integrates the multi-scale spatio-temporal feature parameters generated by the GCN module into this attention mechanism.

Firstly, the spatio-temporal feature parameters are obtained through the GCN module and are, respectively, used as the parameters required for the key unit K and the value unit V mappings in the attention mechanism. The specific operation is as follows:

$$Q_l^p={\widehat{H}}_l$$

(13)

$$K_l^p=\left({\widehat{H}}_l{W}_{l,k}^p+b_{l,k}^p\right){\tilde{X}}_l$$

(14)

$$V_l^p=\left({\widehat{H}}_l{W}_{l,v}^p+b_{l,v}^p\right){\tilde{X}}_l$$

(15)

where $W_{l,k}^p$ and $W_{l,v}^p$ respectively represent the weight matrices for the key unit $K$ and the value unit $V$ in the temporal attention of the l-th module, while $b_{l,k}^p$ and $b_{l,v}^p$ denote the corresponding bias terms.

Subsequently, the module incorporates a temporal decay factor $\beta$, By learning this hyperparameter, the model dynamically adapts to varying time scales. The coefficient governs the influence of time intervals on attention weights, while the calculation of attention mechanisms explicitly accounts for the impact of the time interval $△t$ is fully considered when calculating the attention coefficient, as detailed in the following formula:

$${\mathrm{Score}}_l^p=\tanh \left({\displaystyle \frac{\sigma \left(\frac{Q_l^p{K}_l^{p^T}}{\sqrt{d_c}}\right)}{1+\sigma \left(\sigma \left(\frac{Q_l^p{K}_l^{p^T}}{\sqrt{d_c}}\right)\ast \sigma (\beta )\ast \Delta t\right)}}\right)$$

(16)

where ${\mathrm{Score}}_l^p$ represents the attention coefficient matrix at the time scale of the l-th module, and $\sigma \left(\cdot \right)$ denotes the activation function sigmoid. The function $\sigma \left(\beta \right)$ naturally regularizes the hyperparameter $\beta$ to the range (0, 1).

Finally, multi-head attention is introduced to learn the temporal correlations present in the data from multiple subspaces, with the specific operation as follows:

$${\widehat{H}}_l^p=ML{P}_l^p\left(\mathrm{Concat}\left(hea{d}_{l,1}^p,hea{d}_{l,2}^p,\dots ,hea{d}_{l,\kappa }^p\right)\right)$$

(17)

$$hea{d}_{l,i}^p=\mathrm{Attention}\left(Q_{l,i}^p,K_{l,i}^p,V_{l,i}^p\right)=\mathrm{Softmax}\left({\mathrm{Score}}_{l,i}^p\right)V_{l,i}^p$$

(18)

where $hea{d}_{l,i}^p$ represents the computation of the i-th attention head, $\kappa$ denotes the number of heads in the multi-head attention, $\mathrm{Concat}\left(\cdot \right)$ is the concatenation operation, and ${\widehat{H}}_l^p\in R^{U_l\times N\times d_C}$.

(3)

Spatial Enhancement Attention Mechanism

In the actual urban road network structure, the spatial correlations between provinces cannot be fully reflected solely through the physical connectivity relationships between them. In addition to considering the inherent direct road connections between nodes, the spatial correlations should also take into account the spatial homogeneity connections between nodes. Therefore, we have defined multiple types of spatial connection relationships to consider different levels of spatial connectivity. Specifically, ① the inherent connections in the urban road network structure are defined as strong connection relationships, where inherent connections refer to the direct physical road links between two provinces. ② The spatial homogeneity connections and spatial indirect adjacency connections are defined as weak connection relationships.

The strong connection relationship includes the spatial direct adjacency connection relationship, which means that two nodes are directly connected by a physical road. This is represented using the graph adjacency matrix $A^{\prime }\in R^{N\times N}$, where each element value $A_{ij}^{\prime }$ in the matrix is calculated as follows:

$$A_{ij}^{\prime }=\left\{\begin{array}{rr}1,& \mathrm{Cities} v_i \mathrm{and} v_j \mathrm{are} \mathrm{directly} \mathrm{adjacent}\\ 0,& \mathrm{Otherwise}\end{array}\right.$$

(19)

The weak connection relationship includes spatial homogeneity connection relationships and spatial indirect adjacency connection relationships. Specifically, given the time series data $Z_1=\left(Z_{11},Z_{12},\dots ,Z_{1n}\right)$ and $Z_1=\left(Z_{21},Z_{22},\dots ,Z_{2n}\right)$ for two province nodes, the distance between any two time points in the sequences is initialized through $D_{a,b}=\left|Z_{1a}-Z_{2b}\right|$, defining the initial distance matrix $D\in R^{n\times m}$, where $D_{a,b}=\left|Z_{1a}-Z_{2b}\right|$ is the element value at the indices $a,b$ of the distance matrix. Based on this distance matrix, the cumulative distance between the sequences can be calculated, with the specific calculation formula as follows:

$$M_{a,b}=D_{a,b}+\min \left(M_{a-1,b},M_{a,b-1},M_{a-1,b-1}\right)$$

(20)

where $M_{a,b}$ represents the total distance accumulated up to the a-th row and b-th column of the cumulative distance matrix $M$. $M_{n,m}$ is then used to denote the DTW (Dynamic Time Warping) distance between the time series data $Z_1$ and $Z_2$.

Then, based on the above calculation rules, the DTW distance for the time series data between each pair of nodes can be determined, and based on this distance, the spatial homogeneity connection graph $A^{″}\in R^{N\times N}$ is defined, where each element value $A_{ij}^{″}$ is calculated as follows:

$$A_{ij}^{″}=\left\{\begin{array}{rr}1,& \mathrm{DTW}(v_i,v_j)>\sigma \\ 0,& Otherwise\end{array}\right.$$

(21)

where $\mathrm{DTW}(v_i,v_j)$ represents the DTW distance between provinces $v_i$ and $v_j$, and the threshold $\delta$ is used to determine which pairs of nodes are considered to have spatial homogeneity connectivity.

Spatial indirect adjacency connections refer to the absence of a direct connection between nodes. Since the graph adjacency matrix $A_1^{\prime }$ is an unweighted graph, for an unweighted graph, the power of the matrix $A_{ij}^m$ indicates how many paths of length m exist from node $i$ to node $j$. Based on this property, the second-order neighbors of the graph adjacency matrix $A$ can be represented as $A^2=A\times A$. Therefore, the second-order connection relationship matrix $A^{‴}\in R^{N\times N}$ is specifically defined as follows:

$$A_{ij}^{‴}=\left\{\begin{array}{rr}1,& A_{ij}^2\ge 1\\ 0,& \mathrm{Otherwise}\end{array}\right.$$

(22)

We introduce spatial connection relationship impact factors and integrate them into the attention mechanism. On one hand, it allows for the calculation of the relevance of all node pairs without using a stacking approach, and on the other hand, it enables the setting of different impact factors for different types of neighbors to dynamically and jointly regulate the capture of spatial correlations at various levels. Specifically, the spatial connection relationship impact factors are the strong connection relationship impact factor $\alpha$, the homogeneity connection impact factor $\gamma \in \left[0,0.1\right]$, and the second-order connection impact factor $\epsilon \in \left[0,0.1\right]$. In this paper, three distinct adjacency matrices $A^{\prime },A^{″},A^{‴}$ are constructed to characterize different types of spatial connectivity relationships between nodes. The factors $\alpha$, $\gamma$, and $\epsilon$ are introduced to dynamically and differentially adjust the importance weights of these diverse relationships in the final model. Therefore, the final spatial correlation modeling matrix ${\widehat{A}}^{\prime }$ is calculated as follows:

$${\widehat{A}}_{IJ}^{\prime }=\mathrm{Max}\left(\alpha A_{ij}^{\prime },\gamma A_{ij}^{″},\epsilon A_{ij}^{‴}\right)$$

(23)

where ${\widehat{A}}_{IJ}^{\prime }$ is the element value in the i-th row and j-th column of the spatial correlation modeling matrix ${\widehat{A}}^{\prime }$ and $\mathrm{Max}\left(\cdot \right)$ is the function that takes the maximum value.

The spatial enhancement attention mechanism captures the implicit spatial dependencies present in the data across multiple different subspaces by utilizing the multi-head attention mechanism in the spatial dimension. It takes the output of the temporal attention mechanism as input to indirectly capture local spatio-emporal correlations. Moreover, it integrates the weight matrix ${\widehat{A}}^{\prime }$ into the attention mechanism to regulate the importance of various connection relationships based on specific weight values, thereby ensuring that nodes aggregate information from other nodes differentially. The calculation formula is as follows:

$$Q_l^q={\widehat{H}}_l^p$$

(24)

$$K_l^q={\widehat{H}}_l^p{W}_{l,k}^q+b_{l,k}^q$$

(25)

$$V_l^q={\widehat{H}}_l^p{W}_{l,v}^q+b_{l,v}^q$$

(26)

$${\widehat{H}}_l^q=ML{P}_l^q\left(\mathrm{Concat}\left(hea{d}_{l,1}^q,hea{d}_{l,2}^q,\dots ,hea{d}_{l,\kappa }^q\right)\right)$$

(27)

$$hea{d}_{l,i}^q=\mathrm{Attention}\left(Q_{l,i}^q,K_{l,i}^q,V_{l,i}^q\right)=\mathrm{Softmax}\left({\displaystyle \frac{Q_l^q{K}_l^{q^T}}{\sqrt{d_c}}}{\widehat{A}}^{\prime }\right)V_{l,i}^p$$

(28)

where $hea{d}_{l,i}^q$ represents the computation of the i-th attention head, $\mathcal{K}$ denotes the number of heads in the multi-head attention, and ${\widehat{H}}_l^q\in R^{U_l\times N\times d_C}$.

Finally, we aggregate the output of the spatial enhancement attention mechanism along the time dimension to generate the final output for the current window. Simultaneously, we use the sliding mechanism to capture the outputs of multiple windows and take them as the output of this layer. The specific operation is as follows:

$$H_{l+1}=\left(\mathrm{Agg}\left(ML{P}_l({\widehat{H}}_{l,1}^q)\times {\widehat{H}}_{l,1}^q\parallel \dots \parallel \mathrm{Agg}\left(ML{P}_l({\widehat{H}}_{l,T_l}^q)\times {\widehat{H}}_{l,T_l}^q\right)\right)\right)$$

(29)

where ${\widehat{H}}_{l,1}^q$ represents the output of the first window generated by the sliding window under the l-th module after the strided long-term temporal attention and spatial attention mechanisms. $T_l$ denotes the total number of windows after the sliding window mechanism at the current time scale, and its value will also serve as the time scale size for the next layer’s stacked module, with $T_l={\displaystyle \frac{T_{l-1}-U_l}{s_l}}$ and $T_0=T$. $\mathrm{Agg}\left(\cdot \right)$ indicates the summation operation along the time dimension, and $H_{l+1}\in R^{T_l\times N\times d_C}$.

3.2.3. Output Layer

During the model training process, this model employs a multilayer perceptron to integrate spatio-temporal features obtained from multiple time scales, ultimately performing output prediction, which can be specifically represented as follows:

$$\widehat{Y}=MLP\left({\displaystyle \sum _{l=1}^L\left(H_l{W}_l+b_l\right)}\right)$$

(30)

where $L$ represents the number of layers, and $\widehat{Y}$ is the final prediction output. We complete the learning of the model by optimizing the following loss function:

$$\mathcal{L}=\left\{\begin{array}{ll}{\displaystyle \frac{1}{2\times \delta }}{(Y-\widehat{Y})}^2|Y-X|\le \delta \hfill & \hfill \\ |Y-\widehat{Y}|-{\displaystyle \frac{1}{2}}\delta \hfill & otherwise\hfill \end{array}\right.$$

(31)

where δ is a threshold parameter used to control the squared error loss.

4. Results and Scenario Analysis

4.1. Experimental Results and Analysis

Our paper utilizes panel data from Chinese provinces (municipalities and autonomous regions) from 2001 to 2022 as the dataset for constructing the prediction model, totaling 660 data entries. First, all urban carbon emission data are chronologically ordered (from the earliest to the most recent records), with the first 60% used as the training set, the middle 20% as the validation set, and the final 20% as the test set. Three commonly used evaluation metrics are selected to assess the model we proposed. The three metrics are: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE). The calculations for these evaluation metrics are as follows:

Mean Absolute Error (MAE):

$$MAE={\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n|}{Y}_i-{\widehat{Y}}_i|$$

(32)

Root Mean Square Error (RMSE):

$$RMSE=\sqrt{{\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^n{(Y_i-{\widehat{Y}}_i)}^2}}$$

(33)

Mean Absolute Percentage Error (MAPE):

$$MAPE={\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^n\left(\frac{|Y_i-{\widehat{Y}}_i|}{Y_i}\right)}$$

(34)

where $N$ is the number of data samples, $Y_i$ represents the true value of carbon emissions, and ${\widehat{Y}}_i$ represents the predicted value of carbon emissions.

The experiments in this paper are implemented based on the Pytorch framework, version 1.11.0, using Python version 3.8. The graphics card used is a 12 G NVIDIA GeForce GTX3060, which is a company headquartered in Santa Clara, CA, USA. Adam is one of the most commonly used and highly efficient optimization algorithms in deep learning, which is used as the optimizer, with the total number of training epochs set to 200. An early stopping strategy is adopted during the experiments to avoid overfitting, meaning that training is halted if there is no improvement in prediction performance for 70 consecutive epochs. All datasets are treated with one year as one time step, to hierarchically integrate local, regional, and global spatial information, forming a multi-scale feature representation for each region, while also mitigating overfitting issues associated with deeper networks and aligning with baseline studies. As shown in Figure 2, this experiment sets one step as one time step, with a window size of two time steps for the sliding window to preprocess the original dataset. The experiment sets the number of layers for the stack of spatio-temporal perception modules based on the sliding window to three.

To validate the performance of our proposed MC-STAPM model, representative methods were selected, which are mainly divided into three categories: methods based on statistical analysis, methods based on machine learning, and methods based on deep learning. The specific methods are: Lasso_SVM, XGBoost, BP, MLP, LSTM, CEEMD-IWOA-KELM, CNN-LSTM, and DCRNN.

To evaluate the training effectiveness of MC-STAPM, the training loss results based on the dataset are plotted as shown in Figure 3. It can be observed from the figure that as the number of training steps increases, the total loss continues to decrease, indicating that the model undergoes a positive learning process on the dataset. For this dataset, approximately 100 training iterations are required to reach the specified loss threshold.

Table 5. Error simulation results of the prediction models.

Model	MAE	MAPE	RMSE
Lasso_SVM	0.0626	0.6648	0.0847
XGBoost	0.0720	0.8670	0.1180
BP	0.0756	0.7935	0.0983
MLP	0.0849	0.6930	0.1011
LSTM	0.0682	0.3680	0.0920
CEEMD-IWOA-KELM	0.0615	0.3900	0.0824
CNN-LSTM	0.0611	0.3390	0.0758
DCRNN	0.0583	0.3128	0.0736
MC-STAP	0.0557	0.2895	0.0717

presents the predictive performance of the MC-STAP model proposed in this paper and the baseline methods on the dataset. The evaluation metrics used include MAE, RMSE, and MAPE. It is important to note that the smaller the evaluation metric values, the better the model performance. The results of the optimal model are highlighted in bold, and it can be seen that the MC-STAP model proposed in this paper performs the best across all three evaluation metrics, specifically: MAE of 0.0557, RMSE of 0.0717, and MAPE of 0.28950. Furthermore, an analysis of the results in the table reveals the following:

• The performance of models based on traditional statistics and machine learning is suboptimal. The reason for this is that transportation carbon emission data typically involve complex spatio-temporal relationships and are significantly affected by various factors in each province. Traditional statistical and machine learning models often struggle to capture spatio-temporal dependencies. Additionally, transportation carbon emissions usually exhibit non-stationary features, meaning their patterns dynamically change over time, and this non-stationarity also makes it difficult for traditional models to capture and model these patterns.
• Compared to models based on traditional statistics and machine learning, deep learning methods exhibit superior performance. For instance, Long Short-Term Memory (LSTM) networks outperform traditional statistical and machine learning models due to their ability to effectively capture long-term dependencies in time-series data. However, compared to other deep learning models, LSTM fails to account for spatial correlation modeling. As a model specifically designed for spatiotemporal sequence data, DCRNN’s core strength lies in its use of diffusion convolution to capture spatial dependencies, while leveraging recurrent structures to model temporal dynamics. This enables it to capture complex spatiotemporal relationships in traffic carbon emission data, resulting in more advanced performance. Yet, compared to MC-STAP, DCRNN does not capture multi-scale spatial features as effectively, and it is slightly inferior in terms of model complexity and adaptability to specific scenarios. This further underscores the critical importance of modeling multi-scale spatiotemporal correlations in traffic carbon emission forecasting.

4.2. Scenarios Design

Based on the multi-province transportation carbon emission collaborative prediction model incorporating the spatio-temporal attention mechanism described above, this study employs scenario analysis to forecast the future trends of transportation carbon emissions across multiple provinces under varying policy and development scenarios. By considering the current levels of transportation carbon emissions in each province as well as relevant national and provincial policy documents, three distinct future development scenarios are established: the baseline scenario, the technological innovation scenario, and the regional economic differentiation scenario. These scenarios serve as the foundation for predicting future transportation carbon emissions.

(1) Base scenario (S1). In accordance with China’s current overarching working principle of “seeking progress while maintaining stability,” the base scenario assumes that national and urban socioeconomic development will proceed at the current pace and scale. Transportation demand is expected to remain in equilibrium with economic growth. While taking into account the developmental objectives outlined in relevant policies as a guiding framework, no supplementary measures are introduced to accelerate carbon emission reductions in the transportation sector. This scenario serves as a robust reference benchmark for forecasting future transportation carbon emissions.

(2) Technological Innovation Scenario (S2). Guided by China’s strategic objective of a great power in science and technology, technological innovation serves as the core driving force, providing critical support for the high-quality development of the transportation system. Breakthroughs in technology facilitate the comprehensive modernization of the transportation infrastructure, leading to substantial reductions in operational costs and energy consumption, while promoting the harmonious co-development of economic growth and environmental protection. The synergistic integration of intelligent technologies and green practices establishes a robust framework for the technological innovation scenario, which underpins the prediction of transportation carbon emissions.

(3) Regional economic differentiation Scenario (S3). In line with China’s strategic deployment of scientifically and orderly promoting industrial transfer, the central and western regions are poised to capitalize on their geographical advantages and policy incentives, thereby unlocking new opportunities for economic growth. As these regions optimize their industrial structures and enhance infrastructure development, a transformative advancement in the logistics and transportation sectors is anticipated. This progression establishes a robust framework for regional economic differentiation scenarios, which serves as a critical basis for predicting transportation carbon emissions.

To justify the subsequent configuration of scenario parameters,

Table 6. The setting of standards and policy sources for each period in the baseline scenario.

Features	T1 Trends and Goals	T2 Trends and Setting Ideas	T3 Trends and Setting Ideas	Main Policies
P	Maintain a low level of negative growth.	The trend of slow negative growth is expected to continue, with the annual growth rate remaining in a low negative range.	The negative growth rate is accelerating, and population aging is intensifying.	policy1
UL	The growth rate is declining. The 2025 objective is to continue advancing, with increased emphasis on quality.	After passing Northam’s second inflection point—where urbanization exceeds 66% and growth slows—greater focus should be on quality and urban-rural integration.	The urbanization process has reached a mature and stable phase, characterized by an exceptionally low average annual growth rate.	policy1
MVO	Growth has slowed. The 2025 target projects new energy vehicle sales to reach about 20% of total sales.	As new energy vehicles become more common and shared mobility expands, vehicle ownership growth is expected to slow further.	Vehicle stock may approach its peak or enter a low-growth plateau.	policy2
PTPC	By 2025, the province aims to have 72% of its public transport vehicles run on new energy.	As bus-oriented province development advances, improving service quality and efficiency becomes more critical.	Public transportation has become the dominant form of urban mobility, marked by advanced intelligence and integration.	policy3
CT	Growth is slowing, pointing to a 2025 focus on optimizing transport structure and efficiency.	Logistics efficiency has improved, freight intensity per unit GDP has declined, and multimodal transport’s share has risen.	Smart logistics and efficient supply chains are now dominant, but growth has slowed.	policy3
PT	Structural changes are clear, with high-speed rail and aviation gaining share.	Travel structure is being optimized, with convenience and efficiency as key features.	Transportation efficiency is high, with little room for further growth.	policy3
IS	By 2025, the non-fossil energy consumption target is set at 20%.	The share of green, low-carbon, and high-tech industries continues to grow.	Establish a green, low-carbon, and circular economic system.	policy4, policy2
RL	By 2025, targets include 5.5 million km of road networks and 190,000 km of expressways.	Road network density and accessibility have improved, albeit more slowly.	Road network growth is stabilizing, with focus shifting to maintenance, upgrades, and smart infrastructure.	policy3
TEC	By 2025, energy use per unit of GDP is expected to drop by 13.5% from 2020 levels, with increased promotion of new energy vehicles.	Energy efficiency and structure improvements have progressed significantly, with total consumption possibly leveling off or declining.	Low-carbon energy’s share has grown significantly, as total energy use is expected to decline notably.	policy2
TECP	By 2025, new energy buses are expected to make up 72% of the bus fleet, with continued rapid growth projected thereafter.	As electric vehicles spread and railway electrification progresses, electricity’s share in energy consumption is expected to rise steadily and rapidly.	Electricity has become a key energy source in transportation, with its share rising sharply.	policy2, policy3

presents the growth rates of each indicator under the benchmark scenario along with the corresponding policy references. For simplicity, the periods from 2023 to 2030, 2030 to 2035, and 2035 to 2040 are designated as T1, T2, and T3, respectively. The main policies include “The Northerm Turning Point, Urbanization Trends and Response Strategies for the 15th Five-Year Plan”, “Comprehensive Work Plan for Energy Conservation and Emission Reduction during the 14th Five-Year Plan”, “Development Plan for a Modern Comprehensive Transportation System during the 14th Five-Year Plan”, “Action Plan for Carbon Peaking Before 2030”, etc. For convenience, we, respectively, label them as policy1, policy2, policy3, and policy4. In the subsequent parameter setting phase, we focus on comparing longitudinal data across scenarios to provide robust support for analyzing the differences in carbon emission pathways among them.

4.3. Scenario Parameter Setting

To predict the future transportation carbon emission of various provinces under three distinct development scenarios, this section establishes the growth rates of key feature parameters related to transportation carbon emission under each scenario, as outlined in

Table 7. Setting baseline growth rates for various features under multiple scenarios.

Scenario	Years	Region	P/%	UL/%	MVO/%	PTPC/%	CT/%	PT/%	IS/%	RL/%	TEC/%	TECP/%
S1	T1	\	−0.5	1.2	5.8	4.1	4.5	3.2	1.5	6.2	2.7	12.0
	T2	\	−0.12	0.7	3.5	2.9	3.0	1.8	0.9	4.0	1.2	8.0
	T3	\	−0.25	0.3	1.2	1.5	1.8	−0.5	0.4	2.3	−0.3	5.0
S2	T1	\	−0.03	1.5	7.5	9.2	6.8	−2.1	2.8	3.0	−3.5	25.0
	T2	\	−0.08	1.0	4.0	6.3	5.2	−4.5	2.1	1.2	−6.2	18.0
	T3	\	−0.18	0.7	−1.2	3.8	3.5	−6.0	1.5	−0.5	−8.0	12.0
S3	T1	E	−0.30	0.50	1.20	7.80	2.00	−4.00	3.00	1.80	−5.20	30.00
		MW	0.40	2.00	9.50	3.20	8.50	3.20	0.80	7.20	2.10	10.00
	T2	E	−0.50	0.20	−2.00	4.50	1.20	−6.00	2.50	0.70	−7.00	25.00
		MW	0.20	1.20	4.50	5.80	5.00	1.50	1.50	3.50	−0.50	18.00
	T3	E	−0.70	0.10	−3.50	2.00	0.50	−8.00	1.80	0.30	−9.00	15.00
		MW	−0.10	0.50	1.80	3.50	2.50	−0.50	2.00	1.20	−2.00	12.00

. In the table, E and MW denote the eastern region and the central-western region, respectively. The specific feature encompasses population size, urbanization level, motor vehicle ownership, per capita public transport availability, cargo turnover, passenger turnover, industrial structure, road length, transportation energy consumption, and the proportion of electricity consumption within the transportation sector.

For the base scenario, based on the “Statistical Communiqué on the National Economic and Social Development of the People’s Republic of China in 2024”, “The Outline of the 14th Five-Year Plan for National Economic and Social Development of the People’s Republic of China”, as well as existing policies and domestic/international references [40,41], the following trends are expected in future economic and social development: The GDP growth rate will stabilize between 4% and 5%, consistent with the current policy intensity. Regarding population dynamics and urbanization, the accelerating aging process will lead to a decline in the peak total population, with the average annual population growth rate decreasing from −0.05% during T1 to −0.25% during T3. Additionally, the diminishing household registration dividend in third- and fourth-tier provinces will result in a gradual slowdown in urbanization growth, from 1.2% to 0.3%. In the transportation sector, structural adjustments will emerge. The gradual phase-out of subsidies for fuel vehicle replacements will cause the growth in vehicle ownership to peak, declining from 5.8% to 1.2%. The slowing expansion of new transit lines will reduce the growth rate of per capita public transportation volume from 4.1% to 1.5%. The efficiency limits of traditional logistics models will be reached, causing freight volume growth to drop from 4.5% to 1.8%. The saturation of the high-speed rail network and the widespread adoption of remote work will lead to negative passenger volume growth during T3. The development of the service sector will be constrained by rising costs associated with digital transformation, resulting in a gradual slowdown in the growth rate of the tertiary industry’s share, from 1.5% to 0.4%. In terms of infrastructure, the backbone road network in central and western regions is expected to be largely completed by 2030, leading to a gradual decrease in road length growth from 6.2% during T1 to 2.3% during T3. The transition in energy consumption structure will lag behind, causing transportation energy consumption growth to fall from 2.7% to −0.3%. However, the increasing penetration of electric vehicles and charging infrastructure will drive a rapid increase in the proportion of electricity consumption in transportation, maintaining growth rates between 12% and 5%.

For the technological innovation scenario, based on existing policies such as the “Opinions on Promoting the Innovative Development of Future Industries by the Ministry of Industry and Information Technology and Other Seven Departments”, the “Outline for Building a Strong Transportation Nation”, and the “Outline for the Development of Digital Transportation”, as well as breakthroughs in AI and new energy technologies, the GDP growth rate is projected to rise to 5.5–6.5%, driving profound transformations across all sectors of the economy and society. From a demographic and urbanization perspective, fertility encouragement policies will partially mitigate the effects of aging, narrowing the population growth rate from −0.03% during T1 to −0.18% during T3. Concurrently, the widespread adoption of smart city technologies will reduce migration costs for rural populations, maintaining urbanization growth at a relatively high level of 1.5–0.7%. In the transportation domain, disruptive changes are anticipated. Breakthroughs in shared autonomous vehicle technology and on-demand responsive bus systems will sharply reduce the growth rate of civilian vehicles from 7.5% to −1.2%, while per capita public transportation growth will gradually decline from 9.2% to 3.8%. Automation in port operations will enhance throughput efficiency by 50%, supporting a moderated freight volume growth rate that drops from 6.8% to 3.5%. The advancement of virtual reality meeting technology will lead to sustained declines in passenger volume, with its growth rate falling from −2.1% to −6%. Energy consumption structures will see significant optimization: the increasing penetration of hydrogen fuel cell heavy trucks will drive an annual decline in transportation energy consumption from 3.5% to 8.0%. Meanwhile, the development of new energy electric vehicles and the commercialization of wireless charging roads will enable the proportion of electricity consumption in transportation to maintain robust growth rates of 25–12%. Additionally, the ongoing maturation of vehicle-road coordination technologies will reduce lane requirements, causing road length growth to transition from 3.0% in T1 to negative growth (−0.5%) in T3. The service sector will benefit significantly from technological advancements, with the added value of digital services surpassing 60%. The growth rate of the tertiary industry’s share will decelerate from 2.8% to 1.5%, yet remain markedly higher than under the baseline scenario.

For the scenario of regional economic differentiation, the future development trajectories of the eastern and central-western regions exhibit significant disparities. The eastern region transitions toward a service-oriented economy, maintaining a GDP growth rate of 4–5%, while the central-western regions, benefiting from industrial relocation, sustain a higher GDP growth rate of 6–7%. From the perspective of population dynamics and urbanization, stricter settlement criteria in the eastern region lead to sustained population outflows, exacerbating the decline in population growth rates. Conversely, the central-western regions experience gradual population increases due to the return of industries. Smart provinces in the eastern region approach their capacity thresholds, prompting a shift toward optimization and efficiency improvements, whereas the central-western regions focus on revitalizing underutilized urban areas due to high vacancy rates in newly developed zones. The ban on fuel vehicle sales in the eastern region drives a continuous reduction in civilian vehicle ownership, while the central-western regions witness steady growth in vehicle ownership at a rate of 9.5–1.8%. The completion of the air-rail network in the eastern region is expected to initially boost per capita public transportation usage by 7.8%, while fiscal subsidies in the central-western regions ensure stable growth in this domain. The expansion of cross-border e-commerce in the eastern region results in a relatively low freight volume growth rate of 0.5%, whereas manufacturing upgrades in the central-western regions propel freight volume growth to 8.5–2.5%. The widespread adoption of holographic conferencing technology causes passenger volumes to decline continuously in both regions, albeit with slower declines in the central-western regions compared to the eastern region. Financial openness in the eastern region has reached its peak, leading to a deceleration in tertiary industry growth to 1.8%, while data center investments in the central-western regions drive tertiary industry growth to 2.0%. Infrastructure development trends also diverge: the eastern region shifts focus to underground utility tunnels, reducing road length growth to 0.3%, while road expansion in the central-western regions slows as highway density approaches saturation. In the energy sector, the penetration rate of battery-swap heavy trucks in the eastern region reaches 90%, driving an expanded annual decline in transportation energy consumption to 9.0%, whereas carbon taxes halt coal-to-hydrogen projects in the central-western regions, resulting in negative growth rates for transportation energy consumption. Finally, wireless charging roads cover 60% of main roads in the eastern region, boosting electricity consumption proportions with growth rates of 30–15%, while declining photovoltaic hydrogen production costs in the central-western regions elevate electricity consumption growth rates to 12%.

4.4. Analysis of Scenario Prediction Results

As described in Section 4.1, a high-performing MC-STAPM prediction model has been successfully trained. Subsequently, the study performed a collaborative predictive analysis of transportation carbon emission for 30 Chinese provinces from 2023 to 2040 under the baseline scenario, the technological innovation scenario, and the regional economic differentiation scenario. Figure 4 presents the predicted results for selected provinces—namely Shanghai, Guangzhou, Guizhou, and Chongqing—under three different scenarios. Due to variations in the initial levels of transportation carbon emissions across these provinces, the projected peak values of transportation carbon emissions also differ. The results indicate that the trends in transportation carbon emission for the same province vary across different scenarios, while the trends in transportation carbon emission for different provinces under the same scenario exhibit significant heterogeneity.

In terms of the average transportation carbon emissions in China’s eastern and central-western regions, the evolution paths of transportation carbon emissions under the three scenarios demonstrate significant spatial heterogeneity. Figure 5 illustrates the comparative analysis of average transportation carbon emissions in the eastern and central-western regions across the three scenarios. Specifically, in the eastern region, the average peak transportation carbon emissions under the three scenarios are 32.75 Mt, 20.28 Mt, and 17.83 Mt, respectively; in the central-western regions, the corresponding values are 22.41 Mt, 13.93 Mt, and 21.40 Mt. It is evident that, regardless of whether in the eastern or central-western regions, the baseline scenario exhibits the lowest transportation carbon emission efficiency. However, there are notable differences in the transportation carbon emission paths between the regional differentiation scenario (S3) and the technological innovation scenario (S2). To investigate the underlying drivers of the regional disparities in the aforementioned emission pathways, Figure 6 compares the differences in core parameter settings between the eastern region and the central-western regions under S3. For the eastern region, the emission reduction efficiency under the regional differentiation scenario is significantly higher than that under the technological innovation scenario. Specifically, the eastern region is assumed to experience a more aggressive industrial structure adjustment (with an IS growth rate of 3.0% at T1 and 0.8% in MW), along with a significantly faster adoption of electric vehicles (as indicated by the notable decline in TEC and rise in TECP). This may be attributed to the eastern region entering the phase of transportation system transformation, where it leads the national average in terms of transportation technology application, transportation structure optimization, and the promotion of new energy vehicles, thereby achieving more pronounced emission reduction efficiency. In contrast, the central-western regions are experiencing accelerated industrialization and urbanization, as reflected in our model by higher motor vehicle growth assumptions (MVO growth of 9.5% compared to 1.2% in the eastern region at T1), along with relatively slower advancements in structural optimization and electrification. The growing transportation demand and infrastructure expansion impose a “rigid constraint” on carbon emissions. Under such conditions, only through the widespread adoption of new energy transportation tools and the implementation of intelligent transportation management and other technological innovations can the incremental emissions associated with the development stage be effectively mitigated. Consequently, the technological innovation scenario demonstrates more substantial emission reduction efficiency in the central-western regions.

The overall trend of China’s transportation carbon emissions under the three development scenarios is presented in Figure 7. It can be observed that, under all three scenarios, China’s total transportation carbon emissions are expected to peak around 2030, with the respective peak values being 786.13 Mt, 487.90 Mt, and 602.85 Mt. An analysis of the national trend reveals that the benchmark scenario exhibits the lowest efficiency in reducing transportation carbon emissions, followed by the regional differentiation scenario, while the technological innovation scenario achieves the highest reduction efficiency. This suggests that, from a national perspective, enhancing new energy transportation technologies and promoting technological innovation represent the most effective strategies for reducing transportation-related carbon emissions.

5. Discussion

5.1. Linking the Discussion of Model Findings to Real-World Policy Challenges

Our scenario simulation results indicate that China requires regionally differentiated strategies to effectively reduce transportation-related carbon emissions [42,43,44,45]. This section will explore the underlying real-world rationale behind this finding.

In the regional differentiation scenario (S3), the eastern region demonstrates advantages that are primarily attributable to the more advanced stage of its overall development. The specific transformation characteristics are outlined as follows:

• First, the industrial structure is deeply oriented toward the service sector. In provinces such as Guangdong and Zhejiang, the service industry has become the dominant economic component, with high-value-added sectors such as financial technology and modern logistics serving as key growth engines. The carbon emission intensity associated with transportation in these sectors is markedly lower than that of the industrial sector, thereby offering structural support for the decoupling of economic growth and carbon emissions.
• Second, transportation infrastructure has transitioned toward efficiency optimization. Following the completion of the core network construction, the eastern region has redirected its policy focus toward system upgrades. This includes enhancing road network efficiency through intelligent transportation systems, developing integrated multimodal transport solutions to reduce empty-load rates, and implementing energy-efficient retrofits on existing infrastructure. These initiatives align closely with the core objectives of the RED scenario.
• Third, there is a trend of collaborative innovation in market-oriented policy instruments. The relatively high per capita income level and well-established charging infrastructure network in the eastern region have contributed to the formation of a mature market for new energy vehicles. Policy mechanisms have evolved from direct purchase subsidies to more targeted incentives, such as preferential road access and parking fee discounts. This shift not only alleviates fiscal burdens but also enhances the long-term sustainability of policy implementation.

Collectively, these transformations illustrate that the eastern region has established a synergistic framework encompassing “low-carbon industries, intelligent infrastructure, and precise policy interventions,” thereby offering a replicable and actionable model for the effective implementation of the S3 scenario.

The model demonstrates that technological innovation plays a pivotal role in the transformation of the central and western regions, revealing three key contradictions:

• First, the carbon lock-in risk associated with infrastructure expansion. In provinces such as Henan and Sichuan, large-scale construction of highways, railways, and related infrastructure has generated significant demand for high-carbon transportation of building materials and construction equipment. A singular focus on short-term economic growth may result in long-term fossil fuel dependency, thereby increasing the risk of carbon lock-in within the transportation system.
• Second, the challenge of path selection under fiscal constraints. Local governments face a strategic dilemma: while traditional infrastructure projects can rapidly stimulate GDP growth, emerging infrastructure such as electric vehicle charging networks typically entails a longer return on investment. Although technological innovation requires substantial upfront investment, it can effectively mitigate the high costs associated with future low-carbon transitions.
• Third, the imperative for leapfrog development. The central and western regions must avoid replicating the “pollute first, clean up later” model previously followed by the eastern regions. Technological innovation offers a viable pathway to directly adopt green and intelligent transportation systems. To realize this potential, proactive policy guidance is essential to ensure that new infrastructure adheres to low-carbon standards from the earliest stages of planning and implementation.

Therefore, the findings of this research model underscore the significant heterogeneity in regional development across China, yielding clear and actionable policy implications: the eastern region should prioritize the deepening of low-carbon optimization within its mature economic system, whereas the central and western regions must proactively avoid the risks of high-carbon lock-in and pursue a leapfrog transformation of their transportation systems through strategically directed green investment.

5.2. Management Implications and Practical Application Value

The predictive model developed in this study significantly improves the accuracy of provincial-level traffic carbon emission forecasts by integrating machine learning techniques with spatial synergy effects [46]. This research offers the following key management implications:

• Providing reliable decision-making benchmarks: High-precision forecasting establishes a more robust benchmark for evaluating the effectiveness of various emission reduction policies—such as new energy vehicle promotion and congestion charging—as well as large-scale transportation infrastructure projects. This enhances the scientific rigor of benefit assessments.
• Enabling precise allocation of emission reduction resources: The model allows for the early identification of future emission hotspots and critical provincial nodes. This enables policy-makers to direct limited regulatory attention, financial subsidies, and infrastructure investments toward regions with the highest emission reduction potential or the greatest risk exposure, thereby optimizing cost-effectiveness.
• Supporting the formulation of regionally coordinated strategies: Accurate estimation of regional emission totals facilitates a macro-level assessment of the gap between current development trajectories and carbon neutrality targets. This provides essential data support for the development of coordinated low-carbon transportation strategies and regional energy distribution plans, helping to prevent systemic inefficiencies that may arise from fragmented, uncoordinated actions.

In summary, the value of this research lies in generating a clearer “map of future carbon emissions,” which enables policy-makers to transition from reactive responses to proactive planning, ultimately supporting more forward-looking and evidence-based decision-making.

6. Conclusions

As a critical domain of carbon emissions in China, the transportation sector necessitates the development of scientifically robust and rational carbon emission prediction models and scenario analysis techniques to achieve carbon emission reduction targets. This study integrates the carbon emissions from transportation-related electricity consumption into the transportation carbon emission measurement framework and subsequently develops a multi-province collaborative prediction model for transportation carbon emissions based on a spatio-temporal attention mechanism (MC-STAPM). Furthermore, three future development scenarios are designed to forecast carbon emission trajectories. The key findings of this study are summarized as follows:

(1) With the advancement and widespread adoption of new energy vehicles and other emerging technologies, the share of electricity consumption in the transportation sector has been steadily increasing over the years, resulting in a substantial transformation of the transportation energy structure. The conventional carbon emission measurement method for transportation, which is predominantly based on fossil fuels, can no longer adequately capture the actual carbon emissions within the transportation domain. By incorporating the carbon emissions from transportation-related electricity consumption into the calculation process of transportation carbon emissions, the accuracy of the accounting can be significantly enhanced, thereby offering robust support for the transportation sector to achieve its emission reduction objectives.

(2) To address the complex spatio-temporal dependency issues associated with carbon emissions in multi-province transportation systems, this study integrates the spatio-temporal attention mechanism with the graph convolutional neural network (GCN) to develop the MC-STAPM. This model constructs a multi-source heterogeneous transportation network topology and incorporates an adaptive spatio-temporal attention weight allocation mechanism, enabling multi-scale capture and fine-grained analysis of the dynamic correlation characteristics of cross-regional transportation carbon emissions. The model exhibits superior feature representation capabilities in cross-province carbon emission correlation modeling, thereby offering a novel technical approach for the precise collaborative prediction of multi-province transportation carbon emissions.

(3) Three distinct scenarios were established using the scenario analysis approach, and the future transportation carbon emissions of 30 provinces were successfully predicted using the MC-STAPM. The results indicate significant spatial heterogeneity in the evolution paths of transportation carbon emissions across eastern and western regions under the three scenarios. In the eastern region, the regional differentiation scenario demonstrates higher emission reduction efficiency compared to the technological innovation scenario, whereas in the central and western regions, the technological innovation scenario exhibits more pronounced emission reduction efficiency. From a national perspective, transportation carbon emissions under all three scenarios are projected to peak around 2030. The benchmark scenario shows the lowest emission reduction efficiency, while the technological innovation scenario achieves the highest efficiency, suggesting that enhancing new energy transportation technologies and fostering innovation represent the optimal pathways for reducing national transportation carbon emissions.

(4) In terms of policy formulation, the eastern region should deepen the transformation of its transportation system by further enhancing investment in transportation technology application, optimizing transportation structure, and promoting new energy vehicles, while strengthening regional collaborative emission reduction. The central and western regions need to prioritize the deployment of technological innovation projects such as hydrogen fuel cell heavy trucks and intelligent transportation systems, while improving relevant infrastructure, and establish a coupling monitoring mechanism for carbon emissions and economic development. At the national level, it is essential to reinforce policy guidance by integrating emission reduction targets into the evaluation system, enhancing supervision mechanisms, promoting technological innovation through collaboration among industry, academia, and research institutions, establishing incentive mechanisms, and conducting public education campaigns to enhance environmental awareness, encourage green travel, and foster a favorable environment where society collectively participates in transportation emission reduction.