Science of the Total Environment 927 (2024) 172319
Contents lists available at ScienceDirect
Science of the Total Environment
journal homepage: www.elsevier.com/locate/scitotenv
Review
Carbon emission prediction models: A review
Yukai Jin a, b, Ayyoob Sharifi c, d, *, Zhisheng Li b, Sirui Chen b, Suzhen Zeng b, e, Shanlun Zhao b
a
Urban Environmental Science Lab (URBES), Graduate School of Innovation and Practice for Smart Society, Hiroshima University, Higashi-Hiroshima, 739-8529, Japan
School of Civil and Transportation Engineering, Guangdong University of Technology, Guangdong, 510006, China
c
The IDEC Institute, Hiroshima University, Higashi-Hiroshima, 739-8529, Japan
d
School of Architecture and Design, Lebanese American University, Beirut, Lebanon
e
School of Ocean Engineering and Technology, Sun Yat-sen University, Guangdong, 519000, China
b
H I G H L I G H T S
G R A P H I C A L A B S T R A C T
• We conduct a comprehensive review of
147 carbon emission prediction models.
• Examined models include prediction,
optimization, and prediction factor
selection.
• We analyze the advantages and disad
vantages of each model.
• We compare the prediction performance
of models in existing studies.
• We analyze the temporal and spatial
distribution patterns of carbon emission
prediction models.
A R T I C L E I N F O
A B S T R A C T
Editor: Yuyu Zhou
Amidst growing concerns over the greenhouse effect, especially its consequential impacts, establishing effective
Carbon Emission Prediction Models (CEPMs) to comprehend and predict CO2 emission trends is imperative for
climate change mitigation. A review of 147 Carbon Emission Prediction Model (CEPM) studies revealed three
predominant functions—prediction, optimization, and prediction factor selection. Statistical models, comprising
75 instances, were the most prevalent among prediction models, followed by neural network models at 21.8 %.
The consistent rise in neural network model usage, particularly feedforward architectures, was observed from
2019 to 2022. A majority of CEPMs incorporated optimized approaches, with 94.4 % utilizing metaheuristic
models. Parameter optimization was the primary focus, followed by structure optimization. Prediction factor
selection models, employing Grey Relational Analysis (GRA) and Principal Component Analysis (PCA) for sta
tistical and machine learning models, respectively, filtered factors effectively. Scrutinizing accuracy, preoptimized CEPMs exhibited varied performance, Root Mean Square Error (RMSE) values spanned from 0.112
to 1635 Mt, while post-optimization led to a notable improvement, the minimum RMSE reached 0.0003 Mt, and
Keywords:
Carbon emission
Climate change mitigation
Prediction model
Machine learning
Neural network
Artificial intelligence
* Corresponding author.
E-mail address: sharifi@hiroshima-u.ac.jp (A. Sharifi).
https://doi.org/10.1016/j.scitotenv.2024.172319
Received 27 November 2023; Received in revised form 26 March 2024; Accepted 6 April 2024
Available online 9 April 2024
0048-9697/© 2024 Elsevier B.V. All rights reserved.
Y. Jin et al.
Science of the Total Environment 927 (2024) 172319
the maximum was 95.14 Mt. Finally, we summarized the pros and cons of existing models, classified and counted
the factors that influence carbon emissions, clarified the research objectives in CEPM and assessed the applied
model evaluation methods and the spatial and temporal scales of existing research.
1. Introduction
- Establish a classification framework for the three main models in
CEPM: factor selection models, prediction models, and optimization
models, elucidating the strengths and weaknesses of each.
- Classify and discuss the influencing factors, research objectives, and
performance evaluation methods of CEPM.
- Conduct a comprehensive comparison of reviewed CEPM using the
three main error methods: RMSE, MAE, and MAPE.
- Analyze the temporal and spatial distribution patterns of CEPM and
discuss carbon emission trends in major countries.
Carbon dioxide (CO2) has existed since the formation of the Earth.
Plant photosynthesis and volcanic eruptions produce carbon dioxide in
the environment (Rae et al., 2021). With the emergence of humans,
carbon emissions are increasing (Gasser et al., 2015). This has increased
the frequency and intensity of natural hazards. Some studies (Smith
et al., 2016; van Hooidonk et al., 2016) have shown that by 2100, if
emissions do not decrease, some tropical coastal areas will experience
floods, fires, storms, and other disasters. To control carbon emissions
and mitigate climate change, reducing carbon emissions has emerged as
a worldwide objective for development (Murshed et al., 2023). In 1997,
the ‘Kyoto Protocol’ stipulated carbon emission reduction as a legal
obligation of developed countries (Duus-Otterström, 2023; Nations, U,
1998). More recently, the Paris Agreement, adopted in December 2015,
led most countries to establish targets to mitigate global warming
(Meinshausen et al., 2022; Nations, U, 2016). Setting goals and targets to
reduce carbon emissions effectively is an essential part of climate action
(Aboagye and Sharifi, 2023, 2024). An important initial step towards
this is modeling carbon emissions (Han et al., 2023). Accordingly, in the
realm of artificial intelligence, which is widely used today, carbon
emission prediction has become a popular issue (Cabaneros et al., 2019).
Precise forecasting of carbon emissions is essential to guide policy
makers effectively in mitigating the greenhouse effect (Hsu et al., 2022;
Yin et al., 2022).
Several review studies on CEPM have been published in recent years.
The following is a summary of the main findings from these reviews.
Wang et al. (Wang et al., 2019a) discussed the historical factors influ
encing carbon emissions and projected future emissions within China’s
industrial sector. Grubb et al. (Grubb et al., 2015) reviewed 164 studies
involving carbon emissions in China. They concluded that technological
substitution is critical to reducing emissions to the lower end of the
acceptable range. Dong et al. (Dong et al., 2018) elaborated on the
current status of China’s carbon emission peak, analyzing it from
regional and sectoral perspectives and summarized various predictions.
Hewage et al. (Abeydeera et al., 2019) reviewed carbon emissions
literature from 1981 to 2019. They highlighted key research areas:
carbon capture and storage, trend analysis for future carbon emission
prediction models, evaluating carbon reduction performance, identi
fying emission reduction opportunities, and attaining zero carbon
emission goals. In another study, Huang et al. (Huang et al., 2019a)
examined 807 papers on China’s sectoral carbon emissions. They found
five common methods: extended environmental input-output analysis,
index decomposition analysis, econometrics, evaluating carbon emis
sion control efficiency, and simulation approaches.
Above studies provided a partial overview of methods applied in
CEPM, yet they lacks comprehensiveness. Firstly, predicting carbon
emissions often necessitates integrating various types of models. Simply
classifying CEPM into statistical or artificial intelligence methods may
lead to inaccurate descriptions. Secondly, the frequency and advantages
and disadvantages of different CEPM remain unclear, impeding further
innovation. Lastly, existing research has not thoroughly compared the
predictive performance of CEPM. Therefore, a new systematic summary
is required to provide detailed classification and in-depth analysis of
CEPM.
To address these issues, this study conducted a comprehensive re
view of CEPM research from 2011 to August 2022, aiming to achieve the
following objectives:
This paper is organized as follows: Section 2 delineates the research
methodology and elucidates the search strategy employed. In Section 3,
we provide a comprehensive categorization and review of CEPM,
highlighting its strengths and limitations. Within Section 3, we initially
categorize predictive models in CEPM, followed by an examination of
optimization models used for carbon emissions modeling, including
programming and meta-heuristic analyses. Furthermore, the last part of
Section 3 involves an analysis of reported prediction factor selection
models. Section 4 engages in a discussion of CEPM, exploring the spe
cifics of carbon emission models, including the classification of influ
encing factors, relevant statistical analyses, study targets, prediction
model evaluation methods, considerations of time and spatial scales in
various papers, future carbon emission trends in different countries, the
limitations, and potential solutions for these approaches. Section 5
concludes the paper.
2. Materials and methods
2.1. Boundary definition of carbon emissions
Generally, carbon emissions refer to carbon dioxide emissions, which
are also the primary contributor to global warming (Pachauri et al.,
2014). As a result, this review does not include articles centered on other
greenhouse gases. Furthermore, the concept of carbon intensity, which
signifies the volume of carbon emissions produced per unit of GDP, has
garnered increased scholarly attention in recent years and has thus been
incorporated into this review (Acheampong and Boateng, 2019; Ye
et al., 2018; Wang et al., 2018).
2.2. Search strategies
We followed a rigorous article screening method. First, we searched
all published studies in the core database of the Web of Science to
summarize the research trends of the most authoritative papers. The
main search terms included different combinations of words such as
‘carbon emission’, ‘prediction’, ‘forecasting’, ‘modeling’, and ‘machine
learning’. Word list for literature search is presented in Table 1. The
search process iterated until no further pertinent citations were ac
quired. Additionally, the reference lists of the chosen papers underwent
Table 1
Word list for literature search.
2
Method
Model
Objective
Geographical
extent
Time
horizon
Forecasting
Modeling
Global
Short
Projection
Machine
learning
Deep learning
Statistics
Carbon
emission
Carbon
intensity
Regional
Medium
Country
Long
Y. Jin et al.
Science of the Total Environment 927 (2024) 172319
PR. Song et al. (Wang et al., 2019b) used input–output analysis and PR
to predict trade-related carbon emissions in China and Australia through
2025. The findings indicated a growth in net carbon export from China
to Australia, ranging between 2.2 and 15.5 Mt of carbon. Yuan et al.
(Yang et al., 2018) used dynamic panel data model to predict the carbon
intensity in China. By 2020, the carbon intensity in China decreased by
33 % from the 2005 level. Zhao et al. (Zhao and Du, 2015) used PR to
forecast the carbon emissions of 90 OECD countries and China. The
results showed that CO2 emissions are expected to be approximately 25
% lower than 2010 levels by 2050. Du et al. (Du et al., 2012) utilized PR
to predict provincial carbon emissions in China spanning from 2010 to
2020.
In the papers studied, two univariate time series methods are used:
Autoregressive integrated moving average (ARIMA) and Autoregressive
distributed lag (ARDL) models. The general form of ARIMA, symbolized
as ARIMA (p, d, q), involves ‘p’ indicating autoregressive order, ‘d’
representing difference order, and ‘q’ indicating moving average order
(Kaur et al., 2023). The four reviewed papers use the ARIMA method. Li
et al. (Li et al., 2020a) concluded, based on the ARIMA model’s pre
dictions, that China’s anticipated peak emissions would reach 96.3
million tons by the year 2021. Malik et al. (Malik et al., 2020) employed
the ARIMA model to forecast carbon dioxide emissions in Pakistan up to
2030. The model demonstrated good accuracy, with a generally low
Mean Absolute Percentage Error (MAPE) value consistently below 10 %.
Yang et al. (Yang and O’Connell, 2020) utilized the ARIMA model to
forecast carbon emissions within the Shanghai aviation industry. Their
findings revealed a projected continual increase in carbon emissions. Lin
et al. (Lin and Agyeman, 2019) formulated an ARDL model to forecast
carbon emissions in Ghana spanning from 2017 to 2030.
In 1982, Deng Julong (Deng, 1982) proposed the grey system theory,
tailored for modeling systems with partial unknowns, small samples,
and limited information. The grey prediction model, a pivotal compo
nent of this theory, finds applicability across diverse fields, including
CEPM. With continuous applications and updates, many improved
traditional grey models have been proposed and used for CEPM. A total
of 43 of the studies reviewed applied GMs and their variants, accounting
for approximately 29.1 % of all prediction models reviewed. Grey
models can be divided into univariate and multivariate models, which
analysis to identify crucial references. Ultimately, only literature pub
lished in English was included. The earliest CEPM research we indexed
was from 2011, and the literature search continued until the day before
the article was written. Therefore, the scope of the review papers was
from 2011 to August 2022. Finally, 147 papers that met the established
requirements were included in this study. The Online Supplementary
Appendix I lists studies on CEPMs published between 2011 and 2022
that were used in this review. The literature search strategy and meth
odology for literature selection are illustrated in Fig. 1.
3. Results
3.1. Prediction models
After a detailed and comprehensive literature survey, as shown in
Fig. 2, the prediction models used in CEPM were divided into four cat
egories. We counted the main models used for analysis in each paper.
The number of papers using statistical models was the largest (75 pa
pers), with neural network models (31 papers), and shallow intelligent
models in the field of machine learning (27 papers) being the secondand third-most common, respectively. A combined prediction model
was defined as a combination of two or more prediction models, and
these models were least common (20 papers) among the 147 papers
reviewed.
3.1.1. Statistical models
Fig. 2 shows the classification and quantitative statistics of statistical
models for emission prediction. The grey model (GM) was implemented
43 times, and 7 papers applied the classic regression model. In addition,
19 papers used other statistical models. Among the GMs considered, GM
(1, 1) was most used (26 papers), even more often than classic statistical
methods. Classic regression methods include linear regression and panel
regression, which were used in 5 and 2 papers, respectively.
Regression analysis is an iterative process. In CEPM, two regression
methods are used: panel regression (PR) and linear regression (LR). One
reviewed paper used an LR approach. Shweta et al. (Singh and Kennedy,
2015) proposed a regression model to develop a tool for predicting
carbon emissions in 3646 cities worldwide. In addition, four papers used
Fig. 1. (a) Approach for literature search and selection - adapted from Moher et al. (2009). (b) The methodology employed in selecting the reviewed papers.
3
Y. Jin et al.
Science of the Total Environment 927 (2024) 172319
Fig. 2. (a) Proportion of the four prediction models among all models, with blue indicating statistical prediction models, green for shallow intelligent prediction
models, red for neural network prediction models, and purple for combined prediction models. (b) Classification of prediction models. (c) Time series distribution of
different statistical prediction models, color-coded as shown in (b). (d) Time series distribution of different combined prediction models. (e) Time series distribution
of different neural network prediction models. (f) Time series distribution of different combined prediction models.
are denoted as GM (1, 1) and GM (1, N), respectively. Among them, a
total of 26 papers used the univariate grey model, and 17 papers used
the multivariate grey model. GM (1, 1) and GM (1, N) models are two
basic models in grey theory (Xie and Wang, 2017). The number in the
first bracket indicates the order of the grey differential equation, while
the second number signifies the count of variables involved (Liu et al.,
2016). This model aims to employ an Additive Generation Operator
(AGO) on the initial data, constructing a new sequence based on expo
nential rules. Subsequently, by fitting this new data sequence with an
exponential model, the Inverse Additive Generation Operator (IAGO) is
applied to forecast a value for the original sequence (Liu and Yang,
2017). Moreover, the GM (1, 1) model has been used in predicting
carbon emissions. (Yu and Xu, 2019; Yu et al., 2018; Li et al., 2018a).
Various iterations of GMs have found application in CEPM. These
models can be categorized into equigap grey models and non-equigap
grey models based on their model spacing (Wu et al., 2020). These
improved methods mainly include processes such as initial condition
optimization, background value optimization, and the whitening of grey
derivatives (Xie, 2022; Wang et al., 2011). In addition, GMs are char
acterized by strong generalization ability, so they can be easily com
bined with other models, such as the grey Verhulst model and nonlinear
grey Bernoulli model (NGBM) (Duan and Luo, 2020; Pao et al., 2012).
The grey Verhulst model stands as one of the frequently employed
models in predicting within grey systems. This method is particularly
effective for data that exhibit a relatively unimodal trend (initially
increasing or decreasing before reversing). In 2020 (Duan and Luo,
2020), the Verhulst (1, 1, β) model was introduced, leveraging the
particle swarm optimization (PSO) algorithm to optimize the back
ground variable β. The outcomes illustrated the superiority of this pro
posed model over traditional grey models in predicting carbon emissions
from coal-related sources. Another advancement in 2018 (Wang and Li,
2019), introduced the non-equigap grey Verhulst model, utilizing the
PSO algorithm to optimize its structural parameters. Jiang (Jiang et al.,
2021a) proposed an enhanced grey multivariate Verhulst model (GMVM
(1, N)). This model integrates the grey Verhulst (1, N) model with a
residual correction model while preserving the original parameters.
Additionally, the grey control parameters from GM (1, 1) were incor
porated into the model to enhance prediction accuracy. Chen (Chen,
2008) coined the term “nonlinear grey Bernoulli model” (NGBM). The
NGBM-OP model, introduced by Pao et al. (Pao et al., 2012), is distin
guished by its robust predictive capacity, showcasing an MAPE (Mean
Absolute Percentage Error) of under 6.3 %. Employing NGBM-OP, pre
dictions were made for carbon emissions, energy trends, and outputs
spanning 2009 to 2020.
3.1.2. Shallow intelligent models
We classify models with lower structural complexity and smaller
parameter scales within intelligent models as shallow intelligent models.
This classification aims to distinguish them from neural network models.
In the 147 papers reviewed, there are two main shallow intelligent
prediction models: support vector machine (SVM) and decision tree
model. The SVM model can be further divided into LSSVM (least squares
support vector machine) and SVM models, which are used in 9 and 7
papers, respectively. The decision tree model only appears once in our
literature review. The classification and quantitative statistics of the said
models are given in Fig. 2.
SVM stands as a robust classification and regression tool, rooted in
the Vapnik-Chervonenkis (VC) dimension theory, and guided by the
principle of structural risk minimization. These models are character
ized by fast learning speed and good generalization ability (Ahmad
et al., 2014; Chauhan et al., 2019). Wang et al. (Wang et al., 2020)
compared backpropagation neural network (BPNN), Gaussian process
regression (GPR), and SVM, among which the SVM yielded better ac
curacy than the other models. The hyperparameters of SVM highly affect
4
Y. Jin et al.
Science of the Total Environment 927 (2024) 172319
the accuracy of modeling. The most important is C, which is the penalty
coefficient, or the tolerance for error. A higher ‘C’ signifies increased
intolerance to errors, potentially leading to overfitting. Conversely,
smaller ‘C’ values may lead to underfitting. Additionally, excessively
large, or small ‘C’ values can compromise the model’s generalization
ability. Therefore, many studies involving CEPM are based on optimized
SVM models (Wen and Cao, 2020a; Wang et al., 2019c; Sutthichaime
thee and Kubaha, 2018). In addition to PSO, optimization models are
based on other metaheuristic methods, as will be discussed in the next
section. The LSSVM was first proposed by Suykens et al. and is applied
for problem-solving in pattern classification and function estimation
tasks (Suykens and Vandewalle, 1999). A least squares linear system
replaces the quadratic programming approach typically used in SVMs.
This approach simplifies the calculational complexity and improves the
operating speed of the model. Like that of artificial neural networks
(ANN) and other intelligent algorithms, the performance of LSSVM
models depends on the selected inputs and parameters (Sun and Zhang,
2020; Wen and Cao, 2020b). Zhu et al. (Zhu et al., 2019) predicted
energy consumption and carbon intensity by introducing a new LSSVM
with a mixed kernel function. The results indicated that the proposed
model outperformed the original one in terms of accuracy. Sun et al.
(Sun and Liu, 2016) proposed an LSSVM to predict the CO2 emissions of
different industries in China. Their findings revealed that the LSSVM
exhibited superior accuracy compared to BPNN, GM (1, 1), and logistic
models.
GBDT (gradient boosting decision tree) is an ensemble algorithm
based on a decision tree. This ensemble boosting method uses a classi
fication and regression tree (CART) as the basic learner (Cheng et al.,
2019). GBDT can be used to solve most regression problems by gener
ating multiple weak prediction models and combining them, finally
generating an optimal performance model in an iterative manner (Li
et al., 2020b). Cui et al. (Cui et al., 2021) developed a GBDT prediction
model for forecasting carbon emissions in China. They optimized the
four parameters of the GBDT using the MWOA.
model can be applied to original nonlinear data without any parameter
restrictions required to ensure stability. Cristiana (Tudor, 2016) pre
dicted carbon emissions in Bahrain using an automated prediction
method. Seven different prediction models were considered. According
to the RMSE metric, NNAR yielded the highest accuracy.
From 1986 to 1988, psychologists L.L. McClelland and D.E. Rumel
hart proposed the famous multilayer neural network with an error
backpropagation algorithm, namely, the BPNN (Rumelhart et al., 1986).
In BPNN, carbon emission data flow forward through the network from
the input layer to the output layer, while error messages propagate
backward from the output layer to the hidden layer. During training, the
network computes the error between the actual output and the expected
output, then adjusts the connection weights between neurons to mini
mize this error. The backpropagation algorithm involves both forward
propagation and backward propagation. During forward propagation,
input data is fed through the network and outputs are computed. During
backward propagation, errors propagate backward through the
network, and gradient descent is used to optimize the adjustment of the
connection weights accordingly (Lillicrap et al., 2020; Wright et al.,
2022). Wen et al. (Wen and Yuan, 2020) used BPNN optimized with RF
and PSO to forecast the CO2 emissions of the commercial sector of
China. Zhou et al. (Zhou et al., 2017a) utilized an enhanced BPNN
optimized by PSO to forecast carbon emissions within the thermal power
industry in the Beijing-Tianjin-Hebei region of China. Their findings
indicated that BPNN achieved an error rate of <6 %.
GRNN represents Generalized Regression Neural Network. Unlike
traditional feedforward neural networks with layered organization of
neurons, GRNN has a radial basis function (RBF) architecture (AlMahasneh and Anavatti, 2023). In GRNN, each neuron in the hidden
layer is associated with a radial basis function, which computes the
similarity between the input data and prototype vectors. The output of
each neuron is weighted by this similarity measure. Training of GRNN
involves prototype selection and parameter estimation. Prototype se
lection entails determining prototype vectors from the training data,
which represent the distribution of input data in the feature space.
Parameter estimation involves estimating the width of the radial basis
functions and the weights associated with each neuron. (Zhang et al.,
2023). In 2020, Niu et al. proposed an improved GRNN-based CEPM for
predicting TCE and CEI in China until 2040. (Niu et al., 2020).
The fundamental architecture of a CNN consists of distinct layers: an
input layer, a convolutional layer, a pooling layer, a fully connected
layer, and an output layer (Zhou, 2020; Gu et al., 2018). Typically,
parameters for the convolutional and pooling layers are set through
iterative experimentation. Neurons at the output feature surface of the
convolutional layer are locally connected to the input of that layer. The
corresponding connection weight and local input weight are calculated,
added to the bias value, and used to determine the neuron’s input value.
This process allows CNNs to automatically learn hierarchies of features
from input carbon emission data (Basha et al., 2020; Kiranyaz et al.,
2021). In CEPM, Hien et al. (Hien and Kor, 2022) utilized a CNN model
to forecast Canadian fuel-related carbon emissions. Their findings
highlighted the CNN model’s superior stability compared to a regression
model in this domain.
All parameters of the traditional FNN need to be adjusted, so there
are dependencies (weights and deviations) among different parameter
layers. The gradient descent method is often applied to various FNNs
(Chen et al., 2019). However, these methods based on gradient descent
usually take a long time to run and easily fall to local optima. Unlike the
traditional FNN, ELMs are new learning algorithms based on a singlehidden layer-FNN, which is used to randomly select hidden nodes and
analyze the subsequent output weights (Ma and Dai, 2016). In the field
of CEPM, Wei et al. (Sun et al., 2018) proposed an ELM model based on
RF selection prediction factors and moth flame algorithm optimization.
Their results demonstrated its superior prediction accuracy over BPNNs
in similar conditions. Additionally, Sun et al. (Sun et al., 2017) utilized
PSO to optimize input weights and deviation thresholds within an ELM.
3.1.3. Neural network models
In the models reviewed, 32 papers applied different forms of neural
networks. In the 147 CEPM studies reviewed, two types of neural net
works emerged: feedforward neural networks (FNN), and feedback
neural networks. The feedback neural networks are mainly RNN models.
The feedforward neural networks can be divided into single-hiddenlayer neural networks and multiple-hidden-layer neural networks
based on the number of hidden layers. The classification and quantita
tive statistics of different types of neural networks are shown in Fig. 2.
ANN is a mathematical model algorithm that mimics the distributed,
parallel information processing seen in animal behavior (Hubara et al.,
2018). This network accomplishes information processing by adjusting
the connections between numerous internal nodes. In an ANN network,
each neuron receives input signals, processes them using an activation
function, and generates output signals. The connections between neu
rons have weights, which determine the strength of influence of one
neuron on another. During training, these weights are adjusted based on
specified learning algorithms, such as backpropagation, to minimize the
difference between the actual output and the expected output (Mocanu
et al., 2018; Samek et al., 2021). Notably, Ghalandari et al. (Ghalandari
et al., 2021) employed a Multiple-Hidden-Layer Artificial Neural
Network, known as an MLP model, for predicting carbon emissions in
the UK, Germany, Italy, and France. Their findings indicated that the
MLP model exhibited lower error rates compared to a GMDH model.
NNAR (p, k), short for Nonlinear AutoRegressive Neural Network, is
a type of feedforward neural network designed specifically for time se
ries analysis. In NNAR (p, k), the input to the neural network consists of
lagged values of the time series data. It comprises a single hidden layer
with p nodes, each representing a lagged input, and k nodes in the
output layer. This architecture allows the model to capture nonlinear
relationships within the time series data (Daniyal et al., 2022). The
5
Y. Jin et al.
Science of the Total Environment 927 (2024) 172319
the Kaya identity, the logarithmic mean Divisia index (LMDI) model,
and a scenario analysis method. Their combined model predicted
China’s CO2 emissions to reach 112.89 million tons by 2030. Cui et al.
(Cui et al., 2018) predicted carbon dioxide emissions from China’s
power sector using hybrid PLS grey Markov model, foreseeing an in
crease to 51.029 million tons by 2025. Moreover, Meng et al. (Meng
et al., 2014) combined the GM (1, 1) prediction equation with a linear
model, creating a mixed equation. Their analysis, focusing on fore
casting China’s carbon emissions from 1992 to 2011, demonstrated that
the hybrid model outperformed the traditional linear model and GM (1,
1) in accuracy.
Statistical-intelligent models merge classical statistical prediction
models with intelligent models to enhance forecasting accuracy and
stability. This approach aims to leverage the strengths of statistical and
machine learning models. Within this framework, some foundational
prediction models are grounded in statistical theory, while others stem
from machine learning algorithms (e.g., support vector machines, neural
networks). This approach combined the interpretability of statistical
models and the ability of intelligent models to handle nonlinearity and
high-dimensional data. In response to the random oscillation sequence
(ROS) pattern observed in industrial carbon emission processes, Hu et al.
(Hu and Lv, 2020) developed the ROGM-AFSA-GVM model. This model
combined a ROS-based Grey Model (GM) with a General Vector Machine
(GVM) optimized using the artificial fish swarm algorithm (AFSA). Li
et al. (Li et al., 2018b) employed the GM (1, 1) predicted output as input
for the SVM-ELM model to forecast carbon emissions in the BeijingTianjin-Hebei region, revealing potential emissions control below 96.9
million tons by 2030. Wang et al. (Wang et al., 2018) combined a GM
and a GRNN model to predict China’s primary energy consumption from
2017 to 2030, utilizing the GRA to allocate weights between the models.
Zhao et al. (Zhao et al., 2018a) constructed a mixed frequency data
sampling (MIDAS) regression model and a hybrid BPNN (MIDAS-BP)
approach to predict CO2 emissions in the United States. The results
showed that the prediction ability of MIDAS-BP is significantly better
than MIDAS, and the other four individual prediction models. Zhou et al.
(Zhou et al., 2017b) created a hybrid GNNM model, amalgamating GM
and BPNN models for carbon emission prediction. Their findings
demonstrated that the GNNM model offered improved predictions for
carbon emissions, effectively capturing the nonlinear and nonstationary
nature of carbon emissions.
Intelligent-intelligent models involve merging diverse shallow
intelligent prediction models and neural network prediction models to
capitalize on their respective strengths. By combining predictions from
various intelligent models, this approach reduces prediction errors and
improves precision. Leveraging the advantages of different machine
learning models in handling nonlinearity, high-dimensional data, and
capturing intricate relationships is the cornerstone of this approach. Yet,
handling complex numerical features may lead to overfitting issues. Li
(Li, 2020) constructed a KLS algorithm by combining Kalman filter (KF),
LSTM and SVM to predict carbon emissions in China from 2015 to 2030.
The results indicated that the carbon equivalent in China will continue
to increase, peak in 2024, and then gradually decline. In addition,
Acheampong et al. (Acheampong and Boateng, 2019) utilized an ANNfeedforward multilayer perceptron (FFMLP) model to project carbon
emissions across Australia, Brazil, China, India, and United States. Their
outcomes highlighted an exceedingly minimal prediction error between
the forecasted and actual carbon emissions.
Subsequently, this optimized ELM was employed to forecast carbon
emissions in Hebei, China.
Unlike feedforward neural networks above, RNN is a typical feed
back neural network. Their cyclic structure enables the transmission of
information from preceding moments to subsequent ones, facilitating
specific actions at each moment (Sherstinsky, 2020). The hidden layer of
a RNN contains recurrent connections, enabling the network to capture
temporal dependencies in carbon emission data. At each time step, the
hidden layer receives input from both the current input data and the
previous hidden state, allowing it to retain information over time. The
output layer generates predictions or outputs based on the information
learned by the hidden layer (Sherstinsky, 2020). Building upon this
concept, Mason et al. (Mason et al., 2018) proposed an RNN model in
which the parameters were optimized with the CMA-ES evolutionary
algorithm. Finally, 2.5-hour-ahead carbon emissions and energy con
sumption in Ireland were successfully predicted.
An RNN’s hidden state relies on the previous moment’s hidden layer
and the current input, lacking long-term memory advantage. Hence,
Sepp Hochreiter et al. introduced LSTM as an enhanced RNN model
(Hochreiter and Schmidhuber, 1997). LSTM is a type of RNN designed to
handle sequential data and capture long-term dependencies. Unlike
RNNs, LSTM addresses the vanishing gradient problem by introducing
memory units and three gate mechanisms, input gate, forget gate, and
output gate. The memory units of LSTM maintain a constant state over
time, allowing them to retain information in long sequences. The input
gate controls the flow of new information into the units, while the forget
gate regulates the retention of previous state information. The output
gate determines which parts of the unit’s state are output to the next
layer. These gate mechanisms controlled by activation functions enable
LSTM to selectively update and forget information based on input and
past states, effectively capturing complex temporal dependencies (Van
Houdt et al., 2020). Bismark et al. (Ameyaw et al., 2020) predicted
carbon emissions in West Africa from 2015 to 2030 by using a bidirec
tional long short-term memory (BiLSTM) model. Huang et al. (Huang
et al., 2019b) applied LSTM for predictions in China, revealing an
anticipated 30 % decrease in carbon emissions per unit of GDP between
2015 and 2020. Similarly, Bismark et al. (Ameyaw and Yao, 2018) uti
lized BiLSTM to forecast carbon emissions in Ghana, Nigeria, Burkina
Faso, Senegal, and Benin spanning from 2015 to 2020.
3.1.4. Combined prediction models
Combined prediction models hold significance in CEPM. They
involve crafting multiple individual prediction models and integrating
them strategically to create a unified model (Tascikaraoglu and Uzu
noglu, 2014; Wang and Srinivasan, 2017). The reviewed combined
models can be divided into three groups: statistical-statistical models,
statistical-intelligent models, and intelligent-intelligent models, as
shown in Fig. 2. The intelligent models include shallow intelligent pre
diction models and neural network prediction models. Based on our
review, combined statistical-intelligent prediction models are the most
common among researchers (9 papers). The comparative analysis of
combined models is shown in the Online Supplementary Appendices (II,
and III).
Statistical-statistical models involve combining multiple classic sta
tistical prediction models to enhance forecasting. Its objective is to
amalgamate the strengths of various statistical models, thereby
improving prediction accuracy and stability. These models, based on
statistical theory, encompass linear regression, polynomial regression,
exponential smoothing, among others. By integrating the predictions of
multiple foundational models, Statistical-statistical models mitigates
forecasting errors and heightens precision. Additionally, it counters
limitations in individual models, thus bolstering overall forecast stabil
ity. Nonetheless, determining combination weights relies on experi
mentation and experience, and for intricate problems like nonlinearity
or non-stationarity, these models may encounter limitations. Wang et al.
(Wang et al., 2019d) established a combined prediction model based on
3.1.5. Summary of the prediction models
In statistical prediction models, regression models serve as concise
and intuitive statistical tools with good interpretability, especially the
linear regression. However, their adaptability to nonlinear relationships
is limited, and when dealing with complex, high-dimensional carbon
emission issues, regression models may struggle to comprehensively
consider various influencing factors. Univariate time series methods,
such as ARIMA, perform well in handling trend and seasonality in data
6
Y. Jin et al.
Science of the Total Environment 927 (2024) 172319
but may be influenced by fluctuations when applied to carbon emissions
data affected by multiple factors. Grey models (GM), particularly GM (1,
1), are widely used in CEPM, constituting 29.1 % of all reviewed models.
While GM excels in handling small sample sizes and incomplete infor
mation, it has notable drawbacks. GM often relies on assumptions about
initial values during the modeling process, heavily dependent on the
researcher’s experience. Although enhanced grey models improve pre
dictive accuracy, they introduce more complexity, requiring additional
parameter tuning and optimization.
We conducted a survey on two shallow intelligent prediction models:
Support Vector Machine and Decision Tree. SVM is characterized by its
advantages, including a faster learning speed and robust generalization
ability. However, its effectiveness is highly dependent on hyper
parameters, particularly the penalty coefficient ‘C.’ LSSVM, an improved
version of SVM, simplifies computations by replacing quadratic pro
gramming with least squares. While this simplification enhances
computational efficiency, LSSVM carries the risk of overfitting, espe
cially when the penalty coefficient is too high. On the other hand,
Gradient Boosting Decision Tree (GBDT), as an ensemble model using
decision trees as basic learners, demonstrates high accuracy. GBDT en
hances precision by iteratively combining weak prediction models.
However, similar to LSSVM, GBDT requires careful control of the
number of iterations to prevent overfitting.
There are 32 papers discussing different neural networks, which are
categorized into feedforward neural networks and feedback neural
networks. ANN, as the most basic feedforward neural network, excels in
approximating complex nonlinear relationships in static data and is one
of the most popular models in CEPM research. BPNN, as a multilayer
neural network, demonstrates good generalization ability and adap
tivity; however, it has a slow convergence speed and is prone to local
minima. In contrast, ELM uses the analytical solution method of
randomly selecting hidden layer nodes, eliminating the need for itera
tive tuning and exhibiting efficient training speed on large datasets.
However, ELM is not suitable for handling long-term carbon emission
time series data, and its interpretability is relatively poor due to its
stochastic nature. Feedback neural networks mainly include RNN and
LSTM. RNN captures data features over time through its cyclic structure,
but its hidden state relies heavily on the previous moment’s output and
current input, lacking a long-term memory advantage. LSTM, as an
improved model of RNN, introduces gates and memory units, over
coming the limitations of traditional RNNs. This enables LSTM to
perform well in predicting long-term carbon emission data, especially
when considering economic changes and policy adjustments.
Combined prediction models include statistical-statistical models,
statistical-intelligent models, and intelligent-intelligent models.
Statistical-statistical models integrate classical statistical prediction
models such as linear regression and exponential smoothing, offering
advantages of simplicity, interpretability, and low complexity. They
perform well in linear carbon emission datasets with clear relationships,
mitigating errors through the combination of multiple base models.
However, these models are less proficient in handling nonlinear and
high-dimensional carbon emission data. Statistical-intelligent models
combine classical statistical models with intelligent models like support
vector machines and neural networks, capable of handling relatively
complex carbon emission data. They leverage the strengths of both
statistical and machine learning methods, suitable for a wide range of
prediction tasks. The limitation lies in the determination of combination
weights, which heavily relies on experience or other models. Intelligentintelligent models, combining various shallow intelligent prediction
models and neural networks, maximize prediction accuracy and stabil
ity. Nevertheless, these models exhibit high computational complexity,
poor interpretability, and a susceptibility to overfitting. Due to their
numerous parameters, optimization often requires the assistance of
other models.
3.2. Optimization models
In CEPM, optimization is the process of finding the best possible
solution to one or more models. As shown in Fig. 3, optimization models
are broadly categorized into programming models and metaheuristic
analysis. Metaheuristic analysis can be subdivided into four categories
based on their intended purpose. As a result, the metaheuristic analysis
used for parameter optimization accounts for more than half of these
models (55 %). Among all the optimization models, the number of
programming models is small, being used in only three papers. The
Online Supplementary Appendix III shows the metaheuristic optimiza
tion results in CEPM in the research published during 2011–2022.
3.2.1. Programming models
Programming models originated from mathematical methods and
are widely used in the field of model optimization. According to their
structure, programming models can be divided into linear programming
and nonlinear programming groups. In terms of optimization objectives,
programming models can be divided into single-objective programming
and multi-objective programming. In 2017 (Wang and Ye, 2017), Wang
et al. introduced the power exponent of the relevant variable as an
exogenous factor in the multivariate grey model. To minimize the mean
absolute percentage error, they developed two nonlinear programming
models, aimed at determining the power exponent for the nonlinear
multivariate model.
3.2.2. Metaheuristic analysis (MA)
MA differs from traditional optimization methods by not relying on
gradients, ensuring randomness and ease of use, while effectively
bypassing local minima (Hussain et al., 2019). Typically, the optimiza
tion parameters of the MA model are defined initially, along with the
objective function (Katebi et al., 2020). The search initiates with
randomly generated candidate solutions, evolving to generate the sub
sequent generations. Within the 147 CEPM papers scrutinized, a pre
dominant approach involves metaheuristic algorithms that emulate
natural biological or physical phenomena, structuring mathematical
models to tackle problems.
MA can be combined with predictive models, to rapidly select the
most accurate parameters or parameter combinations, for the prediction
model. Wen et al. (Wen and Cao, 2020b) proposed the enhanced but
terfly optimization algorithm (EBOA) to optimize the parameters of an
LSSVM. The results show that compared with the BOA-LSSVM, PSOLSSVM, GA-LSSVM, CS-LSSVM and SVM models, the EBOA-LSSVM
model yields the lowest error value. Sun et al. (Sun and Zhang, 2020)
established the bacterial foraging optimization (BFO) method to opti
mize the regularization parameters and kernel parameters of LSSVM and
used the subsequent model for prediction. The robustness and parameter
optimization performance of the LSSVM were enhanced. Dai et al. (Dai
et al., 2018) proposed the modified shuffled frog leaping algorithm
(MSFLA) to optimize the regularization parameter and kernel function
width of an LSSVM model. The prediction accuracy of the proposed
MSFLA-LSSVM model was better than that of the considered SFLALSSVM, LSSVM and BPNN Models. Wu et al. (Wu et al., 2020) estab
lished a metaheuristic ant-lion optimizer (ALO) to optimize the frac
tional order α in the heterogeneous grey model CFNGM.
Structural optimization is defined as the optimization of the structure
of a prediction model by using the performance characteristics of a
metaheuristic model, such as the number of hidden layers in the neural
network. The CS-PSO method was proposed by Chiroma et al. (Chiroma
et al., 2015) to optimize the number of hidden layer neurons in ANN.
Sun et al. (Sun et al., 2017) applied PSO to optimize the weight of the
input layer and the deviation in the hidden layer in ELMs, thereby
obtaining an optimal PSO-ELM network. The prediction accuracy was
better than that of the traditional ELM model and BPNN model. Niu et al.
(Niu et al., 2020) introduced the improved fireworks optimization al
gorithm (IFWA) to optimize the smoothing factor in GRNN, which
7
Y. Jin et al.
Science of the Total Environment 927 (2024) 172319
Fig. 3. (a) Proportion of the two optimization models among all models, with blue indicating mataheuristic models, green for programming models. (b) Classifi
cation of optimization models. (c) Time series distribution of different mataheuristic models, color-coded as shown in (b). (d) Time series distribution of different
programming models.
improved the prediction accuracy. In the models reviewed, weight
optimization was often the focus of structural optimization. A total of 9
papers (14, 51) applied weight optimization, and threshold optimization
(Chai et al., 2022; Zhang et al., 2022) and learning rate optimization
(Zhao et al., 2022) were also common.
There are many ways to construct a combined prediction model. The
most important step is to ensure that the advantages of each prediction
model are maximized. One such approach is to obtain the optimal
weight coefficient combination for the applied prediction model. MA is
used to search the optimal weights of each model to achieve the optimal
prediction results for CEPMs. Liu et al. (Liu et al., 2014) constructed a
combined prediction model with GM (1, 1), ARIMA and second-order
polynomial regression (SOPR). PSO was used to optimize the com
bined weights of the above models, and the prediction accuracy was
improved.
Performance improvements were discussed in some of the reviewed
papers, especially in cases that involved optimizing shallow intelligent
models. Wen et al. (Wen and Cao, 2020b) proposed the butterfly opti
mization algorithm (BOA) to optimize the performance of an LSSVM.
Compared with the PSO-LSSVM, GA-LSSVM, CS-LSSVM and SVM
models, the BOA-LSSVM model displayed the best performance. A
similar performance improvement method was adopted by Wen et al.
(Wen and Cao, 2020a), who used ICSO to improve the performance of a
SVM.
3.2.3. Summary of optimization models
In conclusion, optimization models play a crucial role in CEPM,
particularly in finding the optimal parameters for predictive models.
This study categorizes optimization models into programming models
and MA. Programming models, rooted in mathematical methods, are
divided into linear and nonlinear programming groups, addressing sin
gle or multi-objective optimization. Metaheuristic Analysis is an
advancement over traditional methods, utilizing algorithms inspired by
natural phenomena. Among all the CEPM papers applying optimization
models, MA is widely used, accounting for 94.4 %, with a focus on al
gorithms mimicking biological or physical processes. MA is applied for
parameter optimization in predictive models. Examples include the
Enhanced Butterfly Optimization Algorithm (EBOA) for optimizing
LSSVM parameters, and the Ant-Lion Optimizer (ALO) for optimizing
parameters in the grey model CFNGM. Structural optimization involves
optimizing the structure of predictive models, such as the number of
hidden layers in neural networks. Various Metaheuristic algorithms are
employed for structural optimization, including CS-PSO for optimizing
hidden layer neurons in ANN. Weight optimization is a common focus,
with methods like PSO used to optimize weights in combined prediction
models. Performance improvements were discussed in most of the
reviewed papers, especially in cases that involved optimizing shallow
intelligent models. In the majority of papers, optimization models
enhance the performance and accuracy of predictive models in CEPM.
8
Y. Jin et al.
Science of the Total Environment 927 (2024) 172319
3.3. Prediction factor selection models
conducted an analysis on factors influencing carbon emissions in the
crop industry using multiple linear regression, pinpointing social
development and energy inputs as the predominant factors. Meanwhile,
Zhao et al. (Zhao et al., 2022) proposed QAP regression analysis, and the
results showed that the dominant factor leading to differences in carbon
emissions in the Yellow River Basin from 2000 to 2010 was per capita
GDP. After 2010, the main factor affecting carbon emissions in the
Yellow River Basin was population. Ridge regression (RR) is another
method for selecting the significantly correlated variables. RR can
effectively avoid multicollinearity and is used by many scholars. The
largest difference between RR and normal linear regression is the loss
function. The regularization parameter lambda is introduced in RR to
make the coefficients of unimportant prediction factors very small (Li,
2020). Using an RR model, Yu et al. (Yu et al., 2018) utilized an RR
model to emphasize the notable impact of population aging on CO2
emissions. They highlighted a direct correlation between the aging
population and the scale of CO2 emissions in their study.
Indeed, alongside grey prediction models like GM (1, 1) and GM (1,
N), GRA holds significance within the grey system. It serves to gauge the
strength and scale of relationships among system factors. Utilized for
assessing connections between pairs of options through remote mea
surements, GRA assists in identifying quantitative correlations among
intricate factors within a system. For small sample sets and irregular
data, GRA has displayed good analytical performance in the field of
Selecting the most appropriate prediction factor for a prediction
problem is a key step (Khaki and Wang, 2019). The traditional predic
tion model operates without preconceived assumptions about the dis
tribution of the involved factors. However, carbon emission prediction is
generally a macro-prediction task with many influencing factors (Cab
aneros et al., 2019). Therefore, the robustness of a prediction model
depends largely on the model form and the way prediction factors are
input into the model. Additionally, prediction factor selection models
can be used to extract the factors suitable for prediction, and the degree
of correlation between different factor pairs can be determined. Pre
diction factor selection models can be divided into two categories,
namely, statistical models and machine learning models, which are like
prediction models. Some prediction models can also be used as predic
tion factor selection models, such as classical regression models. More
over, machine learning models can be divided into supervised learning
models and unsupervised learning models. The classification and
quantitative of said models are given in Fig. 4.
3.3.1. Statistical models
In addition to forecasting, multiple linear regression serves as a
pivotal tool for determining quantitative interdependencies among
variables in prediction factor selection. Tian et al. (Tian et al., 2016)
Fig. 4. (a) Proportion of the two prediction factor selection models among all models, with blue indicating statistical models, and green for machine learning models.
(b) Classification of prediction factor selection models. (c) Time series distribution of different statistical models, color-coded as shown in (b). (d) Time series
distribution of different machine learning models.
9
Y. Jin et al.
Science of the Total Environment 927 (2024) 172319
CEPM. Ding et al. (Ding et al., 2020) employed GRA to identify factors
potentially exerting strong nonlinear influences on carbon emissions.
This insight into nonlinear effects led to the design of a novel discrete
grey model. Chiu et al. (Chiu et al., 2020) developed a new multivariate
grey prediction model (MGPM) for CO2 emissions. GRA was used to
filter the relevant features with weak correlations with carbon emissions
and finally improve the prediction accuracy. Based on GRA, Zhou et al.
(Zhou et al., 2017a) concluded that the factors that influence carbon
emissions from strong to weak are the installed thermal power capacity,
thermal power generation, urbanization rate, GDP and unit utilization.
FA is a statistical method employed to uncover underlying common
factors within sets of variables (Jones et al., 2015). FA reduces large sets
of prediction factors to a few factors by calculating the correlations
between prediction factors. The extracted common factors serve to
substitute the original prediction factors, preserving essential informa
tion and efficiently representing the intricate interrelations among the
prediction factors. Sun et al. (Sun et al., 2017) integrated FA with an
ELM for carbon emission forecasting, utilizing FA to identify crucial
input prediction factors.
Sensitivity analysis (SA) assesses the influence of each input variable
on carbon emission intensity by calculating the partial rank correlation
coefficient (PRCC) between these variables. Acheampong et al.
(Acheampong and Boateng, 2019) performed sensitivity analyses for
carbon emissions in four countries. In Australia, R&D displayed the
highest sensitivity weight. Urbanization exhibited the highest sensitivity
weight in Brazil. Population size showed the highest sensitivity weight
in China, and energy consumption demonstrated the highest sensitivity
weight in India.
3.3.3. Summary of the prediction factor selection models
In conclusion, the selection of prediction factors is a critical aspect in
addressing macro-prediction tasks such as carbon emission prediction,
characterized by numerous influencing factors. Prediction factor selec
tion models fall into two main categories: statistical models and machine
learning models. In the realm of statistical models, methods like MLR,
RR, GRA, FA, and SA are prominent. MLR elucidates interdependencies
between variables, RR effectively handles multicollinearity, GRA is
suitable for small sample sets and irregular data to identify strong
nonlinear influences, and SA assesses the impact of each input variable
on carbon emission intensity. On the other hand, machine learning
models such as PCA and RF provide alternative approaches. PCA ach
ieves dimension reduction to enhance information while reducing
complexity, and RF excels in revealing complex features and quantifying
their importance. Despite the strengths of these models, they also have
limitations. For instance, statistical models like MLR may struggle to
capture complex nonlinear relationships, and the applicability of GRA
may vary based on data characteristics. Machine learning-based selec
tion models also have limitations; PCA may lose important information
during dimension reduction, and RF, due to its relatively complex model
structure, may pose challenges in interpretability.
4. Discussion
4.1. Influencing factors and research objectives
4.1.1. Influencing factors
The influential factors and the classification and quantitative statis
tics of influencing factors in CEPMs are shown in Fig. 5. All factors were
divided into five sectors: energy (92 times), economy (111 times),
population (68 times), environment (24 times) and industry (46 times).
Energy factors generally include energy consumption and energy
structure, with 79 and 13 associated papers, respectively. Li et al. (Li
et al., 2018b) extensively investigated carbon emission’s primary energy
sources in the Beijing-Tianjin-Hebei region, focusing on coal, gasoline,
natural gas, and coal power. Their analysis emphasized the significant
impact of energy consumption proportions on carbon emissions, fore
seeing a 45 % projection in power and natural gas consumption by 2030.
Additionally, Zhao et al. (Zhao et al., 2018b) scrutinized the multifac
eted influences of GDP, population, energy consumption, economic
structure, energy structure, urbanization rate, and energy intensity on
carbon emissions. They revealed a negative correlation between the
energy structure and carbon emissions.
Most studies have focused on the impacts of economic factors on
carbon emissions. Economic factors can be divided into three categories:
the overall economy, investment, and urbanization. In 67 papers, overall
economic factors were used in CEPMs, although energy consumption
factors were used in more studies. The main index used was GDP. Cao
et al. (Cao et al., 2016) observed that GDP growth positively correlates
with increased carbon emissions. In another study, Sun et al. (Sun et al.,
2021) highlighted the positive influence of per capita GDP in Belt and
Road countries on carbon emissions. Investment variables were also
linked to carbon emissions, evidenced by Jiang et al. (Jiang et al.,
2021a) identifying a nonlinear association between CO2 equivalent and
foreign direct investment, with outward foreign direct investment
amplifying CO2 emissions in China. Urbanization, the transition from
traditional agricultural rural settings to modern urban societies, was
identified as a significant factor affecting carbon emissions in China by
Ma et al. (Ma et al., 2020) using the Apriori algorithm.
Population factors include the total population, policy, consumption
level and education level. The total population is the most frequently
used population factor. A total of 36 papers used the total population as
an independent variable. Most papers were published in 2018, with a
total of 9 papers discussing the impact of population on carbon emis
sions in that year. Cui et al. (Cui et al., 2018) explored the driving forces
behind CO2 emissions in China’s power industry, attributing increased
3.3.2. Machine learning
PCA is a prevalent method for dimensionality reduction and feature
extraction in carbon prediction. It generates new orthogonal mappings
to represent the data in a lower-dimensional space, capturing the
maximum variance (Zhang et al., 2019). This reduction in dimension
ality helps eliminate potential basic directions with limited information.
PCA is commonly used to process and analyze various types of data,
reducing high-dimensional and complex prediction factors to those
suitable for analysis, and improving prediction quality. Scholars in
CEPMs widely apply PCA in their studies (Wen and Cao, 2020a; Sun
et al., 2019; Sun and Sun, 2017; Sun and Wang, 2021). For example,
Huang et al. (Huang et al., 2019b) developed the grey-PCA-LSTM model.
PCA is used for feature extraction, and 15 prediction factors were
transformed into 4 prediction factors, ultimately improving the predic
tion accuracy of the model. Kernel principal component analysis (KPCA)
was proposed by Wen et al. (Wen and Cao, 2020b). Initially, prediction
factors with the most significant contributions to the selected function
were extracted using KPCA. Subsequently, these prediction factors were
combined with an EBOA-LSSVM prediction model to forecast carbon
emissions in the residential sector of the Yangtze River Delta. The model
with KPCA demonstrated superior performance compared to the original
prediction model.
RF stands as an integrated machine learning approach widely
employed in prediction and feature selection. Its strength lies in
unraveling complex interactive features and quantifying feature
importance. RF possesses robustness against noise, handles data gaps
proficiently, and showcases swift learning abilities. Consequently, RF
serves as a pivotal tool for feature selection, particularly in complex
domains like CEPM. In their work Sun et al. (Sun et al., 2018) introduced
a hybrid model merging RF and ELM techniques for CEPM. Remarkably,
RF was utilized to scrutinize emission-influencing factors. Empirical
simulations highlighted the superiority of the proposed RF-MFO-ELM
model over parallel models like MFO-ELM, RF-PSOELM, RF-ELM, and
RF-BP, showcasing substantial advantages in predictive fitting and
accuracy.
10
Y. Jin et al.
Science of the Total Environment 927 (2024) 172319
Fig. 5. (a) Classification of influencing factors. (b) Quantitative statistics of influencing factors.
emissions to human activities. Zhao et al. (Zhao et al., 2017), employing
Grey Relational Analysis (GRA), revealed a robust correlation between
China’s population and carbon emissions, recording a correlation coef
ficient of 0.6798. Additionally, Niu et al. (Niu et al., 2020) conducted
scenario analyses to explore China’s commitment goals for 2030. They
concluded that policy dynamics significantly influence carbon emis
sions, where policy-driven scenarios directly impact achieving these
goals. The analysis by Sun et al. (Sun and Zhang, 2020) indicated that
education plays a pivotal role as one of the main drivers of energyrelated carbon emissions. Furthermore, consumption levels, another
population factor, were discussed in 15 papers.
In 24 papers, the impact of environmental factors on carbon emis
sions was analyzed. Environmental factors, including climate factors,
greenhouse gases (Yu et al., 2018; Sutthichaimethee and Kubaha, 2018;
Tang et al., 2016), and temperature (Singh and Kennedy, 2015) were
identified as the most commonly considered factors. Yu (Hao and Wei,
2015), after building an economic model, found that carbon efficiency is
the most important factor in determining whether there is a turning
point in total CO2 emissions. Singh and Kennedy (Singh and Kennedy,
2015) demonstrated, via scenario analysis, that climate change can have
a marginal impact on the growth trajectory of energy-related carbon
emissions.
In the reviewed papers, the main industry-related factors are the
industrial structure, technological level, and transportation. Notably,
Zhu et al. (Zhu et al., 2019) utilized a Markov chain model and scenario
analysis to forecast industrial and energy structures. Their findings
highlighted that adjusting the industrial structure could potentially have
a greater impact on achieving carbon intensity targets than modifying
the energy structure. Transportation-related factors involved metrics
like the number of motor vehicles and transport mileage. Lin et al. (Lin
et al., 2018) conducted a grey correlation analysis, revealing that the
highest correlation with carbon emissions was observed with the num
ber of registered motor vehicles. Additionally, the impact of scientific
and technological progress on carbon emissions was explored in 12
studies.
4.1.2. Research objectives
The prediction models studied have different goals. In 147 papers, six
targets were identified, including total population-related carbon
emissions, economic sector carbon emissions, and industry carbon and
energy-related carbon emissions. The classification and quantitative
statistics of study targets are given in Fig. 6.
The overall targets include total carbon emissions (TCE) and the
carbon emission intensity (CEI). There were 94 studies of total carbon
Fig. 6. (a) Classification of study targets. (b) Quantitative statistics of study targets.
11
Y. Jin et al.
Science of the Total Environment 927 (2024) 172319
emissions in the analysis, which is the most studied target, accounting
for 63.5 % of all the papers. Additionally, there were far more TCE
studies than CEI studies, and some studies focused on both TCE and the
CEI. For example, Zhou et al. (Zhou et al., 2021) forecasted TCE and the
CEI in China, and the results showed that both are projected to decrease
from 2018 to 2030. Niu et al. (Niu et al., 2020) showed that in certain
scenarios, China’s TCE will gradually decline after reaching a peak in
2030. Additionally, there is still some pressure to achieve CEI reduction
targets. Population-related carbon emissions are mainly carbon emis
sions in the residential sector. Four papers have studied carbon emis
sions in the residential sector. Wen et al. (Wen and Cao, 2020b)
predicted the emissions from the residential sector in the Yangtze River
Delta region and concluded that the carbon emissions from the resi
dential sector increased from 55.86 Mt in 2000 to 192.36 Mt in 2017.
In the papers reviewed, the relevant carbon emissions from four
sectors, namely, the industrial, agricultural, construction and trans
portation sectors, were analyzed. Among them, the transportation sector
was most explored, with a total of seven papers discussing this sector in
the context of carbon emissions. Huang et al. (Huang et al., 2022)
studied carbon emissions from the transportation sector. Between 2019
and 2025, the average increases in carbon emissions from the trans
portation sector in China and the United States were 2.837 % and 2.394
%, respectively, and a decreasing trend was observed in Japan, with an
average decline of 1.2231 %. Industrial sector carbon emissions were
explored in a total of six papers. Yu et al. (Yu and Xu, 2019) used the GM
(1, 1) model to conclude that industrial carbon emissions will peak in
2030, suggesting that the emission reduction targets will be met. In
addition, the industrial sector includes the cement industry (OfosuAdarkwa et al., 2020) and the ferrous metal industry (Huang et al.,
2019c). Two papers studied the steel industry in the ferrous metal in
dustry (Gao et al., 2015; Li et al., 2022a). Carbon emissions in the
construction industry were studied in four papers. Sutthichaimethee
et al. (Sutthichaimethee and Kubaha, 2018) used a combined model to
study the carbon emissions from the Thai construction industry over 20
years (2019–2038). The results showed that the carbon emissions from
the construction industry will increase by 37.88 % in this period. In
addition, one paper studied agricultural carbon emissions (Jiang et al.,
2021b) and interindustry carbon emission transfer (Hu and Lv, 2020).
One paper studied trade-related carbon emissions in the economic
sector. Wang et al. (Wang et al., 2019b) delved into trade-related carbon
emissions within the economic sector, simulating and predicting carbon
emissions resulting from bilateral trade. Their findings indicated a sig
nificant increase in net carbon outflow from China to Australia, rising
from 2.2 million tons to 15.5 million tons between 2000 and 2014. We
divided energy-related carbon emissions into total energy carbon
emissions, power-related carbon emissions and fossil energy carbon
emissions. Fifteen studies focused on total energy carbon emissions, 5
papers focused on power-related carbon emissions, and 8 papers focused
on fossil energy carbon emissions. Fossil energy carbon emissions
include coal-related carbon emissions (Duan and Luo, 2020; Duan et al.,
2020).
4.2. Model evaluation
4.2.1. Overview
Model evaluation methods are used to evaluate the performance of
prediction models. The three main indicators include accuracy, error,
and complexity. Among them, error is used the most frequently, and
many derivative indicators have been applied, including the MAPE (70
papers), RMSE (51 papers) and MAE (28 papers). The specific classifi
cation and use frequency of these indexes are shown in Fig. 7. The
evaluation results of prediction models in CEPM in the research pub
lished during 2011–2022 are shown in the Online Supplementary Ap
pendix II.
A positive correlation exists between the accuracy index and the
prediction accuracy of a model. In the 147 CEPM papers reviewed, three
accuracy indexes were used: R, R2 and IA. The R value, also known as
the Pearson correlation coefficient, was proposed by statistician Carl
Pearson. R2, or the coefficient of determination, quantifies the propor
tion of the variance in the dependent variable y that’s explained by the
independent variable x in a regression model. This metric measures how
well the model fits the observed data. Among the accuracy indicators, R2
is the most used (16 papers), followed by R (5 papers). In addition, Duan
(Duan and Luo, 2020) used the index of agreement (IA) to evaluate the
accuracy of a prediction model.
Unlike accuracy index, error index indicates good model perfor
mance when the corresponding values are small. CEPM research mainly
involves absolute error, relative error, and square error. Absolute errors
are based on the absolute difference between the actual and modeled
outputs. Indicators of this type include the mean absolute percentage
error (MAPE) and mean absolute error (MAE). Sun et al. (Sun and Zhang,
2020) used the maximum absolute percent error (MaxAPE) and median
absolute percent error (MdAPE) to evaluate models. The square error is
based on the square between the actual value and the model output.
Common indicators include the standard error (STD) and mean square
error (MSE). The commonly used root mean square error (RMSE) is an
Fig. 7. (a). Classification and (b). Quantitative analysis of evaluation methods.
12
Y. Jin et al.
Science of the Total Environment 927 (2024) 172319
extension of MSE and was applied in 51 CEPM studies. Different from
the absolute error, the traditional relative error refers to the ratio of the
predicted value to the true value. Common examples of this type include
the mean relative error (MRE), which was applied in three selected
CEPM papers.
Complexity indicators include information standards, such as the
Akaike information criterion (AIC). In addition to modeling error, the
AIC considers the complexity of the model. Ning et al. (Ning et al., 2021)
used the AIC to evaluate the ARIMA model. The results showed that the
AIC values of the ARIMA (1,2,0) and ARIMA (2,2,0) models were
5.947830 and 5.643812, respectively.
thirteen models exceeding the maximum value of the neural network
prediction model (163.68 Mt). All models were evaluated using MAPE,
and the MAPE for all models ranged from 0.008 % to 37.14 %.
Further analysis of the optimized model errors showed a reduction in
RMSE for all models, with the minimum and maximum values being
0.0003 Mt and 95.14 Mt, corresponding to the WOA-ELM (Sun and
Huang, 2022) and BA-BPNN (Li et al., 2022a) models, respectively. The
MAE for the optimized models also decreased, with median values of
0.24 Mt for neural network prediction models and 30.9 Mt for statistical
prediction models. For MAPE, the optimized models showed a signifi
cant decrease, ranging from 0.002 % to 6.54 %.
4.2.2. Comparative analysis
As shown in Fig. 8, we counted the three most frequently used error
evaluation metrics, RMSE, MAE, and MAPE. For comparison purposes,
we focused solely on carbon emission studies with units in Mt. For the
baseline CEPMs, the RMSE ranged from 0.11 to 1635 Mt. Among them,
the Shallow Intelligent Prediction Model and the Ensemble Prediction
Model performed well, with median RMSE values of 1.45 Mt and 0.42
Mt, respectively. Out of the nine Ensemble Prediction Models consid
ered, eight had RMSE values <1 Mt, demonstrating stable performance.
Statistical prediction models showed significant variability, with the
highest and lowest RMSE models among all baseline CEPMs being sta
tistical prediction models, specifically ARIMA (2, 1) (Ding et al., 2020)
(1635 Mt) and EDGM (1, 1) (Guo et al., 2021) (0.11 Mt). This indicates
the need for improved stability in statistical models. For neural network
prediction models, the RMSE ranged from 0.30 to 212.51 Mt. Some
studies evaluated the MAE error of the models, with the MAE for all
models distributed between 0.09 and 452.57 Mt. Comparing neural
network prediction models and statistical prediction models, their
respective median MAE values were 0.45 Mt and 104.22 Mt. However,
some statistical prediction models had large MAE errors, with six out of
4.3. Temporal and spatial distributions
4.3.1. Temporal distribution
a. Overall temporal trend
Fig. 9 shows the temporal distribution of review papers and the
utilization frequency of prediction models in papers by year. The first
prediction model was published in 2011. Initially, a few models were
developed, with 2, 3, 1 and 3 in 2011, 2012, 2013 and 2014, respec
tively. The number of papers involving CEPM was 7 in 2015, increasing
to 6 and 10 in 2016 and 2017, respectively. Perhaps affected by the
coronavirus epidemic, the number of research papers decreased from 18
to 12 from 2018 to 2019. In the past five years, 78.4 % of all the CEPM
papers included in this study were published. From 2019 to 2020, the
number of CEPM increased from 12 to 26, an increase of 115.4 %.
Similarly, in 2020, China proposed a carbon neutrality plan, and there
may be a certain correlation between this plan and the publication
trend. In the next year, 2021, the largest number of papers (31 papers)
was published. The scope of this review extended through August 2022,
Fig. 8. (1) Accuracy of baseline CEPMs. (2) Accuracy of optimized CEPMs.
13
Y. Jin et al.
Science of the Total Environment 927 (2024) 172319
Fig. 9. (a) Temporal distribution of review papers. (b) Utilization frequency of prediction models in papers by year.
so the number of papers on CEPM is likely to continue to rise in 2022,
and even for the foreseeable future.
methods began to increase in popularity. However, combined model use
displays no obvious trend, and the fluctuation range is large. In 2018, 6
papers used combined prediction methods, making these models 4 times
more popular than shallow intelligent models (2 papers). 2019–2021,
the utilization frequency decreases to less than four for combined
models. The reason for this trend may be that combined models are more
computationally complex and take longer to run than individual models.
b. Model comparison
Among the forecasting methods, the first to be used in the study
period were statistical models, of which GM was the first (Pao and Tsai,
2011). Since 2014, statistical models have been used every year, with an
overall upward trend peaking in 2021, with a total of 18 papers that used
statistical models. Since 2015, the use of shallow intelligent models and
neural network models has increased significantly. Shallow intelligent
models reached a peak (5 papers) in 2020 and 2021 and have since
declined in use, while neural network models have continued to grad
ually increase in use, with 3, 7, and 10 corresponding papers published
in 2020, 2021 and 2022, respectively. After 2017, combined prediction
4.3.2. Spatial distribution
As shown in Fig. 10a, 83 countries were investigated 282 times in the
studied papers. Most of the papers focused on the carbon emissions of
individual countries, and 12 papers considered carbon emissions in
multiple countries, including those from interstate organizations, on five
continents. In addition, 7 papers focused on global carbon emissions,
and in 1 paper, carbon emissions throughout Europe were predicted.
Fig. 10. (a). Frequency statistics of countries studied on five continents. (b). Frequency statistics for the Asian countries studied. (c). Frequency statistics for the
American countries studied. (d). Frequency statistics for the African countries studied. (e). Frequency statistics for the European countries studied. (f). Frequency
statistics for the Oceanian countries studied.
14
Y. Jin et al.
Science of the Total Environment 927 (2024) 172319
Among the five continents, Asia was most studied, and 22 countries
were mentioned 159 times, with 114 in reference to China. Among all
the papers discussing China, 38.6 % studied different regions of China,
including cross-provincial regions (6 papers), provincial regions (33
papers) and municipal regions (5 papers). In addition to China, Turkey
and Japan were studied 6 times each, followed by South Korea and India
at 4 times each. The frequency statistics for the Asian countries studied is
given in Fig. 10b.
As shown in Fig. 13c., a total of 9 countries in the Americas were
studied, of which the United States was the most frequently studied
country in North America, with a total of 9 papers, followed by Canada
(5 papers). In South America, Brazil was most studied (4 papers), mainly
because Brazil is a member of the BRICS organization, which was
studied three times.
Oceania displayed the lowest number of countries studied at 3,
namely, Australia (6 papers), New Zealand (3 papers) and Papua New
Guinea (1 paper). The frequency statistics for the Oceanian countries
studied are given in Fig. 10 f.
As shown in Fig. 10e, Europe was second only to Asia in terms of
mentions across all papers reviewed, with 58 mentions across 27
countries and Europe as a whole. In terms of the frequency distribution,
the countries displayed relatively similar trends, with 8 countries stud
ied once and 13 countries studied twice. Russia was the most studied
country (5 papers), followed by the UK, France and Italy (4 papers each).
As shown in Fig. 10d, Similar to the findings for Europe, the fre
quency of study of African countries was relatively consistent, ranging
from 1 to 3 times. Most countries were studied once, with only Nigeria
being studied three times.
4.4. Future trends in carbon emissions
In this section, we examine the countries with large numbers of
studies and their future carbon emissions predicted with CEPMs. The
four selected countries are China, the USA, Japan, and Russia. As shown
in Fig. 11, according to the previous model classification, we rank the
carbon emission prediction values. Models with high levels of intelli
gence are shown in darker colors, and shallow intelligent models and
statistical models are illustrated in light colors.
4.4.1. China
Fig. 11a shows the future predictions of carbon emissions in China in
review papers. Most of the models provide consistent assessments of
future carbon emissions in China, with carbon emissions likely to in
crease through 2025 or 2030. Notably, the MWOA-GBDT forecast for
2020–2030 is above average. It is worth noting that a SVM model pre
diction suggests that carbon emissions in China peaked in 2013 and
declined year by year from 2013 to 2018. The predicted value in 2018
was 7841.6 Mt, which is lower than the predicted value of all other
models in 2018. The maximum value in 2018 was 11,081.9 Mt predicted
by FAGM (1, 1), which is 41.3 % higher than the prediction of a SVM. In
addition, the Verhulst model predictions fluctuate significantly from
2008 to 2018. AVGM (1, 1) also displays similar fluctuations. Overall,
for CEPM in China, highly intelligent models tend to yield higher pre
dicted values, but overall fluctuations are stable. In addition, researchers
prefer to use intelligent prediction models to forecast large-span time
series. In contrast, shallow intelligent models and statistical models,
especially various types of GMs, have displayed good prediction per
formance; thus, GMs are the most widely used prediction models.
However, models with low intelligence are prone to large fluctuations in
predicted values.
Fig. 11. (a). Future predictions of carbon emissions in China in review papers (Wu et al., 2020; Chiu et al., 2020; Qiao et al., 2020; Chang et al., 2013; Xu et al., 2021;
Ikram et al., 2021; Gao et al., 2021; Tong et al., 2021; Huang et al., 2021; Qiao et al., 2021; Gao et al., 2022; Li et al., 2022b). (b). Future predictions of carbon
emissions in the Americas in review papers (Qiao et al., 2020; Chang et al., 2013; Xu et al., 2021; Ikram et al., 2021; Gao et al., 2021; Huang et al., 2021). (c). Future
predictions of carbon emissions in Japan in review papers (Qiao et al., 2020; Chang et al., 2013; Xu et al., 2021; Gao et al., 2021; Huang et al., 2021). (d). Future
predictions of carbon emissions in Russia in review papers (Wu et al., 2020; Chang et al., 2013; Xu et al., 2021; Tong et al., 2021; Gao et al., 2022).
15
Y. Jin et al.
Science of the Total Environment 927 (2024) 172319
4.4.2. USA
As shown in Fig. 11b, we found that 11 public prediction models
were reported. Different models displayed obvious differences. Three
models predicted that the carbon emissions of the United States would
plummet in 2016 to 2017 and will approach zero in 2025. The predicted
values of VSSGM (1, 1), NDGM (1, 1) and ODGM (1, 1) for the United
States in 2025 are 965.5855, 675.90985 and 642.1143575, respectively,
which are much lower than those of other models. Among the other
eight models, MLANN has the shortest prediction span of two years. In
addition, the predicted trends of 8 models are stable, with 3 models
indicating a slow increasing trend and 5 models indicating a slow
decreasing trend.
5. Conclusion
To explore CEPMs, we summarize 147 relevant papers published
between 2012 and August 2022 and classify the models into three main
groups: prediction models, optimization models, and prediction factor
selection models. In addition, we discuss other indispensable compo
nents of CEPM, such as influencing factors, research objectives, model
evaluation, spatial and temporal scales, and future trends. We draw the
following conclusions from this review.
Among all the prediction models, statistical models were the most
prevalent, accounting for 75 instances—more than half—followed by
neural network models at 21.8 %. Notably, the use of neural network
models has consistently risen from 2019 to 2022, predominantly fa
voring feedforward architectures. A substantial portion of CEPMs
incorporated optimized approaches, with metaheuristic models consti
tuting 94.4 %. Within this category, parameter optimization took pre
cedence, followed by structure optimization. Prediction factor selection
models played a role in filtering factors, employing Grey Relational
Analysis (GRA) and Principal Component Analysis (PCA) as represen
tatives for statistical and machine learning models, respectively.
We scrutinized the accuracy of both pre-optimized and postoptimized CEPMs. In the pre-optimized stage, RMSE values spanned
from 0.112 to 1635 Mt, showcasing a polarized performance among
statistical prediction models—some highly accurate, others less so.
Combined prediction models, however, displayed a relatively consistent
performance. Post-optimization, there was a noticeable improvement
across the board, with RMSE, MAE, and MAPE values reducing. The
minimum RMSE reached 0.0003 Mt, and the maximum was 95.14 Mt.
CEPMs typically factored in five influences: energy, economy, pop
ulation, environment, and industry. Energy and economic consider
ations dominated, with 92 and 111 times, respectively.
Drawing insights from the reviewed literature, we outlined pro
spective carbon emission trajectories for China, the United States,
Japan, and Russia. China exhibited a continuous upward trend until
2025 in six models, with only one suggesting a minor decline. Pro
jections indicate China’s total carbon emissions may surpass 10,000 Mt
by 2025. Conversely, the majority of models for the United States
indicated a decline, with certain results suggesting the U.S. might
approach carbon neutrality by 2025, as suggested by VSSGM (1, 1),
NDGM (1, 1), and ODGM (1, 1).
4.4.3. Japan
Among the four major countries reviewed, Japan has the lowest
forecasted values of total carbon emissions, with an average value of
1215.66 Mt. The maximum value predicted by the combined DMSFE
model for 2012 was 1321.4904 Mt. The two minimum values are pre
dicted after 2020: 1085.31 Mt in 2022 predicted by CFNGM (1, 1) and
1091.78 Mt in 2025 predicted by the LSO-GA-LSSVM. The future pre
dictions of carbon emissions in Japan in review papers is given in
Fig. 11c.
4.4.4. Russia
As shown in Fig. 11d, the overall future predictions for Russia are
like those for China, with 18.2 % of the models, such as the Verhulst
model and the AVGM (1, 1) model, displaying significant fluctuations. A
total of 45.5 % of the model prediction curves are similar, and the range
of carbon emissions is between 1400 and 1600 Mt. The corresponding
models are NGM (1, 1), NGMO (1, 1), FAGM (1, 1), FANGM (1, 1), and
CFNGM (1, 1).
4.5. Limitations of existing research and future suggestions
Firstly, we should consider the influence of significant events, such as
COVID-19, in the CEPM domain. The novel coronavirus pandemic has
had a profound impact on various fields of research around the world.
However, few studies have considered the impact of novel coronavirus
epidemic factors, such as the number of novel coronavirus infections or
mortality, on future carbon emissions. Secondly, some studies have
explored multiple regions in one paper, but only individual regional
predictions were provided, without considering the geographical links
among regions. Like other gases in the atmosphere, CO2 can be trans
ported, so the links between carbon emissions and geographic and other
factors need to be further considered. Addressing this issue requires
additional scientific data collection and in-depth data analysis. Thirdly,
Machine learning techniques, including shallow intelligent models and
neural network models, have been proven to be suitable for big data
forecasting and have extremely high accuracy, including in stock fore
casting and building energy consumption forecasting (energy con
sumption forecasting 1). However, in the field of carbon emissions
forecasting, there are not significantly more studies that use machine
learning techniques than statistical models. The reason for this phe
nomenon may be that most of the carbon emissions forecasting data are
annual data (134 papers), so the amount of data is small, and it is
difficult to fully train machine learning models. Determining how to
overcome the inherent disadvantages of the data is worth studying in the
future. Finally, all the prediction models reviewed predict carbon
emissions at the city level, provincial level, cross-provincial level, na
tional level and global level to help control carbon emissions at the
macroscales, but these models are difficult to implement at the building
level. Due to the lack of research on specific buildings, large amounts of
data are not available to statistically analyze, undoubtedly hindering the
development of CEPM.
CRediT authorship contribution statement
Yukai Jin: Writing – original draft, Formal analysis, Conceptuali
zation. Ayyoob Sharifi: Writing – review & editing, Supervision.
Zhisheng Li: Writing – review & editing, Formal analysis. Sirui Chen:
Writing – review & editing, Formal analysis. Suzhen Zeng: Writing –
review & editing, Data curation. Shanlun Zhao: Writing – review &
editing.
Declaration of Generative AI and AI-assisted technologies in the
writing process
During the preparation of this work, the author(s) used AI-assisted
tools to check the grammar, punctuation, and clarity. After using the
tools, the author(s) reviewed and edited the content as needed and take
(s) full responsibility for the content of the publication.
Declaration of competing interest
The authors declare that they have no known competing financial
interests or personal relationships that could have appeared to influence
the work reported in this paper.
16
Science of the Total Environment 927 (2024) 172319
Y. Jin et al.
Data availability
Acknowledgment
Data will be made available on request.
This work was supported by the SmaSo-X Challenge Project Graduate
Students Research Grant.
Appendix A
List of abbreviations, including units and nomenclature.
Abbreviations
Nomenclature
CEPM
TCE
CEI
GM
PR
LR
GPR
EKC
NGBM
ARMA
ARIMA
ARDL
SVM
EKC
ELC
GSM
BPNN
GPR
LSSVM
GBDT
CART
NN
FNN
ANN
NNAR
GRNN
RBFNN
CNN
ELM
RNN
LSTM
BiLSTM
LMDI
ROS
GVM
MIDAS
OLS
PDL
ADL
DMSE
QHS
KF
MA
AFSA
KF
BOA
EBOA
PSO
BFO
MSFLA
CS
ALO
FWA
IFWA
SOPR
GRA
SA
FA
PCA
RELM
RF
LCA
R2
CO2 emission prediction model
Total carbon emissions
Carbon emission intensity
Grey model
Panel regression
Linear regression
Gaussian process regression
Environmental Kuznets curve
Nonlinear grey Bernoulli model
Autoregressive moving average
Autoregressive integrated moving average
Autoregressive distributed lag
Support vector machines
Environmental Kuznets curve
Environmental logic curve
Green-Solow model
Backpropagation neural network
Gaussian process regression
Least squares support vector machine
Gradient boosting decision tree
Classification and regression tree
Neural network
Feedforward neural networks
Artificial neural network
Neural network autoregressive
Generalized Regression Neural Network
Radial Basis Function Neural Network
Convolutional neural network
Extreme learning machine
Recurrent neural network
Long short-term memory
Bidirectional long short-term memory
Logarithmic mean Divisia index
Random oscillation sequence
General vector machine
Mixed frequency data sampling
Ordinary least squares
Polynomial distribution lag
Autoregressive distribution lag
Discounted mean square error
Quantum harmony search
Kalman filter
Metaheuristic analysis
Artificial fish swarm algorithm
Kalman filter
Butterfly optimization algorithm
Enhanced butterfly optimization algorithm
Particle swarm optimization
Bacterial foraging optimization
Modified shuffled frog leaping algorithm
Cuckoo Search
Ant-lion optimizer
Fireworks optimization algorithm
Improved fireworks optimization algorithm
Second-order polynomial regression
Grey relation analysis
Sensitivity analysis
Factor analysis
Principal component analysis
Regularized extreme learning machine
Random Forest
Life cycle assessment
The coefficient of determination
(continued on next page)
17
Y. Jin et al.
Science of the Total Environment 927 (2024) 172319
(continued )
Abbreviations
Nomenclature
IA
MAPE
MAE
MaxAPE
MdAPE
STD
RMSE
MRE
AIC
IPCC
The index of agreement
Mean absolute percentage error
Mean absolute error
Maximum absolute percent error
Median absolute percent error
Standard error
Root mean square error
Mean relative error
Akaike information criterion
Intergovernmental Panel on Climate Change
Appendix B. Supplementary data
Supplementary data to this article can be found online at https://doi.org/10.1016/j.scitotenv.2024.172319.
References
Dai, S., Niu, D., Han, Y., 2018. Forecasting of energy-related CO2 emissions in China
based on GM(1,1) and least squares support vector machine optimized by modified
shuffled frog leaping algorithm for sustainability. Sustainability 10 (4).
Daniyal, M., et al., 2022. Comparison of conventional modeling techniques with the
neural network autoregressive model (NNAR): application to COVID-19 data.
J. Healthc. Eng. 2022.
Deng, J.L., 1982. Control-problems of grey systems. Syst. Control Lett. 1 (5), 288–294.
Ding, S., et al., 2020. Estimating Chinese energy-related CO2 emissions by employing a
novel discrete grey prediction model. J. Clean. Prod. 259.
Dong, F., Hua, Y.F., Yu, B.L., 2018. Peak carbon emissions in China: status, key factors
and countermeasures-a literature review. Sustainability 10 (8).
Du, L., Wei, C., Cai, S., 2012. Economic development and carbon dioxide emissions in
China: provincial panel data analysis. China Econ. Rev. 23 (2), 371–384.
Duan, H.M., Luo, X.L., 2020. Grey optimization Verhulst model and its application in
forecasting coal-related CO(2)emissions. Environ. Sci. Pollut. Res. 27 (35),
43884–43905.
Duan, H.M., et al., 2020. A novel forecasting approach based on multi-kernel nonlinear
multivariable grey model: a case report. J. Clean. Prod. 260.
Duus-Otterström, G., 2023. Subsistence emissions and climate justice. Br. J. Polit. Sci. 53
(3), 919–933.
Gao, C., et al., 2015. Analyzing and forecasting CO2 emission reduction in China’s steel
industry. Front. Earth Sci. 9 (1), 105–113.
Gao, M.Y., et al., 2021. A novel fractional grey Riccati model for carbon emission
prediction. J. Clean. Prod. 282.
Gao, M.Y., et al., 2022. A novel method for carbon emission forecasting based on
Gompertz’s law and fractional grey model: evidence from American industrial
sector. Renew. Energy 181, 803–819.
Gasser, T., et al., 2015. Negative emissions physically needed to keep global warming
below 2 ◦ C. Nat. Commun. 6.
Ghalandari, M., et al., 2021. Energy-related carbon dioxide emission forecasting of four
European countries by employing data-driven methods. J. Therm. Anal. Calorim.
144 (5), 1999–2008.
Grubb, M., et al., 2015. A review of Chinese CO2 emission projections to 2030: the role of
economic structure and policy. Clim. Pol. 15, S7–S39.
Gu, J.X., et al., 2018. Recent advances in convolutional neural networks. Pattern Recogn.
77, 354–377.
Guo, J.L., et al., 2021. Forecasting carbon dioxide emissions in BRICS countries by
exponential cumulative grey model. Energy Rep. 7, 7238–7250.
Han, Y.M., et al., 2023. Novel economy and carbon emissions prediction model of
different countries or regions in the world for energy optimization using improved
residual neural network. Sci. Total Environ. 860.
Hao, Y., Wei, Y.-M., 2015. When does the turning point in China’s CO2 emissions occur?
Results based on the Green Solow model. Environ. Dev. Econ. 20 (6), 723–745.
Hien, N.L., Kor, A.L., 2022. Analysis and prediction model of fuel consumption and
carbon dioxide emissions of light-duty vehicles. Appl. Sci. Basel 12 (2).
Hochreiter, S., Schmidhuber, J., 1997. Long short-term memory. Neural Comput. 9 (8),
1735–1780.
Hsu, A.E., et al., 2022. Predicting European cities’ climate mitigation performance using
machine learning. Nat. Commun. 13 (1).
Hu, Y., Lv, K.J., 2020. Hybrid prediction model for the interindustry carbon emissions
transfer network based on the Grey model and general vector machine. IEEE Access
8, 20616–20627.
Huang, L., et al., 2019a. A systematic review of empirical methods for modelling sectoral
carbon emissions in China. J. Clean. Prod. 215, 1382–1401.
Huang, Y.S., Shen, L., Liu, H., 2019b. Grey relational analysis, principal component
analysis and forecasting of carbon emissions based on long short-term memory in
China. J. Clean. Prod. 209, 415–423.
Huang, J.B., Luo, Y.M., Feng, C., 2019c. An overview of carbon dioxide emissions from
China’s ferrous metal industry: 1991-2030. Resour. Policy 62, 541–549.
Huang, H.F., Wu, X.M., Cheng, X.F., 2021. The prediction of carbon emission information
in Yangtze River economic zone by deep learning. Land 10 (12).
Abeydeera, L., Mesthrige, J.W., Samarasinghalage, T.I., 2019. Global research on carbon
emissions: a scientometric review. Sustainability 11 (14).
Aboagye, P.D., Sharifi, A., 2023. Post-fifth assessment report urban climate planning:
Lessons from 278 urban climate action plans released from 2015 to 2022. Urban
Clim. 49, 101550 https://doi.org/10.1016/j.uclim.2023.101550.
Aboagye, P.D., Sharifi, A., 2024. Urban climate adaptation and mitigation action plans: A
critical review. Renew. Sustain. Energy Rev. 189, 113886 https://doi.org/10.1016/j.
rser.2023.113886.
Acheampong, A.O., Boateng, E.B., 2019. Modelling carbon emission intensity:
application of artificial neural network. J. Clean. Prod. 225, 833–856.
Ahmad, A.S., et al., 2014. A review on applications of ANN and SVM for building
electrical energy consumption forecasting. Renew. Sustain. Energy Rev. 33,
102–109.
Al-Mahasneh, A.J., Anavatti, S.G., 2023. Novel general regression neural networks for
improving control accuracy of nonlinear MIMO discrete-time systems. IEEE Trans.
Cybern. 53 (10), 6122–6132.
Ameyaw, B., Yao, L., 2018. Analyzing the impact of GDP on CO2 emissions and
forecasting Africa’s total CO2 emissions with non-assumption driven bidirectional
long short-term memory. Sustainability 10 (9).
Ameyaw, B., et al., 2020. West Africa’s CO2 emissions: investigating the economic
indicators, forecasting, and proposing pathways to reduce carbon emission levels.
Environ. Sci. Pollut. Res. 27 (12), 13276–13300.
Basha, S.H.S., et al., 2020. Impact of fully connected layers on performance of
convolutional neural networks for image classification. Neurocomputing 378,
112–119.
Cabaneros, S.M., Calautit, J.K., Hughes, B., 2019. A review of artificial neural network
models for ambient air pollution prediction. Environ. Model Softw. 119, 285–304.
Cao, Y., et al., 2016. A comprehensive study on low-carbon impact of distributed
generations on regional power grids: a case of Jiangxi provincial power grid in
China. Renew. Sustain. Energy Rev. 53, 766–778.
Chai, Z.Y., et al., 2022. Carbon emissions index decomposition and carbon emissions
prediction in Xinjiang from the perspective of population-related factors, based on
the combination of STIRPAT model and neural network. Environ. Sci. Pollut. Res. 29
(21), 31781–31796.
Chang, H., Sun, W., Gu, X., 2013. Forecasting energy CO2 emissions using a quantum
harmony search algorithm-based DMSFE combination model. Energies 6 (3),
1456–1477.
Chauhan, V.K., Dahiya, K., Sharma, A., 2019. Problem formulations and solvers in linear
SVM: a review. Artif. Intell. Rev. 52 (2), 803–855.
Chen, C.I., 2008. Application of the novel nonlinear grey Bernoulli model for forecasting
unemployment rate. Chaos, Solitons Fractals 37 (1), 278–287.
Chen, M.R., et al., 2019. A two-layer nonlinear combination method for short-term wind
speed prediction based on ELM, ENN, and LSTM. IEEE Internet Things J. 6 (4),
6997–7010.
Cheng, J., Li, G., Chen, X.H., 2019. Research on travel time prediction model of freeway
based on gradient boosting decision tree. IEEE Access 7, 7466–7480.
Chiroma, H., et al., 2015. Global warming: predicting OPEC carbon dioxide emissions
from petroleum consumption using neural network and hybrid cuckoo search
algorithm. PLoS One 10 (8).
Chiu, Y.J., et al., 2020. A multivariate Grey prediction model using neural networks with
application to carbon dioxide emissions forecasting. Math. Probl. Eng. 2020.
Cui, H., Wu, R., Zhao, T., 2018. Decomposition and forecasting of CO2 emissions in
China’s power sector based on STIRPAT model with selected PLS model and a novel
hybrid PLS-Grey-Markov model. Energies 11 (11).
Cui, X.W., et al., 2021. Forecasting of carbon emission in China based on gradient
boosting decision tree optimized by modified whale optimization algorithm.
Sustainability 13 (21).
18
Y. Jin et al.
Science of the Total Environment 927 (2024) 172319
Huang, S.Y., Xiao, X.P., Guo, H., 2022. A novel method for carbon emission forecasting
based on EKC hypothesis and nonlinear multivariate grey model: evidence from
transportation sector. Environ. Sci. Pollut. Res. 29 (40), 60687–60711.
Hubara, I., et al., 2018. Quantized neural networks: training neural networks with low
precision weights and activations. J. Mach. Learn. Res. 18.
Hussain, K., et al., 2019. Metaheuristic research: a comprehensive survey. Artif. Intell.
Rev. 52 (4), 2191–2233.
Ikram, M., et al., 2021. Assessment and prediction of environmental sustainability: novel
grey models comparative analysis of China vs. the USA. Environ. Sci. Pollut. Res. 28
(14), 17891–17912.
Jiang, H., et al., 2021a. Forecasting China’s CO2 emissions by considering interaction of
bilateral FDI using the improved grey multivariable Verhulst model. Environ. Dev.
Sustain. 23 (1), 225–240.
Jiang, J.Q., Zhao, T., Wang, J., 2021b. Decoupling analysis and scenario prediction of
agricultural CO2 emissions: an empirical analysis of 30 provinces in China. J. Clean.
Prod. 320.
Jones, R.V., Fuertes, A., Lomas, K.J., 2015. The socio-economic, dwelling and appliance
related factors affecting electricity consumption in domestic buildings. Renew.
Sustain. Energy Rev. 43, 901–917.
Katebi, J., et al., 2020. Developed comparative analysis of metaheuristic optimization
algorithms for optimal active control of structures. Eng. Comput. 36 (4), 1539–1558.
Kaur, J., Parmar, K.S., Singh, S., 2023. Autoregressive models in environmental
forecasting time series: a theoretical and application review. Environ. Sci. Pollut.
Res. 30 (8), 19617–19641.
Khaki, S., Wang, L.Z., 2019. Crop yield prediction using deep neural networks. Front.
Plant Sci. 10.
Kiranyaz, S., et al., 2021. 1D convolutional neural networks and applications: a survey.
Mech. Syst. Signal Process. 151.
Li, Y., 2020. Forecasting Chinese carbon emissions based on a novel time series
prediction method. Energy Sci. Eng. 8 (7), 2274–2285.
Li, F., Xu, Z., Ma, H., 2018a. Can China achieve its CO2 emissions peak by 2030? Ecol.
Indic. 84, 337–344.
Li, M., et al., 2018b. Forecasting carbon emissions related to energy consumption in
Beijing-Tianjin-Hebei region based on Grey prediction theory and extreme learning
machine optimized by support vector machine algorithm. Energies 11 (9).
Li, Y., Wei, Y.G., Dong, Z., 2020a. Will China achieve its ambitious goal?-forecasting the
CO2 emission intensity of China towards 2030. Energies 13 (11).
Li, L., et al., 2020b. Using improved gradient-boosted decision tree algorithm based on
Kalman filter (GBDT-KF) in time series prediction. J. Supercomput. 76 (9),
6887–6900.
Li, W., Zhang, S.H., Lu, C., 2022a. Research on the driving factors and carbon emission
reduction pathways of China’s iron and steel industry under the vision of carbon
neutrality. J. Clean. Prod. 357.
Li, W., Zhang, S.H., Lu, C., 2022b. Exploration of China’s net CO2 emissions evolutionary
pathways by 2060 in the context of carbon neutrality. Sci. Total Environ. 831.
Lillicrap, T.P., et al., 2020. Backpropagation and the brain. Nat. Rev. Neurosci. 21 (6),
335–346.
Lin, B.Q., Agyeman, S.D., 2019. Assessing Ghana’s carbon dioxide emissions through
energy consumption structure towards a sustainable development path. J. Clean.
Prod. 238.
Lin, C.-C., He, R.-X., Liu, W.-Y., 2018. Considering multiple factors to forecast CO2
emissions: a hybrid multivariable Grey forecasting and genetic programming
approach. Energies 11 (12).
Liu, S.F., Yang, Y.J., 2017. Explanation of terms of grey forecasting models. Grey Syst.
Theory Appl. 7 (1), 123–128.
Liu, L., et al., 2014. Can China realize its carbon emission reduction goal in 2020: from
the perspective of thermal power development. Appl. Energy 124, 199–212.
Liu, S.F., et al., 2016. On the new model system and framework of Grey system theory.
J. Grey Syst. 28 (1), 1–15.
Ma, Z.C., Dai, Q., 2016. Selected an stacking ELMs for time series prediction. Neural.
Process. Lett. 44 (3), 831–856.
Ma, X.J., Jiang, P., Jiang, Q.C., 2020. Research and application of association rule
algorithm and an optimized grey model in carbon emissions forecasting. Technol.
Forecast. Soc. Chang. 158.
Malik, A., et al., 2020. Forecasting CO2 emissions from energy consumption in Pakistan
under different scenarios: the China-Pakistan Economic Corridor. Greenh. Gases Sci.
Technol. 10 (2), 380–389.
Mason, K., Duggan, J., Howley, E., 2018. Forecasting energy demand, wind generation
and carbon dioxide emissions in Ireland using evolutionary neural networks. Energy
155, 705–720.
Meinshausen, M., et al., 2022. Realization of Paris Agreement pledges may limit warming
just below 2 ◦ C. Nature 604 (7905), 304-+.
Meng, M., Niu, D., Shang, W., 2014. A small-sample hybrid model for forecasting energyrelated CO2 emissions. Energy 64, 673–677.
Mocanu, D.C., et al., 2018. Scalable training of artificial neural networks with adaptive
sparse connectivity inspired by network science. Nat. Commun. 9.
Moher, D., Liberati, A., Tetzlaff, J., Altman, D.G., The, P.G., 2009. Preferred Reporting
Items for Systematic Reviews and Meta-Analyses: The PRISMA Statement. Plos
Medicine 6 (7), e1000097. https://doi.org/10.1371/journal.pmed.1000097.
Murshed, M., et al., 2023. Can energy productivity gains harness the carbon dioxideinhibiting agenda of the Next 11 countries? Implications for achieving sustainable
development. Sustain. Dev. 31 (1), 307–320.
Nations, U, 1998. Kyoto Protocol to the United Nations Framework Convention on
Climate Change.
Nations, U, 2016. The Paris Agreement.
Ning, L.Q., Pei, L.J., Li, F., 2021. Forecast of China’s carbon emissions based on ARIMA
method. Discret. Dyn. Nat. Soc. 2021.
Niu, D.X., et al., 2020. Can China achieve its 2030 carbon emissions commitment?
Scenario analysis based on an improved general regression neural network. J. Clean.
Prod. 243.
Ofosu-Adarkwa, J., Xie, N.M., Javed, S.A., 2020. Forecasting CO2 emissions of China’s
cement industry using a hybrid Verhulst-GM(1,N) model and emissions’ technical
conversion. Renew. Sust. Energ. Rev. 130.
Pachauri, R.K., et al., 2014. In: Pachauri, R.K., Meyer, L. (Eds.), Climate Change 2014:
Synthesis Report. Contribution of Working Groups I, II and III to the Fifth Assessment
Report of the Intergovernmental Panel on Climate Change. IPCC, Geneva,
Switzerland, p. 151.
Pao, H.-T., Tsai, C.-M., 2011. Modeling and forecasting the CO2 emissions, energy
consumption, and economic growth in Brazil. Energy 36 (5), 2450–2458.
Pao, H.-T., Fu, H.-C., Tseng, C.-L., 2012. Forecasting of CO2 emissions, energy
consumption and economic growth in China using an improved grey model. Energy
40 (1), 400–409.
Qiao, W.B., et al., 2020. A hybrid algorithm for carbon dioxide emissions forecasting
based on improved lion swarm optimizer. J. Clean. Prod. 244.
Qiao, Z.R., Meng, X.M., Wu, L.F., 2021. Forecasting carbon dioxide emissions in APEC
member countries by a new cumulative grey model. Ecol. Indic. 125.
Rae, J.W.B., et al., 2021. Atmospheric CO<sub>2</sub> over the past 66 million years
from marine archives. In: Jeanloz, R., Freeman, K.H. (Eds.), Annual Review of Earth
and Planetary Sciences, vol. 49, pp. 609–641, 2021.
Rumelhart, D.E., Hinton, G.E., Williams, R.J., 1986. Learning representations by backpropagating errors. Nature 323 (6088), 533–536.
Samek, W., et al., 2021. Explaining deep neural networks and beyond: a review of
methods and applications. Proc. IEEE 109 (3), 247–278.
Sherstinsky, A., 2020. Fundamentals of recurrent neural network (RNN) and long shortterm memory (LSTM) network. Phys. D-Nonlinear Phenom. 404.
Singh, S., Kennedy, C., 2015. Estimating future energy use and CO2 emissions of the
world’s cities. Environ. Pollut. 203, 271–278.
Smith, P., et al., 2016. Biophysical and economic limits to negative CO<sub>2</sub>
emissions. Nat. Clim. Chang. 6 (1), 42–50.
Sun, W., Huang, C.C., 2022. Predictions of carbon emission intensity based on factor
analysis and an improved extreme learning machine from the perspective of carbon
emission efficiency. J. Clean. Prod. 338.
Sun, W., Liu, M., 2016. Prediction and analysis of the three major industries and
residential consumption CO2 emissions based on least squares support vector
machine in China. J. Clean. Prod. 122, 144–153.
Sun, W., Sun, J., 2017. Prediction of carbon dioxide emissions based on principal
component analysis with regularized extreme learning machine: the case of China.
Environ. Eng. Res. 22 (3), 302–311.
Sun, W., Wang, Y.W., 2021. Prediction and analysis of CO2 emissions based on
regularized extreme learning machine optimized by adaptive whale optimization
algorithm. Pol. J. Environ. Stud. 30 (3), 2755–2767.
Sun, W., Zhang, J.J., 2020. Analysis influence factors and forecast energy-related CO(2)
emissions: evidence from Hebei. Environ. Monit. Assess. 192 (10).
Sun, W., Wang, C., Zhang, C., 2017. Factor analysis and forecasting of CO2 emissions in
Hebei, using extreme learning machine based on particle swarm optimization.
J. Clean. Prod. 162, 1095–1101.
Sun, W., Wang, Y., Zhang, C., 2018. Forecasting CO2 emissions in Hebei, China, through
moth-flame optimization based on the random forest and extreme learning machine.
Environ. Sci. Pollut. Res. 25 (29), 28985–28997.
Sun, W., Jin, H.Y., Wang, X.M., 2019. Predicting and analyzing CO2 emissions based on
an improved least squares support vector machine. Pol. J. Environ. Stud. 28 (6),
4391–4401.
Sun, L.L., Cui, H.J., Ge, Q.S., 2021. Driving factors and future prediction of carbon
emissions in the ‘Belt and Road Initiative’ countries. Energies 14 (17).
Sutthichaimethee, J., Kubaha, K., 2018. Forecasting energy-related carbon dioxide
emissions in Thailand’s construction sector by enriching the LS-ARIMAXi-ECM
model. Sustainability 10 (10).
Suykens, J.A.K., Vandewalle, J., 1999. Least squares support vector machine classifiers.
Neural. Process. Lett. 9 (3), 293–300.
Tang, D., et al., 2016. Trend prediction and decomposed driving factors of carbon
emissions in Jiangsu Province during 2015-2020. Sustainability 8 (10).
Tascikaraoglu, A., Uzunoglu, M., 2014. A review of combined approaches for prediction
of short-term wind speed and power. Renew. Sustain. Energy Rev. 34, 243–254.
Tian, J.X., et al., 2016. Drivers of agricultural carbon emissions in Hunan Province,
China. Environ. Earth Sci. 75 (2).
Tong, M.Y., Duan, H.M., He, L.Y.H., 2021. A novel Grey Verhulst model and its
application in forecasting CO2 emissions. Environ. Sci. Pollut. Res. 28 (24),
31370–31379.
Tudor, C., 2016. Predicting the evolution of CO2 emissions in Bahrain with automated
forecasting methods. Sustainability 8 (9).
van Hooidonk, R., et al., 2016. Local-scale projections of coral reef futures and
implications of the Paris agreement. Sci. Rep. 6.
Van Houdt, G., Mosquera, C., Nápoles, G., 2020. A review on the long short-term
memory model. Artif. Intell. Rev. 53 (8), 5929–5955.
Wang, Z.-X., Li, Q., 2019. Modelling the nonlinear relationship between CO2 emissions
and economic growth using a PSO algorithm-based grey Verhulst model. J. Clean.
Prod. 207, 214–224.
Wang, Z.Y., Srinivasan, R.S., 2017. A review of artificial intelligence based building
energy use prediction: contrasting the capabilities of single and ensemble prediction
models. Renew. Sustain. Energy Rev. 75, 796–808.
19
Y. Jin et al.
Science of the Total Environment 927 (2024) 172319
Yang, Y., Zhang, J., Wang, C., 2018. Forecasting China’s carbon intensity: is China on
track to comply with its Copenhagen commitment? Energy J. 39 (2), 63–86.
Ye, F., et al., 2018. An improved Grey model and scenario analysis for carbon intensity
forecasting in the Pearl River Delta Region of China. Energies 11 (1).
Yin, L., et al., 2022. Urban carbon accounting: an overview. Urban Clim. 44, 101195.
Yu, Y., Xu, W., 2019. Impact of FDI and R&D on China’s industrial CO2 emissions
reduction and trend prediction. Atmos. Pollut. Res. 10 (5), 1627–1635.
Yu, Y., Deng, Y.-r., Chen, F.-f., 2018. Impact of population aging and industrial structure
on CO2 emissions and emissions trend prediction in China. Atmos. Pollut. Res. 9 (3),
446–454.
Zhang, Y.G., et al., 2019. A novel hybrid model based on VMD-WT and PCA-BP-RBF
neural network for short-term wind speed forecasting. Energy Convers. Manag. 195,
180–197.
Zhang, S.S., Huo, Z.G., Zhai, C.C., 2022. Building carbon emission scenario prediction
using STIRPAT and GA-BP neural network model. Sustainability 14 (15).
Zhang, J.X., et al., 2023. Shear strength prediction and failure mode identification of
beam-column joints using BPNN, RBFNN, and GRNN. Arab. J. Sci. Eng. 48 (4),
4421–4437.
Zhao, X., Du, D., 2015. Forecasting carbon dioxide emissions. J. Environ. Manag. 160,
39–44.
Zhao, H., Guo, S., Zhao, H., 2017. Energy-related CO2 emissions forecasting using an
improved LSSVM model optimized by whale optimization algorithm. Energies 10
(7).
Zhao, X., et al., 2018a. Forecasting carbon dioxide emissions based on a hybrid of mixed
data sampling regression model and back propagation neural network in the USA.
Environ. Sci. Pollut. Res. 25 (3), 2899–2910.
Zhao, H., Huang, G., Yan, N., 2018b. Forecasting energy-related CO2 emissions
employing a novel SSA-LSSVM model: considering structural factors in China.
Energies 11 (4).
Zhao, J.J., et al., 2022. Carbon emission prediction model and analysis in the Yellow
River Basin based on a machine learning method. Sustainability 14 (10).
Zhou, D.X., 2020. Universality of deep convolutional neural networks. Appl. Comput.
Harmon. Anal. 48 (2), 787–794.
Zhou, J., et al., 2017a. Carbon emissions scenario prediction of the thermal power
industry in the Beijing-Tianjin-Hebei region based on a back propagation neural
network optimized by an improved particle swarm optimization algorithm. Pol. J.
Environ. Stud. 26 (4), 1895–1904.
Zhou, J., Guang, F., Gao, Y., 2017b. Prediction of CO2 emissions based on the analysis
and classification of decoupling. Pol. J. Environ. Stud. 26 (6), 2851–2860.
Zhou, Y., Zhang, J.Y., Hu, S.Y., 2021. Regression analysis and driving force model
building of CO2 emissions in China. Sci. Rep. 11 (1).
Zhu, B.Z., et al., 2019. Achieving the carbon intensity target of China: a least squares
support vector machine with mixture kernel function approach. Appl. Energy 233,
196–207.
Wang, Z.-X., Ye, D.-J., 2017. Forecasting Chinese carbon emissions from fossil energy
consumption using non-linear grey multivariable models. J. Clean. Prod. 142,
600–612.
Wang, Z.X., Dang, Y.G., Li, X.M., 2011. On whitening the Grey parameters of Grey power
model. J. Grey Syst. 23 (3), 223–228.
Wang, Y., et al., 2018. Can China achieve the 2020 and 2030 carbon intensity targets
through energy structure adjustment? Energies 11 (10).
Wang, J., et al., 2019a. The evolution of Chinese industrial CO2 emissions 2000-2050: a
review and meta-analysis of historical drivers, projections and policy goals. Renew.
Sustain. Energy Rev. 116.
Wang, S., Zhao, Y.H., Wiedmann, T., 2019b. Carbon emissions embodied in ChinaAustralia trade: a scenario analysis based on input-output analysis and panel
regression models. J. Clean. Prod. 220, 721–731.
Wang, J.M., Yang, F., Chen, K.K., 2019c. Regional carbon emission evolution mechanism
and its prediction approach: a case study of Hebei, China. Environ. Sci. Pollut. Res.
26 (28), 28884–28897.
Wang, D., He, W., Shi, R.Y., 2019d. How to achieve the dual-control targets of China’s
CO2 emission reduction in 2030? Future trends and prospective decomposition.
J. Clean. Prod. 213, 1251–1263.
Wang, L.Q., et al., 2020. Finding the de-carbonization potentials in the transport sector:
application of scenario analysis with a hybrid prediction model. Environ. Sci. Pollut.
Res. 27 (17), 21762–21776.
Wen, L., Cao, Y., 2020a. Influencing factors analysis and forecasting of residential
energy-related CO2 emissions utilizing optimized support vector machine. J. Clean.
Prod. 250.
Wen, L., Cao, Y., 2020b. A hybrid intelligent predicting model for exploring household
CO2 emissions mitigation strategies derived from butter fly optimization algorithm.
Sci. Total Environ. 727.
Wen, L., Yuan, X.Y., 2020. Forecasting CO2 emissions in Chinas commercial department,
through BP neural network based on random forest and PSO. Sci. Total Environ. 718.
Wright, L.G., et al., 2022. Deep physical neural networks trained with backpropagation.
Nature 601 (7894), 549-+.
Wu, W.Q., et al., 2020. A novel conformable fractional non-homogeneous grey model for
forecasting carbon dioxide emissions of BRICS countries. Sci. Total Environ. 707.
Xie, N.M., 2022. A summary of grey forecasting models. Grey Syst. Theory Appl. 12 (4),
703–722.
Xie, N.M., Wang, R.Z., 2017. A historic review of Grey forecasting models. J. Grey Syst.
29 (4), 1–29.
Xu, Z.C., Liu, L.Y., Wu, L.F., 2021. Forecasting the carbon dioxide emissions in 53
countries and regions using a non-equigap grey model. Environ. Sci. Pollut. Res. 28
(13), 15659–15672.
Yang, H.J., O’Connell, J.F., 2020. Short-term carbon emissions forecast for aviation
industry in Shanghai. J. Clean. Prod. 275.
20