Science Heritage Journal (GWS) 8(1) (2024) 13-21

Science Heritage Journal (GWS)

DOI: [Link]

ISSN: 2521-0858 (Print)

ISSN: 2521-0866 (Online)
CODEN: SHJCAS

RESEARCH ARTICLE
PREDICTION OF CARBON EMISSIONS FROM TRANSPORTATION IN CHINA BASED
ON THE ARIMA-LSTM-BP COMBINED MODEL
Sheng Kai*, Ye Shanli

School of Science, Zhejiang University of Science and Technology, Hangzhou 310023, China.
*Corresponding Author Email: 18858238294@[Link]

This is an open access article distributed under the Creative Commons Attribution License CC BY 4.0, which permits unrestricted use, distribution, and
reproduction in any medium, provided the original work is properly cited.

ARTICLE DETAILS ABSTRACT

Article History: Transportation is not only a significant force in promoting economic and social development but also one of
the primary industries that consume energy and emit greenhouse gas emissions. In order to achieve China's
Received 18 November 2023 overall goal of reaching the carbon peak by 2030, this paper selects six influencing factors, such as population,
Revised 20 December 2023 GDP and urbanization rate, and proposes a combined prediction model based on ARIMA-LSTM-BP, which
Accepted 28 January 2024
Available online 30 January 2024 predicts transportation carbon emissions in China from 2022 to 2050 under three scenarios of low carbon,
benchmark and high carbon. The results show that the peak emissions of transportation in low-carbon,
benchmark and high-carbon scenarios are 1624.7732 million tons, 1478.1694 million tons and 1367.5417
million tons, respectively, reaching the peak in 2031, 2034 and 2039. It can be seen that in China, the
transportation industry alone cannot achieve the goal of reaching the peak by 2030, and more measures need
to be taken to achieve the carbon peak of the transportation industry as soon as possible.
KEYWORDS
Carbon emission prediction; combination model; transportation industry; scenario analysis method

1. INTRODUCTION correlation between each decomposition index and transportation carbon

emissions by using the grey correlation model (Jia, 2020). Commoner
The document of the 20th National Congress of the CPC Central Committee proposed the classic IPAT model and used it to explain the relationship
proposed that "actively and steadily promote carbon to peak and carbon between population, wealth, and environmental pressure (Commoner,
neutrality" and achieve the goal of "double carbon" is not only the 1990). Because the IPAT model only explores the relationship between
inevitable need for China to seek development at a new level of population, wealth and environment, it has certain limitations, so Dietz et
harmonious coexistence between man and nature but also an essential al. proposed the STIRPAT model, which has also become an essential
part of promoting Chinese style modernization. As one of the three means for scholars to study the influencing factors of traffic carbon
significant areas of energy consumption and carbon dioxide emissions, emissions in recent years (Dietz, 1997). Andrés et al. took energy intensity,
transportation is an important source of carbon dioxide emissions (Yang, per capita freight volume and other factors as influencing factors of traffic
2023). The transportation industry is developing rapidly and will become carbon emissions based on the STIRPAT model (Andrés et al., 2018). Liu
the key industry to achieve carbon emission reduction goals in the future. et al. used the STIRPAT and GTWR model to reveal the impact of
However, with the popularity and rapid increase of the number of population size, urbanization rate, the number of private cars and other
transportation tools, the transportation industry's contribution to factors on traffic carbon emissions from the perspective of time and space
environmental pollution is also increasing, which will not only affect (Liu et al., 2012). Zhang et al. pointed out through the expanded STIRPAT
human health and the ecological environment but also hurt the model that the population size and per capita GDP will promote the rise of
development of the transportation industry. Therefore, it is of great traffic carbon emissions, while the energy intensity has a restraining effect
practical significance to predict the peak carbon emissions of the (Zhang et al., 2020). Yan et al. selected urbanization rate, subway
transportation industry and formulate appropriate carbon emission passenger volume and energy consumption per unit GDP as the main
reduction policies. influencing factors of traffic carbon emissions by building the STIRPAT
model (Yan et al., 2020). In the existing research, scholars usually choose
Scholars at home and abroad have carried out a series of research studies factors such as population size, gross national product, energy intensity,
on traffic carbon emissions, among which research on the influencing and energy structure to study (Dong et al., 2020; Liu et al., 2023; Han et al.,
factors and prediction of traffic carbon emissions is a hot spot in this field. 2022; Wang et al., 2021). It can be found that the STIRPAT model has
obvious advantages in selecting factors. Therefore, this paper uses the
In the selection of influencing factors of traffic carbon emissions, scholars STIRPAT model to select the influencing factors of transportation carbon
usually use logarithmic mean decomposition analysis (LMDI), Kaya emissions.
identity, IPAT model, and extensible stochastic environmental impact
assessment model (STIRPAT) to select influencing factors. Solaymani used In the field of transportation carbon emission prediction, scholars at home
the LMDI decomposition method to conduct in-depth research on seven and abroad have conducted much research. The prediction models mainly
major carbon emission countries and analyzed the driving factors of include long-term energy alternative planning system (LEAP),
carbon emissions in the transportation sector (Solaymani, 2019). Jia et al. autoregressive moving average model (ARIMA), STIRPAT model, long and
decomposed the factors of transportation carbon emissions in Hebei short-term memory model (LSTM), BP neural network, etc. Fang et al.
Province through the improved Kaya identity and calculated the predicted the carbon emissions in the field of transportation in Hunan

Quick Response Code Access this article online

Website: DOI:
[Link] 10.26480/gws.01.2024.13.21

Cite The Article: Sheng Kai, Ye Shanli (2024). Prediction of Carbon Emissions from Transportation in
China Based on The Arima-Lstm-Bp Combined Model. Science Heritage Journal, 8(1): 13-21.
 Science Heritage Journal (GWS) 8(1) (2024) 13-21

Province from 2022 to 2035 through the leap model (Fang et al., 2023). two account for a relatively small proportion, and most of the energy
The results showed that carbon emissions were expected to peak in 2033 consumption of the storage and post and Telecommunications industries
and 2029 under the low-carbon and enhanced low-carbon scenarios. is used for transportation, the consolidated data can also reflect the level
Hassouna et al. used the ARIMA model to predict China's aviation carbon of energy consumption in transportation to a large extent, so this study
emissions (Hassouna et al., 2020). Bian et al. estimated the future carbon uses this consolidated data to measure the total carbon emissions
emissions of the Qinghai Province's transportation system Based on the generated by energy consumption in transportation. The carbon emission
STIRPAT model simulation results. They believed improving the technical factors of different energy sources used in the accounting are determined
level was the key to reducing carbon emissions from transportation in by the correlation coefficients published by the IPCC (United Nations
Qinghai Province (Bian et al., 2019). In recent years, scholars have Intergovernmental Panel on Climate Change) and the National
gradually applied machine learning methods to predict carbon emissions Development and Reform Commission. The calculation results of China's
from transportation. Ağbulut used an artificial neural network model transportation carbon emissions and per capita transportation emissions
(ANN), support vector machine (SVM), and deep learning (DL) to predict from 1990 to 2021 are shown in Figure 1.
the carbon emissions of transportation in Turkey. The research showed
that the three methods had achieved ideal results (Ağbulut, 2022). Huang The results show that from 1990 to 2021, China's total carbon emissions
et al. used LSTM to predict China's carbon emissions (Huang et al., 2019). from transportation showed an upward trend. By 2021, the carbon
Pan et al. established a BP neural network model and predicted CO2 emissions from transportation were 946.667 million tons, with an average
emissions in Gansu Province from 2021 to 2030, with a prediction error of annual growth rate of 2.87%. At the same time, China's per capita
2 × 10-4, with high accuracy (Pan et al., 2023). Zhao et al. combined the transportation carbon emissions are also rising yearly. The annual
lasso regression model with the BP neural network to predict the carbon increase in total and per capita carbon emissions shows that the
emissions of Henan Province from 2021 to 2035 (Zhao et al., 2022). Hu et transportation industry has a high potential for low-carbon emission
al. predicted China's carbon emission intensity based on the ARIMA-BP reduction. Therefore, predicting transportation carbon emissions is
model, and the results showed that the prediction effect was excellent (Hu significant to realizing China's dual carbon goal.
et al., 2022). Shao et al. constructed the grey neural network combination
model of the grey prediction model BP neural network model
combination. They verified that the combination model could improve the
model's prediction accuracy (Shao et al., 2016).

There has been much research on predicting carbon emissions from

transportation, and the research and improvement methods have their
characteristics. In this paper, the ARIMA model is selected to better fit the
linear characteristics in the prediction time series, LSTM has obvious
advantages for long-term prediction, the BP neural network has self-
adaptive solid ability and can deal with nonlinear characteristics well
(Zhang et al., 2022).

Now, ARIMA, LSTM and BP models are combined, and a weighted

combination model based on ARIMA-LSTM-BP is proposed. First, the
ARIMA model is used to fit the linear characteristics in the sequence. Then,
the LSTM neural network is used to correct the ARIMA model prediction
residual to fit the nonlinear characteristics in the sequence, and then
combined with the BP neural network prediction results, the error (a) Carbon emissions and per capita carbon emissions
reciprocal method is used to determine the weight for combined
prediction to reduce the error. Finally, combined with the scenario
analysis method, this paper predicts the changing trend of China's
transportation carbon emissions from 2022 to 2050 under the three
baseline scenarios: low-carbon and high-carbon.

2. DATA AND METHODS

2.1 Carbon Emission Measurement Of China's Transportation
Industry

Regarding measuring carbon dioxide emissions from transportation, the

IPCC 2006 guidelines for national greenhouse gas inventories provide
measurement methods based on two perspectives. The first is the "top-
down" method, which calculates the carbon emissions of the
transportation industry according to the conversion coefficient of energy
consumption and vehicles' energy carbon emission coefficient. The second
is the "bottom-up" method, which calculates the energy consumption of
(b) Carbon emissions
the transportation industry according to the data of different types of
vehicles and the fuel consumption per unit mileage to calculate the carbon
dioxide emissions (Fang et al., 2017). This study simulates the future trend
of carbon emissions in the transportation industry through prediction, so
the "top-down" method is adopted for calculation. The calculation formula
is shown in formula (1).
44
𝐶𝑂2 = ∑𝑖 𝐶𝑖 = ∑𝑖 𝐸𝑖 × 𝐹𝑖 = ∑𝑖 (𝐸𝑖 × 𝑁𝐶𝑉𝑖 × 𝐶𝐶𝑖 × 𝐶𝑂𝐹𝑖 × ) (1)
12

Where C is the carbon emission of the transportation industry; i refers to

fuel type (such as gasoline, diesel, fuel oil, etc.); Ci is the amount of carbon
dioxide produced by the ith energy consumption; Ei is the consumption of
the ith energy; Fi is the carbon emission coefficient of the ith energy; Ncvi
is the average low calorific value of the ith energy source; CCi is the carbon
content per unit calorific value of the ith energy; COFI is the carbon
oxidation rate of the ith energy; 44/12 is the conversion coefficient
between carbon and carbon dioxide.

At present, the energy consumption data of the transportation industry (c) Per capita carbon emissions
are not counted separately in the China Energy Statistical Yearbook and
the China Statistical Yearbook, but are combined with the transportation, Figure 1: Emissions from China's transportation industry and per capita
storage, and post and telecommunications industries. Because the latter carbon emissions from transportation.

Cite The Article: Sheng Kai, Ye Shanli (2024). Prediction of Carbon Emissions from Transportation in
China Based on The Arima-Lstm-Bp Combined Model. Science Heritage Journal, 8(1): 13-21.
 Science Heritage Journal (GWS) 8(1) (2024) 13-21

2.2 Selection Of Influencing Factors transportation efficiency, travel intensity of urban residents and per
capita economic activity level, will affect the carbon emissions of the
The IPAT model first proposed comprehensively reflects the relationship transportation industry to varying degrees (Yang, 2023). Chen et al.
among population, per capita wealth, technology and environment. It is a selected seven indicators such as vehicle ownership, carbon emission
widely recognized model, and its general form is: intensity and urbanization rate as the influencing factors of regional traffic
carbon emissions in China (Chen et al., 2018). Liu et al. found in the
𝐼 =𝑃×𝐴×𝑇 (2) research on the influencing factors of carbon emissions from the
Where I refers to environmental pollution; P is the population; A is GDP transportation industry in a province in Southwest China that the
per capita; T indicates technical status. population, per capita GDP and vehicle ownership are the main reasons
for the increase of carbon emissions from transportation in the province
STIRPAT model is proposed based on the IPAT model, which can better (Liu et al., 2023).
explore the relationship between various influencing factors and carbon
emissions. Its standard form is (Jiao et al., 2012): After summarizing the existing literature, this paper expands the basic
STIRPAT model and selects six influencing factors, including population,
𝐼 = 𝑎𝑃 𝑏 𝐴𝑐 𝑇 𝑑 𝑒 (3) GDP, energy intensity, urbanization rate, vehicle ownership and passenger
turnover, to establish the extended STIRPAT model of the transportation
Where, I represents environmental factors; a represents model coefficient; industry. The model expression is shown in formula (4):
P、A and T respectively represent the population, financial status and
scientific development technology; b、c and d respectively represent the 𝐶 = 𝑎 × 𝑃 𝑏 × 𝐴𝑐 × 𝑇 𝑑 × 𝑈𝑒 × 𝐶𝑃 𝑓 × 𝑅𝑃 𝑔 (4)
changes of these reference values; e is random error.
Where C is the total carbon emission; P stands for population (ten
Gu found that clean energy, transportation energy intensity, freight thousand); A stands for GDP (trillion yuan); T stands for energy intensity
turnover, passenger turnover, urbanization level, per capita GDP and (tons of standard coal/ ten thousand yuan); U represents urbanization
other factors significantly impact the transportation industry's carbon rate (%); CP stands for vehicle ownership (billion); RP represents
emissions (Gu, 2023). Zuo et al. summarized the existing research and passenger turnover (billion person kilometers).
found that many complex factors affect transportation's carbon emissions,
which can be summarized into five aspects: transportation technology and When measuring the level of the above six influencing factors, GDP adopts
management, economic development, residents' travel behavior, urban the constant price in 1990 to consider the actual GDP of each year to avoid
morphology and environmental factors (Zuo et al., 2023). Liu et al. the impact of price. The above-detailed data are mainly from each year's
believed that the per capita level of economic activity was crucial in China Statistical Yearbook, and the descriptive statistical results of each
improving carbon emissions in the transportation industry (Liu et al., variable are shown in Table 1.
2015). Yang found that many variables, including energy intensity,

Table 1: Descriptive Statistics of Variables

Variable Unit Minimum Maximum Mean Standard deviation
population ten thousand 114333 141389 130221.44 7978.01
GDP Trillion yuan 1.89 54.16 20.05 15.99
urbanization rate % 26.41 64.19 44.19 12.28
Energy intensity Tons of standard coal/10000 yuan 0.51 1.56 0.88 0.27
Vehicle ownership Billion 0.15 3.79 1.55 1.17
Passenger turnover billion person kilometers 5628.40 35349.24 18946.32 9554.26

𝑝 𝑞
Pearson correlation is used to analyze the correlation between various 𝑌𝑡 = 𝜇 + ∑𝑖=1 𝛼𝑖 𝑌𝑡−1 + ∑𝑖=1 𝜃𝑖 𝜀𝑡−1 + 𝜀𝑡 (5)
influencing factors and carbon emissions from transportation, and the
correlation coefficient is calculated. The analysis results are shown in Where Yt is the data involved in the calculation at time t; 𝜇 is a constant; p
Table 2. It can be seen from the table that there is a significant positive is the number of autoregressive terms; q is the number of moving average
correlation between population, GDP, urbanization rate, vehicle items; αi is the data autocorrelation coefficient; θi is the error
ownership, passenger turnover and transportation carbon emissions, and autocorrelation coefficient; εt is the white noise sequence.
a significant negative correlation between energy intensity and
transportation carbon emissions. These six factors are strongly correlated The modeling steps of the ARIMA model are as follows: 1) check the
with transportation carbon emissions. stationarity of the time series, and if the series fails the test, carry out the
difference operation. In general, the first-order difference can get the
Table 2: Correlation Analysis Of Influencing Factors stationary series, and the high order of the difference will lead to the loss
Influencing factors Transportation carbon emissions of data information, which is not conducive to the establishment of the
model; 2) Draw the autocorrelation function diagram and partial
population 0.988*** autocorrelation function diagram to find the possible value range of
GDP 0.991*** parameters p and q, and then determine the optimal combination
of p and q in the model through Akaike information criterion (AIC) and
urbanization rate 0.988***
Bayesian information criterion (BIC). The more pronounced the AIC and
energy intensity -0.864*** BIC values are, the better the model effect is; 3) Check the residual error
vehicle ownership 0.998*** to see if it conforms to the normal distribution. If it does not conform to
the normal distribution, determine the p and q values and the order again;
passenger turnover 0.843*** 4) The prediction and error analysis are carried out after the test.
1 Note: * * *, ***, * respectively represent 1%, 5% and 10% significance 2.4. Long Short-Term Memory
levels
Long short-term memory networks (LSTM) is a variant of recurrent neural
2.3 ARIMA Model networks (RNN) proposed by Hochreiter et al., which can remember long
and short-term information (Hochreiter et al., 1997). The model solves the
The autoregressive moving average model is a time series prediction and
problem that RNN has limited processing ability for too long or too short
analysis method widely used in time series data analysis (Wang et al.,
sequences, overcomes the influence of gradient disappearance in the
2022). The model is expressed as ARIMA (p, d, q), where the parameters
p, d and q represent the structure of the prediction model, AR represents process of RNN training, and has good processing ability for time series
autoregression, and p is the number of autoregressive terms; MA stands data. Compared with RNN, each neuron in the LSTM model contains three
for moving average, q is the number of moving average items; d represents basic cell gates, which are divided into input gate it, output gate Ot and
the difference order of the stationary sequence, and the calculation forgetting gate ft according to different functions. The cell structure of
formula is shown in formula (5): LSTM is shown in Figure 2.

Cite The Article: Sheng Kai, Ye Shanli (2024). Prediction of Carbon Emissions from Transportation in
China Based on The Arima-Lstm-Bp Combined Model. Science Heritage Journal, 8(1): 13-21.
 Science Heritage Journal (GWS) 8(1) (2024) 13-21

Figure 2: LSTM Hidden Layer Structure Diagram.

The steps of LSTM network prediction are as follows: Step 3: Calculate the output gate. The output gate determines how much
information is extracted from the memory unit and uses the following step
Step 1: Calculate the forgetting gate. The forgetting gate controls the to predict. First, control how many memory units will be output according
information deleted or forgotten in the memory unit. ht-1 represents the to the sigmoid function, then use the tanh function to scale the value,
hidden layer information of the previous time, and Xt represents the data calculate the value to be output, and finally multiply the two results to
input at the current time. The amount of information to be forgotten and determine the output information ht. The calculation formula is as follows:
retained through the forgetting gate can be determined. The calculation
formula is as follows: 𝑂𝑡 = 𝜎(𝑊𝑜 × [ℎ𝑡−1 , 𝑋𝑡 ] + 𝑏𝑜 ) (10)

𝑓𝑡 = 𝜎(𝑊𝑓 × [ℎ𝑡−1 , 𝑋𝑡 ] + 𝑏𝑓 ) (6) ℎ𝑡 = 𝑂𝑡 × tanh(𝐶𝑡 ) (11)

Where ft is the output value of the forgetting gate at time t, σ is the sigmoid Where Ot is the output information of the output gate at time t; Wo is the
activation function, Wf is the weight matrix of the forgetting gate, and bf is weight; bo is the offset coefficient of the output gate; σ activates the
the bias coefficient of the forgetting gate. function for sigmoid.

Step 2: Calculate the input gate. The input gate controls the amount of new 2.5 Back-Propagation Neural
information flowing into the control unit. First, the sigmoid activation
function takes the input data and the output of the previous time step as As a widely used depth method, the BP neural network is famous for
the input and then creates a candidate value C through the tanh layer. effectively storing and understanding many data association
Combine these two parts to update the cell state. The calculation formula characteristics. It is a multilayer feedforward neural network trained
is as follows: according to the error backpropagation algorithm. Firstly, the features of
the learning samples are input from the input layer through forward
𝑖𝑡 = 𝜎(𝑊𝑖 × [ℎ𝑡−1 , 𝑋𝑡 ] + 𝑏𝑖 ) (7) propagation, then processed by each hidden layer, and finally output from
the output layer. For the error between the actual output and the expected
𝐶̃𝑡 = tanh(𝑊𝑐 × [ℎ𝑡−1 , 𝑋𝑡 ] + 𝑏𝑐 ) (8) output of the network, the weight and threshold of the model need to be
continuously adjusted by backpropagation, and the gradient descent
Where it is the output information of the input gate at time t; Wi is the method is used to optimize the model parameters until the final fitting
weight; bi is the input gate offset coefficient; σ is the sigmoid activation value error is the minimum, so as to end the training. A neural network
function; bC is the bias coefficient of the cell state gating unit. generally comprises input, hidden, and output layers. The number of
neurons in each layer varies according to different situations.
The input gate will continuously update the state and update Ct-1 to Ct. The
memory information Ct-1 at the previous time multiplied by ft indicates the Taking the historical value of carbon emissions of China's transportation
information to be discarded and retained. Then, add the candidate value industry every three years as the input, and through continuous training,
𝐶̃𝑡 to obtain the new cell state Ct, and the calculation formula is as follows: a BP neural network prediction model with a topology of 3-6-1, which is
influenced by the historical value, is finally constructed, as shown in Figure
𝐶𝑡 = 𝑓𝑡 × 𝐶𝑡−1 + 𝑖𝑡 × 𝐶̃𝑡 (9) 3. Where Xt-1、Xt-2、Xt-3 is the carbon emission at time t-1、t-2、t-3 , and
Yt is the predicted carbon emission at time t.

Figure 3: BP Neural Network Structure.

Cite The Article: Sheng Kai, Ye Shanli (2024). Prediction of Carbon Emissions from Transportation in
China Based on The Arima-Lstm-Bp Combined Model. Science Heritage Journal, 8(1): 13-21.
 Science Heritage Journal (GWS) 8(1) (2024) 13-21

2.6 Combined Forecasting Model 1

𝑅𝑀𝑆𝐸 = √ ∑𝑛𝑖=1(𝑦𝑖 − 𝑦̂𝑖 )2 (17)
𝑛
In order to solve the problem of the accuracy of a single model not being
high in the prediction and improve the prediction accuracy and 1
𝑀𝐴𝐸 = ∑𝑛𝑖=1|(𝑦𝑖 − 𝑦̂𝑖 )| (18)
𝑛
generalization ability of the model, the single model is combined in
different combinations. The advantages of each model are used to make Where Yi is the real value at time i; 𝑦̂𝑖 is the predicted value at time i; n is
up for the defects of the single model to improve the prediction's stability the number of samples.
and accuracy. This study mainly adopts the combination method based on
residual sequence and error weight. The structure of the combination 3. RESULTS AND DISCUSSION
model is shown in Figure 4.
3.1 ARIMA Model Construction

ADF test is used to judge whether the time series data is stable, and the
result shows that p = 0.152 > 0.05, indicating that the original series is a
non-stationary time series. The difference operation of d = 1 on the
original sequence shows that p = 0.039 < 0.05. The significance test shows
that the first-order difference sequence is stable. Therefore, the first-order
difference sequence can be modeled.

The autocorrelation function (ACF) and partial autocorrelation function

(PACF) are used to determine the order of the ARIMA model, and the
autocorrelation function diagram and partial correlation function diagram
are drawn. It can be seen from Figures 5 and 6 that the first-order
difference sequence ACF and PACF diagrams fall within the confidence
interval after the second order, so p and q can be taken as 0, 1 and 2.

Figure 4: Structure diagram of composite model.

Combination method based on residual sequence: To compensate for the

disadvantage that the ARIMA model is challenging to extract nonlinear
components of time series, the ARIMA-LSTM combination model is
established by combining the advantages of the ARIMA model and LSTM
neural network model. Firstly, the ARIMA model is used to predict the
carbon emissions of the transportation industry, and the predicted
value X1 of the ARIMA model is obtained, and the error sequence e is
obtained. Then, taking the error sequence e as the input, the LSTM neural Figure 5: Autocorrelation function diagram.
network is used to predict the error, and the error prediction value e1 is
obtained. Finally, the ARIMA-LSTM combined model prediction value Y1 is
obtained by adding the ARIMA model prediction value and the error
prediction value;

𝑌1 = 𝑋1 + 𝑒1 (12)

Combination method based on error weight: To reduce the prediction

error of the combination model, the model with more minor errors is given
an enormous weight. Firstly, the prediction value Y2 of the BP neural
network model is obtained, and the prediction error e2 of the model is
calculated. Similarly, the prediction error e3 of the ARIMA-LSTM model is
obtained. Secondly, the weight l of the prediction error of the two models
is calculated respectively 𝑙1 、𝑙2 . The predicted value y of the ARIMA-LSTM
-BP combined model is obtained by weighting the predicted value
according to the error weight.
𝑒2
𝑙1 = (13)
𝑒2 +𝑒3

𝑒3
𝑙2 = (14)
𝑒2 +𝑒3

𝑌 = 𝑙1 × 𝑌1 + 𝑙2 × 𝑌2 (15)
Figure 6: Partial autocorrelation function diagram.

2.7 Model Evaluating Indicator The available values of ARIMA (p, d, q) obtained from ACF and PACF charts
in the previous step are selected as the optimal model using AIC, BIC and
This study mainly uses the average absolute percentage error (MAPE), HQIC criteria. The results are shown in Table 3. It can be seen from Table
root mean square error (RMSE), and average absolute error (MAE) to 3 that when p=1, d=1 and q=1, AIC, BIC and HQIC are the smallest,
evaluate the model's prediction results. The smaller the indicators, the indicating that ARIMA (1, 1, 1) has the best fitting effect on the data. After
more accurate the prediction results are. The specific formula is as follows: the ARIMA (1, 1, 1) model is established, the first 23 years of data are used
1 𝑦𝑖 −𝑦̂𝑖
as training data to predict the carbon emissions in the next nine years
𝑀𝐴𝑃𝐸 = ∑𝑛𝑖=1 | | × 100% (16) (2013-2021).
𝑛 𝑦𝑖

Cite The Article: Sheng Kai, Ye Shanli (2024). Prediction of Carbon Emissions from Transportation in
China Based on The Arima-Lstm-Bp Combined Model. Science Heritage Journal, 8(1): 13-21.
 Science Heritage Journal (GWS) 8(1) (2024) 13-21

Table 3: AIC、BIC and HQIC Results. Then, the LSTM output data is de-normalized to obtain the residual
prediction value of the LSTM model. Add the ARIMA model prediction
（p，d，q） AIC BIC HQIC value obtained in 3.1 and the LSTM model residual prediction value to
obtain the ARIMA-LSTM combined model prediction value.
（0，1，0） 305.66 307.09 306.12
（0，1，1） 282.04 284.90 282.97 3.3 Back-Propagation Neural

（0，1，2） 280.63 284.93 282.03 First, normalize the data, and select the first 23 years of the data set as the
training set of the model, and the last 9 years of the data set as the test set.
（1，1，0） 264.09 266.96 265.02
The normalized historical carbon emissions every three years are used as
（1，1，1） 261.46 265.76 262.86 input, and the carbon emissions affected by historical values are used as
output. Levenberg-Marquardt backpropagation algorithm is selected as
（1，1，2） 262.75 268.48 264.62 the BP neural network training algorithm; Set the maximum number of
（2，1，0） 264.06 268.36 265.46 training times as 1000, determine the learning rate as 0.01; Set the number
of hidden layer neurons to 1-20, and finally get the BP neural network
（2，1，1） 262.96 268.69 264.83 model with the optimal number of hidden layer neurons of 6.
（2，1，2） 262.03 269.20 264.37 3.4 ARIMA-LSTM-BP Combined Model

3.2 ARIMA-LSTM Combined Model According to the predicted values of the ARIMA-LSTM combined model
and BP neural network model obtained in 3.1, 3.2 and 3.3, the model is
The residual sequence of 23 years from 1990 to 2012 is selected as the recombined using the combination method based on error weight to
training sample, and the residual sequence of 9 years from 2013 to 2021 obtain the predicted results of the ARIMA-LSTM-BP combined model.
is selected as the test sample. The maximum minimum method is used to
normalize it, and the residual predicted by the ARIMA model is converted 3.5 Analysis of Prediction Results
into a value between 0 and 1. Set the data set time step to 3 (that is, use
the carbon emissions of the previous three years to predict the carbon In order to verify the prediction accuracy and effectiveness of the
emissions of the following year), set the number of hidden layers of the combined model for carbon emissions from the transportation industry,
LSTM neural network to 1, and the number of training rounds is the ARIMA model, LSTM model, BP neural network model, ARIMA-LSTM
epoch=600. After repeated training, we can get the best number of training model and ARIMA-LSTM-BP combined model were respectively
times and the number of hidden nodes. Among them, Batch_Size=9, constructed for comparison. The comparison between the predicted value
Hidden_Size=9. The loss function is the mean square error function MAE, and the actual value of each model is shown in Figure 7, and the
and the model optimizer selects Adam. comparison results of the evaluation indicators of each model are shown
in Table 4:

Figure 7: Comparison between predicted and actual values of various models.

Table 4: Comparison Of Evaluation Indicators Of Various Models. in the prediction model are taken as the research object, and three
scenarios of high carbon, benchmark, and low carbon are set. The
Model RMSE MAE MAPE prediction period of scenario analysis is 2022-2050. Considering the long
ARIMA 52.75 41.11 4.65% prediction period, in order to ensure the rationality and scientificity of the
prediction results, three prediction frequency bands will be set for each
LSTM 24.99 18.53 2.11% influencing factor, as shown in Table 5.
BP 24.41 21.36 2.53%
Population. Historical data shows that China's population growth has
ARIMA-LSTM 18.92 16.90 2.01% slowed significantly in the past six years. From 2016 to 2021, the
ARIMA-LSTM-BP 13.04 12.12 1.42% population increased by only 21.57 million, with an average annual
growth rate of about 0.25%. The seventh national census bulletin shows
that China's total fertility rate is only 1.3, and the trend of negative
It can be seen from the chart that the ARIMA-LSTM-BP combined model
population growth is visible to the naked eye. According to the Research
has the best prediction accuracy and the most stable prediction result. It
Report on the national population development strategy, China's
shows that the combined model can thoroughly combine the advantages
population will peak at 1.442 billion in 2028, showing a negative growth
of a single model and learn from each other. Therefore, the ARIMA-LSTM-
trend after that. Therefore, the growth rate of the population in different
BP combined model can be used to predict the future carbon emissions of
periods under the benchmark scenario is set to be 0.15% from 2022 to
China's transportation industry.
2028, -0.1% from 2029 to 2038 and -0.25 from 2039 to 2050.
3.6 Multi Scenario Prediction
GDP. In 2021, facing the double test of the ups and downs of the epidemic
3.6.1 Scenario Settings situation in the century and the complex and severe external environment,
China effectively coordinated the epidemic prevention and control and
Through the development of the transportation industry in China and economic and social development, strengthened the cross-cycle
referring to a large number of policy documents, the six influencing factors adjustment of macro policies, and the national economy continued to

Cite The Article: Sheng Kai, Ye Shanli (2024). Prediction of Carbon Emissions from Transportation in
China Based on The Arima-Lstm-Bp Combined Model. Science Heritage Journal, 8(1): 13-21.
 Science Heritage Journal (GWS) 8(1) (2024) 13-21

recover, with an increase of 7.81% over the previous year at constant compared with 2021. From 2021 to 2050, energy intensity decreased by
prices. The research on China's long-term low-carbon development about 3.5% annually. Hu et al. pointed out that with the increasing
strategy and transformation path mentioned that during the "14th five- difficulty of technical emission reduction, the energy consumption growth
year plan" period, the average annual GDP growth rate will exceed 5.00%. rate per unit transportation turnover will slow down (Hu et al., 2022).
The "14th five-year plan" and the proposal of the long-term goal of 2035 Based on this view and the analysis of the actual situation, the growth rate
also clearly put forward the goal of doubling the economic aggregate by of energy intensity in different periods under the benchmark scenario is
2035 compared with 2020. Using the quantitative economic model, Xiao set as -3% from 2022 to 2028, -4% from 2029 to 2038 and -4.5% from
estimated that the average annual growth rate of China's GDP will reach 2039 to 2050.
about 5.3% from 2023 to 2025 (Xiao, 2023). Therefore, the growth rate of
GDP in different periods under the benchmark scenario is set to be 5% Vehicle ownership. With the continuous development and expansion of
from 2022 to 2028, 4% from 2029 to 2038, and 3.5% from 2039 to 2050. the overall urban scale in China, the demand for automobiles continues to
grow. By 2021, the number of motor vehicles was 394 million, an increase
Urbanization rate. In recent years, the level of urbanization in China has of 104 million compared with 2016, with an average annual growth of
developed rapidly. In 2021, the urbanization rate reached 64.19%, and the about 4.4%. Zhou et al. predicted that by 2028, the number of motor
urbanization rate of developed countries in the world has stabilized at vehicles would be about 440 million (Zhou et al., 2023). Therefore, the
about 85%. Ou et al. have studied China's urbanization rate, which is growth rate of vehicle ownership in different periods under the
expected to reach about 67.45% by 2025 (Ou et al., 2021). Hu predicted benchmark scenario is set to be 4% from 2022 to 2028, 3% from 2029 to
that China's urbanization rate would reach 73.41% -74.53% in 2035 2038, and 1.5% from 2039 to 2050.
through multiple model screening. It is expected to reach 80% in 2050
(Hu, 2023). Therefore, the growth rate of urbanization rate in different Passenger turnover. In the ten years before the outbreak, from 2010 to
periods under the benchmark scenario is set to be 1.1% from 2022 to 2019, the average annual increase in passenger turnover was 2.11%.
2028, 0.7% from 2029 to 2038, and 0.35% from 2039 to 2050. Affected by the epidemic, compared with 2019, the passenger turnover in
2020 and 2021 decreased significantly to 1.29 trillion person kilometers.
Energy intensity. In recent years, China's energy intensity has shown a However, with the epidemic's end, passenger turnover will gradually
downward trend, decreasing by 17.65% from 2016 to 2021, with an warm up, and it is expected to return to its average level in 2028.
average annual decline of about 3%. The comprehensive work plan for Therefore, the growth rate of passenger turnover in different periods
energy conservation and emission reduction during the 14th five-year under the benchmark scenario is set to be 10% from 2022 to 2028, 2.5%
plan puts forward the goal of reducing energy intensity by 10% in 2025 from 2029 to 2038, and 1.5% from 2039 to 2050.

Table 5: Scenario Settings For Various Parameters In The Year.

Growth rate of influencing factors
Scene Time
Population GDP Urbanization rate Energy intensity Vehicle ownership Passenger turnover
2022-2028 0.25% 5.5% 1.3% -2% 5% 11.5%
high
2029-2038 0.05% 4.5% 0.9% -3% 4% 3.5%
carbon
2039-2050 -0.15% 4% 0.55% -3.5% 2.5% 2.5%
2022-2028 0.15% 5% 1.1% -3% 4% 10%
benchmark 2029-2038 -0.1% 4% 0.7% -4% 3% 2.5%
2039-2050 -0.25% 3.5% 0.35% -4.5% 1.5 1.5%
2022-2028 0.05% 4.5% 0.9% -4% 3% 8.5%
low carbon 2029-2038 -0.2% 3.5% 0.5% -5% 2% 1.5%
2039-2050 -0.35% 3% 0.15% -5.5% 0.5% 0.5%

3.6.2 Multi Scenario Prediction and Analysis of 678.1062 million tons compared with 2021, and 125.7585 million tons
in 2050, 1.32 times higher than 2021. Under the benchmark scenario, the
Based on good fitting and validity test results of the ARIMA-LSTM-BP transportation industry will reach a peak of 1478169400 tons in 2034, an
combined model, carbon emissions from 2022 to 2050 are predicted. The increase of 531.524 million tons compared with 2021, and will drop to
prediction results are shown in Figure 8. 113.7572 million tons in 2050, 1.20 times higher than 2021. Under the
low-carbon scenario, the transportation industry will peak at 1367.5417
The prediction results show that the peak value of carbon emissions from million tons in 2031, an increase of 420.8747 million tons compared with
transportation in China is different under the three scenarios of high 2021, and drop to 980.7545 million tons in 2050, 1.03 times higher than
carbon, benchmark and low carbon, and the time to reach the peak value 2021. By comparing the above three scenarios, in order to better keep up
is also different, which is 2031, 2034 and 2039 respectively. The with the country's goal of reaching the carbon peak by 2030, we should
transportation industry alone cannot achieve the goal of reaching the peak choose the low-carbon development scenario. Compared with the baseline
carbon by 2030. Under the high carbon scenario, the transportation and high-carbon scenarios, the time of reaching the carbon peak has
industry will reach a peak of 1624.7732 million tons in 2039, an increase advanced by 3 and 8 years, respectively.

Figure 8: Prediction Results of Carbon Emissions from Transportation in China.

Cite The Article: Sheng Kai, Ye Shanli (2024). Prediction of Carbon Emissions from Transportation in
China Based on The Arima-Lstm-Bp Combined Model. Science Heritage Journal, 8(1): 13-21.
 Science Heritage Journal (GWS) 8(1) (2024) 13-21

4. CONCLUSIONS AND SUGGESTIONS algorithms. Sustainable Production and Consumption, 29(1), Pp. 141-
157.
4.1 Conclusions
Bian, L.; Ji, M. 2019. Research on Influencing Factors and Prediction of
This study follows the logic of "Measurement Research - influencing Traffic Carbon Emissions in Qinghai. Ecological Economy, 35 (2), Pp.
factors - prediction model selection - scenario analysis" to conduct a 35–39 (in chinese).
specific study of China's transportation carbon emissions and draws the
following conclusions: Commoner, B., 1990. Making Peace with the Planet. Pantheon Books, New
York.
Firstly, from 1990 to 2021, China's total emissions from transportation
showed a gradual upward trend. By 2021, the total carbon emissions were Chen, L., Wang, J., He, T., et al. 2018. Research on Prediction of Regional
946.6672 million tons, with an average annual growth rate of 7.37%. Traffic Carbon Emissions Based on SVR. Transp. Syst. Eng. Inform.,
18(2), Pp. 14–19 (in chinese).
Secondly, compared with ARIMA, LSTM, BP, and ARIMA-LSTM, the ARIMA-
LSTM-BP combined model shows better prediction ability, higher Dietz, T., Rosa, E.A., 1997. Effects of Population and Affluence on CO2
accuracy, and more robust stability. Emissions. PNAS, 94(1), Pp. 175–179.

Thirdly, the ARIMA-LSTM-BP combined model is used to predict the Dong, M., Han, Z., Guo, J., 2020. Carbon Emission Efficiency Measurement
carbon emissions of transportation in the three scenarios. It can be seen and Influencing Factors Analysis of China's Marine Transportation
that the carbon peak time is 2031, 2034, and 2039, respectively, and the Industry. Ocean Bull., 39, Pp. 169–177 (in chinese).
peak value is 1624773200 tons, 1478169400 tons, and 1367541700 tons.
Fang, X., Luo, Y., 2017. Comparative Study on Calculation Methods of Urban
Fourthly, if the transportation industry wants to achieve the goal of a Traffic Carbon Emissions. Transportation Energy Conservation and
carbon peak by 2030, it also needs to strengthen low-carbon measures and Environmental Protection, 13(4), Pp. 81–83 (in chinese).
promote low-carbon emission reduction.
Fang, H., Liu, C., Jiang, K., et al. 2023. Research on the Peak Path of Carbon
4.2 Suggestions Emissions in the Field of Transportation in Hunan Province. Transp.
Syst. Eng. Inf., 23 (04), Pp. 61–69 (in chinese).
Based on the above research and scenario analysis, in order to achieve the
carbon peak of transportation in China as soon as possible, the following Gu, D., 2023. Regional Correlation Analysis of Carbon Emissions from
suggestions are put forward: China's Transportation Industry. J. Shanghai Marit. Univ., 44(03), Pp.
64–70+118 (in chinese).
Firstly, optimize the energy structure. China's transportation industry is
still an oil-based energy structure, but if we want to achieve a carbon peak Hochreiter, S., Schmidhuber, J., 1997. Long Short-Term Memory. Neural
as soon as possible, optimizing the energy structure is the only way. Computation, 9, Pp. 1735–1780.
Therefore, the government should increase financial subsidies to support
the development of new clean energy for the transportation industry. Huang, Y.S., Shen, L., Liu, H., 2019. Grey Relational Analysis, Principal
Actively promote new energy vehicles, reduce energy consumption, and Component Analysis and Forecasting of Carbon Emissions Based on
vigorously develop "zero emission" and low-energy trams. Long Short-Term Memory in China. Journal of Cleaner Production, 209
(10), Pp. 415–423 (in chinese).
Secondly, improve the quality of the population. To improve the
population's quality and promote social technology development, we must Hassouna, F., Alsahili, K., 2020. Environmental Impact Assessment of the
strengthen energy management, vigorously promote a conservation- Transportation Sector and Hybrid Vehicle Implications in Palestine.
oriented economy, strengthen the publicity and education of the Sustainability, Pp.12(19).
environment, and popularize a conservation-oriented culture to create a
healthy, safe, and comfortable ecological civilization. In order to protect Han, R.; Li, L.; Zhang, X.; et al. 2022. Spatial-Temporal Evolution
our environment, we must actively promote resource conservation and Characteristics and Decoupling Analysis of Influencing Factors of
encourage people to take adequate measures in daily life to reduce China’s Aviation Carbon Emissions. Chinese Geographical Science,
dependence on resources to achieve sustainable development. 32(2), Pp. 218–236 (in chinese).

Thirdly, improve the level of technological innovation. In order to achieve Hu, J., Luo, Z., Li, F., 2022. Prediction of China's Carbon Emission Intensity
green and sustainable development, we should strengthen technological under the "Carbon Peak" Target: Analysis Based on LSTM and ARIMA-
innovation, improve energy efficiency, vigorously promote the low-carbon BP Model. Financial Science, (2), Pp. 89–101 (in chinese).
circular economy, and achieve green and sustainable economic
development to save energy, reduce pollution, and reduce emissions. Hu, M., Zhen, Y., Li, Y., 2022. Research on Prediction of Peak Carbon
Emissions from Transportation in Hubei Province Under Multi
Through the above suggestions, it is hoped that the transportation carbon Scenario. J. Environ. Sci., 42, Pp. 464–472 (in chinese).
peak can be achieved as soon as possible and that the national goal of
reaching the carbon peak by 2030 can be achieved. Hu, A., 2023. China's Urbanization Rate and the Main Spatial Form of Urban
Agglomeration in 2035. Technical Economy, 42(05), Pp. 174-188 (in
AUTHOR CONTRIBUTIONS chinese).
S.K.: methodology, software, formal analysis, writing—original draft. Y.S.: Jiao, W., Chen, X., 2012. Environmental Impact Analysis of Gansu Province
conceptualization, investigation, validation, supervision. The statement Based on STIRPAT Model -- Taking Energy Consumption from 1991 to
that all authors have approved the final article should be true and included 2009 as an Example. Resources and Environment in the Yangtze River
in the disclosure. Basin, 21 (01), Pp. 105–110 (in chinese).
FUNDING Jia, J., 2020. Research on Influencing Factors and Countermeasures of
Transportation Carbon Emissions in Hebei Province Based on GRA.
This research was funded by the Natural Science Foundation of Zhejiang
Hebei Enterprise, (9), Pp. 72–73 (in chinese).
Province (Grant No. LY23A010003).
Liu, Z., Guan, D., Wei, W., et al., 2015. Reduced Carbon Emission Estimates
DECLARATION OF INTEREST from Fossil Fuel Combustion and Cement Production in China. Nature,
The authors declare no conflict of interest. 524, Pp. 335–338 (in chinese).

REFERENCES Liu, J., Li, S., Ji, Q., 2021. Regional Differences and Driving Factors Analysis
of Carbon Emission Intensity from Transport Sector in China. Energy,
Andrés, L., Padilla, E., 2018. Driving factors of GHG emissions in the EU 224 (in chinese).
transport activity. Transport Policy, 61, Pp. 60-74.
Liu, C., Qu, J., Ge, Y., 2023. Carbon Emission Prediction of China's
Ağbulut, Ü., 2022. Forecasting of transportation-related energy demand Transportation Industry Based on LSTM Model. China Environmental
and CO2 emissions in Turkey with different machine learning Science, 43 (05), Pp. 2574–2582 (in chinese).

Cite The Article: Sheng Kai, Ye Shanli (2024). Prediction of Carbon Emissions from Transportation in
China Based on The Arima-Lstm-Bp Combined Model. Science Heritage Journal, 8(1): 13-21.
 Science Heritage Journal (GWS) 8(1) (2024) 13-21

Liu, J., Huang, F., Chen, B., 2023. Research on Influencing Factors and Yan, S., Chen, L., 2020. Analysis on Influencing Factors of Traffic Carbon
Emission Reduction Strategies of Carbon Emissions in Transportation Emissions: A Case Study of Xi'an. Statistics and Decision Making,
Industry. Highway, 68(05), Pp. 252–259 (in chinese). 36(04), Pp. 62–66 (in chinese).

Ou, Y.H., Li, Z., Li, P.L., 2021. The change trend and policy implications of Yang, J., 2023. Research on Carbon Emission Measurement and Emission
China's urbanization rate during the 14th Five Year Plan period. Urban Reduction Path of China's Transportation Industry. Ph.D. thesis,
Development Research, 28(6), Pp. 1-9 (in chinese). Jiangxi University of Finance and Economics, Jiangxi, China.
Pan, S.Y., Zhang, M.L., 2023. Prediction and Influencing Factors of Carbon Zhang, G., Su, Z., 2020. Decomposition of Influencing Factors and Scenario
Dioxide Emissions in Gansu Province Based on BP Neural Network.
Prediction of Transportation Carbon Emissions in the Yellow River
Environmental Engineering, 41(07), Pp. 61-68+85 (in chinese).
Basin. Management Review, 32(12), Pp. 283–294 (in chinese).
Shao, M., Cheng, T., Ma, X., 2016. Combined Forecasting of Railway Freight
Volume Based on Grey Neural Network. J. Transp. Eng. Inf., 14 (3), Pp. Zhang L., Li P., Pang W., et al. 2022. Precipitation Prediction Based on
129–135 (in chinese). Seasonal Autoregressive Integral Moving Average and Deep Learning
Long-Term and Short-Term Memory Neural Network. Sci. Technol.
Solaymani, S., 2019. CO2 emissions patterns in 7 top carbon emitter Eng., 22(9), Pp. 3453-3463 (in chinese).
economies: The case of transport sector. Energy, 168, Pp. 989-1001.
Zhao, J.H., Li, J.S., Wang, P.L., et al., 2022. Research on the Path of Carbon
Wang, Y., Li, H., Guo, X., et al., 2021. Research on Influencing Factors of Peak in Henan Province Based on Lasso BP Neural Network Model.
Carbon Dioxide Emission from Railway Operation in China. Journal of Environmental Engineering, 40 (12), Pp. 151-156,164 (in chinese).
Railways, 43 (6), Pp. 189-195 (in chinese).
Zhou, J., Li, N., Feng, W., et al., 2023. Estimation of Carbon Dioxide
Wang, X., Li, A., Li, Y., et al., 2022. Integrated Energy System Load and Wind Emissions from Road Sources and Prediction of Future Scenarios in
and Solar Resources Prediction Based on ARIMA-LSTM Model. J. Xi'an China. J. Environ. Sci., Pp. 1-12 (in chinese).
Univ. Archit. Technol. (Nat. Sci. Ed.), 54(05), Pp. 762-769 (in chinese).
Zuo, D., Zhao, L., Xiong, Q., et al., 2023. Review of Research on Traffic
Xiao, H., 2023. Progress Evaluation and Prospect of Main Indicators in the Carbon Emissions: Accounting Methods, Influencing Factors and
Outline of the 14th Five-Year Plan. Economic Aspect, (07), Pp. 27-33 Mechanism. J. Transp. Eng. Inf., Pp. 1-22 (in chinese).
(in chinese).

Cite The Article: Sheng Kai, Ye Shanli (2024). Prediction of Carbon Emissions from Transportation in
China Based on The Arima-Lstm-Bp Combined Model. Science Heritage Journal, 8(1): 13-21.