Papermaking process carbon footprint prediction method based on two-stage machine learning

Abstract

The invention discloses a papermaking process carbon footprint prediction method based on two-stage machine learning, and belongs to the technical field of industrial environmental protection and machine learning crossing. Aiming at the problems that the conventional life cycle evaluation is difficult to process multivariable nonlinear coupling of the papermaking process, the prediction precision is insufficient and the interpretability is weak, the method comprises the following steps of: constructing a feature data set covering a full life cycle link and preprocessing the feature data set; training by adopting a random forest algorithm to obtain an initial prediction model and outputting a prediction result; calculating a prediction residual error; training an XGBoost model based on the residual error and the original features to construct a residual error prediction model; and superposing prediction results of the two models to obtain a final carbon footprint prediction value. According to the method, the prediction precision and the model interpretability of the high emission interval are remarkably improved, and the method is suitable for carbon emission accounting and process optimization in the papermaking industry.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of industrial environmental protection and machine learning, and in particular relates to a method for predicting the carbon footprint of papermaking processes based on two-stage machine learning. Background Technology

[0002] Against the backdrop of global climate change and national carbon neutrality goals, the scientific quantification and prediction of carbon emissions from industrial products has become a critical issue. The paper industry, as a typical energy- and resource-intensive process manufacturing sector, primarily relies on traditional life cycle assessment (Life Cycle Assessment) methods for its life cycle carbon footprint assessment. This method systematically quantifies greenhouse gas emissions from raw material acquisition, production, energy supply, transportation, and waste disposal, laying a theoretical foundation for understanding industry emission patterns. However, in actual paper production scenarios, due to long process chains, numerous process variables, and significant differences between plants, traditional Life Cycle Assessment methods have obvious limitations. They typically rely on pre-set scenarios and fixed parameters, making it difficult to absorb and process massive amounts of continuous process data, and even more difficult to effectively characterize the complex nonlinear coupling effects between various process parameters. This limits its potential for accurate prediction and process optimization under actual production conditions involving multiple plants and multiple process combinations.

[0003] With the development of machine learning technology in the environmental field, algorithms such as random forests and XGBoost have been attempted for carbon emission prediction. These data-driven models have shown advantages in handling high-dimensional inputs and nonlinear relationships. However, in specific applications to complex process industries such as papermaking, existing technologies still have shortcomings. First, most studies use single-stage machine learning models for prediction, lacking systematic diagnosis and correction of the model's residual structure, resulting in insufficient prediction accuracy under extreme conditions or in high-emission ranges. Second, the interpretability of the models is often weak, with correlation analyses mostly remaining at the level of variable importance ranking, failing to deeply integrate with the specific physicochemical mechanisms of the papermaking process, thus making it difficult to extract actionable emission reduction recommendations that can guide production practices from the prediction results. Finally, when constructing prediction models, the choice of data preprocessing methods has a significant impact on performance, but existing technologies lack a consensus on standardized preprocessing strategies for the characteristics of papermaking carbon footprint data, which brings difficulties to the stable reproduction and reliable application of the models.

[0004] Therefore, existing technologies face multiple challenges in dealing with the carbon footprint prediction problem of the complex system of papermaking, such as limited model prediction accuracy, insufficient adaptability to key operating conditions, weak interpretability, and inconsistent preprocessing methods. There is an urgent need for a prediction method that can deeply integrate process mechanisms, has residual learning capabilities, and is more robust. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a two-stage machine learning-based method for predicting the carbon footprint of the papermaking process, thereby resolving the issues present in the existing technologies.

[0006] Firstly, to achieve the above objectives, this invention provides a method for predicting the carbon footprint of a papermaking process based on two-stage machine learning, comprising the following steps: Construct a feature dataset covering all stages of the papermaking lifecycle and perform preprocessing; The random forest algorithm is used to train the preprocessed dataset to obtain the initial carbon footprint prediction model and output the prediction results; Calculate the prediction residual of the initial carbon footprint prediction model; Based on the predicted residuals and the original feature dataset, train the XGBoost model to construct a residual prediction model; The prediction results of the initial carbon footprint prediction model are superimposed with the output of the residual prediction model to obtain the final carbon footprint prediction value.

[0007] Optionally, a feature dataset covering all stages of the papermaking lifecycle is constructed and preprocessed, including: Integrate raw material structure, energy structure, process parameters, and waste treatment data to construct a feature matrix; One-hot encoding is performed on categorical variables, and logarithmic transformation is performed on continuous long-tailed variables; The dataset is divided into a training set and a test set, and all preprocessing operations are performed only on the training set.

[0008] Optionally, training the preprocessed dataset using the random forest algorithm includes: Construct multiple decision trees with the constraint that the number of leaf nodes is 1 and the tree depth is not limited; When splitting at a node of each decision tree, a subset of features is randomly selected to determine the optimal splitting method; Based on the trained decision tree set, the initial prediction results are output through an ensemble method.

[0009] Optionally, calculating the prediction residuals of the initial carbon footprint prediction model includes: Subtract the predicted value from the initial carbon footprint prediction model from the actual calculated carbon footprint value to obtain the original residual; Analyze the correlation between the original residuals and the predicted values ​​to confirm whether there are learnable systematic biases in the residuals.

[0010] Optionally, training the XGBoost model to build a residual prediction model includes: Use the preprocessed original feature dataset as input features; The predicted residuals are used as training targets, and the XGBoost algorithm is used for fitting. The XGBoost algorithm is guided by the residuals of the previous fitting and gradually approximates the objective function through an additive tree model.

[0011] Optionally, after training with the random forest algorithm, the process may further include: Based on the results of the random forest algorithm, a permutation importance analysis is performed. Based on the importance of each feature, the key process parameters affecting carbon footprint are identified.

[0012] Optionally, after obtaining the final carbon footprint prediction, the following may be further included: The performance of the final carbon footprint prediction was evaluated on an independent test set using the coefficient of determination and root mean square error metrics. The final carbon footprint prediction is then applied to the carbon emission trend analysis of a specific type of paper.

[0013] Optionally, the specific type of paper is cultural paper.

[0014] Secondly, the present invention also provides a computer terminal device, comprising: One or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the papermaking process carbon footprint prediction method based on two-stage machine learning in the first aspect described above.

[0015] Thirdly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the steps of the papermaking process carbon footprint prediction method based on two-stage machine learning in the first aspect described above.

[0016] Compared with the prior art, the present invention has the following advantages and technical effects: This invention provides a two-stage machine learning-based method for predicting the carbon footprint of papermaking processes, effectively improving the overall accuracy of carbon footprint prediction, particularly in high-emission ranges. By introducing a residual relearning mechanism, this invention enhances the modeling ability for complex nonlinear relationships and extreme operating conditions, improving the model's robustness and stability. Simultaneously, the deep integration of feature importance analysis based on random forests and process mechanisms strengthens the interpretability of the model's predictions, clearly identifying key controllable parameters and providing direct evidence for emission reduction diagnosis and process optimization in actual production. Furthermore, through systematic preprocessing and parameter optimization experiments, this invention clarifies the processing paradigm applicable to this type of data, simplifying the modeling process. Ultimately, this method is not only suitable for macro-level industry predictions but can also accurately depict the production process of a single paper type, demonstrating significant engineering application value and potential for widespread adoption. Attached Figure Description

[0017] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a schematic diagram illustrating the improvement of different standardization methods relative to the baseline in embodiments of the present invention; Figure 2 This is a flowchart of the random forest model according to an embodiment of the present invention; Figure 3 This is a flowchart of the stacking model according to an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the effects of different parameter combinations relative to a baseline in an embodiment of the present invention; Figure 5 This is a schematic diagram showing the ranking of factors in the production process according to an embodiment of the present invention; Figure 6 This is a schematic diagram comparing the prediction performance of two models in an embodiment of the present invention, where (a) is a schematic diagram of the prediction results of the random forest model and (b) is a schematic diagram of the prediction results of the stacked model with residual compensation. Detailed Implementation

[0018] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0020] Example 1 This embodiment provides a method for predicting the carbon footprint of the papermaking process based on two-stage machine learning, including: constructing a feature dataset covering all stages of the papermaking life cycle and performing preprocessing; The random forest algorithm is used to train the preprocessed dataset to obtain the initial carbon footprint prediction model and output the prediction results; Calculate the prediction residual of the initial carbon footprint prediction model; Based on the predicted residuals and the original feature dataset, train the XGBoost model to construct a residual prediction model; The prediction results of the initial carbon footprint prediction model are superimposed with the output of the residual prediction model to obtain the final carbon footprint prediction value.

[0021] Furthermore, a feature dataset covering all stages of the papermaking lifecycle is constructed and preprocessed, including: Integrate raw material structure, energy structure, process parameters, and waste treatment data to construct a feature matrix; One-hot encoding is performed on categorical variables, and logarithmic transformation is performed on continuous long-tailed variables; The dataset is divided into a training set and a test set, and all preprocessing operations are performed only on the training set.

[0022] Specifically, the implementation process of this embodiment includes: The dataset is derived from a lifecycle database. A feature matrix X and a target vector y are constructed using ecoinvent v3.11 and FisherSolve databases as data sources: ; In the formula: y is the CO2 equivalent intensity (kg CO2e / t paper) calculated by openLCA, and X is the feature matrix, which includes the raw material structure, energy structure, process flow, transportation and packaging intensity, and waste treatment after normalization to functional units. All raw flow rates are first unified in units and converted to functional units. For processes that do not exist, the value is retained as 0 and missing indicator variables are added. For measurement missing values, quantiles (median first) of the same process / paper type / region are used for imputation. To suppress the influence of low-impact factors, sparse processes with a total contribution of <1% are merged. Long-tail variables (such as chemical dosage and transportation intensity) are transformed using log1P; categorical variables (region, paper type, process route) are one-hot encoded.

[0023] The dataset contains 1832 samples. To ensure the rigor of the evaluation, the training and test sets are randomly divided in an 8:2 ratio. All preprocessing (standardization, encoding, and optional feature selection) is performed only on the training set. Subsequently, the test set is transformed using a uniform pipeline to avoid information leakage. This process ensures that the key stages of the entire papermaking lifecycle are fully represented in the feature space and provides stable and reproducible input for subsequent grid search and interpretable modeling.

[0024] Although random forest models are generally considered insensitive to the scale of variables and do not rely on feature standardization, this assumption does not always hold true in practical applications. On the one hand, some implementation mechanisms (such as continuous variable processing based on data partitioning thresholds) may still be affected by the feature scale distribution; on the other hand, different preprocessing methods may introduce additional biases when the model is used for tasks such as feature selection and interpretive analysis. Furthermore, for complex datasets involving multiple preprocessing steps such as encoding, missing value handling, or sample sampling, differences in standardization methods may indirectly affect model performance. Therefore, it is necessary to systematically evaluate the specific impact of different standardization strategies on model performance on a real-world papermaking carbon footprint dataset to determine the importance and necessity of the "standardization" question in this embodiment's scenario.

[0025] Therefore, to quantify the effect of preprocessing on the performance of the tree model, this embodiment, under the same 5-fold partition and the same model / hyperparameters, only changes the normalization method and uses the paired ΔRMSE of the same fold as the effect size (a negative value indicates better performance than the "unscaled" baseline, such as...). Figure 1 (Points represent the mean, and the error bar is 95% CI).

[0026] Subsequent experiments uniformly adopted "none" as the default preprocessing for three reasons: first, it maintains a natural fit with the tree model; second, it avoids introducing interference to the upstream feature selection; and third, it simplifies the process and reduces the parameter space and result variance.

[0027] Furthermore, training the preprocessed dataset using the random forest algorithm includes: Construct multiple decision trees with the constraint that the number of leaf nodes is 1 and the tree depth is not limited; When splitting at a node of each decision tree, a subset of features is randomly selected to determine the optimal splitting method; Based on the trained decision tree set, the initial prediction results are output through an ensemble method.

[0028] Specifically, the implementation process of this embodiment includes: Random Forest, a typical ensemble learning method proposed by Breiman, significantly reduces model variance and improves generalization ability by training multiple decision trees on different bootstrap sampling subsets and averaging their predictions. Compared to traditional decision trees, Random Forest randomly selects a subset of features for optimal splitting when nodes split. This "feature randomization" mechanism effectively enhances the diversity between tree models, thereby reducing the risk of overfitting. Since the training processes of each tree are independent, Random Forest also possesses good parallel computing characteristics, making it suitable for modeling high-dimensional, noisy industrial process data.

[0029] In this embodiment, random forests can simultaneously handle multi-dimensional features such as energy consumption, process parameters, and raw material structure in the papermaking process prediction. Furthermore, it has no special requirements for feature distribution, making it suitable for scenarios like the papermaking industry where there are strong nonlinear relationships and multi-source data. During model training, this embodiment optimized the performance of key hyperparameters (such as the number of trees, maximum depth, and minimum leaf node) to achieve better prediction accuracy. The model performance is demonstrated through... The flowchart for RMSE indicator evaluation is as follows: Figure 2 As shown.

[0030] While random forests offer advantages such as robustness and ease of use, they also have limitations. On one hand, the prediction mechanism based on multi-tree averaging may underperform in extreme samples or boundary conditions; on the other hand, feature importance primarily reflects the contribution of variables to node purity, making it difficult to reveal deep interactions between variables and limiting interpretability. Therefore, in this embodiment, random forests are mainly used for key feature identification and baseline prediction model construction, and in subsequent analysis, they are combined with more expressive models (such as gradient boosting models) to further optimize prediction performance.

[0031] A random forest regression model was constructed using the Scikit-learn library in Python. To systematically evaluate the impact of hyperparameters on model error, this embodiment performed a grid search on four key dimensions—maximum tree depth, sampling subset, number of leaves, and number of trees—while keeping the data partitioning and feature selection fixed. The evaluation metric was 5-fold cross-validation RMSE. The grid search results for different hyperparameters are shown in Table 1.

[0032] Table 1

[0033] To avoid random fluctuations, this embodiment pairs each set of hyperparameters with the baseline model at the same fold, calculates its ΔRMSE = RMSE(combined) - RMSE(baseline), and calculates the mean and 95% confidence interval at 5 folds. Figure 3(Negative values ​​are better). This "same-fold pairing" can significantly offset the noise brought about by data partitioning, making the differences between items more reflective of the influence of the hyperparameter itself; and whether the confidence interval crosses 0 reflects statistical significance.

[0034] Results of the paired experiment (Table 1) Figure 4 As can be seen, the model performs best when the number of leaf nodes is 1, which effectively degenerates the random forest into an ensemble model composed of a large number of weak learners (decision stubs). Its advantage lies in avoiding overfitting and enhancing the model's stability in high-dimensional sparse variable ranges. This embodiment also conducted experiments with restricted tree depth, but the results showed no significant decrease in RMSE (the maximum improvement was only 0.01), indicating that the main interpretable structure of the papermaking carbon footprint is better captured by a large number of shallow trees. This is consistent with the characteristic that some industrial process variables exhibit linear or piecewise linear relationships. Therefore, this embodiment chooses a leaf node count of 1, does not restrict tree depth, selects logarithmic transformation as the subset sampling method, and sets the number of trees to 500.

[0035] Furthermore, calculating the prediction residuals of the initial carbon footprint prediction model includes: Subtract the predicted value from the initial carbon footprint prediction model from the actual calculated carbon footprint value to obtain the original residual; Analyze the correlation between the original residuals and the predicted values ​​to confirm whether there are learnable systematic biases in the residuals.

[0036] Specifically, the implementation process of this embodiment includes: After constructing the Random Forest (RF) model and obtaining preliminary prediction results, this embodiment systematically analyzes its error characteristics to evaluate its fitting ability in predicting the carbon footprint of papermaking. The residual is defined as: ; In the formula: y represents the baseline model residuals, and y represents the calculated carbon emissions. This represents the carbon emission prediction results from the random forest model.

[0037] The calculation results show that the Pearson correlation coefficient (Corr(|ε|, The error margin is as high as 0.42, indicating that the model error is significantly affected by the predicted values ​​to some extent. Although the random forest model can fit most samples well, its residuals show systematic bias on high-emission samples, which contains structural information that the random forest model fails to capture.

[0038] Based on Friedman's phased approximation approach, if the residuals of the first-stage model are still learnable, a new learner can be used to model them to further improve the overall prediction accuracy. However, since random forests essentially reduce variance by averaging multiple trees, their ability to characterize extreme samples is limited, which may also cause strong nonlinear relationships to remain in the residuals. Therefore, this embodiment proposes using XGBoost to relearn the residuals of the random forest model, constructing a two-stage stacked model to improve the overall expressive power of the model.

[0039] Furthermore, training the XGBoost model to build a residual prediction model includes: Use the preprocessed original feature dataset as input features; The predicted residuals are used as training targets, and the XGBoost algorithm is used for fitting. The XGBoost algorithm is guided by the residuals of the previous fitting and gradually approximates the objective function through an additive tree model.

[0040] Specifically, the implementation process of this embodiment includes: In the second stage, the XGBoost model takes the original feature matrix X as input and uses the random forest residuals... As a new learning objective, a residual learning function is constructed: ; In the formula: These are the predicted residual values. is the XGBoost learner, where X is the feature matrix of the dataset.

[0041] Because XGBoost approximates complex functions progressively through an additive tree model and focuses on learning hard-to-fit regions in each round using residuals as a guide, it is particularly suitable for capturing nonlinear structures and local mutation patterns in RF residuals. The final prediction form of the two-stage model is: ; The first item Provides overall trends, the second item This structure is used to compensate for systematic errors in RF models in high-emission regions. It avoids the overfitting risk that may occur when directly training a single complex model, while leveraging XGBoost's strong learning capabilities on complex residual structures to effectively improve the model's prediction accuracy under extreme conditions.

[0042] Furthermore, after training using the random forest algorithm, the process further includes: Based on the results of the random forest algorithm, a permutation importance analysis is performed. Based on the importance of each feature, the key process parameters affecting carbon footprint are identified.

[0043] Specifically, the implementation process of this embodiment includes: Analysis of existing research indicates that steam consumption, energy structure, and process technology are generally considered the main sources of carbon footprint in papermaking. To further clarify the relative importance of these factors from a data perspective, this embodiment constructs a random forest model based on the dataset and ranks the influence of each parameter by permutation importance (see...). Figure 5 ).

[0044] The results show that: Among all process variables, the dominant ones are "drying section steam intensity" and "grid emission factor," both with an importance exceeding 1.9. The drying section is the part of the papermaking process with the highest heat consumption, while the "grid emission factor" directly determines the carbon emission level under different energy structures. Therefore, it is reasonable that these two are the most important process factors affecting the carbon footprint.

[0045] The importance of energy-related variables such as "fossil fuel percentage", "electricity consumption per unit product", and "CO2 emissions from lime kiln calcination" is also quite prominent. Together, they reflect the core impact of the plant's energy system on carbon emissions: the higher the use of fossil fuels and the greater the electricity consumption per unit product, the higher the corresponding carbon footprint level; while lime kilns are a typical unavoidable carbon source in the process, and their emissions show a relatively independent and stable contribution in the model.

[0046] Besides factors directly related to energy, variables such as "recycled fiber percentage," "pulping route," "dryness after pressing," and "in-plant self-generation / purchased electricity ratio" are also highly important. These factors have clear correspondences with the "process flow," "raw material structure," and "energy structure" components of the life cycle. For example, the recycled fiber percentage affects the energy consumption of the upstream pulping stage; the dryness after pressing determines the downstream steam demand and is a key factor influencing the drying load; while the self-generation ratio reflects changes in electricity sources and complements the trend of grid emission factors.

[0047] Many variables (such as chlorine dioxide dosage, hydrogen peroxide dosage, and ozone bleaching intensity) are mostly related to chemical dosing, material handling, and auxiliary system operation. Although their direct contribution to the carbon footprint is not as great as the aforementioned energy-related variables, the model results show that these factors still have a certain impact on emissions, especially the significant correlation between chemical dosing and wastewater treatment, which may have an indirect impact on carbon emissions under certain operating conditions.

[0048] The variables ranked lower have a relatively weaker impact, indicating that the end-of-pipe treatment stage accounts for a limited proportion of the company's overall emissions, which is consistent with the structure of "lower contribution of end-of-pipe emissions" in the life cycle.

[0049] Overall, the importance ranking of random forest echoes the life cycle analysis in the previous section: the most critical factors in the life cycle assessment are energy-related factors, which also show the highest importance in the model; while pulping conditions, recycled fiber ratio, and auxiliary system operation form the second tier of influence structure. Through this model screening, this embodiment further clarifies the key controllable parameters affecting the carbon footprint of papermaking, providing direct data basis for subsequent predictive model construction and process emission reduction optimization.

[0050] Furthermore, after obtaining the final carbon footprint prediction, the process further includes: The performance of the final carbon footprint prediction was evaluated on an independent test set using the coefficient of determination and root mean square error metrics. The final carbon footprint prediction is then applied to the carbon emission trend analysis of a specific type of paper.

[0051] Specifically, the implementation process of this embodiment includes: like Figure 6 The experimental results shown are illustrated in (a) and (b) respectively. (a) is a schematic diagram of the prediction results from the random forest model, and (b) is a schematic diagram of the prediction results from the stacked model with residual compensation. It can be seen that the random forest model is able to capture the overall trend of the carbon footprint of papermaking quite well, and its predicted values ​​have a high degree of consistency with the calculated values. =0.921, RMSE = 4.498). However, in the high carbon emission range (such as 1700–2000 kgCO2eq / t), random forests still exhibit a certain degree of systematic bias, with some samples showing predicted values ​​that are much lower than the calculated values. This result is consistent with the previously mentioned excessively high Pearson correlation coefficient between predicted values ​​and absolute residual values.

[0052] In contrast, introducing the XGBoost model to address the two-stage stacked model after residual relearning significantly improves this bias, increasing its coefficient of determination to [missing value]. =0.947, RMSE decreased to 3.68, and overall prediction accuracy was significantly improved. The scatter plot shows that the stacked model's points in the high-emission region are closer to the baseline, indicating that the second-stage XGBoost effectively captured the residual structure that the random forest failed to fit, especially in high-emission samples, where this improvement was more pronounced.

[0053] As shown in Table 2, compared to RF, the stacking scheme reduces RMSE by 18.19%; compared to XGB alone, the RMSE is reduced by 19.72%. Unexplained variance (1- The unexplained variance decreased from 0.0792 to 0.0531, representing a reduction of approximately 32.95%. This performance improvement is not a simple averaging effect, but rather stems from the targeted correction of model bias structure by the residual relearning method. The first-stage random forest model captures stable trends in the dataset, while the second-stage model utilizes the strong nonlinear expressive power of XGBoost to specifically correct systematic errors in local regions. This results in a final prediction with not only higher overall fit but also more reliable performance on extreme samples. This result verifies that the proposed "random forest + XGBoost residual relearning" stacked strategy is highly effective for predicting the carbon footprint of complex industrial processes.

[0054] Table 2

[0055] Unlike traditional single-stage machine learning models, this embodiment proposes a residual decomposition-based two-stage learning framework for predicting industrial carbon footprint. The method first utilizes random forests to capture the dominant structural relationships of carbon emissions from the papermaking process and then uses residual diagnostics to identify systematic biases in high-emission ranges. To address this deficiency, this embodiment proposes using the residuals as the learning objective in the second stage, introducing XGBoost to construct a structured residual learner to compensate for nonlinear interactions that the baseline random forest fails to characterize.

[0056] Unlike direct stacking models or simple weighted models, the improvements proposed in this embodiment focus on: determining whether nonlinear coupling is learnable through residual structure diagnosis; constructing a bias-oriented correction module to achieve targeted fitting for extreme high-emission samples; and maintaining the interpretability of a single model while significantly improving overall prediction accuracy.

[0057] To verify the applicability of the model to specific paper types, this embodiment selected 126 samples belonging to cultural paper from the dataset for subset prediction analysis. The results show that the two-stage stacking model maintains stable prediction performance on this paper type, with a determination coefficient R0. 2 The value reached 0.953, and the RMSE was 3.21 kgCO2eq / t.

[0058] From the perspective of characteristic contributions, the carbon emissions of this paper type are mainly affected by drying steam intensity, grid emission factor, and the proportion of fossil fuels, and are related to... Figure 5The order of importance is consistent. Notably, when the proportion of fossil fuels in cultural paper production increases from 40% to 70%, the model predicts an increase in carbon emissions of approximately 140 kg CO2eq / t; while increasing the post-press dryness from 43% to 46% can reduce the carbon footprint by approximately 50 kg CO2eq / t.

[0059] This case demonstrates that the proposed residual decomposition two-stage model can not only achieve high-precision predictions at the industry-wide level, but also accurately characterize carbon emission trends under the operating conditions of a single paper type, proving its engineering application value at the factory and paper type levels.

[0060] Example 2 In this embodiment, a computer terminal device is provided, including: One or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the above-described method for predicting the carbon footprint of the papermaking process based on two-stage machine learning.

[0061] In this embodiment, a computer-readable storage medium is also provided, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the above-described method for predicting the carbon footprint of the papermaking process based on two-stage machine learning.

[0062] This invention provides a two-stage machine learning-based method for predicting the carbon footprint of papermaking processes, effectively improving the overall accuracy of carbon footprint prediction, particularly in high-emission ranges. By introducing a residual relearning mechanism, this invention enhances the modeling ability for complex nonlinear relationships and extreme operating conditions, improving the model's robustness and stability. Simultaneously, the deep integration of feature importance analysis based on random forests and process mechanisms strengthens the interpretability of the model's predictions, clearly identifying key controllable parameters and providing direct evidence for emission reduction diagnosis and process optimization in actual production. Furthermore, through systematic preprocessing and parameter optimization experiments, this invention clarifies the processing paradigm applicable to this type of data, simplifying the modeling process. Ultimately, this method is not only suitable for macro-level industry predictions but can also accurately depict the production process of a single paper type, demonstrating significant engineering application value and potential for widespread adoption.

[0063] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting the carbon footprint of a papermaking process based on two-stage machine learning, characterized in that, Includes the following steps: Construct a feature dataset covering all stages of the papermaking lifecycle and perform preprocessing; The random forest algorithm is used to train the preprocessed dataset to obtain the initial carbon footprint prediction model and output the prediction results; Calculate the prediction residual of the initial carbon footprint prediction model; Based on the predicted residuals and the original feature dataset, train the XGBoost model to construct a residual prediction model; The prediction results of the initial carbon footprint prediction model are superimposed with the output of the residual prediction model to obtain the final carbon footprint prediction value.

2. The method according to claim 1, characterized in that, Construct a feature dataset covering all stages of the papermaking lifecycle and perform preprocessing, including: Integrate raw material structure, energy structure, process parameters, and waste treatment data to construct a feature matrix; One-hot encoding is performed on categorical variables, and logarithmic transformation is performed on continuous long-tailed variables; The dataset is divided into a training set and a test set, and all preprocessing operations are performed only on the training set.

3. The method according to claim 1, characterized in that, Training the preprocessed dataset using the random forest algorithm includes: Construct multiple decision trees with the constraint that the number of leaf nodes is 1 and the tree depth is not limited; When splitting at a node of each decision tree, a subset of features is randomly selected to determine the optimal splitting method; Based on the trained decision tree set, the initial prediction results are output through an ensemble method.

4. The method according to claim 1, characterized in that, Calculating the prediction residuals of the initial carbon footprint prediction model includes: Subtract the predicted value from the initial carbon footprint prediction model from the actual calculated value of the carbon footprint to obtain the original residual; Analyze the correlation between the original residuals and the predicted values ​​to confirm whether there are learnable systematic biases in the residuals.

5. The method according to claim 1, characterized in that, Training an XGBoost model to build a residual prediction model includes: Use the preprocessed original feature dataset as input features; The predicted residuals are used as training targets, and the XGBoost algorithm is used for fitting. The XGBoost algorithm is guided by the residuals of the previous fitting and gradually approximates the objective function through an additive tree model.

6. The method according to claim 1, characterized in that, After training using the random forest algorithm, the following further steps are included: Based on the results of the random forest algorithm, a permutation importance analysis is performed. Based on the importance of each feature, the key process parameters affecting carbon footprint are identified.

7. The method according to claim 1, characterized in that, After obtaining the final carbon footprint prediction, the following further steps are included: The performance of the final carbon footprint prediction was evaluated on an independent test set using the coefficient of determination and root mean square error metrics. The final carbon footprint prediction is then applied to the carbon emission trend analysis of a specific type of paper.

8. The method according to claim 7, characterized in that, The specific type of paper mentioned is cultural paper.

9. A computer terminal device, characterized in that, include: One or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors perform the steps of the method as described in any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-8.

Images