To formulate targeted emissions reduction policies and achieve the expected results, China must understand the factors that lead to the increase of carbon dioxide emissions, in order to effectively reduce carbon dioxide emissions by controlling these factors. Many scholars have studied the influencing factors of carbon dioxide emissions. Dong Feng [
12] applied the LMDI method to decompose the influencing factors of carbon emissions into economic scale, industrial structure, technological progress, and energy structure, and found that economic scale is the main driving factor of carbon emissions growth, while technological progress plays a negative role. Based on the LMDI factor decomposition method, Zhou Yannan [
13] analyzed the changes of carbon emissions in different provinces and regions from three aspects: economic scale, structural transformation, and technological progress. The study showed that the leading effects of carbon emissions in different periods varied between different provinces and regions. Yan Qingyou [
14] used the GDIM model to study the driving factors of CO
2 emissions change of thermal power generation in China from 2000 to 2016. The decomposition results show that economic activities, power demand, and energy use play a positive role in driving carbon emissions of thermal power generation in China. Gao Sihui et al. [
15] and Liu Xiaodie [
16] used a Lasso regression model to screen out eight important carbon emissions influencing factors, simulated and set the value of each influencing factor from 2019 to 2030, and established a BP neural network model to predict the carbon emissions and peak time of Jiangsu Province from 2019 to 2030. Therefore, in terms of analysis of the influencing factors of carbon emissions from energy consumption, there are many mature research methods, and there is no final conclusion on which method can accurately analyze the influencing factors. However, based on a range of research, combined with the empirical research and comparative analysis of the decomposition methods of carbon emissions by experts and scholars in the energy field in recent years, the XG-Boost method selected in this paper can achieve the desired effect on the decomposition of the factors affecting carbon emissions.
Global warming has attracted increasing attention, and scholars have made rich research achievements in carbon emissions prediction. At present, the research mainly focuses on the prediction of carbon emissions and their peaks. Scholars have studied carbon emissions using different types of models based on different theoretical bases and different research perspectives. Shi Shaoqing [
17] and others used the multi-resolution time series neural network (MTNN) model to achieve the power consumption prediction task. Sun Xiujuan [
18] applied the information entropy theory to traffic flow forecasting, and established a variable weight combination forecasting model based on the entropy weight method by using the hidden information contained in the objective data. Dong Xiaogang [
19] used the DTGARCH Laplace model to forecast the return on assets, providing investors with a more accurate decision-making basis. Duan Fumei [
20] used the improved BP neural network model to predict peak carbon emissions in China under eight development models. The research shows that China can achieve the goal of reaching peak carbon emissions by 2030 under the economic recession model, energy-saving model, and other models. Wang Hehe [
21] used the whale optimization algorithm to improve the limit learning machine model, overcoming the deficiency of the limit learning machine that is easy to fall into the local optimal solution, and used the improved model to predict China’s carbon emissions and intensity from 2019 to 2040, accurately reflecting China’s future carbon emissions trend. Chiun SinLin [
22] used the gray prediction model to predict the trend of Taiwan’s carbon dioxide emissions from 2010 to 2012. Cao Zhengsheng [
23] used the autoregressive integral moving average ARIMA prediction model to predict carbon emissions, and calculated the regression results and predicted values using temperature and global greenhouse gas emissions. After calculation, they completed 16% carbon emissions reduction annually, and predicted the global average temperature to increase by 1.6 °C by 2050, reaching the goal set by the United Nations. Zhao Jinyuan [
24] established a BP neural network and multiple linear regression model to predict and analyze carbon emissions of iron and steel enterprises, and found that the BP neural network model is superior to the multiple linear regression model in predicting carbon emissions. Hu Jianbo [
25] deduced and predicted the change trend of China’s carbon emissions intensity based on a LSTM neural network model and certain economic growth expectations. At the same time, they established an ARIMA-BP neural network model as a verification model to directly predict the carbon emissions intensity, which verified the robustness and accuracy of the prediction model. Yaqin Hu [
26] estimated the CO
2 emissions of large-scale coal-fired power plants using Orbiting Carbon Observatory-2 (OCO-2) satellite data based on remote sensing inversions and bottom-up methods, Machine learning algorithms extract key information and the main data features from known datasets to make accurate identification and prediction of future data. Tong Xin [
27] used the gray model GM(1,1) to predict China’s carbon emissions from 2012 to 2020. According to the predicted results, China still faces great pressure to reduce emissions. Meng [
28] proposed a new prediction model based on the GM(1,1) model by introducing a linear equation, and predicted China’s carbon emissions in the past 20 years. The results showed that the improved model had a better prediction effect. Song [
29] proposed an improved gray multivariate model, which avoids the inaccuracy and poor adaptability of the traditional gray model, and used the improved model to predict China’s carbon emissions, providing reasonable guidance for the formulation of energy consumption plans. Liu [
30] combined the gray model with the autoregressive comprehensive moving average model and the second-order polynomial regression model, and optimized the model parameters by using the PSO method. The results show that if the government did not control carbon emissions, it was difficult for China to achieve the set goal in 2020.The gray prediction model is simple in application, with few parameters, is easy to train, and has high prediction accuracy. However, the fitting ability of the model is inferior to that of the machine learning algorithm. A large number of studies have shown that machine learning can better deal with prediction problems, and also proved that the idea of applying machine learning to prediction in the energy field is feasible.
In summary, in the relevant literature on carbon emissions of coal-fired power plants, there are many studies on the establishment of carbon emissions prediction models of coal-fired power plants using traditional algorithms. The selection of influencing factors for predicting the carbon emissions of coal-fired power plants mostly focuses on direct carbon emissions influencing factors, and there are few studies on the optimization and improvement of models and the dimensional expansion of influencing factors. In terms of the selection of influencing factors, this paper uses the XG-Boost algorithm to analyze the feature importance of influencing factors on the basis of statistical analysis. Compared with principal component analysis, analytic hierarchy process, and other subjective methods, the XG-Boost algorithm can objectively analyze various influencing factors of carbon emissions of coal-fired power plants, making the selection of influencing factors more scientific. In terms of establishing the prediction model, inspired by the sparrow’s foraging and anti-predation behavior, this paper selects the sparrow search algorithm (SSA) with its strong optimization ability and fast convergence speed to optimize the super parameters of the LSTM algorithm, and proposes to use the SSA-LSTM algorithm to establish the carbon emissions regression prediction model of coal-fired power plants. The advantage of the SSA-LSTM algorithm is that it can effectively simplify the super parameter selection process of the LSTM algorithm, effectively solve the global optimization problem, prevent the model from falling into overfitting and local optimization, and enable the carbon emissions regression prediction model of coal-fired power plants to achieve a better fitting effect. Through multiple screenings of influencing factors, this paper determines that the relevant influencing factors of the boiler feed water system have a strong correlation and characteristic importance with the carbon emissions results of coal-fired power plants. While previous studies have ignored this more important influencing factor, this paper proposes to introduce it as an auxiliary variable into the model, which can more reasonably reflect the carbon emissions results of coal-fired power plants, and then introduce the SSA-LSTM algorithm, The optimized carbon emissions regression prediction model is established to improve the prediction performance of the model on carbon emissions of coal-fired power plants.