Coupling LSTM and CNN Neural Networks for Accurate Carbon Emission Prediction in 30 Chinese Provinces

Abstract

Global warming is a major environmental issue facing humanity, and the resulting climate change has severely affected the environment and daily lives of people. China attaches great importance to and actively responds to climate change issues. In order to achieve the “dual carbon” goal, it is necessary to clearly define the emission reduction path and scientifically predict future carbon emissions, which is the basis for setting emission reduction targets. To ensure the accuracy of data, this study applies the emission coefficient method to calculate the carbon emissions from the energy consumption in 30 provinces, regions, and cities in China from 1997 to 2021. Considering the spatial correlation between different regions in China, we propose a new machine learning prediction model that incorporates spatial weighting, namely, an LSTM-CNN combination model with spatial weighting. The spatial weighting explains the spatial correlation and the combined model is used to analyze the carbon emissions in the 30 provinces, regions, and cities of China from 2022 to 2035 under different scenarios. The results show that the LSTM-CNN combination model with four convolutional layers performs the best. Compared with other models, this model has the best predictive performance, with an MAE of 8.0169, an RMSE of 11.1505, and an $R^2$ of 0.9661 on the test set. Based on different scenario predictions, it is found that most cities can achieve carbon peaking before 2030. Some cities need to adjust their development rates based on their specific circumstances in order to achieve carbon peaking as early as possible. This study provides a research direction for deep learning time series forecasting and proposes a new predictive method for carbon emission forecasting.

1. Introduction

Global climate change is a significant environmental issue facing humanity, and the resulting climate change has had a severe impact on the environment and daily lives of people. Research has found that the substantial increase in emissions of greenhouse gases, such as carbon dioxide, caused by human activities is a major factor in global warming [1]. In order to mitigate the trend of global warming, the United Nations has successively formulated the United Nations Framework Convention on Climate Change (1992) and the Kyoto Protocol (1997), which provide legally binding quantitative carbon reduction and emission limitation targets for the contracting parties. The Paris Agreement (2015) made arrangements for global actions to address climate change after 2020.

China is a developing country with rapid urbanization and industrialization, and it is also the world’s largest manufacturing nation, known as the “world’s factory”. Currently, China is still in a phase of rapid economic development, which relies heavily on energy consumption. In the coming years, energy consumption and carbon emissions are expected to continue to increase. Resolving the contradiction between economic development and rising carbon emissions is an urgent issue that the Chinese government needs to address. In 2020, China proposed the “dual carbon” target, which aims to peak carbon dioxide emissions before 2030 and strive to achieve carbon neutrality by 2060. Achieving carbon peaking before 2030 is a major challenge that China is currently facing. With its vast territory, China has significant regional disparities in resource endowment, population size, economic development level, and industrial structure, leading to variations in energy structure across different regions. The levels of development and the structure of carbon emissions also differ across regions, requiring tailored and reasonable carbon reduction policies. According to the “2006 IPCC Guidelines for National Greenhouse Gas Inventories” and the “Guidelines for Provincial-Level Greenhouse Gas Inventories (Trial)” compiled by China in 2011, energy activities are usually the most important sector in greenhouse gas emission inventories. In developed countries, energy activities generally contribute over 90% of carbon dioxide emissions and 75% of total greenhouse gas emissions. Energy activities are also significant sources of greenhouse gas emissions in China.

To achieve China’s “dual carbon” goal, it is essential to clarify the emission reduction pathway and scientifically predict future carbon emissions. This article first applies the emission coefficient method to calculate the carbon emissions from energy consumption in 30 provinces and municipalities in China from 1997 to 2021. Considering the spatial correlation among different regions in China, we propose a new machine learning prediction model that incorporates spatial weights, namely, an LSTM-CNN combination model with spatial weights. We conduct a data ablation experiment using historical data (from 1997 to 2021) to test the predictive capability of the model. Then, we use the aforementioned combination model to analyze the scenario of carbon emissions in 30 provinces and municipalities in China from 2022 to 2035 and provide differentiated carbon emission reduction policy recommendations for different regions.

In this paper, the combined prediction model of LSTM-CNN is constructed to predict China’s carbon emission, with innovation in the following three main areas:

(1)

The existing literature generally predicts a certain city or a certain region, but this paper studies the prediction in 30 provinces and cities in China.

(2)

In the existing literature, the spatial autocorrelation between regions is not considered in the prediction of regional carbon emission. In this paper, the spatial autocorrelation weights are introduced to the input variables.

(3)

In a large number of papers, machine learning models are used to predict the carbon emissions inside samples. On this basis, this paper uses the trained models to predict the carbon peaks under different scenarios.

2. Related Works

The existing research focuses on predicting carbon emissions in individual provinces, cities, and regions. Zhang et al. [2] proposed a machine learning method using BPNN to predict carbon emissions in urban blocks. Yan et al. [3] used a convolutional neural network model to predict carbon emissions in the preliminary design stage of residential buildings in Beijing, which could effectively reflect changes in carbon emissions. The authors believed that predictions for the midterm and late stages should also be included to compare the generalization of the model. Su et al. [4] predicted carbon emissions in large-scale green buildings in Dalian using data-driven models to explore energy consumption patterns in commercial buildings. Combination models are widely favored to improve prediction accuracy. Zhang et al. [5] used ANN-CA and IWOA-LSTM models to explore the correlation between carbon emissions and surface temperature in Wuhan. Chen et al. [6] studied greenhouse gas emissions on the Chengdu Metro Line 18 by establishing a WOA-DELM model. The authors suggested reporting the greenhouse gas situation on other lines to enhance the model’s generalizability. Wang et al. [7] proposed a method that combined machine learning with DDF to estimate the carbon emission efficiency and reduction potential in Chinese cities. Sun et al. [8] fitted carbon emission data for 281 cities in China using the sparrow optimization neural network algorithm. Yan et al. [3] proposed a real-time actionable carbon emission prediction method based on a CNN and predicted a typical northern case, Beijing. The authors suggested extending carbon emission predictions to all provinces and cities nationwide for internal comparisons.

In addition to predicting individual provinces and regions, carbon emission research in a specific spatial area has been effective in recent years. Zhao et al. [9] proposed a combined model of SSA-LSTM and applied it to carbon emission forecasting in the Yellow River basin. The results showed an increasing trend in carbon emissions in the Yellow River basin with significant interprovincial differences. The authors suggested that in addition to predicting this particular region, factors such as spatial autocorrelation among variables within the region should also be considered. Qin et al. [10] used machine learning methods such as panel data and random forest to determine the factors influencing CO₂ emissions and found that areas with higher levels of economic development in the eastern region had the highest CO₂ emissions. The authors suggested that causal inference methods could be introduced to validate the causal relationship between influencing factors and carbon emissions in order to identify more significant influencing factors. Aryai et al. [11] proposed a PSO-ERT regression model for predicting the emission intensity in Australia’s regional electricity market. Sarwar et al. [12] selected a suitable model for predicting electricity prices and carbon emissions in the eastern region of Saudi Arabia. Xu et al. [13] conducted a study on carbon sequestration in mangroves based on the coastal areas south of the Yangtze River in China, using the RF and gradient boost models. The authors suggested that in addition to predicting a specific region, other regions should also be considered in order to compare the forecast results and examine the generalizability of the models.

Machine learning methods have shown good performance in energy and carbon emission prediction and analysis, both for in-sample fitting and out-of-sample prediction. Han et al. [14] proposed an ISM-ELM model for energy and carbon emission prediction and analysis. Anthony et al. [15] introduced a tool called Carbontracker for tracking and predicting the energy and carbon footprints of training deep learning models. This tool reports the energy and carbon footprints of model development and training together with performance indicators. In addition to optimizing the models, we can also preprocess the raw data to achieve better fitting and prediction results. Han et al. [14] proposed an improved integrated structure model based extreme learning machine (ISM-ELM) for energy and carbon emission analysis and prediction. This model not only denoised and reduced the dimensionality of the raw data, reducing training time and errors of ELM prediction models, but also exhibited a robustness and effectiveness in terms of model accuracy and training time. Furthermore, a research trend was sparked by comparing the results of various models and identifying the best model. To compare the effects of multiple machine learning models, Zhao et al. [16] conducted a comparative analysis of the main carbon emission prediction models, including backpropagation neural network, support vector machine, long short-term memory neural network, random forest, and extreme learning machine. Based on this analysis, they proposed research ideas for future machine learning-based carbon emission prediction models. Considering the characteristics of carbon emission data, Yang et al. [17] proposed a combined carbon emission ensemble prediction model that integrated a singular spectrum analysis and a variational mode decomposition. Nadirgil et al. [18] developed 48 hybrid machine learning models using a complete ensemble empirical mode decomposition and compared their performance. The authors suggested that carbon peak prediction scenarios under different conditions should be introduced in the out-of-sample prediction to compare the peak time between different regions in the future. Zhang et al. [19] proposed a comprehensive carbon emission prediction model called SSA-FAGM-SVR for predicting the carbon emissions of the G20 countries in the next 10 years.

The reason why this study adopted the construction of an LSTM-CNN model is that the CNN-LSTM model has shown significant performance in fitting and predicting time series data, especially in areas such as gold prices [20], stock prices [21], and residential energy consumption [22]. Therefore, compared to some classical time series models, the LSTM-CNN model not only has better in-sample fitting performance but also a higher prediction accuracy for out-of-sample data. The choice of this combined model allows us to take advantage of the LSTM’s ability to fit and predict time series data and the CNN’s ability to extract more comprehensive features from the data.

3. Methodology

3.1. STIRPAT

The STIRPAT model [23] is an extension of the IPAT model by Dietz and Rosa, designed as a scalable and stochastic environmental impact assessment model. Its standard form is shown in Equation (1):

$$I=a{P}^b{A}^c{T}^{d}e,$$

(1)

In the equation, I, P, A, and T represent the environmental impact, population, affluence, and technological level, respectively; a is a model coefficient; b, c, and d are estimated coefficients; and e is the error term.

Taking the logarithm of both sides of the equation:

$$lnI=lna+blnP+clnA+dlnT+lne,$$

(2)

3.2. Global Moran’s Index

The global Moran index primarily reflects the overall spatial autocorrelation of a certain indicator in different regions. The specific calculation formula is shown in Equation (3):

$$I_m=\frac{n{\sum }_{i=1}^n{\sum }_{j=1}^n{w}_{ij}(x_i-\overline{X})(x_j-\overline{X})}{{\sum }_{i=1}^n{\sum }_{j=1}^n{w}_{ij}{\sum }_{i=1}^n(x_i-\overline{X})},$$

(3)

In the equation, $I_m$ is the global Moran index, n represents the total number of provinces or regions, $x_i,x_j$ represent the carbon emissions of individual provinces i or regions j, $\overline{X}$ is the average carbon emissions of all provinces or regions, $w_{ij}$ represents the element (i,j) of the spatial weight matrix W, and the spatial weight matrix W is based on the Rook distance matrix derived from adjacency relationships. The significance of the value I in the equation is assessed using a value Z test:

$$Z=\frac{I_m-E\left(I_m\right)}{\sqrt{Var\left(I_m\right)}},$$

(4)

where $E\left(I_m\right)$ represents the expected value of Moran’s I, and $Var\left(I_m\right)$ represents the variance of Moran’s I.

3.3. Spatial Durbin Model

The spatial Durbin model is an extended form of the spatial lag model and the spatial lag of the independent variable model, simultaneously considering the autocorrelation of both the dependent and independent variables.

$$y=\lambda W_{1}y+X\beta +W_{2}X\delta +ϵ,$$

(5)

In the equation, $W_1$, $W_2$ represents the spatial weight matrix, $W_{1}y$ is the spatial lag of the dependent variable y, $W_{2}X$ is the spatial lag of the independent variable X, $\lambda$ is the spatial autocorrelation coefficient, and $ϵ$ is the error term. The equation can be transformed into:

$$y={(E-\lambda W_1)}^{-1}(X\beta +W_{2}X\delta +ϵ),$$

(6)

In the equation, E represents the identity matrix.

3.4. CNN

A CNN [20] has a strong capability for processing spatial grid data, primarily composed of convolutional layers, pooling layers, and fully connected layers. The convolutional layers are mainly responsible for feature extraction; the pooling layers perform dimensionality reduction and sampling without compromising the recognition results; and the fully connected layers connect all neurons and output results through hidden layers. The specific flowchart of a simple CNN is shown in Figure 1.

3.5. LSTM

LSTM [24], which stands for long short-term memory, is a specialized form of recurrent neural networks (RNNs). It is primarily used to address the issues of vanishing or exploding gradients that occur during training long time sequences. Compared to traditional RNNs, an LSTM network provides better analytical results for long time sequence dependencies. The internal structure of the LSTM network is illustrated in Figure 2.

3.6. Gaussian Noise

In order to prevent overfitting during the model training process, this study adopted a noise injection approach on the original data for adversarial training. This was achieved by using the random.gauss ($\mu$, $\sigma$) function from the random library, where the parameter mu represents the mean of the Gaussian distribution and the parameter sigma represents the standard deviation of the Gaussian distribution.

3.7. Model Construction

Taking into account the spatial correlation of carbon emissions, this study incorporated spatial weights into the model. LSTM performs well in handling time series data, while a CNN is effective in feature extraction from spatial grid data. A combined LSTM-CNN model was constructed based on LSTM and CNN models to predict carbon emission intensity at the provincial level in China. The model structure is illustrated in Figure 3.

The model consists of an input layer, a feature extraction layer, and an output layer. In the input layer, data are processed with spatial weights and transformed into tensors suitable for neural network training. The feature extraction layer includes LSTM and convolutional layers, with the convolutional layer comprising a one-dimensional convolutional neural network layer, an activation layer, and a pooling layer. The output layer consists of a fully connected layer that outputs the prediction results.

4. Data Source and Characteristics Analysis

4.1. Data Source

The data utilized in this study primarily included energy data, carbon emission factors, and socio-economic data. Specifically, energy data consisted of energy consumption; carbon emission factors encompassed net calorific value, carbon content per unit of heat value, and carbon oxidation rate; socio-economic data mainly comprised regional GDP, total population, energy intensity, industrial structure, and urbanization rate for various provinces in China.

Energy data were sourced from the “China Energy Statistical Yearbook” (1997–2021), and missing data were filled using interpolation methods.

Carbon emission factors were drawn from the “Guidelines for Compilation of Provincial Greenhouse Gas Inventories (Trial)” (2011), “China Energy Statistical Yearbook”, “2006 IPCC National Greenhouse Gas Inventories Guidelines”, and “Data Descriptor: China CO₂ emission accounts 1997–2015”. The calculation methods and data sources for the socio-economic data are outlined in

Table 1. Socio-economic data.

The Index	Calculation Method	Connotation	Unit	Time	Data Source
Total population at year end	Year-end population by region	Population	Thousands of people	From 1997 to 2020	China Population and Employment Statistical Yearbook (1997–2021)
Per capita GDP	Gross regional product/year-end population of subregion	economy	Yuan	From 1997 to 2020	National Bureau of Statistics
Energy intensity	Total energy consumption/GDP	Technology	Ten thousand tons of standard coal/CNY 100 million	From 1997 to 2020	China Energy Statistical Yearbook (1997–2019), Statistical Yearbook of Provinces and Municipalities (2020–2021), Statistical Yearbook of Ningxia (2001)
Industrial structure	Value added of secondary industry/gross regional product	Technology	%	From 1997 to 2020	China Statistical Yearbook (1997–2021)
Urbanization rate	Urban population/year-end population by region	Urbanization level	%	From 1997 to 2020	China Statistical Yearbook (2005–2021), China Population and Employment Statistical Yearbook (1997–2004)

.

4.2. Carbon Emission Accounting

This study employed the emission coefficient method proposed by the IPCC to estimate carbon emissions. Due to the lack of publicly available data or significant missing data on energy consumption for the Tibet Autonomous Region, Hong Kong Special Administrative Region, Macau Special Administrative Region, and Taiwan Province, this study calculated carbon emissions for the 30 provinces and municipalities in China excluding the aforementioned regions from 1997 to 2021. The reference method outlined in the “2006 IPCC National Greenhouse Gas Inventories Guidelines” is based on the apparent consumption of various fossil fuels, net calorific value, carbon content per unit of heat value for different fuel types, and carbon oxidation rate in major equipment for burning various fuels. It comprehensively calculates these parameters and deducts non-energy-related carbon sequestration of fossil fuels to derive the estimation. The specific calculation formula is shown in Equation (7):

$$CE=\sum _i(A{D}_i·NC{V}_i·C{C}_i·CO{F}_i·\frac{44}{12}),$$

(7)

In the equation, CE represents the calculated carbon emissions; $A{D}_i$ stands for the final consumption of fossil fuels i; $NC{V}_i$ denotes the net calorific value of fossil fuels i; $C{C}_i$ represents the carbon content per unit of heat value for fossil fuels i; $CO{F}_i$ signifies the carbon oxidation rate of fossil fuels i; and $\frac{44}{12}$ represents the ratio of molecular weights between carbon dioxide and carbon.

4.3. Spatial Self-Correlation Analysis

Table 2. Global Moran’s index.

Year	Moran’s I	Z-Value	p-Value
1997	0.172482	1.578783	0.070
1998	0.203977	1.855481	0.040
1999	0.225506	2.132896	0.023
2000	0.228301	2.206887	0.021
2001	0.260459	2.392598	0.018
2002	0.254042	2.301015	0.023
2003	0.184919	1.802718	0.046
2004	0.244378	2.387470	0.013
2005	0.272978	2.676280	0.007
2006	0.283878	2.711853	0.009
2007	0.284480	2.772128	0.008
2008	0.246667	2.384451	0.012
2009	0.230676	2.100769	0.027
2010	0.226314	2.319527	0.018
2011	0.222190	2.076378	0.024
2012	0.184055	1.876683	0.041
2013	0.188710	1.879829	0.042
2014	0.195953	1.979889	0.035
2015	0.171137	1.747687	0.054
2016	0.153856	1.590229	0.078
2017	0.118852	1.237058	0.113
2018	0.124516	1.371906	0.092
2019	0.133253	1.411485	0.082
2020	0.151197	1.570206	0.080

shows the global Moran index of carbon emissions of 30 provinces, autonomous regions, and municipalities in China from 1997 to 2020.

Except for the year 1997 and the years 2016–2020, the global Moran I index for the other years passed the significance test, indicating a spatial autocorrelation in carbon emissions among various provinces in China. Moreover, Moran’s I index values were all greater than zero, suggesting a positive spatial correlation in carbon emissions across the country. Provinces with higher (lower) carbon emissions are relatively adjacent to those with higher (lower) carbon emissions.

Given the spatial autocorrelation in carbon emissions among different provinces in China and in combination with the STIRPAT model, the following model was established:

$$\begin{array}{cc}\hfill y_t=& \sum _{n=t-3}^{t-1}(W{y}_n+\lambda W{y}_n+ln{Z}_{n,1}{\beta }_1+ln{Z}_{n,2}{\beta }_2+ln{Z}_{n,3}{\beta }_3+ln{Z}_{n,4}{\beta }_4+ln{Z}_{n,5}{\beta }_5\hfill \\ \hfill & +Wln{Z}_{n,1}{\delta }_1+Wln{Z}_{n,2}{\delta }_2+Wln{Z}_{n,3}{\delta }_3+Wln{Z}_{n,4}{\delta }_4+Wln{Z}_{n,5}{\delta }_5+ϵ),\hfill \end{array}$$

(8)

In the equation, $\lambda$ represents the spatial autocorrelation coefficient, W stands for the spatial weight matrix, $y_t$ represents carbon emissions in the ith year, $Z_{n,1}$ is the population at the end of year n after feature extraction by a neural network for 30 provinces in China, $Z_{n,2}$ is the per capita GDP for 30 provinces in China in year n after feature extraction by a neural network, $Z_{n,3}$ is the energy intensity for 30 provinces in China in year n after feature extraction by a neural network, $Z_{n,4}$ is the industrial structure for 30 provinces in China in year n after feature extraction by a neural network, $Z_{n,5}$ is the urbanization rate for 30 provinces in China in year n after feature extraction by a neural network, and ${\beta }_1$, ${\beta }_2$, ${\beta }_3$, ${\beta }_4$, ${\beta }_5$, ${\delta }_1$, ${\delta }_2$, ${\delta }_3$, ${\delta }_4$, and ${\delta }_5$ represent the regression coefficients of respective variables.

4.4. Carbon Emission Trend Analysis of Provinces and Cities in China

The 30 provinces and municipalities in China are divided into three regions: North, Central, and South, each containing 10 provinces and municipalities. The carbon emissions’ variations in the northern provinces and municipalities of China from 1997 to 2021 are shown in Figure 4; the carbon emissions’ variations in the central region are shown in Figure 5; and the carbon emissions’ variations in the southern region are shown in Figure 6. In Figure 4, “Shan1Xi” represents Shanxi Province, and “Shan3Xi” represents Shaanxi Province. From Figure 4, it can be observed that Hebei Province has the highest carbon emissions. From Figure 5, it can be seen that Shandong Province has the highest carbon emissions. From Figure 6, it can be observed that Guangdong Province has the highest carbon emissions.

4.5. Variable Correlation Analysis

In order to explore the linear relationship between carbon emissions y and the selected five original independent variables $X_1$–$X_5$, the calculation of the correlation coefficients between variables was adopted to reflect the degree of linear correlation between variables, as shown in Figure 7. From the perspective of linear relationships with carbon emissions y, a strong linear relationship was observed between carbon emissions and the year-end population $X_1$, reaching a correlation of 0.646. However, the linear relationships with per capita GDP $X_2$, industrial structure $X_4$ and urbanization rate $X_5$ were relatively weak, with correlations of 0.339, 0.297, and 0.235, respectively. Additionally, there existed a weak negative correlation between carbon emissions y and energy intensity $X_3$, with a correlation coefficient of −0.311.

5. Empirical Analysis

5.1. Datasets and Preprocessing

The experimental data consisted of previously calculated carbon emissions and their relevant influencing factors, including carbon emissions, year-end population, per capita GDP, energy intensity, industrial structure, and urbanization rate. The dataset underwent preprocessing by adding Gaussian noise with a mean of zero and a standard deviation of 0.05. The data were then divided into windows of a time step of three, resulting in 630 sets of data. The first 70% of the data were used as the training set, while the remaining 30% were designated as the testing set, thus creating the initial dataset.

5.2. Parameter Setting

The experimental parameter settings were as follows: In the model parameters, a unidirectional LSTM network was used with a hidden layer output dimension of 32 and a stack of three layers. In the CNN convolutional layers, the linear activation function used was ReLU, and the convolutional kernel size was three. In the fully connected layers, the linear activation function used was also ReLU. The input and output dimensions of each convolutional layer and fully connected layer are shown in the following table,

Table 3. Parameter Settings.

Layer	Input Channel	Output Channel
Convolution layer 1	3	64
Convolution layer 2	64	64
Convolution layer 3	64	128
Convolution layer 4	128	128
Convolution layer 5–10	128	128
Fully connected layer 1	3072	32
Fully connected layer 2	32	16
Fully connected layer 3	16	1

.

For the setting of hyperparameters in the model, we refer to the relevant conclusions and pre-experimental rules from Bengio Y et al. [25], and through continuous optimization adjustments, the final settings were as follows: the loss function was set to MSE (mean squared error); the optimization algorithm was AdamW, with a learning rate of 0.0001 and weight decay of 0.001; the number of iterations was set to 50,000.

5.3. Evaluation Criteria

In this paper, three types of metrics—mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R-squared)—were chosen as evaluation criteria to measure errors. The calculation methods are shown in Equations (9)–(11):

$$MAE=\frac{1}{n}\sum _{i=1}^n\mid \widehat{y_i}-y_i\mid ,$$

(9)

$$RMSE=\sqrt{\frac{1}{n}\sum _{i=1}^n{(y_i-\widehat{y_i})}^2},$$

(10)

$$R^2=\frac{{\sum }_{i=1}^n{(\widehat{y_i}-\overline{y})}^2}{{\sum }_{i=1}^n{(y_i-\overline{y})}^2},$$

(11)

In the equation, $\widehat{y_i}$ represents the predicted value, $y_i$ represents the true value, and $\overline{y}$ represents the mean of $y_i$. The ranges of MAE and RMSE are $[0,+\infty )$, where smaller values indicate smaller errors and a better accuracy of the model. The range of R-squared is $[0,1]$, where larger values indicate smaller errors and a better accuracy of the model.

5.4. Results of the Analysis

5.4.1. The Relationship between the Number of Convolution Layers and the Training Results

In order to explore the relationship between the training performance of LSTM-CNN and different numbers of convolutional layers, we conducted 50,000 iterations of training on convolutional networks ranging from 1 to 10 layers. The results are shown in

Table 4. Training results of LSTM-CNN model under different convolution layers.

Number of CNN Convolution Layers	Training Set (MAE)	Training Set (RMSE)	Training Set (R-Squared)	Testing Set (MAE)	Testing Set (RMSE)	Testing Set (R-Squared)
1	12.6518	17.3916	0.9772	9.5291	12.8854	0.9548
2	9.3253	14.3121	0.9846	8.4488	11.7271	0.9625
3	9.4626	14.4317	0.9843	9.0037	12.0501	0.9605
4	10.1382	15.2988	0.9824	8.0169	11.1505	0.9661
5	19.1646	25.0227	0.9529	11.4450	14.8190	0.9402
6	10.9222	16.7670	0.9788	8.7979	12.2485	0.9591
7	13.1721	19.2177	0.9722	8.6127	12.1395	0.9599
8	9.8745	15.3462	0.9823	9.0593	12.6999	0.9561
9	16.8803	25.6881	0.9503	10.0847	13.7042	0.9489
10	9.7553	14.9535	0.9832	8.3071	11.4421	0.9643

. To exclude the interference of other influencing factors on the model training results, we ensured that the hyperparameters of these 10 models were set according to Section 5.2. In addition, our models were experimentally conducted in the environment of Google Colab using a Tesla T4.

From the table, it can be seen that in the training set, when the number of convolutional layers reaches two, the MAE reaches its minimum value of 9.3253, and the RMSE reaches its minimum value of 14.3121, with a maximum value of 0.9846. In the test set, when the number of convolutional layers reaches four, the MAE reaches its minimum value of 8.0169, and the RMSE reaches its minimum value of 11.1505, with a maximum value of 0.9661. Therefore, it can be considered that the LSTM-CNN model performs optimally when the number of convolutional layers reaches four.

5.4.2. Ablation Experiment

To verify the relative effectiveness of each part of the LSTM-CNN model, we conducted ablation experiments and the results are shown in

Table 5. Ablation results of LSTM-CNN model.

Model	Training Set (MAE)	Training Set (RMSE)	Training Set (R-Squared)	Testing Set (MAE)	Testing Set (RMSE)	Testing Set (R-Squared)
LSTM (no weighting)	10.4812	15.1635	0.9827	9.1512	11.5155	0.9639
CNN (no weighting)	23.8648	32.9929	0.9181	14.6824	21.2849	0.8766
CNN-LSTM (no weighting)	10.6959	15.1564	0.9827	8.3908	11.6647	0.9629
LSTM-CNN (no weighting)	15.4686	22.3543	0.9624	9.1497	12.6901	0.9561
LSTM (weighting)	8.6944	13.4054	0.9865	8.4607	11.5644	0.9636
CNN (weighting)	7.8887	12.0049	0.9892	9.7619	13.9977	0.9466
CNN-LSTM (weighting)	8.3815	12.3646	0.9885	10.2428	12.9617	0.9543
LSTM-CNN (weighting)	10.2301	15.7271	0.9814	7.8500	10.6967	0.9688

.

From the table, it can be seen that for the model without spatial weights, in the training set, the LSTM model achieved a minimum MAE of 10.4812, and the CNN-LSTM model achieved a minimum RMSE of 15.1564, with a maximum value of 0.9827. In the testing set, the CNN-LSTM model achieved a minimum MAE of 8.3908, and the RMSE reached a minimum value of 11.6647, with a maximum value of 0.9629. Figure 8 shows that the CNN-LSTM model had the best fitting performance for the real data without considering spatial weights in both the training and testing sets. This indicates that in the model without weights, the CNN-LSTM model had the best fitting effect on the data.

For the model with spatial weights, in the training set, the CNN-LSTM model achieved a minimum MAE of 8.3815 and the RMSE reached a minimum value of 12.3646, with a maximum value of 0.9885. In the testing set, the LSTM-CNN model achieved the minimum MAE of 7.8500 and the RMSE reached the minimum value of 10.6967, with a maximum value of 0.9688. Figure 9 shows that the LSTM-CNN model had the best fitting performance for the real data considering spatial weights in both the training and testing sets. By comparing the results, it can be concluded that after adding spatial weights, the model’s performance was improved in both the testing and training sets. Therefore, this ablation experiment verified that each part of the model had its own contribution to the model fitting.

6. Scenario Analysis

6.1. Scenario Setting

Based on the “14th Five-Year Plan and Long-Term Goals for 2035 of the People’s Republic of China’s National Economic and Social Development” and the relevant policy documents of “14th Five-Year Plans” for various provinces and regions, this paper sets the growth rates of year-end population, per capita GDP, energy intensity, industrial structure, and urbanization rate, with three levels of rates: low, medium, and high. Among them, the medium rate was set to meet the minimum requirements of government documents for each indicator. The growth rate settings for each indicator are provided in

Table A1. Setting of provincial influencing factor change rates in different scenarios.

Province/City	Year	Development Rate	Annual Total Population Change Rate	Per Capita GDP Change Rate	Energy Intensity Change Rate	Industrial Structure Change Rate	Urbanization Rate Change Rate
Beijing	2021–2025	Low	−0.0476	4.6	−3.4	−0.7	0.1
Beijing	2021–2025	Medium	−0.0476	5.1	−3	−0.6	0.2
Beijing	2021–2025	High	−0.0476	5.6	−2.6	−0.5	0.3
Beijing	2026–2030	Low	−0.0476	4	−3.1	−0.4	0.05
Beijing	2026–2030	Medium	−0.0476	4.5	−2.7	−0.3	0.1
Beijing	2026–2030	High	−0.0476	5	−2.3	−0.2	0.2
Beijing	2031–2035	Low	−0.0476	3.6	−2.9	−0.2	0
Beijing	2031–2035	Medium	−0.0476	4.1	−2.5	−0.1	0.05
Beijing	2031–2035	High	−0.0476	4.6	−2.1	−0.05	0.1
Tianjing	2021–2025	Low	0.9524	4.5	−3.4	−4	0.15
Tianjing	2021–2025	Medium	0.9524	5	−3	−3.5	0.25
Tianjing	2021–2025	High	0.9524	5.5	−2.6	−3	0.35
Tianjing	2026–2030	Low	0.9524	4	−3.1	−4.5	0.1
Tianjing	2026–2030	Medium	0.9524	4.5	−2.7	−4	0.2
Tianjing	2026–2030	High	0.9524	5	−2.3	−3.5	0.3
Tianjing	2031–2035	Low	0.9524	3.8	−2.9	−4.8	0.05
Tianjing	2031–2035	Medium	0.9524	4.2	−2.5	−4.3	0.15
Tianjing	2031–2035	High	0.9524	4.7	−2.1	−3.8	0.25
Hebei	2021–2025	Low	0.21	5.3	−3.4	−3.3	1.1
Hebei	2021–2025	Medium	0.21	5.8	−3	−2.8	1.6
Hebei	2021–2025	High	0.21	6.3	−2.6	−2.3	2.1
Hebei	2026–2030	Low	0.21	4.7	−3.1	−3.5	1
Hebei	2026–2030	Medium	0.21	5.2	−2.7	−3	1.5
Hebei	2026–2030	High	0.21	5.7	−2.3	−2.5	2
Hebei	2031–2035	Low	0.21	4.3	−2.9	−3.7	0.9
Hebei	2031–2035	Medium	0.21	4.8	−2.5	−3.2	1.4
Hebei	2031–2035	High	0.21	5.3	−2.1	−2.7	1.9
Shan1xi	2021–2025	Low	−0.093	7.6	−3.5	−1.5	1.2
Shan1xi	2021–2025	Medium	−0.093	8.1	−3.1	−1	1.7
Shan1xi	2021–2025	High	−0.093	8.6	−2.7	−0.5	2.2
Shan1xi	2026–2030	Low	−0.093	7	−3.4	−2	1.1
Shan1xi	2026–2030	Medium	−0.093	7.5	−2.8	−1.5	1.6
Shan1xi	2026–2030	High	−0.093	8	−2.4	−1	2.1
Shan1xi	2031–2035	Low	−0.093	6.6	−2.9	−2.5	1
Shan1xi	2031–2035	Medium	−0.093	7.1	−2.5	−2	1.5
Shan1xi	2031–2035	High	−0.093	7.6	−2.1	−1.5	2
Neimenggu	2021–2025	Low	−0.4739	5	−3.6	−0.4	0.5
Neimenggu	2021–2025	Medium	−0.4739	5.5	−3.2	−0.3	1
Neimenggu	2021–2025	High	−0.4739	6	−2.8	−0.2	1.5
Neimenggu	2026–2030	Low	−0.4739	4.5	−3.2	−0.3	0.4
Neimenggu	2026–2030	Medium	−0.4739	5	−2.8	−0.2	0.9
Neimenggu	2026–2030	High	−0.4739	5.5	−2.4	−0.1	1.4
Neimenggu	2031–2035	Low	−0.4739	4	−2.9	−0.2	0.3
Neimenggu	2031–2035	Medium	−0.4739	4.5	−2.5	−0.1	0.8
Neimenggu	2031–2035	High	−0.4739	5	−2.1	0	1.3
Liaoning	2021–2025	Low	0.8563	4.6	−3.5	−2	0.6
Liaoning	2021–2025	Medium	0.8563	5.1	−3.1	−1.5	1.1
Liaoning	2021–2025	High	0.8563	5.6	−2.7	−1	1.6
Liaoning	2026–2030	Low	0.8563	4	−3.4	−1.5	0.5
Liaoning	2026–2030	Medium	0.8563	4.5	−2.8	−1	1
Liaoning	2026–2030	High	0.8563	5	−2.4	−0.5	1.5
Liaoning	2031–2035	Low	0.8563	3.6	−2.9	1	0.4
Liaoning	2031–2035	Medium	0.8563	4.1	−2.5	−0.5	0.9
Liaoning	2031–2035	High	0.8563	4.6	−2.1	0	1.4
Jiling	2021–2025	Low	−2.2936	8.5	−3.5	−2.5	0.7
Jiling	2021–2025	Medium	−2.2936	9	−3.1	−2	0.8
Jiling	2021–2025	High	−2.2936	9.5	−2.7	−1.5	0.9
Jiling	2026–2030	Low	−2.2936	9.5	−3.4	−3	0.6
Jiling	2026–2030	Medium	−2.2936	10	−2.8	−2.5	0.7
Jiling	2026–2030	High	−2.2936	10.5	−2.4	−2	0.8
Jiling	2031–2035	Low	−2.2936	10.5	−2.9	−3.5	0.5
Jiling	2031–2035	Medium	−2.2936	11	−2.5	−3	0.6
Jiling	2031–2035	High	−2.2936	11.5	−2.1	−2.5	0.7
Heilongjiang	2021–2025	Low	−1.4019	6.5	−3.5	−3.5	0.15
Heilongjiang	2021–2025	Medium	−1.4019	7	−3.1	−3	0.16
Heilongjiang	2021–2025	High	−1.4019	7.5	−2.7	−2.5	0.17
Heilongjiang	2026–2030	Low	−1.4019	6	−3.4	−2.5	0.14
Heilongjiang	2026–2030	Medium	−1.4019	6.5	−2.8	−2	0.15
Heilongjiang	2026–2030	High	−1.4019	7	−2.4	−1.5	0.16
Heilongjiang	2031–2035	Low	−1.4019	5.5	−2.9	−1.5	0.13
Heilongjiang	2031–2035	Medium	−1.4019	6	−2.5	−1	0.14
Heilongjiang	2031–2035	High	−1.4019	6.5	−2.1	−0.5	0.15
Shanghai	2021–2025	Low	0.0953	4.4	−3.4	−2.6	0.1
Shanghai	2021–2025	Medium	0.0953	4.9	−3	−2.1	0.2
Shanghai	2021–2025	High	0.0953	5.4	−2.6	−1.6	0.3
Shanghai	2026–2030	Low	0.0953	3.9	−3.1	−2	0.05
Shanghai	2026–2030	Medium	0.0953	4.4	−2.7	−1.5	0.1
Shanghai	2026–2030	High	0.0953	4.9	−2.3	−1	0.2
Shanghai	2031–2035	Low	0.0953	3.4	−2.9	−1.4	0
Shanghai	2031–2035	Medium	0.0953	3.9	−2.5	−0.9	0.05
Shanghai	2031–2035	High	0.0953	4.4	−2.1	−0.4	0.1
Jiangsu	2021–2025	Low	1.0536	3.9	−3.4	−1.9	0.35
Jiangsu	2021–2025	Medium	1.0536	4.4	−3	−1.4	0.45
Jiangsu	2021–2025	High	1.0536	4.9	−2.6	−0.9	0.55
Jiangsu	2026–2030	Low	1.0536	3.4	−3.1	−1.8	0.3
Jiangsu	2026–2030	Medium	1.0536	3.9	−2.7	−1.3	0.4
Jiangsu	2026–2030	High	1.0536	4.4	−2.3	−0.8	0.5
Jiangsu	2031–2035	Low	1.0536	2.9	−2.9	−1.7	0.25
Jiangsu	2031–2035	Medium	1.0536	3.4	−2.5	−1.2	0.35
Jiangsu	2031–2035	High	1.0536	3.9	−2.1	−0.7	0.45
Zhejiang	2021–2025	Low	0.0949	4.9	−3.5	−4	0.7
Zhejiang	2021–2025	Medium	0.0949	5.4	−3.1	−3.5	0.8
Zhejiang	2021–2025	High	0.0949	5.9	−2.7	−3	0.9
Zhejiang	2026–2030	Low	0.0949	4.4	−3.4	−4.5	0.6
Zhejiang	2026–2030	Medium	0.0949	4.9	−2.8	−4	0.7
Zhejiang	2026–2030	High	0.0949	5.4	−2.4	−3.5	0.8
Zhejiang	2031–2035	Low	0.0949	3.9	−2.9	−5	0.5
Zhejiang	2031–2035	Medium	0.0949	4.4	−2.5	−4.5	0.6
Zhejiang	2031–2035	High	0.0949	4.9	−2.1	−4	0.7
Anhui	2021–2025	Low	−1.1142	7.2	−3.4	−3	1.2
Anhui	2021–2025	Medium	−1.1142	7.7	−3	−2.5	1.3
Anhui	2021–2025	High	−1.1142	8.2	−2.6	−2	1.4
Anhui	2026–2030	Low	−1.1142	6.7	−3.1	−2.5	1.1
Anhui	2026–2030	Medium	−1.1142	7.2	−2.7	−2	1.2
Anhui	2026–2030	High	−1.1142	7.7	−2.3	−1.5	1.3
Anhui	2031–2035	Low	−1.1142	6.2	−2.9	−2	1
Anhui	2031–2035	Medium	−1.1142	6.7	−2.5	−1.5	1.1
Anhui	2031–2035	High	−1.1142	7.2	−2.1	−1	1.2
Fujian	2021–2025	Low	0.8539	4.9	−3.5	−3	0.7
Fujian	2021–2025	Medium	0.8539	5.4	−3.1	−2.5	0.8
Fujian	2021–2025	High	0.8539	5.9	−2.7	−2	0.9
Fujian	2026–2030	Low	0.8539	4.4	−3.4	−2.5	0.6
Fujian	2026–2030	Medium	0.8539	4.9	−2.8	−2	0.7
Fujian	2026–2030	High	0.8539	5.4	−2.4	−1.5	0.8
Fujian	2031–2035	Low	0.8539	3.9	−2.9	−2	0.5
Fujian	2031–2035	Medium	0.8539	4.4	−2.5	−1.5	0.6
Fujian	2031–2035	High	0.8539	4.9	−2.1	−1	0.7
Jiangxi	2021–2025	Low	1.9048	4.5	−3.5	−3	1.1
Jiangxi	2021–2025	Medium	1.9048	5	−3.1	−2.5	1.2
Jiangxi	2021–2025	High	1.9048	5.5	−2.7	−2	1.3
Jiangxi	2026–2030	Low	1.9048	5.9	−3.4	−2.5	1
Jiangxi	2026–2030	Medium	1.9048	6.4	−2.8	−2	1.1
Jiangxi	2026–2030	High	1.9048	6.9	−2.4	−1.5	1.2
Jiangxi	2031–2035	Low	1.9048	7.3	−2.9	−2	0.9
Jiangxi	2031–2035	Medium	1.9048	7.8	−2.5	−1.5	1
Jiangxi	2031–2035	High	1.9048	8.3	−2.1	−1	1.1
Shandong	2021–2025	Low	0.572	4.4	−3.8	−3.2	0.6
Shandong	2021–2025	Medium	0.572	4.9	−3.4	−2.7	0.7
Shandong	2021–2025	High	0.572	5.4	−3	−2.2	0.8
Shandong	2026–2030	Low	0.572	3.9	−3.4	−2.7	0.5
Shandong	2026–2030	Medium	0.572	4.4	−3	−2.2	0.6
Shandong	2026–2030	High	0.572	4.9	−2.6	−1.7	0.7
Shandong	2031–2035	Low	0.572	3.4	−3.1	−2.2	0.4
Shandong	2031–2035	Medium	0.572	3.9	−2.7	−1.7	0.5
Shandong	2031–2035	High	0.572	4.4	−2.3	−1.2	0.6
Henan	2021–2025	Low	0.3788	5.1	−3.5	−4.5	1.5
Henan	2021–2025	Medium	0.3788	5.6	−3.1	−4	1.6
Henan	2021–2025	High	0.3788	6.1	−2.7	−3.5	1.7
Henan	2026–2030	Low	0.3788	6.1	−3.4	−4	1.4
Henan	2026–2030	Medium	0.3788	6.6	−2.8	−3.5	1.5
Henan	2026–2030	High	0.3788	7.1	−2.4	−3	1.6
Henan	2031–2035	Low	0.3788	7.1	−2.9	−3.5	1.3
Henan	2031–2035	Medium	0.3788	7.6	−2.5	−3	1.4
Henan	2031–2035	High	0.3788	8.1	−2.1	−2.5	1.5
Hubei	2021–2025	Low	0.4717	5.5	−3.5	−5	0.6
Hubei	2021–2025	Medium	0.4717	6	−3.1	−4.5	0.7
Hubei	2021–2025	High	0.4717	6.5	−2.7	−4	0.8
Hubei	2026–2030	Low	0.4717	5	−3.4	−4.5	0.5
Hubei	2026–2030	Medium	0.4717	5.5	−2.8	−4	0.6
Hubei	2026–2030	High	0.4717	6	−2.4	−3.5	0.7
Hubei	2031–2035	Low	0.4717	4.5	−2.9	−4	0.4
Hubei	2031–2035	Medium	0.4717	5	−2.5	−3.5	0.5
Hubei	2031–2035	High	0.4717	5.5	−2.1	−3	0.6
Hunan	2021–2025	Low	0.0944	5.4	−3.4	−2.9	1.4
Hunan	2021–2025	Medium	0.0944	5.9	−3	−2.4	1.5
Hunan	2021–2025	High	0.0944	6.4	−2.6	−1.9	1.6
Hunan	2026–2030	Low	0.0944	4.9	−3.1	−2.4	1.3
Hunan	2026–2030	Medium	0.0944	5.4	−2.7	−1.9	1.4
Hunan	2026–2030	High	0.0944	5.9	−2.3	−1.4	1.5
Hunan	2031–2035	Low	0.0944	4.4	−2.9	−1.9	1.2
Hunan	2031–2035	Medium	0.0944	4.9	−2.5	−1.4	1.3
Hunan	2031–2035	High	0.0944	5.4	−2.1	−0.9	1.4
Guangdong	2021–2025	Low	0.9615	3.5	−3.5	−2.9	0.9
Guangdong	2021–2025	Medium	0.9615	4	−3.1	−2.4	1
Guangdong	2021–2025	High	0.9615	4.5	−2.7	−1.9	1.1
Guangdong	2026–2030	Low	0.9615	3	−3.4	−2.4	0.8
Guangdong	2026–2030	Medium	0.9615	3.5	−2.8	−1.9	0.9
Guangdong	2026–2030	High	0.9615	4	−2.4	−1.4	1
Guangdong	2031–2035	Low	0.9615	2.5	−2.9	−1.9	0.7
Guangdong	2031–2035	Medium	0.9615	3	−2.5	−1.4	0.8
Guangdong	2031–2035	High	0.9615	3.5	−2.1	−0.9	0.9
Guangxi	2021–2025	Low	0.9479	5	−3.4	−3.5	1
Guangxi	2021–2025	Medium	0.9479	5.5	−2.8	−3	1.1
Guangxi	2021–2025	High	0.9479	6	−2.4	−2.5	1.2
Guangxi	2026–2030	Low	0.9479	4.5	−2.9	−3	0.9
Guangxi	2026–2030	Medium	0.9479	5	−2.5	−2.5	1
Guangxi	2026–2030	High	0.9479	5.5	−2.1	−2	1.1
Guangxi	2031–2035	Low	0.9479	4	−2.7	−2.5	0.8
Guangxi	2031–2035	Medium	0.9479	4.5	−2.3	−2	0.9
Guangxi	2031–2035	High	0.9479	5	−1.9	−1.5	1
Hainan	2021–2025	Low	1.196	8.2	−3.4	−4	1.5
Hainan	2021–2025	Medium	1.196	8.7	−2.8	−3.5	1.6
Hainan	2021–2025	High	1.196	9.2	−2.4	−3	1.7
Hainan	2026–2030	Low	1.196	7.7	−2.9	−1.5	1.4
Hainan	2026–2030	Medium	1.196	8.2	−2.5	−1	1.5
Hainan	2026–2030	High	1.196	8.7	−2.1	−0.5	1.6
Hainan	2031–2035	Low	1.196	7.2	−2.7	−1	1.3
Hainan	2031–2035	Medium	1.196	7.7	−2.3	−0.5	1.4
Hainan	2031–2035	High	1.196	8.2	−1.9	0	1.5
Chongqing	2021–2025	Low	0.3788	5.1	−3.4	−2.9	0.9
Chongqing	2021–2025	Medium	0.3788	5.6	−3	−2.4	1
Chongqing	2021–2025	High	0.3788	6.1	−2.6	−1.9	1.1
Chongqing	2026–2030	Low	0.3788	4.6	−3.1	−2.4	0.8
Chongqing	2026–2030	Medium	0.3788	5.1	−2.7	−1.9	0.9
Chongqing	2026–2030	High	0.3788	5.6	−2.3	−1.4	1
Chongqing	2031–2035	Low	0.3788	4.1	−2.9	−1.9	0.7
Chongqing	2031–2035	Medium	0.3788	4.6	−2.5	−1.4	0.8
Chongqing	2031–2035	High	0.3788	5.1	−2.1	−0.9	0.9
Sichuan	2021–2025	Low	0.3788	5.1	−3.4	−3.4	1.1
Sichuan	2021–2025	Medium	0.3788	5.6	−3	−2.9	1.2
Sichuan	2021–2025	High	0.3788	6.1	−2.6	−2.4	1.3
Sichuan	2026–2030	Low	0.3788	4.6	−3.1	−2.9	1
Sichuan	2026–2030	Medium	0.3788	5.1	−2.7	−2.4	1.1
Sichuan	2026–2030	High	0.3788	5.6	−2.3	−1.9	1.2
Sichuan	2031–2035	Low	0.3788	4.1	−2.9	−2.4	0.9
Sichuan	2031–2035	Medium	0.3788	4.6	−2.5	−1.9	1
Sichuan	2031–2035	High	0.3788	5.1	−2.1	−1.4	1.1
Guizhou	2021–2025	Low	−1.382	8	−3.4	−1.8	1.7
Guizhou	2021–2025	Medium	−1.382	8.5	−2.8	−1.3	1.8
Guizhou	2021–2025	High	−1.382	9	−2.4	−0.8	1.9
Guizhou	2026–2030	Low	−1.382	7.5	−2.9	−1.4	1.6
Guizhou	2026–2030	Medium	−1.382	8	−2.5	−0.9	1.7
Guizhou	2026–2030	High	−1.382	8.5	−2.1	−0.4	1.8
Guizhou	2031–2035	Low	−1.382	7	−2.7	−1	1.5
Guizhou	2031–2035	Medium	−1.382	7.5	−2.3	−0.5	1.6
Guizhou	2031–2035	High	−1.382	8	−1.9	0	1.7
Yunnan	2021–2025	Low	0.4673	6.5	−3.4	−1.8	3.2
Yunnan	2021–2025	Medium	0.4673	7	−2.8	−1.3	3.7
Yunnan	2021–2025	High	0.4673	7.5	−2.4	−0.8	4.2
Yunnan	2026–2030	Low	0.4673	6	−2.9	−1.4	2.7
Yunnan	2026–2030	Medium	0.4673	6.5	−2.5	−0.9	3.2
Yunnan	2026–2030	High	0.4673	7	−2.1	−0.4	3.7
Yunnan	2031–2035	Low	0.4673	5.5	−2.7	−1	2.2
Yunnan	2031–2035	Medium	0.4673	6	−2.3	−0.5	2.7
Yunnan	2031–2035	High	0.4673	6.5	−1.9	0	3.2
Shan3xi	2021–2025	Low	−0.4695	6	−3.5	−3	0.7
Shan3xi	2021–2025	Medium	−0.4695	6.5	−2.9	−2.5	0.8
Shan3xi	2021–2025	High	−0.4695	7	−2.5	−2	0.9
Shan3xi	2026–2030	Low	−0.4695	5.5	−3	−2.5	0.6
Shan3xi	2026–2030	Medium	−0.4695	6	−2.6	−2	0.7
Shan3xi	2026–2030	High	−0.4695	6.5	−2.2	−1.5	0.8
Shan3xi	2031–2035	Low	−0.4695	5	−2.8	−2	0.5
Shan3xi	2031–2035	Medium	−0.4695	5.5	−2.4	−1.5	0.6
Shan3xi	2031–2035	High	−0.4695	6	−2	−1	0.7
Gansu	2021–2025	Low	−0.1874	6.2	−3.3	−3.5	2.1
Gansu	2021–2025	Medium	−0.1874	6.7	−2.7	−3	2.2
Gansu	2021–2025	High	−0.1874	7.2	−2.3	−2.5	2.3
Gansu	2026–2030	Low	−0.1874	5.7	−2.8	−2.5	2
Gansu	2026–2030	Medium	−0.1874	6.2	−2.4	−2	2.1
Gansu	2026–2030	High	−0.1874	6.7	−2	−1.5	2.2
Gansu	2031–2035	Low	−0.1874	5.2	−2.6	−1.5	1.9
Gansu	2031–2035	Medium	−0.1874	5.7	−2.2	−1	2
Gansu	2031–2035	High	−0.1874	6.2	−1.8	−0.5	2.1
Qinghai	2021–2025	Low	0.4762	4.5	−3.3	−0.6	0.6
Qinghai	2021–2025	Medium	0.4762	5	−2.7	−0.1	0.7
Qinghai	2021–2025	High	0.4762	5.5	−2.3	0.4	0.8
Qinghai	2026–2030	Low	0.4762	4	−2.8	−1.1	0.5
Qinghai	2026–2030	Medium	0.4762	4.5	−2.4	−0.6	0.6
Qinghai	2026–2030	High	0.4762	5	−2	−0.1	0.7
Qinghai	2031–2035	Low	0.4762	3.5	−2.6	−1.6	0.4
Qinghai	2031–2035	Medium	0.4762	4	−2.2	−1.1	0.5
Qinghai	2031–2035	High	0.4762	4.5	−1.8	−0.6	0.6
Ningxia	2021–2025	Low	0.3788	5.1	−3.6	−0.8	0.1
Ningxia	2021–2025	Medium	0.3788	5.6	−3.2	−0.3	0.2
Ningxia	2021–2025	High	0.3788	6.1	−2.8	0.2	0.3
Ningxia	2026–2030	Low	0.3788	4.6	−3.2	−1.3	0.2
Ningxia	2026–2030	Medium	0.3788	5.1	−2.8	−0.8	0.3
Ningxia	2026–2030	High	0.3788	5.6	−2.4	−0.3	0.4
Ningxia	2031–2035	Low	0.3788	4.1	−2.9	−1.8	0.3
Ningxia	2031–2035	Medium	0.3788	4.6	−2.5	−1.3	0.4
Ningxia	2031–2035	High	0.3788	5.1	−2.1	−0.8	0.5
Xingjiang	2021–2025	Low	1.1342	5.3	−3.7	−2	1.1
Xingjiang	2021–2025	Medium	1.1342	5.8	−3.3	−1.5	1.2
Xingjiang	2021–2025	High	1.1342	6.3	−2.9	−1	1.3
Xingjiang	2026–2030	Low	1.1342	4.8	−3.3	−1.5	1
Xingjiang	2026–2030	Medium	1.1342	5.3	−2.9	−1	1.1
Xingjiang	2026–2030	High	1.1342	5.8	−2.5	−0.5	1.2
Xingjiang	2031–2035	Low	1.1342	4.3	−3	−1	0.9
Xingjiang	2031–2035	Medium	1.1342	4.8	−2.6	−0.5	1
Xingjiang	2031–2035	High	1.1342	5.3	−2.2	0	1.1

. Four carbon emission scenarios were established: the baseline scenario, low-economic-development scenario, and high-economic-development scenario, as detailed in

Table 6. Scenario setting of carbon emission.

Scenario	Year-End Population	Per Capita GDP	Energy Intensity	Industrial Structure	Urbanization Rate
Baseline	Intermediate	Intermediate	Intermediate	Intermediate	Intermediate
Industrial structure optimizing	Intermediate	Intermediate	Intermediate	High	Intermediate
Extensive	Intermediate	High	Low	Low	High
Emission reduction	Intermediate	Intermediate	High	Intermediate	Intermediate

.

Baseline scenario: The year-end population was set based on the annual average growth rate in different periods for each province and city, while the rest of the influencing factors were determined according to the relevant policy documents of each province’s “14th Five-Year Plan”, meeting the minimum development requirements. This scenario can to a certain extent represent the development situation of China’s 30 provinces and regions from 2021 to 2035.

Industrial structure optimizing scenario: The industrial structure was set to develop at a high rate, while the change rates of the other influencing factors remained the same as in the baseline scenario. This scenario aimed to explore the impact of optimizing the industrial structure further under the existing economic development model on carbon emissions for different regions.

Extensive scenario: The year-end population was the same as in the baseline scenario, while per capita GDP and urbanization rate were set to develop at a high rate, and energy intensity and industrial structure were set to develop at a low rate. This scenario aimed to study the impact of a strong economic development with less focus on carbon emissions in various regions on carbon emissions.

Emission reduction scenario: Energy intensity developed at a high rate, while the change rates of other influencing factors were the same as in the baseline scenario. This model aimed to study the impact of intensifying energy policy implementation and actively implementing energy conservation and emission reduction measures under the current economic development model in various regions on carbon emissions.

6.2. Carbon Emission Prediction

Using the LSTM-CNN model constructed earlier, a scenario analysis of carbon emissions at the provincial level in China was conducted. The predicted results are shown in Figure 10, and specific peak time details are provided in

Table 7. Prediction of carbon peak time.

Province/City	Baseline	Industrial Structure Optimizing	Extensive	Emission Reduction
BeiJing	2031	2027	2025	2025
TianJing	2027	2027	2025	2028
HeBei	2023	2023	2023	2023
Shan1Xi	Underpeak	2025	Underpeak	2024
NeiMengGu	2024	2025	2024	2024
LiaoNing	2024	2024	2024	2024
JiLing	2031	2025	2025	2025
HeiLongJiang	2026	2026	2032	2027
ShangHai	2027	2024	2026	2025
JiangSu	2023	2023	2024	2024
ZheJiang	2024	2024	Underpeak	2027
AnHui	2026	2025	2025	2024
FuJian	Underpeak	2029	2028	2033
JiangXi	2030	2029	2034	2031
ShanDong	2023	2024	2028	2024
HeNan	2033	2024	2026	2027
HuBei	2026	2030	2026	2025
HuNan	2024	2033	2024	2027
GuangDong	2025	2024	2027	2025
GuangXi	2026	2027	2028	2033
HaiNan	2023	2022	2022	2022
ChongQing	2029	2034	2029	2029
SiChuan	2023	2024	2031	2024
GuiZhou	Underpeak	2027	2026	2028
YunNan	2033	2032	2032	2034
Shan3Xi	2027	2031	2032	2028
GanSu	2032	2028	2029	2032
QingHai	Underpeak	Underpeak	2029	2031
NingXia	2029	2026	2030	2025
XingJiang	2024	2024	2027	2026

.

Different provinces have varying development speeds, technological levels, and development strategies, leading to differences in the timing of carbon peaking under different scenarios. It can be observed that, before 2035, Shanxi Province does not reach its peak in both the baseline and extensive development scenarios. Zhejiang Province does not reach its peak in the extensive development scenario. Fujian Province and Guizhou Province do not reach their peaks in the baseline scenario. Qinghai Province does not reach its peak in both the baseline and optimized industrial structure scenarios. Other provinces reach their peaks under different development scenarios.

Comparing the carbon peaking times of various cities under different scenarios, Beijing reaches its peak earliest in both the extensive development and energy-saving and emission reduction scenarios, achieving carbon peaking in 2025. Tianjin, Fujian, Guizhou, and Qinghai provinces reach their peaks the earliest in the extensive development scenario, achieving carbon peaking in 2025, 2028, 2026, and 2029, respectively. Hebei and Liaoning provinces reach their peaks in the same years under all four scenarios, with Hebei achieving carbon peaking in 2023 and Liaoning in 2024. Shanxi, Anhui, Hubei, and Ningxia provinces reach their peaks the earliest in the energy-saving and emission reduction scenario, achieving carbon peaking in 2024. Inner Mongolia Autonomous Region and Chongqing reach their peaks the earliest in the baseline, extensive development, and energy-saving and emission reduction scenarios, with Inner Mongolia achieving carbon peaking in 2024 and Chongqing in 2029. Jilin and Hainan provinces reach their peaks the earliest in the optimized industrial structure, extensive development, and energy-saving and emission reduction scenarios, with Jilin achieving carbon peaking in 2025 and Hainan in 2022.

Heilongjiang, Jiangsu, Zhejiang, and Xinjiang reach their peaks the earliest in the baseline and optimized industrial structure scenarios, achieving carbon peaking in 2026, 2023, 2024, and 2024, respectively. Shanghai, Jiangxi, Henan, Guangdong, and Gansu provinces reach their peaks the earliest in the optimized industrial structure scenario, achieving carbon peaking in 2024, 2029, 2024, 2024, and 2028, respectively. Shandong, Guangxi Zhuang Autonomous Region, Sichuan, and Shaanxi provinces reach their peaks the earliest in the baseline scenario, achieving carbon peaking in 2023, 2026, 2023, and 2027, respectively. Hunan province reaches its peaks the earliest in the baseline and extensive development scenarios, achieving carbon peaking in 2024. Yunnan province reaches its peaks the earliest in the optimized industrial structure and extensive development scenarios, achieving carbon peaking in 2032.

Based on the four development scenarios, an analysis was conducted on the 30 provinces and regions in China. If there were scenarios where the same city reached its peak at the same time, priority was given to economic development speed. In the baseline scenario, Shandong Province, Guangxi Zhuang Autonomous Region, Sichuan Province, and Shaanxi Province reach their peaks the earliest. This indicates that the effects of optimizing industrial structure, extensive development, and energy-saving and emission reduction scenarios on carbon reduction are not significant in these cities. Among them, Shandong Province and Sichuan Province can achieve carbon peaking before 2025 while maintaining their current development speed. Guangxi Zhuang Autonomous Region and Shaanxi Province can achieve carbon peaking before 2030 while maintaining their current development speed.

In the scenario of optimizing industrial structure, Heilongjiang Province, Shanghai, Jiangsu Province, Zhejiang Province, Jiangxi Province, Henan Province, Guangdong Province, Gansu Province, and Xinjiang Uygur Autonomous Region reach their peaks the earliest. This indicates that these cities are more suitable for the different development rates set in the industrial structure optimization scenario. Among them, Shanghai, Jiangsu Province, Zhejiang Province, Henan Province, Guangdong Province, and Xinjiang Uygur Autonomous Region have a rapid development of low-carbon technologies. By maintaining the development speed of the industrial structure optimization scenario, they can achieve carbon peaking before 2025. On the other hand, Heilongjiang Province, Jiangxi Province, and Gansu Province have a slower development of low-carbon technologies. By maintaining the development speed of the industrial structure optimization scenario, they can achieve carbon peaking before 2030.

In the extensive scenario, Beijing, Tianjin, Hebei Province, Inner Mongolia Autonomous Region, Liaoning Province, Jilin Province, Fujian Province, Hunan Province, Hainan Province, Chongqing, Guizhou Province, Yunnan Province, and Qinghai Province achieve carbon peaking the earliest. Among them, Beijing, Tianjin, Jilin Province, and Hunan Province have a rapid development of low-carbon technologies. By following the different development rates set in the extensive scenario, they can not only achieve carbon peaking before 2025 but also ensure maximum economic development. Hainan Province has a distinct industrial structure with a lower energy consumption, resulting in lower carbon emissions. Following the different development rates set in the extensive scenario, Hainan Province can achieve carbon peaking in 2022 while maximizing economic development. Hebei Province, Inner Mongolia Autonomous Region, and Liaoning Province have similar or identical carbon peaking times across the four scenarios. Due to differing development strategies and economic structures from other cities, they are major coal or industrial-heavy provinces, with closely intertwined economic development and carbon emissions. Their slow transformation of industrial structure is notable.

Fujian Province, Guizhou Province, and Qinghai Province have a slower development of low-carbon technologies and relatively lower economic development levels. They primarily rely on high-carbon energy products such as coal. To achieve the different development rates set in the extensive scenario, these cities must increase the use of clean energy, reduce the proportion of coal usage, and promote the development of low-carbon technologies to ensure carbon peaking before 2030. The carbon emissions of Chongqing are significantly influenced by surrounding cities. Dominated by the manufacturing industry and characterized by a large urban–rural development gap, Chongqing’s carbon emissions show an increasing trend. Following the different development rates set in the extensive scenario, Chongqing can achieve carbon peaking in 2029. Although Yunnan Province reaches carbon peaking the earliest in the extensive scenario, it cannot achieve carbon peaking before 2030. It is evident that Yunnan Province also reaches carbon peaking earliest in the scenario where industrial structure is optimized. To advance the carbon peaking time, Yunnan Province could intensify efforts in industrial structural transformation and accelerate the development of low-carbon technologies.

In the energy-saving and emission reduction scenario, Shanxi Province, Anhui Province, Hubei Province, and Ningxia Province achieve carbon peaking the earliest. This indicates that the aforementioned cities are more suitable for promoting low-carbon lifestyles, increasing the proportion of clean-energy usage, and accelerating the development of low-carbon technologies. By following the different development rates set in the energy-saving and emission reduction scenario, maintaining the pace of development in this scenario, these cities can achieve carbon peaking before 2025.

7. Conclusions and Future Research

7.1. Conclusions

This article calculated the carbon emissions from the energy consumption in 30 provinces, regions, and municipalities in China from 1997 to 2021 and analyzed the trends in carbon emissions in each province. From 1997 to 2021, China’s carbon emissions from energy consumption showed an overall upward trend, increasing from 2051.507 million tons to 5738.955 million tons. Among different provinces, Hebei, Shandong, Jiangsu, Liaoning, Guangdong, and Inner Mongolia Autonomous Region had higher carbon emissions, while Hainan, Qinghai, Ningxia, Beijing, and Gansu had relatively lower carbon emissions. The trends in carbon emissions varied among different provinces.

Moran’s index was used to analyze the spatial autocorrelation of carbon emissions among provinces. There was spatial autocorrelation in carbon emissions among provinces in China, and Moran’s index values were all greater than zero, indicating a positive spatial correlation in carbon emissions. Provinces with higher (lower) carbon emissions were relatively adjacent to provinces with higher (lower) carbon emissions.

A new machine learning prediction model incorporating spatial weights, called the LSTM-CNN model with spatial weights, was proposed in this article and applied to carbon emission forecasting in 30 provinces, regions, and municipalities in China. By adjusting parameters, it was found that the model performed optimally with four convolutional layers. Through ablation experiments, it was found that each part of the model contributed relatively independently. Compared to other models, this model achieved an MAE of 8.0169, RMSE of 11.1505, and R2 score of 0.9661 on the test set, which were higher than those of a single LSTM model, a single CNN model, and an LSTM-CNN model without spatial weights. This indicates that incorporating spatial weights in the model and using a combination of LSTM and CNN models can improve the prediction accuracy.

Four different development scenarios were set to analyze the carbon emissions in 30 provinces, regions, and municipalities in China from 2022 to 2035. Local governments can formulate corresponding carbon emission reduction policies based on the peaking situations under different development scenarios. The results showed that except for Yunnan province, all other 29 provinces, regions, and municipalities could achieve carbon peaking before 2030 under suitable development scenarios. Specifically, in the baseline scenario, Shandong, Guangxi Zhuang Autonomous Region, Sichuan, and Shaanxi would reach their peaks the earliest. In the scenario of industrial structure optimization, Heilongjiang, Shanghai, Jiangsu, Zhejiang, Jiangxi, Henan, Guangdong, Gansu, and Xinjiang Uygur Autonomous Region would reached their peaks earliest. In the extensive scenario, Beijing, Tianjin, Hebei, Inner Mongolia Autonomous Region, Liaoning, Jilin, Fujian, Hunan, Hainan, Chongqing, Guizhou, Yunnan, and Qinghai would reach their peaks the earliest. In the energy conservation and emission reduction scenario, Shanxi, Anhui, Hubei, and Ningxia would reach their peaks the earliest.

7.2. Policy Recommendations

For different cities, it is recommended that Shandong Province, Guangxi Zhuang Autonomous Region, Sichuan Province, and Shaanxi Province develop at different rates as set in the baseline scenario. The decarbonization effects are not satisfactory in the optimized industrial structure, extensive development, and energy-saving and emission reduction scenarios in the above-mentioned cities. It is recommended that Heilongjiang Province, Shanghai, Jiangsu Province, Zhejiang Province, Jiangxi Province, Henan Province, Guangdong Province, Gansu Province, and Xinjiang Uyghur Autonomous Region develop at different rates as set in the scenario where industrial structure is optimized. Compared with the other three scenarios, accelerating the transformation of industrial structure in the above-mentioned cities can advance the carbon peaking time. It is recommended that Shanxi Province, Anhui Province, Hubei Province, and Ningxia Province develop at different rates as set in the energy-saving and emission reduction scenario. The above-mentioned cities are suitable for promoting low-carbon lifestyles, increasing the proportion of clean-energy use, and accelerating the development of low-carbon technologies.

It is recommended that Beijing, Tianjin, Jilin Province, Hunan Province, Hainan Province, and Chongqing develop at different rates as set in the extensive development scenario. Except for Chongqing, the development of low-carbon technologies is rapid in these cities, and maintaining a high economic development speed has little impact on the timing of carbon peaking. The carbon emissions of Chongqing are greatly influenced by surrounding cities. Due to different development strategies and economic structures from other cities, Hebei Province, Inner Mongolia Autonomous Region, Liaoning Province, Fujian Province, Guizhou Province, and Qinghai Province need to increase the use of clean energy, reduce the proportion of coal consumption, and strengthen the development of low-carbon technologies in order to develop at different rates as set in the extensive development scenario.

If Yunnan Province wants to advance its timing of carbon peaking, it can consider intensifying the transformation of industrial structure and accelerating the development of low-carbon technologies.

7.3. Insufficiency and Future Prospects

For the limitations and improvements of the model encountered in the research process, this paper provides the following two aspects:

(1) This paper applied deep learning algorithms to predict carbon emissions. The internal mechanism of deep learning algorithms is very complex, and the judgment and inference results of the model are often difficult to explain. This means the model can only be used for predicting carbon emissions but not for analyzing the influencing factors of carbon emissions.

(2) The independent variables in the model are all socio-economic data. In the scenario analysis, it was necessary to first set the rate of change of each independent variable, predict the independent variables, and then predict the carbon emissions. The subjective setting of the rate of change of independent variables was relatively strong, and the longer the prediction time, the greater the deviation between the predicted data and the real data. Based on the development plan released by the country, this paper cautiously set the rate of change of each independent variable to minimize data deviation.

For future work, we could use feature extraction networks with stronger exploratory capabilities, such as Transformer [26] and GNN models [27], for the input variable part of the model. We can also introduce causal inference [28] and other methods to improve the prediction effect of carbon emissions based on influencing factors, and explore the causal relationship between influencing factors. Finally, the selection of original data can also contribute to the improvement and enhancement of the final prediction effect.