https://orcid.org/0009-0003-9903-7187
Figures
Abstract
This study develops a method to predict carbon dioxide (CO2) emissions from the cement industry in the Beijing-Tianjin-Hebei region using artificial intelligence-based neural networks. By analyzing data from the National Bureau of Statistics and the China Statistical Yearbook (2010–2021), we calculated CO2 emissions generated by fossil fuel combustion during cement production. The neural network model achieved robust predictive performance with a root mean square error (RMSE) of 0.05, a mean absolute error (MAE) of 2,640,769 tons, and a coefficient of determination (R2) of 0.9620. These results demonstrate the model’s effectiveness in identifying emission trends and supporting real-time strategies to mitigate CO2 emissions. Future research could expand this approach to other high-emission industries, providing valuable tools for global carbon reduction efforts.
Citation: Liu Y, Xu Q, Wang Z, Qi L, Lu J (2025) Estimation of carbon dioxide emissions from the cement industry in Beijing-Tianjin-Hebei using neural networks. PLOS Clim 4(3): e0000544. https://doi.org/10.1371/journal.pclm.0000544
Editor: Muhammad Irfan Ashraf, University of Sargodha, PAKISTAN
Received: October 29, 2024; Accepted: January 25, 2025; Published: March 17, 2025
Copyright: © 2025 Liu et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: The authors confirm that the data supporting the findings of this study are available within the article.
Competing interests: The authors have declared that no competing interests exist.
Introduction
The escalating threat of global climate change has made the reduction of carbon dioxide (CO2) emissions from industrial processes a critical goal for governments and businesses worldwide. The Sixth Assessment Report by the UN’s Intergovernmental Panel on Climate Change (IPCC) indicated that global greenhouse gas emissions peaked between 2010 and 2019 [1]. The Paris Agreement, signed by 197 countries in 2015, aims to curb the rise in global temperatures through significant CO2 emission reductions [2]. According to the International Energy Agency (IEA), China’s CO2 emissions in 2021 totaled 1.19 × 1010 tons, accounting for about 33.0% of global emissions [3]. The cement industry, being a major consumer of energy and a significant emitter, contributes approximately 8% to global anthropogenic CO2 emissions, with this share expected to grow [4]. Therefore, accurately forecasting CO2 emissions from the cement industry is essential for developing effective mitigation strategies.
Traditional methods for estimating CO2 emissions, such as statistical analyses[5–7] and empirical models[8–11], are limited in their ability to handle the nonlinear relationships and complex dynamics inherent in emission datasets. These methods may not fully capture the long-term dependencies and dynamic changes in CO2 emissions influenced by factors like energy consumption patterns, production processes, and policy interventions[12–19]. In recent years, deep learning techniques, particularly Long Short-Term Memory (LSTM) neural networks[20], have shown significant potential in managing time-series data with complex patterns and nonlinear relationships.
In the realm of time series forecasting, Recurrent Neural Networks (RNN) can process sequences of any length but are sensitive to long-term dependencies[21]. LSTM, a special type of RNN, effectively addresses long-term dependencies through its memory cell mechanism, despite higher computational costs. Gated Recurrent Units (GRU)[22], a simplified version of LSTM, are more effective than standard RNNs in handling long-term dependencies and are less computationally expensive, though they may not be as accurate as LSTM. In contrast, Autoregressive Integrated Moving Average (ARIMA) models are suitable for stationary time series but are sensitive to outliers and cannot capture the sequential nature of the data [23, 24].
This study proposes an LSTM-based method to predict CO2 emissions from the cement industry in the Beijing-Tianjin-Hebei region, utilizing data from the National Bureau of Statistics and the China Statistical Yearbook (2010–2021). Our aim is to enhance the accuracy and reliability of CO2 emission predictions, thereby providing valuable insights for real-time monitoring and mitigation strategies in the cement industry.
Materials and methods
Data collection
The data used in the study are related to energy consumption and cement production in Beijing-Tianjin-Hebei from 2010 to 2021, which are obtained from the National Bureau of Statistics and the China Statistical Yearbook. The data were used to account for and predict the CO2 generated by the cement industry in Beijing-Tianjin-Hebei from 2010 to 2021, and the relevant data are shown in Tables 1, 2, and 3. The dataset includes eight variables (raw coal, coke, gasoline, kerosene, diesel, fuel oil, natural gas, and cement production) in 10,000 tons (natural gas in billion cubic meters), and the number of samples is 36 in total.
China does not directly publish the emissions of CO2 , so it is necessary to use the relevant methods to estimate the emissions of CO2 in the three provinces of Beijing-Tianjin-Hebei, and the specific formula for calculating the emissions of CO2 from fossil energy combustion is as Eq (1) [25]:
where EC denotes the total estimated CO2 emissions from various types of energy consumption;Ei is the total amount of energy consumed in each province for each type of energy, measured in tons for solid and liquid fuels and in billion cubic meters for natural gas;CFi is the calorific value of the energy source, representing the amount of energy released when a unit of fuel is completely burned. It is measured in terajoules per tonne (TJ/tonne) or terajoules per billion cubic meters (TJ/billion m3) for natural gas;CCi is the carbon content of the energy source, indicating the amount of carbon per unit of energy. It is measured in tonnes of carbon per terajoule (t-C/TJ);COFi is the carbon oxidation factor, which accounts for the fraction of carbon in the fuel that is converted to CO2 during combustion. It is a dimensionless quantity, typically close to 1;The constant 3.67 is used to convert the carbon emission factor into CO2 emissions by considering the molecular weight ratio of CO2 to carbon (44/12). CO2 emitted during cement production is calculated as Eq (2):
where CC denotes the total amount of CO2 emitted in the cement production process, and Q denotes the total amount of cement production, and EFcement is the CO2 emission factor for cement production. Table 4 lists the CO2 emission factors for each type of emission source.
Data processing
After filling in the missing values of the data by interpolation, the labeled data are processed by noise reduction filtering, which can improve the speed of model training as well as the prediction accuracy. Considering that SG filtering can retain the change information of the data more effectively while filtering out the noise, and can extract useful data from the noisy data and other advantages, SG filter is used to denoise the data after carbon emission accounting. Fig 1 show CO2 emission trends in the cement industry of Beijing, Tianjin, and Hebei from 2010 to 2021, highlighting the impact of policy changes and economic events, posing challenges for emission accounting.
Pearson’s correlation coefficient, a statistical measure of the degree of linear correlation between two variables, takes values between [–1, 1]. The Pearson’s correlation coefficient between two variables is defined as the quotient of the covariance and standard deviation between the two variables, as shown in Eq (3):
The above equation defines the overall correlation coefficient, which is often represented by the lowercase Greek letter f ( ⋅ ) as representative symbols, as shown in Eq (4). Estimating the covariance and standard deviation of the sample yields the Pearson’s correlation coefficient, which is often represented by a lowercase English letter r as representative symbols:
The characteristic correlation diagram is shown in Fig 2. Except for gasoline, fuel oil and cement industry production CO2, the correlation coefficient is less than 0.5, which is a weak correlation; the correlation coefficients of raw coal, coke, kerosene, diesel oil, natural gas and cement production CO2 are all greater than 0.7, which is a strong correlation. The correlation coefficients did not show any extremely weak correlation features less than 0.1, so the original features were retained as input feature data for the neural network model.
Before constructing the model, the feature data needs to be normalized in order to eliminate the effect of magnitude between features. market_cap is a labeled dataset, and labeled data does not need to be normalized. Normalization also speeds up the processing of the machine learning model, thus shortening the training time. Maximum-minimum normalization is used to map the feature data between 0 and 1. The maximum-minimum normalization as shown in Eq (5):
where x∗ is the normalized data; x is the original data; xmax is the maximum value of the data; xmin is the minimum value of the data.
LSTM model construction
Long Short Term Memory (LSTM) network is chosen as a deep learning model because LSTM can effectively capture long-term dependencies in time-series data, and is suitable for processing time-varying data in the cement production process. Compared to feed-forward neural networks, LSTM can better utilize previous information when processing inputs and is suitable for tasks that require consideration of contextual relationships, such as natural language processing and sequence data analysis. The steps of a time-series prediction problem can be generally divided into: collecting the original dataset, cleaning and visualizing the dataset, dividing the training and testing sets, normalizing the data, setting parameters (sliding window size, input parameters, output parameters, number of neurons in the hidden layer, drop_out, batch_size, optimizer, number of training sessions, etc.), constructing the model, training the model, optimizing the model hyperparameters, predicting the metrics, model prediction, optimizing the model, optimizing the model hyperparameters, optimizing the model hyperparameters, and optimizing the model. model hyperparameters, metrics prediction, model evaluation, result visualization and analysis, etc. When feedforward neural networks are used, the neural network treats the input results at time t and time t+1 as independent, which is not a problem in many occasions (e.g., image classification and recognition), but in some scenarios, such as in Natural Language Processing (NLP, Natural Language Processing) or when we want to do something like taking continuous photos of the data, proper utilization of t or previous inputs for t time can make better use of the inputs.The LSTM cell graph is shown in Fig 3. LSTM cells can be categorized as forget gates, memory gates and output gates. Some of the “information” about the previous cell state in the LSTM may become “outdated” with the passage of time. In order to prevent too much memory from affecting the neural network’s processing of the current input, it should selectively forget some components of the previous unit state - this work is done by the forgetting gate. The schematic diagram of the forgetting gate is shown in Fig 4.
The following Eq (6) can be used to describe the computation of the forgetting gate which is the output vector of the sigmoid neural layer:
Memory gates are control units used to control whether or not to incorporate data from the current moment into the cell state. First, the tanh function layer is used to extract the valid information from the current vector, and then the sigmoid function is used to control “how much” of this memory is put into the unit state. The combination of the two can be done as shown in Fig 5, the memory gate diagram.
The output gate, is the neural layer used by the LSTM unit to compute the output value at the current moment. The output layer will first combine the current input values with the vector integrating the output values of the previous moment, and extract the information in it with the sigmoid function, and then, it will compressively map the current state of the unit into the interval ( − 1, 1) by the tanh function.
A BP neural network is divided into two phases: forward propagation and backward feedback, or BP for short. It learns and saves the mapping relationship between inputs and outputs without having to overtly describe the mathematical equations. The weights and thresholds are adjusted through backward propagation to minimize the sum of squares of the network error. It is characterized by forward signal and backward error. The BP neural network model consists of an input layer, a hidden layer, and an output layer, and the network structure is shown in Fig 6.
In summary, the LSTM model has a memory function and can perform well in dealing with time-series regression problems, so LSTM is more suitable for predicting the CO2 generated in the cement industry.
The model built in this experiment is LSTM neural network regression prediction model, and the model structure contains input layer, activation function layer, regression layer, LSTM layer and fully connected layer. The model structure is shown in Table 5.
The model parameters are set as shown below:
The model parameters were carefully selected based on optimization strategies to ensure the best performance. The dropout rate was set to 0.5 to prevent overfitting, while the sliding window size was set to 2 to capture short-term dependencies. The hidden layer size was optimized to 15 neurons, and the learning rate was set to 0.001 after extensive hyperparameter tuning. The dataset was split into training and testing sets with an 80:20 ratio to validate the model’s robustness.
Results and discussion
The accounting results of CO2 emissions from the cement industry in the Beijing-Tianjin-Hebei region are presented in Figs 7, 8, 9, 10. It is evident that the peak CO2 emissions occurred in Beijing in 2010, at approximately 5.5 million tons, followed by a consistent downward trend, with the lowest emissions recorded in 2021 at about 1.5 million tons. Tianjin’s cement industry saw its highest emissions in 2014, at around 5.7 million tons, which then gradually decreased, reaching the lowest point in 2021 at less than 1.5 million tons. The Hebei region had the highest average annual emissions among the three, peaking in 2011 at about 76 million tons, and experienced a fluctuating trend from 2011 to 2021, with the lowest emissions in 2015, below 50 million tons.
Figs 11 and 12 illustrates the fitting curve graph comparing the real and predicted values, indicating a minimal error between them and a high degree of trend alignment. The model’s performance evaluation indices are an MAE of 264.0769 million tons and an R2 of 0.9620, demonstrating its high accuracy and reliability. Figs 13 and 14 presents a comparison of the model’s predictions on the test set against actual values, highlighting the LSTM network’s strong capability to capture temporal relationships between features and labels, and its enhanced ability to predict CO2 emission trends in the cement industry. This provides a more reliable basis for future CO2 emission forecasts.
Notably, certain prediction errors are observed under specific production conditions influenced by policy changes, suggesting areas for potential model refinement. Future research could further investigate optimization strategies for the model under such conditions. The model proposed in this study has high feasibility in practical application, allowing for timely adjustments to production strategies to mitigate CO2 emissions through real-time monitoring of key production parameters.
Conclusion
In this study, we developed a deep learning model based on LSTM to predict CO2 emissions from the cement industry in the Beijing-Tianjin-Hebei region. This approach demonstrated significant advantages over traditional methods in handling complex data and capturing nonlinear relationships.
(1) The correlation analysis revealed that, except for gasoline, fuel oil, and cement industry CO2 emissions, the correlation coefficients for other variables were less than 0.5, indicating a weak correlation. In contrast, the coefficients for raw coal, coke, kerosene, diesel fuel, natural gas, and cement production CO2 were all greater than 0.7, indicating a strong correlation. Importantly, no features showed an extremely weak correlation of less than 0.1, validating the retention of these features as input data for the LSTM model.
(2) The LSTM model effectively captured the trend of CO2 emissions, with prediction results closely aligning with actual observations. This alignment suggests that the LSTM model is well-suited for predicting CO2 emissions in the cement industry.
(3) Using MAE and R2 as performance metrics, our model achieved an MAE of 2,640,769 tons and an R2 of 0.9620, reflecting high precision and reliability.
(4) Policymakers and industry stakeholders involved in reducing CO2 emissions from the cement sector provide valuable insights. For example, based on projected emission trends, policies can be designed to gradually reduce emissions by optimizing energy consumption or promoting the use of low-carbon fuels. Accurate forecasts can effectively allocate resources to mitigation initiatives. The model can be used as a monitoring tool to track the effectiveness of existing policies and ensure compliance with emission standards.
Looking ahead, future research should focus on refining the model’s architecture and hyperparameters, expanding the dataset, and exploring its application in other high-emission sectors like steel and petrochemical industries. The integration of real-time data and alternative models, such as Transformer models, holds promise for enhancing predictive accuracy further. Overall, this study presents an LSTM-based model with practical applicability for real-time monitoring and adjusting production strategies to reduce CO2 emissions in the cement industry.
References
- 1. Edenhofer P, Pichsmadruga R, Sokona Y, Farahani E, Kadner S, Seyboth K. Climate change 2014: mitigation of climate change. In: Contribution of Working Group III to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change. 2014.
- 2. Blau J. The Paris agreement. Springer; 2017.
- 3. Zhou J, Li N, Feng W, Tang K, Zhu S, Zhu W. Estimation and future scenario prediction of road source carbon dioxide emissions in China. Huanjing Kexue Xuebao/Acta Scientiae Circumstantiae. 2023;43(10):267–78.
- 4. Edwardes-Evans H, Watson F, Elliott S. Global energy CO2 emissions forecast to rise 5 coal burn: IEA. Platts Coal Trader Int. 2021;21(75).
- 5. Hou C. Analysis and prediction of carbon intensity and carbon peak in Gansu province under the dual carbon goals. Gansu Finance. 2021;000(011):44–9.
- 6. Jiang S, Xie Z, Yu B. Prediction of carbon dioxide emissions and carbon reduction paths in the cement industry of Guangdong Province. Cement. 2023;12(12):11–7.
- 7. Rietmann N, Hügler B, Lieven T. Forecasting the trajectory of electric vehicle sales and the consequences for worldwide CO2 emissions. J Clean Prod. 2020;261:121038.
- 8. Duan F. Scenario prediction and peak characteristics of China’s carbon dioxide emissions: analysis based on BP neural network optimized by particle swarm optimization algorithm. J Northeast Univ Finance Econ. 2018:5(2018);9.
- 9. Wu W, Ma X, Zhang Y, Li W, Wang Y. A novel conformable fractional non-homogeneous grey model for forecasting carbon dioxide emissions of BRICS countries. Sci Total Environ. 2020;707:135447. pmid:31863989
- 10. Nguyen DK, Huynh TLD, Nasir MA. Carbon emissions determinants and forecasting: evidence from G6 countries. J Environ Manage. 2021;285:111988. pmid:33561733
- 11. Qiao W, Lu H, Zhou G, Azimi M, Yang Q, Tian W. A hybrid algorithm for carbon dioxide emissions forecasting based on improved lion swarm optimizer. J Clean Prod. 2020;244:118612.
- 12. Tian X, Li Z, Geng Y, Chen W. Resource and environmental impacts of carbon reduction in China’s steel industry under the carbon neutrality goal. Resource Sci. 2024;46(4):700–16.
- 13. Hou J, Wang L, Hong S. Scenario prediction and policy recommendations for carbon peak in Zhejiang province. China Resources Comprehens Utiliz. 2022;40(8):196–201.
- 14. Yue L. Study on the influencing factors of synergistic reduction of air pollutants and carbon dioxide in Hebei Province. Hebei University; 2022.
- 15. Wei H, Wu J. Analysis and prediction of factors affecting carbon emissions in Guangxi. J Nanning Norm Univ: Philos Soc Sci Edn. 2022;43(3):17–30.
- 16. Yang H. Prediction study on the path to carbon peak in the railway industry. Railway Eng Econ. 2023;38(4):45–51.
- 17. Chang S, Feng G, Cui H, Zhang L, Li Q. Analysis of carbon emission characteristics and reduction potential in the construction industry. J Shenyang Jianzhu Univ: Nat Sci Edn. 2023;39(1):139–46.
- 18. Pang L, Weng H, Chang J, Li Y, Cai B, Lei Y, et al. Study on the path to carbon dioxide emission peak in China’s petrochemical and chemical industry. Res Environ Sci. 2022:35(2);12.
- 19. Wei L, Hou Y. Estimation of marginal abatement costs of carbon dioxide in China’s industry and prediction of carbon peak in sectors. Econ Theory Econ Manag. 2023;43(2):63–77.
- 20. Hochreiter S. Long short-term memory. Neural computation. MIT Press; 1997.
- 21. Medsker L, Jain L, et al. Recurrent neural networks. Design Appl. 2001;5(2):64–7.
- 22. Chung J, Gulcehre C, Cho K, Bengio Y. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint. 2014. http://arxiv.org/abs/1412.3555
- 23. Box GEP, Pierce DA. Distribution of residual autocorrelations in autoregressive-integrated moving average time series models. J Am Statist Assoc. 1970;65(332):1509–26.
- 24. Pirani M, Thakkar P, Jivrani P, Bohara MH, Garg D. A comparative analysis of ARIMA, GRU, LSTM and BiLSTM on financial time series forecasting. In: 2022 IEEE International Conference on Distributed Computing and Electrical Circuits and Electronics (ICDCECE). 2022. p. 1–6.
- 25. Du L. Factors affecting carbon dioxide emissions in China: an analysis based on provincial panel data. Southern Econ. 2010;11(11):20–33.