Received January 9, 2020, accepted January 19, 2020, date of publication January 22, 2020, date of current version

January 31, 2020.

Digital Object Identifier 10.1109/ACCESS.2020.2968585

Hybrid Prediction Model for the Interindustry

Carbon Emissions Transfer Network Based on
the Grey Model and General Vector Machine
YING HU 1 AND KANGJUAN LV 2
1 School
of Economics, Shanghai University, Shanghai 200444, China
2 SHU-UTS SILC Business School, Shanghai University, Shanghai 201800, China

Corresponding author: Kangjuan Lv (lvkangjuan@shu.edu.cn)

This work was supported in part by the National Natural Science Foundation of China under Grant 71774108.

ABSTRACT Through analysis of the carbon emissions transfer network formed by the exchange of
intermediate products among industries, we can promote the realization of national carbon emissions
reduction goals. Therefore, it is of great significance to build a prediction model of the carbon emissions
transfer network for more accurate predictions. According to the characteristics of the random oscillation
sequence (ROS) of interindustry carbon emissions transfer, a hybrid prediction model denoted as the ROGM-
AFSA-GVM is proposed based on the grey model (GM) for ROS and the general vector machine (GVM)
optimized by the artificial fish swarm algorithm (AFSA). The proposed model uses the ROGM model to
predict the general ROS trend and relies on the AFSA-GVM model to predict the nonlinear law of ROS. The
predicted values of the two parts are combined to obtain predicted interindustry carbon emissions transfer
values. The proposed model is used to simulate the interindustry carbon emissions transfer network of
China. The simulation results show that the ROGM-AFSA-GVM model can effectively resolve the prediction
problem of ROS. Comparing the predicted networks with the actually measured networks, it is verified that
the proposed model is suitable for simulating the interindustry carbon emissions transfer network and has a
good prediction performance.

INDEX TERMS Hybrid prediction model, grey model, general vector machine, artificial fish swarm
algorithm, carbon emissions transfer network.

I. INTRODUCTION
To mitigate global climate change, it is necessary to control
the carbon emissions of key industries, such as the power,
steel, building materials and chemical industries [1]. Dif-
ferent industries are interconnected in the processes of pro-
duction, distribution and exchange. A network is formed in
terms of the flow and transfer of carbon emissions, which
we call the interindustry carbon emissions transfer network.
To effectively set emissions reduction targets and achieve
effective control of the network, it is very important to predict
the amount of carbon emissions transferred among industries.
Many studies have investigated the prediction of carbon
emissions by establishing mathematical models based on the
influencing factors of carbon emissions. Mladenović et al.
The associate editor coordinating the review of this manuscript and
approving it for publication was Rongbo Zhu . indicated that the performance of the PSO-ELM model is
adopted the urban and rural population growth rates and pro-
portions as influencing factors and entered them as input into
the support vector machine (SVM) to predict carbon emis-
sions. Their results shows revealed that the prediction accu-
racy of the proposed method is better than that of the artificial
neural network (ANN) [2]. Sun and Liu chose corresponding
influencing factors according to the different characteristics
of the carbon emissions of three major industries in China,
and used the least squares SVM to predict carbon emissions.
The results showed that the classification prediction of carbon
emissions can greatly improve the prediction accuracy [3].
Sun et al. selected 22 indexes as influencing factors, and
input 8 of them into their proposed PSO-ELM model based
on extreme learning machine (ELM) and particle swarm
optimization (PSO) to predict carbon emissions in Hebei
Province through factor analysis. Their simulation results

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see http://creativecommons.org/licenses/by/4.0/
20616 VOLUME 8, 2020
Y. Hu, K. Lv: Hybrid Prediction Model for the Interindustry Carbon Emissions Transfer Network Based on the GM and GVM
better than those of the ELM model and the BP neural
network [4]. Ding et al. selected the carbon emissions of
fuel combustion as characteristic system behavior variables
and adopted the actual energy consumption, GDP and urban
population as relevant influencing factors to establish a novel
grey multivariable model based on the trends of the driving
variable model (TDVGM (1, 4)) to predict carbon emissions
in China. The results revealed that the new model is more
accurate than the GM (1, 4) and autoregressive moving aver-
age (ARMA) models [5]. Huang et al. selected 15 influencing
factors that have a strong correlation with carbon emissions,
extracted 4 principal components by principal component
analysis (PCA) and used long short-term memory (LSTM)
neural networks to predict carbon emissions of China. The
results showed that the prediction accuracy of carbon emis-
sions based on LSTM is higher than those based on the BP
neural network and Gaussian process regression (GPR) [6].
Acheampong and Boateng chose economic growth, energy
consumption, research and development, financial develop-
ment, foreign direct investment, trade opening, industrial-
ization and urbanization as influencing factors and entered
them as input into an ANN to predict the carbon emissions
intensity in Australia, Brazil, China, India and the United
States [7]. Liu et al. adopted socioeconomic variables and
landscape indicators as influencing factors for support vec-
tor regression (SVR) to predict the carbon emissions of
Guangzhou [8]. In addition, other studies have also pro-
posed panel regression models [9], logical models [10], [11],
STIRPAT models [12], [13], etc.
The above research is usually carried out on the premise
of defining a unified influencing factor among the different
industries. However, the factors that influence the carbon
emissions transfer among industries are complex and diverse.
Moreover, each industry has its own characteristics with
different influencing factors. Therefore, unified influencing
factors may lead to a decrease in prediction accuracy. Carbon
emissions prediction based on carbon emissions time series
can avoid the above problem. Pao et al. used the nonlinear
grey Bernoulli model (NGBM) to predict carbon emissions,
and proposed a numerical iterative algorithm to optimize
NGBM parameters. Compared with the grey model (GM) and
autoregressive integrated moving average (ARIMA) model,
the mean absolute percentage error of the NGBM has been
significantly improved [14]. Fang et al. proposed a PSO-GPR
model based on the GPR method and an improved PSO algo-
rithm and tested the performance of their model using total
carbon emissions data of the United States, China and Japan.
The results revealed that their proposed model improves the
prediction accuracy of the original GPR method and is supe-
rior to other traditional prediction methods, such as the BP
neural network [15]. Xu et al. developed an adaptive GM and
combined it with the buffer rolling method to improve the
prediction accuracy. Compared with the traditional model,
it has an improved adaptability to data characteristics [16].
The above studies are mainly focused on the prediction of
total carbon emissions time series, which has a significant
VOLUME 8, 2020
increasing trend year by year. However, the time series of
interindustry carbon emissions transfer has the characteris-
tics of nonlinear, nonmonotoney, volatility and randomness,
which is called the random oscillation sequence (ROS) [17],
since it is affected by the industry scale, industrial investment,
differences in energy consumption structure, national policies
and other factors. In addition, due to the impact of national
input-output data statistics, the number of time series samples
of interindustry carbon emissions transfer is relatively small.
Prediction of this kind of small-sample ROS has always been
a difficult problem in the field of prediction research.
In recent years, an increasing number of machine learning
methods have been applied to carbon emissions predic-
tion [18], [19]. The grey prediction model can solve the prob-
lem of insufficient sample data to a certain extent, but it has a
poor prediction performance for nonlinear sequences. Based
on amplitude compression, the ROS can be transformed into
a new relatively smooth sequence, and then a GM for ROS
denoted as the ROGM model can be constructed based on
the classic GM model, which can improve the prediction
accuracy of the GM model for ROS to a certain extent.
On the other hand, the general vector machine (GVM) with
a strong nonlinear sequence prediction ability has been pro-
posed recently. The experimental results show that it can still
obtain a prediction model with a good generalization ability
even if using a small number of training samples or even if
there is a lack of training samples [20]–[22]. However, as a
newly proposed prediction model, there is a lack of research
on the application of the GVM and optimization of network
weights to further improve its prediction performance.
Research shows that the hybrid prediction model based on
error correction technology is an effective method to improve
the prediction model performance [23]. In this paper, the GM
and the GVM are adopted for carbon emissions transfer
prediction. First, the parameters of the GVM are optimized by
the artificial fish swarm algorithm (AFSA), hereafter referred
to as AFSA-GVM, which improves the convergence speed
and prediction performance of the GVM. Then, a hybrid pre-
diction model denoted as ROGM-AFSA-GVM is established
based on ROGM and AFSA-GVM for ROS prediction. In this
model, the ROGM (1, 1) model can predict the general trend
of sample data, and the AFSA-GVM model can effectively
predict the nonlinear law of ROS. In the experimental part,
we use data between 1997 and 2015 to calculate the carbon
transfer among 28 industries in China and use the model to
simulate the carbon emissions transfer network in 2017. The
experimental analysis results show that the proposed model
has a good applicability to small-sample ROS such as the
carbon emissions transfer between industries in China.
II. RELATED WORK
A. ROGM (1, 1) MODEL
If the ROS with a given amplitude is X(0) = (x (0) (1),
x (0) (2), · · · , x (0) (n)), then the sequence X(0) D = (x (0) (1)d,
x (0) (2)d, · · · , x (0) (n − 1)d) is called the first order smooth
20617
 Y. Hu, K. Lv: Hybrid Prediction Model for the Interindustry Carbon Emissions Transfer Network Based on the GM and GVM

sequence of ROS X(0) . Therefore, the GVM is no longer sensitive to a single sam-
 (0) ple, and a small number of samples containing noise will
x (k) + T + x (0) (k + 1) + T
  
(0)
x (k)d = , not significantly affect the prediction performance of the
4 model.
k = 1, 2, · · · , n − 1. (1)
We use the three-layer network model to analyze the train-
If the ROS is X(0) = (x (0) (1), x (0) (2), · · · , x (0) (n + 1)), ing process of the GVM. The structure of the GVM is shown
then its first-order smooth sequence is Y(0) = (y(0) (1), in Fig. 1. We assume that the input layer of the GVM model
y(0) (2), · · · , y(0) (n)), According to Y(0) , the discrete GM has M nodes, the hidden layer has N nodes, while the output
(1, 1) model is established, and ŷ(1) (k + 1) = β1 ŷ(1) (k) + β2 layer has L nodes. The expression of the outputs of the hidden
is obtained. Using the least square method, b̂ can be obtained and output layer nodes is as follows:
as 
 XM
wij · xi + bj )), j = 1, · · · , N

β̂ = (β1 , β2 )T = (BT B)−1 BT Y,
 h = f (β (
(2)  j j


i=1
 (1)   (1) 
N (5)
y (1) 1 y (2) X
vjl · hl , l = 1, · · · , L

 y(1) (2)  y(1) (3)  y =

1  l



where B =  .. .. , and Y =  .. .
   
j=1
 . .   . 
y(1) (n − 1) 1 y(1) (n) In equation (5), xi denotes the i th input feature, wij is the
connection weight
P between input node i and node j of the
For order ŷ(1) (1)(1)
= y (0), the restoring value is obtained
hidden layer, M i=1 wij · xi is the weighted input sum of node
as:
of the hidden layer; bj is the offset of node of the hidden
ŷ(0) (k + 1) = α 1 ŷ(1) (k + 1) = ŷ(1) (k + 1) − ŷ(1) (k), layer, and all the offsets form the offset vector Wb . βj is the
k = 1, 2, · · · , n − 1. (3) control parameter of node of the hidden layer, and all the
control parameters form the control vector Wβ . hj represents
On this basis, x̂ (0) (t) is derived as: the output value of node j of the hidden layer, and vjl is the
x̂ (0) (t) = Fβ1t−3 − (−1)t (Fβ1−1 − C − T ) − T . (4) connection weight between the output node l and node j of
the hidden layer. All hidden nodes are linearly connected to
Equation (4) is called the grey prediction model of ROS, the output nodes through the weight matrix V.
which is referred to as the ROGM (1, 1) model (refer to the The GVM introduces parameters β to control the sensitiv-
literature [24] for details). ity of the model output to the input. By controlling β within a
certain range, the trained GVM can maintain robustness to
B. GVM small fluctuation in the input vector. Therefore, the model
Unlike the SVM, which selects a small number of sample can avoid the problem of a low prediction accuracy resulting
data as the support vector, the GVM makes full use of from adopting too simple a model by using more hidden layer
all training samples to determine the whole model. These nodes to enhance its learning ability.
training samples are called general vectors. In the GVM, To improve the training efficiency of the GVM, the con-
the Monte Carlo (MC) algorithm is used to directly determine nection weights V are fixed, and the model is optimized by
the appropriate project vector in the general vector space. adjusting the connection weights W. In addition, the GVM

FIGURE 1. Structure of the GVM.

20618 VOLUME 8, 2020

Y. Hu, K. Lv: Hybrid Prediction Model for the Interindustry Carbon Emissions Transfer Network Based on the GM and GVM

introduces the MC algorithm to train the model and adjust

weight matrices W,Wb and Wβ . The cost function is defined
as follows:

N
1X
COST = (ŷi − yi )2 . (6)
2
i=1

The main idea of the MC training algorithm is to randomly

select a weight from among all adjustable weights and to
slightly adjust the weight in a small range ε. If the cost func-
tion value of the GVM is reduced due to the change in weight,
the change in weight value is accepted; otherwise, the weight
value remains unchanged. Through such iterations, the GVM
gradually converges on a stable optimal solution, and the
weight value of the GVM is determined [20].

III. ROGM-AFSA-GVM MODEL

A. GVM OPTIMIZED BASED ON THE ARTIFICAL FISH
SWARM ALGORITHM
FIGURE 2. GVM model optimized based on AFSA.
Optimizing the network weights of the GVM is the key to
improving its prediction performance. To achieve an excellent
performance of the GVM, the number of hidden nodes is
usually larger than 10 times that of the BP neural network. in functions AF-Follow, AF-Swarm and AF-Prey, respec-
Moreover, control parameters of the hidden nodes are intro- tively. In these functions, the distance between
two artifi-
duced. These factors increase the training time of the model. cial fish individuals is expressed as dij = Pi − Pj , and
In addition, only one weight of the GVM is optimized in VISUAL, STEP and δ represents the visual range, mov-
one iteration using the MC algorithm to optimize the net- ing step length and congestion factor of the artificial fish,
work weights, and the convergence speed of the model is respectively. In each iteration of AFSA, the network weights
low. represented by each artificial fish individual are passed to
Section II, it indicates that the parameter optimization the GVM for prediction, and then the fitness of the cost
problem of the GVM is actually the optimization prob- function (equation (6)) is calculated. Based on the change in
lem of determining the network parameters that minimize fitness value, the optimal state of each artificial fish in the
the cost function (equation (6)). The intelligent optimiza- current iteration is recorded. When the maximum number
tion algorithm is an effective method to solve this kind of of iterations is reached, the state of all optimal artificial
problem. fish individuals in the population is determined based on the
The AFSA is an intelligent optimization algorithm inspired network weights of the GVM.
by the behavior of fish swarms [25]. The algorithm has the
advantages of a high robustness, good global convergence
Function AF_Follow (FC, NF, NC)
and low sensitivity to the initial value. The AFSA algorithm
1: FC is the food concentration
mainly searches the solution space through three behaviors:
2: NF is the population size of artificial fish
preying, following and swarming. The behavior of prey is
3: NC is the number of companions of artificial fish in the
a kind of random swimming behavior, that maintains the
visual range
diversity of the population. The behaviors of following and
4: for i = 1 to NF
swarming lead the fish to search in a suitable direction to
5: for j = 1 to NC
ensure convergence of the population. Therefore, AFSA imi-
6: if COST(Pj ) < FC
tates the three behaviors of fish swarms to achieve a balance
7: FC = COST(Pj )
between diversity and convergence, and finally determine the
8: Pmin = Pj
global optimal solution.
9: end if
The GVM optimized based on AFSA is shown in Fig. 2.
10: end for
The population individuals of the AFSA algorithm represent
11: if FC · NC < COST(Pi ) · δ
the network weights to be optimized, which is the vector
12: Pinext = Pi + Random(STEP)(Pmin − Pi )/dij
composed of the GVM network weights as shown in Fig. 2.
13: end if
AF-follow ( ), AF-swarm ( ) and AF-prey ( ) operations are
14: end for
used to optimize the position of each artificial fish individ-
15: end function
ual. The pseudocodes of these operations are summarized

VOLUME 8, 2020 20619

 Y. Hu, K. Lv: Hybrid Prediction Model for the Interindustry Carbon Emissions Transfer Network Based on the GM and GVM

TABLE 1. Industry classification.

Function AF_Swarm(NF, NC,Pic ) Function AF_Prey (NF, TN)

1: NF is the population size of artificial fish 1: NF is the population size of artificial fish
2: NC is the number of companions of artificial fish in the 2: TN is the number of random movement trials of the
visual range artificial fish in the visual range
3: Pic is the center position of the companion in the visual 3: for i = 1 to NF
range 4: for j = 1 to TN
4: for i = 1 to NF 5: P0j = Pi + VISUAL · Rand()
NC
5: Pic =
P
Pj /NC 6: if COST(P0j ) < COST(Pi )
j=1 7: Pinext = Pi + Random(STEP)(Pj − Pi )/dij
6: if COST(Pic ) · NC < COST(Pi ) · δ 8: else
7: Pinext = Pi + Random(STEP)(Pic − Pi )/dij 9: Pinext = Pi
8: else 10: end if
9: AF_Prey( ) 11: end for
10: end if 12: end for
11: end for 13: end function
12: end function

The industries are interconnected in their the processes

B. PREDICTION OBJECT
In this paper, the input-output data come from Chinese input-
output tables of 1997, 2002, 2007, 2012 and 2017 and
input-output extension tables of 2000, 2005, 2010, and 2015.
The energy data come from the China Energy Statistics Year-
books 1997-1999, 2000-2002, 2006, 2008, 2011, 2013, 2016,
and 2018. In this study, we incorporate various materials of
different industries and consolidate them into 28 industries.
The industry classification results are listed in Table 1.
of production, distribution and exchange. With the flow and
transfer of carbon emissions, a network is formed, which
we call the interindustry carbon emissions transfer network.
Generally, we refer to the CO2 released by the energy
consumption of an industry as the direct carbon emissions
(DCE) [1], while the carbon emissions of the other industries
caused by intermediate products and services are defined as
the indirect carbon emissions (ICE). Both the DCE and ICE
consist of total carbon emissions (TCE).

20620 VOLUME 8, 2020

Y. Hu, K. Lv: Hybrid Prediction Model for the Interindustry Carbon Emissions Transfer Network Based on the GM and GVM
ICE are transferred to other industries with the transfer of
intermediate products and services. At the same time a con-
siderable part of DCE is also transferred to other industries
through intermediate products. We use the total industry car-
bon emissions to measure the interindustry carbon emissions
transfer and adopt the 28 industries as network nodes in this
study. The interindustry carbon emissions transfer network
can be constructed by considering the interindustry carbon
emissions transfer as the edge and using the interindustry
TCE as weights.
According to 19 kinds of energy consumption data
of 28 industries in China, we use the ‘‘inventory method’’
(IPCC, 2006) [1] to calculate the direct carbon emissions
of each industry. The 19 kinds of energy include raw coal,
cleaned coal, other coal washing, briquette, coke, coke oven
gas, other gas, other coking products, crude oil, gasoline,
kerosene, diesel oil, fuel oil, liquefied petroleum gas, refin-
ery gas, other petroleum products, natural gas, heat and
electricity. First, equation (7) directly calculates the car-
bon dioxide emissions coefficients of 17 kinds of energy
sources except for heat and electricity. Second, the carbon
emissions factors of heat and electricity in China are not
uniform, so they need to be calculated separately. Third,
the carbon emissions generated by thermal power generation
and heat supply can be allocated to the heat and electric-
ity use industries by using the principle of electricity heat
carbon allocation [26]. Finally, the DCE of 17 kinds of
energy in each industry and the DCE of heat and electric-
ity in each industry are added to obtain the DCE of each
industry.
K
X characteristics of the original series. On the other hand, the
DCEi = Eik × NCVk × CCk × COFk × TC
k=1
K it has the disadvantages mentioned in part A of Section III.
X
= Eik × CECk
k=1
where k = 1, 2, · · · K is the type of energy consumption; i is
the type of industry; DCEi is the DCE of the industry i; Eik is
energy consumption type k in the industry i; NCVk is the low
calorific value of energy type k; CCk is the carbon content of
energy type k; COFk is the carbon oxidation rate of energy
type k; TC is the conversion coefficient (the mass change in
carbon dioxide generated during unit carbon combustion is
44/12); and CECk is the carbon dioxide emissions coefficient
of energy type k.
To eliminate the impact of imports, the noncompeti-
tive input-output method is used to calculate the carbon
emissions transfer matrix. The carbon emission transfer
matrix TCE can be calculated using the following
equations.
h i−1
TCE = DCI × I − Ad × Yjd
DCI = [DCI1 , DCI2 · · · DCIn ] , DCIj = DCEj /Xj (9)
VOLUME 8, 2020
 
a11 a12 ··· a1n
 a21 a22 ··· a2n  xijd
Ad =  . .. .. ..  , aij = ,
 
 .. . . .  Xj
an1 an2 · · · ann
i, j = 1, 2, · · · , n (10)
Xi Xi
xijd = xij , d
Yi = Yi (11)
Xi + Mi Xi + Mi
where DCIj is the DCE intensity of industry j, Xj is the total
output of industry j, (I − Ad )−1 is the Leontief inverse matrix
of the intermediate input of national products, Ad is the direct
consumption coefficient matrix of national products, aij is the
direct consumption coefficient of national products, n is the
number of industries, xijd is the intermediate use of national
products of industry i consumed by industry j to produce its
output, Xi and Mi are the total domestic output and import,
respectively, of industry i, Yid is the final use of national
products of industry i, and Yi is the final use of industry i.
C. HYBRID PREDICTION MODEL BASED ON ROGM AND
AFSA-GVM
The interindustry carbon emissions transfer data exhibit
small-sample ROS characteristics. GM and GVM have their
own advantages and disadvantages for small sample data
prediction. The grey prediction model can predict the trend of
small-sample data series, but it cannot predict the volatility of
data series, that is, it has a low performance in simulating non-
linear sequences. Although the ROGM (1, 1) model improves
the smoothness of ROS by the data change method, it may
decrease the prediction accuracy due to damage to the data
GVM has a good fitting effect on small- sample data series
and an excellent nonlinear prediction ability; nevertheless,
The hybrid prediction model can make use of the advan-
(7)
tages of various prediction models that complement each
other and obtain better prediction results than a single predic-
tion model. To improve the prediction accuracy, a hybrid pre-
diction model denoted as ROGM-AFSA-GVM is constructed
as shown in Fig. 3. First, the ROGM (1, 1) model is used to
predict the original sequence of the carbon emissions transfer
among industries to obtain the basic prediction results and
residuals. Then, the prediction residuals are adopted as the
analysis object, and the AFSA-GVM model is established to
obtain the prediction result of the residual sequence. Finally,
the prediction results of TCE transfer can be obtained by
integrating the prediction results of the two parts. The specific
steps of the proposed model are as follows.
Step 1: The ROGM (1, 1) model is used to predict time
series data of the carbon emissions transfer among industries
from 1997 to 2015 to generate basic prediction and residual
series;
Step 2: Residual sequences from 1997 to 2010 are used
(8) as training data and entered as input into the AFSA-GVM
model, and the residual sequence of 2012 is used as
20621
 Y. Hu, K. Lv: Hybrid Prediction Model for the Interindustry Carbon Emissions Transfer Network Based on the GM and GVM

FIGURE 3. Modeling method of ROGM-AFSA-GVM.

training target data to form the input-output relationship of On this basis, we use AFSA to optimize the weight values
the AFSA-GVM model; of W,Wb and Wβ to determine the GVM. In the experiment,
Step 3: The AFSA-GVM prediction model is established the parameters of AFSA are listed in Table 2.
by determining the number of nodes in the hidden layer and
the AFSA algorithm is used to train the parameters of the TABLE 2. Parameters of AFSA.
GVM;
Step 4: The residual sequences from 2000 to 2012 are
adopted as test data, and they are entered as input into
the AFSA-GVM prediction model to obtain the residual
sequence in 2015. The final prediction data can be obtained
by integrating the corresponding basic prediction series
obtained in step 1 with the residual sequence in 2015. Then,
the prediction performance of the model is evaluated by To evaluate the performance of the prediction model from
comparing the prediction data to the test target data; multiple perspectives, two metrics are defined as follows:
Step 5: A time window with a window size of 6 is con- the root mean square error (RMSE) and mean absolute error
structed from the data of 2002, which adopts the input data (MAE):
of the ROGM-AFSA-GVM prediction model in the window
s
PN 2
range and predicts the next sequence adjacent to the window, i=1 (ŷi − yi )
RMSE = (12)
to obtain the prediction of the out-of-sample sequence. N
PN  
i=1 ŷi − yi

IV. EXPERIMENTAL RESULTS AND ANALYSIS
A. EXPERIMENTAL DATA AND PARAMETERS
According to experimental experience, the number of nodes
in the input, hidden and output layers of the GVM prediction
model is set to 6, 150 and 1, respectively. The transfer func-
tion is a Gaussian function. The input data are adjusted to
between −10 and 10. The value of the output layer weight V
is randomly set to 1 or −1, and remains unchanged during
training of the GVM. The value range of input layer weight W
is limited from −1 to 1, and the value range of bias vector Wb
is set to [−10, 10], while the value range of the control param-
eters constituting control vector Wβ is set to [−0.6, 0.6].
20622
MAE = (13)
N
where ŷi represents the predicted value, yi represents the real
value, and N is the number of samples. RMSE indicates the
degree of error dispersion between the predicted and real
values and MAE indicates the average error between the
predicted and real values. By analyzing these two metrics
of the prediction results, a comprehensive evaluation of the
prediction effect of the proposed model can be performed.
B. PREDICTION RESULTS OF CARBON EMISSIONS
TRANSFER BY THE DIFFERENT MODELS
Table 3 and 4 and Fig. 4 show the prediction results and errors
of the ROGM-AFSA-GVM model for industry i = 1 to other
VOLUME 8, 2020
Y. Hu, K. Lv: Hybrid Prediction Model for the Interindustry Carbon Emissions Transfer Network Based on the GM and GVM

TABLE 3. Prediction results of the carbon emissions transfer between industry 1 and the other industries in 2015 for the different models.

TABLE 4. Evaluation metrics of the prediction results of the different models.

industries j = 1, 2, 3, . . . , 28. The results of the ROGM, Fig. 4 shows that although the different models can predict
GVM and ROGM-GVM models are also used to evaluate the the trend of data change, the accuracy of the prediction results
effectiveness of the different models. differs. It is shown that the prediction data of the GVM are

VOLUME 8, 2020 20623

 Y. Hu, K. Lv: Hybrid Prediction Model for the Interindustry Carbon Emissions Transfer Network Based on the GM and GVM

FIGURE 4. Prediction performance comparison of the different models.

closer to the actual data in the change trend and volatility
than those of the ROGM model. In addition, we can also
see that the prediction result of the ROGM-GVM model
is better than that of the GVM model, and the prediction
result of the ROGM-AFSA-GVM model is the best among
all models. It is shown that the prediction performance can
be further improved by the hybrid model and using AFSA
to optimize the GVM network weights. On the other hand,
the black and bold values in Tables 3 and 4 indicate the best
prediction results and evaluation metrics values. Although
the GVM has the best prediction effect among the single
models, the ROGM-AFSA-GVM hybrid model has the best
prediction performance among all models.
To further verify the validity of the ROGM-AFSA-GVM
hybrid model, we also use it to predict the carbon emissions
transfer among the other industries. The carbon emissions
transfer data in 2015 for industries i = 10 and 20 and other
industries j = 1, 2, 3 . . . , 28 are selected as the prediction
target, and the ROGM-AFSA-GVM hybrid model is used for
prediction. Fig. 5 shows a comparison of the measured and
predicted results. We can see that, similar to the prediction
results in Fig. 4, the change trends of the prediction and mea-
sured data are consistent, and the prediction results are closer
to the actual data. All above mentioned results indicate that
the ROGM-AFSA-GVM hybrid model is more suitable for
the prediction of the carbon interindustry emissions transfer.

20624 VOLUME 8, 2020

Y. Hu, K. Lv: Hybrid Prediction Model for the Interindustry Carbon Emissions Transfer Network Based on the GM and GVM

FIGURE 5. Comparison of the prediction results of the ROGM-AFSA-GVM model.

TABLE 5. Comparison of the measured and predicted interindustry carbon emissions transfer networks in 2017.

C. SIMULATION OF THE INTERINDUSTRY CARBON
EMISSIONS TRANSFER NETWORK IN 2017
Based on the comparison and analysis of the experimental
results, the ROGM-AFSA-GVM hybrid model can better pre-
dict the carbon emissions transfer among industries. There-
fore, this model is used to predict the interindustry carbon
emissions transfer in 2017, and on this basis, the interindustry
carbon emissions transfer network is constructed. As the
carbon emissions transfer between two industries is relatively
independent, we enter the 28 × 28 = 784 predicted values of
each year into the carbon emissions transfer matrix, and then
process them following the steps below. First, the elements

VOLUME 8, 2020 20625

 Y. Hu, K. Lv: Hybrid Prediction Model for the Interindustry Carbon Emissions Transfer Network Based on the GM and GVM
along the diagonal represent the carbon emissions transferred
within each industry, so the diagonal elements are set to 0;
second, the average flow values of 2017 can be calculated
which are considered critical values, and the elements below
the critical value are set to 0, which is regarded as weak trans-
fer. The elements above this critical value constitute a stable
transfer path On this basis, the interindustry carbon emissions
transfer network of China in 2017 can be constructed and
visualized, as shown in Fig. 6 and 7.
FIGURE 6. The interindustry carbon emissions transfer network of China
in 2017 (predicted).
FIGURE 7. The interindustry carbon emissions transfer network of China
in 2017 (measured).
Table 5 indicates that the predicted total flow of the carbon
emissions transfer network in 2017 was 374 million tons
larger than the measured total flow, accounting for 6% of the
total measured flow, which is within a reasonable error range.
The network density and local clustering coefficients of the
measured and predicted networks are also very close. In addi-
tion, Table 5 reveals that the top six industries ranked by the
network characteristics are exactly the same, except that the
ranking order is slightly different. These results further prove
the validity and accuracy of the proposed model.
20626
V. CONCLUSION
In this study, prediction of the interindustry carbon emis-
sions transfer network of China is studied by using a
hybrid intelligent prediction model based on the GM and
GVM. To improve the prediction performance of the model,
the AFSA algorithm is adopted to optimize the network
weights of the GVM, and the ROGM-AFSA-GVM model
is constructed on this basis. The experimental results show
that the proposed model is feasible and has better prediction
effect. Using the predicted results, the interindustry carbon
emissions transfer network of China in 2017 is constructed.
By comparing the predicted 2017 carbon emissions transfer
network with the measured network, it is verified that the
proposed prediction model in this paper indeed can achieve
good performance for predicting ROS data such as the data
series of the interindustry carbon emissions transfer in China.
In subsequent studies, we will continue to improve the pro-
posed model to enhance its prediction accuracy at the turning
point of the ROS.
REFERENCES
[1] H. S. Eggleston, L. Buendia, and K. Miwa. IPCC Guidelines for National
Greenhouse Gas Inventories. Accessed: Jul. 16, 2018. [Online]. Available:
https://www.ipccnggip.iges.or.jp/public/2006gl/index
[2] I. Mladenović, S. Sokolov-Mladenović, M. Milovančević, D. Marković,
and N. Simeunović, ‘‘Management and estimation of thermal comfort, car-
bon dioxide emission and economic growth by support vector machine,’’
Renew. Sustain. Energy Rev., vol. 64, pp. 466–476, Oct. 2016.
[3] W. Sun and M. Liu, ‘‘Prediction and analysis of the three major industries
and residential consumption CO2 emissions based on least squares sup-
port vector machine in China,’’ J. Cleaner Prod., vol. 122, pp. 144–153,
May 2016.
[4] W. Sun, C. Wang, and C. Zhang, ‘‘Factor analysis and forecasting of
CO2 emissions in Hebei, using extreme learning machine based on par-
ticle swarm optimization,’’ J. Cleaner Prod., vol. 162, pp. 1095–1101,
Sep. 2017.
[5] S. Ding, Y.-G. Dang, X.-M. Li, J.-J. Wang, and K. Zhao, ‘‘Forecasting Chi-
nese CO2 emissions from fuel combustion using a novel grey multivariable
model,’’ J. Cleaner Prod., vol. 162, pp. 1527–1538, Sep. 2017.
[6] Y. Huang, L. Shen, and H. Liu, ‘‘Grey relational analysis, principal compo-
nent analysis and forecasting of carbon emissions based on long short-term
memory in China,’’ J. Cleaner Prod., vol. 209, pp. 415–423, Feb. 2019.
[7] A. O. Acheampong and E. B. Boateng, ‘‘Modelling carbon emission inten-
sity: Application of artificial neural network,’’ J. Cleaner Prod., vol. 25,
pp. 833–856, Jul. 2019.
[8] X. Liu, J. Ou, Y. Chen, S. Wang, X. Li, L. Jiao, and Y. Liu, ‘‘Sce-
nario simulation of urban energy-related CO2 emissions by coupling the
socioeconomic factors and spatial structures,’’ Appl. Energy, vol. 238,
pp. 1163–1178, Mar. 2019.
[9] X. Zhao and D. Du, ‘‘Forecasting carbon dioxide emissions,’’ J. Environ.
Manage., vol. 160, pp. 39–44, Sep. 2015.
[10] M. Meng and D. Niu, ‘‘Modeling CO2 emissions from fossil fuel combus-
tion using the logistic equation,’’ Energy, vol. 36, no. 5, pp. 3355–3359,
May 2011.
[11] R. Pérez-Suárez and A. J. López-Menéndez, ‘‘Growing green? Forecasting
CO2 emissions with environmental Kuznets curves and logistic growth
models,’’ Environ. Sci. Policy, vol. 54, pp. 428–437, Dec. 2015.
[12] X. E. Wang, Y. X. Wang, and H. Y. Duan, ‘‘Forecasting area’s carbon emis-
sions of energy consumption and controllability study,’’ China Population,
Res. Environ., vol. 24, no. 8, pp. 9–16, Aug. 2014.
[13] R. Huang, Z. Wang, G. Q. Ding, Y. R. Gong, and C. X. Liu, ‘‘Trend predic-
tion and analysis of influencing factors of carbon emissions from energy
consumption in Jiangsu province based on STIRPAT model,’’ Geograph.
Res., vol. 35, no. 4, pp. 781–789, Apr. 2016.
[14] H.-T. Pao, H.-C. Fu, and C.-L. Tseng, ‘‘Forecasting of CO2 emissions,
energy consumption and economic growth in China using an improved
grey model,’’ Energy, vol. 40, no. 1, pp. 400–409, Apr. 2012.
VOLUME 8, 2020
Y. Hu, K. Lv: Hybrid Prediction Model for the Interindustry Carbon Emissions Transfer Network Based on the GM and GVM
[15] D. Fang, X. Zhang, Q. Yu, T. C. Jin, and L. Tian, ‘‘A novel method for car-
bon dioxide emission forecasting based on improved Gaussian processes
regression,’’ J. Cleaner Prod., vol. 173, pp. 143–150, Feb. 2018.
[16] N. Xu, S. Ding, Y. Gong, and J. Bai, ‘‘Forecasting Chinese greenhouse gas
emissions from energy consumption using a novel grey rolling model,’’
Energy, vol. 175, pp. 218–227, May 2019.
[17] S. F. Liu, Y. G. Dang, and Z. G. Fang, Grey System Theory and Application.
Beijing, China: Science Press, 2010, pp. 96–100.
[18] E. Garip and A. B. Oktay, ‘‘Forecasting CO2 emission with machine
learning methods,’’ in Proc. Int. Conf. Artif. Intell. Data Process. (IDAP),
Sep. 2018, pp. 1–4.
[19] Y. Wang and S. Yang, ‘‘The prediction of CO2 emissions from manufactur-
ing industry based on GM(1,N) model and SVM in Chongqing,’’ in Proc.
Int. Conf. Sens., Diagnostics, Prognostics, Control (SDPC), Aug. 2018,
pp. 85–89.
[20] Q. Zhou, H. Chen, H. Zhao, G. Zhang, J. Yong, and J. Shen, ‘‘A local
field correlated and Monte Carlo based shallow neural network model for
nonlinear time series prediction,’’ ICST Trans. Scalable Inf. Syst., vol. 3,
no. 8, Aug. 2016, Art. no. 151634.
[21] B. Yong, F. Li, Q. Lv, J. Shen, and Q. Zhou, ‘‘Derivative-based acceleration
of general vector machine,’’ Soft Comput., vol. 23, no. 3, pp. 987–995,
Feb. 2017.
[22] B. Yong, J. Shen, Z. Shen, H. Chen, X. Wang, and Q. Zhou, ‘‘GVM based
intuitive simulation web application for collision detection,’’ Neurocom-
puting, vol. 279, pp. 63–73, Mar. 2018.
[23] Y. Zhang, H. Sun, and Y. Guo, ‘‘Wind power prediction based on PSO-SVR
and grey combination model,’’ IEEE Access, vol. 7, pp. 136254–136267,
2019.
[24] B. Zeng, X. Y. Yin, and W. Meng, Practical Grey Prediction Modeling
Method and Its MATLAB Program Implementation. Beijing, China: Sci-
ence Press, 2018, pp. 114–118.
[25] L. Zou, H. Li, W. Jiang, and X. Yang, ‘‘An improved fish swarm algorithm
for neighborhood rough set reduction and its application,’’ IEEE Access,
vol. 7, pp. 90277–90288, 2019.
[26] M. Zhang, H. Mu, Y. Ning, and Y. Song, ‘‘Decomposition of energy-related
CO2 emission over 1991–2006 in China,’’ Ecol. Econ., vol. 68, no. 7,
pp. 2122–2128, Jul. 2009.
VOLUME 8, 2020
YING HU received the B.S. degree in electronic
information engineering from Northwestern Poly-
technical University, Xi’an, China, in 2006, and
the M.S. degree in economics from the Southwest
University of Finance and Economics, Chengdu,
China, in 2011. She is currently pursuing the Ph.D.
degree in economics with Shanghai University,
Shanghai, China.
From 2013 to 2017, she was a Lecturer with the
School of Business, Jiangsu University of Tech-
nology, Changzhou, China. Her current research interests include the appli-
cation of machine learning algorithm, artificial intelligence algorithm, and
complex network theory in resource and environment economics.
KANGJUAN LV received the B.S. degree in man-
agement science and engineering and the M.S.
degree in geographic information system from
Northeast Agricultural University, Harbin, China,
in 1997 and 2000, respectively, and the Ph.D.
degree in economics from the Harbin Institute of
Technology, Harbin, in 2003.
From 2009 to 2010, she was an Academic
Visitor with the Department of Land Economics,
University of Cambridge, Cambridge, U.K. Since
2003, she has been a Lecturer, an Associate Professor, and a Professor with
the SHU-UTS SILC Business School, Shanghai University, Shanghai, China.
Her current research interests include research on economic and industrial
complex network and spatial economy.
20627