Process optimization method for preparing light aromatic hydrocarbons from waste plastics by pyrolysis based on machine learning

Abstract

This invention belongs to the field of waste plastic pyrolysis resource utilization, specifically a machine learning-based process optimization method for preparing light aromatics from waste plastic pyrolysis. This method constructs a dataset based on raw material characteristics, pyrolysis conditions, and product data from published literature, performs preprocessing, trains and optimizes various machine learning regression models, and selects the optimal model for predicting light aromatics yield based on evaluation indicators. Then, interpretability analysis is used to identify key factors affecting the yield, and raw material structural characteristics such as saturated C-H bonds, methyl branches, and carbonyl groups are used as inputs to enhance the characterization of raw material differences. Under industrial constraints, with the light aromatics yield as the optimization objective, a particle swarm optimization algorithm is used to globally optimize the reaction condition parameters to obtain the optimal combination of operating parameters. This invention can achieve light aromatics yield prediction and operating parameter recommendation under fluctuating raw material composition, improving prediction stability and generalization ability, reducing experimental iterations, and lowering optimization costs.

Description

Technical Field

[0001] This invention belongs to the field of waste plastic pyrolysis resource utilization, specifically involving a process optimization method for preparing light aromatic hydrocarbons from waste plastic pyrolysis based on machine learning. Background Technology

[0002] With the continuous growth of plastic consumption, the demand for the reduction and resource-based disposal of waste plastics is becoming increasingly prominent. Pyrolysis, as a thermochemical conversion technology, can convert waste plastics into light aromatic hydrocarbons such as benzene and toluene. However, the pyrolysis process of waste plastics is significantly affected by differences in raw materials and reaction conditions, usually requiring a large number of experiments for parameter screening and combination optimization, resulting in long research and development cycles, high costs, and difficulty in reusing optimization experience across raw materials and operating conditions. To reduce the number of experiments and improve optimization efficiency, some studies have attempted to use machine learning to establish predictive models of "raw material characteristics - process parameters - target product indicators" to achieve rapid screening and optimization decisions for parameter combinations.

[0003] However, existing machine learning solutions still have limitations in practical applications: First, they often use single models such as random forests and gradient boosting trees for fitting, which can lead to fluctuations in prediction results and require improved stability when the data distribution varies significantly or when there are a few extreme working conditions. Second, existing models for the pyrolysis of mixed waste plastics often rely on the content of each plastic component as input features, resulting in a relatively singular dimension of raw material representation. This makes it difficult to fully reflect the differences in reaction and synergistic / inhibitory effects of different plastic components during pyrolysis, limiting the model's ability to describe mixed raw materials and thus restricting its generalization ability. Third, the sources of real waste plastics are complex, and the composition of raw materials is difficult to accurately identify and quantitatively characterize, thus limiting the applicability of modeling and optimization methods that rely on component content. Fourth, existing machine learning methods often focus on yield prediction and identification of key influencing factors, lacking reverse optimization methods for process parameters oriented towards target yield, making it difficult to achieve synergistic optimization of parameter combinations under multivariable and multi-constraint conditions.

[0004] Based on the above problems, it is necessary to propose a machine learning optimization method for the preparation of light aromatics from waste plastic pyrolysis that can adapt to fluctuations in incoming materials and improve prediction stability and generalization ability. Summary of the Invention

[0005] In view of this, the present invention provides a process optimization method for the preparation of light aromatics from waste plastic pyrolysis based on machine learning, comprising the following steps:

[0006] S1. Obtain experimental data on the pyrolysis of waste plastics, construct a dataset containing input features and target product information, and preprocess the dataset;

[0007] S2. Divide the dataset into training set and test set, train multiple machine learning regression models based on the training set, and perform weighted fusion of multiple machine learning regression models to construct a corresponding hybrid regression model. Optimize the hyperparameters of the hybrid regression model.

[0008] S3. Substitute the test set into the multiple mixed regression models obtained in step S2 for evaluation, and select the model with the best predictive performance based on the evaluation index.

[0009] S4. Perform interpretability analysis on the optimal model to obtain the key influencing parameters for regulating the yield of light aromatics;

[0010] S5. Based on the above-obtained influence parameters, the actual waste plastic characteristic parameters are used as constraints, and the reaction condition parameters are used as process optimization variables. Within the industrial feasible domain jointly defined by the two, global optimization is performed with the goal of maximizing the yield of light aromatics, thereby obtaining the optimal combination of operating parameters that both meets the requirements of industrial implementation and is conducive to improving the yield of light aromatics.

[0011] Furthermore, waste plastics include one or more of the following four types of plastics: PP, PE, PET, and PS.

[0012] Furthermore, in step S1, the input features include waste plastic characteristic parameters and reaction condition parameters;

[0013] The waste plastic characteristic parameters include at least structural characteristic parameters, which include the relative contents of saturated CH bonds, methyl branches, carbonyl groups, and aromatic CH bonds;

[0014] The waste plastic characteristic parameters also include raw material particle size, the proportion of each type of plastic component, and elemental composition parameters; the elemental composition parameters are the mass fractions of carbon, hydrogen, and oxygen elements.

[0015] The reaction operating parameters include residence time, pyrolysis temperature, and heating rate;

[0016] The target product information is the actual yield of light aromatic hydrocarbons obtained from a publicly available database.

[0017] Furthermore, in step S2, the dataset is randomly divided into a training set and a test set in an 8:2 ratio.

[0018] Furthermore, in step S2, the hybrid regression model is formed by linearly weighting two different regression models; the regression model is selected from one of the following: Random Forest Regression Model (RF), Limit Gradient Boosting Tree Model (XGB), and Support Vector Regression Model (SVR).

[0019] Furthermore, the hybrid regression model includes at least one of RF-XGB, RF-SVR, or XGB-SVR, expressed as:

[0020]

[0021] Where i and j represent any two different regression models among RF, XGB, and SVR, and i ≠ j; α is the optimal weight learned from the training data, and 0 ≤ α ≤ 1, which is optimized on the training set through five-fold cross-validation to ensure that the R-value of the hybrid regression model on the test set is optimal. 2 Maximum and minimum RMSE; Y i and Y j These are the light aromatics yield values ​​predicted by the corresponding regression models, Y and Y. i,j The yield value of light aromatics is predicted by the mixed regression model formed by linear weighting of the two regression models.

[0022] Furthermore, in step S2, the training process employs five-fold cross-validation for training and hyperparameter optimization to improve the model's generalization ability.

[0023] Furthermore, in step S3, the evaluation index is determined by the coefficient of determination R. 2 The root mean square error (RMSE) and mean absolute error (MAE) are used together for evaluation, and the optimal model is determined by the comprehensive evaluation results.

[0024] Furthermore, in step S4, the interpretability analysis includes feature importance analysis and cross-validation in conjunction with SHAP analysis.

[0025] Furthermore, in step S5, the heuristic optimization algorithm is the Particle Swarm Optimization (PSO) algorithm.

[0026] The beneficial effects of this invention are:

[0027] 1) This invention constructs a machine learning prediction model based on waste plastic characteristic parameters, pyrolysis process parameters and light aromatic hydrocarbon yield information in published literature, realizes light aromatic hydrocarbon yield prediction and reverse optimization of process parameters, can automatically complete model screening and parameter recommendation, reduce the number of experimental iterations, shorten the R&D cycle and reduce optimization costs, and is suitable for mixed waste plastic pyrolysis systems under fluctuating raw material composition conditions.

[0028] 2) This invention employs a multi-model construction and combination screening mechanism, selecting R from a variety of candidate learners and their combinations. 2 RMSE and MAE are used as quantitative evaluation indicators to uniformly evaluate and comprehensively rank the prediction effects of each model, thereby selecting the prediction model with the best overall performance, improving the adaptability to changes in data distribution and extreme working conditions, enhancing prediction stability and generalization ability, and providing reliable support for process optimization.

[0029] 3) This invention breaks through the limitation of using only the proportion of each component of waste plastic as a raw material characterization method. It introduces characteristic parameters reflecting the differences in molecular structure of raw materials, such as saturated CH bonds, methyl branches, and carbonyl groups, into the input feature system, realizing the expansion from "component ratio characterization" to "structural feature characterization". This allows the model input features to be more directly associated with key reactions such as main chain breakage, depolymerization, and aromatization during pyrolysis, thereby improving the model's ability to learn the pyrolysis law of mixed waste plastics and improving the accuracy and stability of target product yield prediction.

[0030] 4) This invention introduces the PSO algorithm to perform global optimization of process parameters. It utilizes a group parallel search and iterative update mechanism to efficiently search for optimal combination of operating parameters in the parameter space, thereby obtaining process conditions that are conducive to improving the yield of light aromatics, improving process optimization efficiency and reducing the experimental cost of parameter optimization. Attached Figure Description

[0031] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a schematic diagram of the overall method flow of the present invention. Detailed Implementation

[0034] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of systems consistent with some aspects of the invention as detailed in the appended claims.

[0035] refer to Figure 1 This invention provides a process optimization method for preparing light aromatics from waste plastic pyrolysis based on machine learning, comprising the following steps:

[0036] S1. Obtain experimental data on the pyrolysis of waste plastics, construct a dataset containing input features and target product information, and preprocess the dataset;

[0037] S2. Divide the dataset into training set and test set, train multiple machine learning regression models based on the training set, and perform weighted fusion of multiple machine learning regression models to construct multiple hybrid regression models, and optimize the hyperparameters of the hybrid regression models.

[0038] S3. Substitute the test set into the multiple mixed regression models obtained in step S2 for evaluation, and select the model with the best predictive performance based on the evaluation index.

[0039] S4. Perform interpretability analysis on the optimal model to obtain the key influencing parameters for regulating the yield of light aromatics;

[0040] S5. Based on the above-obtained influence parameters, the actual waste plastic characteristic parameters are used as constraints, and the reaction condition parameters are used as process optimization variables. Within the industrial feasible domain jointly defined by the two, global optimization is performed with the goal of maximizing the yield of light aromatics, thereby obtaining the optimal combination of operating parameters that both meets the requirements of industrial implementation and is conducive to improving the yield of light aromatics.

[0041] The specifics are as follows:

[0042] Step S1: Dataset Construction and Preprocessing

[0043] This example collected 254 sets of data from experimental data on the pyrolysis of waste plastics, based on publicly available literature.

[0044] The input features include:

[0045] 1) Waste plastic characteristic parameters: raw material particle size, elemental composition parameters (mass fraction of carbon, hydrogen and oxygen elements) and structural characteristic parameters (relative content of saturated CH bonds, methyl branches, carbonyl groups and aromatic CH bonds);

[0046] 2) Reaction operating parameters: residence time, pyrolysis temperature, and heating rate.

[0047] The target product information is the yield of light aromatic hydrocarbons.

[0048] Preprocess the dataset, including outlier removal and missing value imputation, to improve data consistency and usability.

[0049] Step S2: Model Building and Hyperparameter Optimization

[0050] The 254 sets of data were randomly divided into a training set and a test set in an 8:2 ratio, with 203 sets in the training set and 51 sets in the test set.

[0051] Based on the training set data, RF, XGB, and SVR regression models were constructed respectively, and further hybrid regression models RF-XGB, RF-SVR, and XGB-SVR were constructed; wherein, the hybrid regression model is a linear weighted combination model of the predicted values ​​of the corresponding two regression models. This can be expressed as:

[0052] ;

[0053] Where i and j represent any two different regression models among RF, XGB, and SVR, and i ≠ j; α is the optimal weight learned from the training data, and 0 ≤ α ≤ 1, which is optimized on the training set through five-fold cross-validation to ensure that the R-value of the hybrid regression model on the test set is optimal. 2 Maximum and minimum RMSE; Y i and Y j These are the light aromatics yield values ​​predicted by the corresponding regression models, Y and Y. i,j The yield value of light aromatics is predicted by the mixed regression model formed by linear weighting of the two regression models.

[0054] The hyperparameters of the three models are optimized by combining grid search with five-fold cross-validation to improve the model's prediction accuracy and generalization ability. Model training and optimization can be implemented in common programming environments and machine learning toolkits.

[0055] Among them, RF is an ensemble learning regression algorithm, which constructs multiple decision trees and integrates their outputs to obtain prediction results. It has good robustness and is suitable for multi-feature input data; XGB is a gradient boosting tree regression algorithm, which integrates multiple weak learners in an iterative manner to improve regression performance. It is suitable for handling nonlinear relationships and has high computational efficiency; SVR is a support vector regression algorithm, which models nonlinear regression problems through kernel function mapping and interval optimization. It is suitable for regression prediction of small to medium-sized datasets.

[0056] Step S3: Model Evaluation and Optimal Model Determination

[0057] Substituting the test set into each optimized model yielded predicted yields of light aromatics, and Ri was used. 2 RMSE and MAE were used as evaluation indicators. Based on a comprehensive comparison of the evaluation indicators, the model with the best predictive performance was determined. In this embodiment, the prediction performance of the mixed regression model was better than that of the single regression model, and the RF-XGB model performed best, as shown in Table 1.

[0058] Table 1. R-values ​​of different models on the test set 2 RMSE and MAE

[0059]

[0060] Step S4: Interpretability Analysis

[0061] Eigenvalue importance analysis and SHAP analysis were used to perform interpretability analysis on the RF-XGB mixture model, identifying key influencing factors on the yield of light aromatics. The results show that:

[0062] 1) Based on the characteristic importance analysis, pyrolysis temperature, relative content of aromatic CH bonds, and mass fraction of carbon elements ranked as the top three important influencing factors.

[0063] 2) According to SHAP analysis, the pyrolysis temperature (SHAP mean = 0.45), the relative content of aromatic CH bonds (SHAP mean = 0.41), and the carbon mass fraction (SHAP mean = 0.33) are the top three key factors.

[0064] 3) PDP analysis showed that the pyrolysis temperature and yield were significantly positively correlated nonlinearly. The yield increased the fastest in the range of 480-550°C and then slowed down after about 525°C. The relative content of aromatic CH bonds had a significant impact on the yield in the range of 24-52%. A higher yield could be obtained when the carbon mass fraction exceeded 60wt%.

[0065] Step S5: Global Optimization under Constraints

[0066] Based on the key influencing parameters (i.e., factors) identified in step S4, with the objective function of maximizing the predicted yield of light aromatics, and setting the waste plastic characteristic parameters among the key influencing factors as constraints, and using the reaction operating parameters as process optimization variables, the PSO algorithm is used to globally optimize the reaction operating parameters within the industrial feasible region jointly defined by both, with the objective of maximizing the yield of light aromatics. It should be noted that the "global search for the influencing factors in step S4" mentioned in step S5 refers to selecting the factors that rank highly and are measurable in the model interpretability analysis as optimization variables or constraints for search optimization and feasible region limitation, and these factors are also constrained by the actual pyrolysis unit.

[0067] In this embodiment, the following constraints are set for the structural characteristics of the raw materials: the relative content of aromatic CH bonds is 28-42%, the relative content of carbonyl groups is 0-5%, and the relative content of methyl branches is 10-25%. Within the above feasible region, the PSO algorithm is used for global optimization, and the optimal combination of pyrolysis parameters predicted by the model is finally obtained as follows: pyrolysis temperature 525℃, heating rate 20℃ / min, and residence time 45min.

[0068] Comparative Example

[0069] The difference between this comparative example and the above embodiment is that in step S1, the waste plastic characteristic parameters are raw material particle size, component ratio, and elemental composition parameters (mass fraction of carbon, hydrogen, and oxygen elements); other steps are the same as in the embodiment.

[0070] In this comparative study, the optimal model remains the RF-XGB mixed regression model, with the model predicting R0. 2 The accuracy was 0.914, and the RMSE and MAE were 4.17 and 2.62, respectively, indicating that the model still has a certain predictive ability; however, its prediction accuracy was lower than that of the example, indicating that the introduction of structural feature parameters such as saturated CH bonds, methyl branches, and carbonyl groups helps to more fully characterize the differences in raw materials, thereby enhancing the nonlinear mapping ability between raw material characteristics and light aromatics yield and improving prediction performance.

[0071] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these changes and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A process optimization method for the preparation of light aromatics from waste plastic pyrolysis based on machine learning, characterized in that, Includes the following steps: S1. Obtain experimental data on the pyrolysis of waste plastics, construct a dataset containing input features and target product information, and preprocess the dataset; the input features include waste plastic characteristic parameters and reaction condition parameters; S2. Divide the dataset into training set and test set, train multiple machine learning regression models based on the training set, and perform weighted fusion of multiple machine learning regression models to construct a hybrid regression model, and optimize the hyperparameters of the hybrid regression model. S3. Substitute the test set into the multiple mixed regression models obtained in step S2 for evaluation, and select the model with the best predictive performance based on the evaluation index. S4. Perform interpretability analysis on the optimal model to obtain the parameters affecting the yield of light aromatics; S5. Based on the influence parameters obtained above, the actual waste plastic characteristic parameters are used as constraints, and the reaction operating condition parameters are used as process optimization variables. Within the industrial feasible domain jointly defined by the two, global optimization is carried out with the goal of maximizing the yield of light aromatics to determine the optimal combination of operating condition parameters and achieve process optimization.

2. The method as described in claim 1, characterized in that, Waste plastics include one or more of the following four types of plastics: PP, PE, PET, and PS.

3. The method as described in claim 1, characterized in that, In step S1, the waste plastic characteristic parameters include at least structural characteristic parameters, which include the relative contents of saturated CH bonds, methyl branches, carbonyl groups, and aromatic CH bonds; The waste plastic characteristic parameters also include raw material particle size, the proportion of each type of plastic component, and elemental composition parameters; the elemental composition parameters are the mass fractions of carbon, hydrogen, and oxygen elements. The reaction operating parameters include residence time, pyrolysis temperature, and heating rate; The target product information is the actual yield of light aromatic hydrocarbons obtained from a publicly available database.

4. The method as described in claim 1, characterized in that, In step S2, the dataset is randomly divided into a training set and a test set in an 8:2 ratio.

5. The method as described in claim 1, characterized in that, In step S2, the hybrid regression model is formed by linearly weighting two different regression models; the regression model is selected from one of the following: Random Forest Regression Model (RF), Limit Gradient Boosting Tree Model (XGB), and Support Vector Regression Model (SVR).

6. The method as described in claim 5, characterized in that, The hybrid regression model includes at least one of RF-XGB, RF-SVR, or XGB-SVR, expressed as: ； where i and j represent any two different regression models among RF, XGB and SVR, and i≠j; a is the optimal weight learned from training data, and 0≤a≤1, which is optimized and determined on training set by five-fold cross-validation to make the hybrid regression model have the maximum R 2 and the minimum RMSE on test set; Y i and Y j are the predicted light aromatic yield values of the corresponding regression model, Y i,j is the light aromatic yield value predicted by the hybrid regression model formed by linear weighting of the two regression models.

7. The method as described in claim 1, characterized in that, In step S2, the training process uses five-fold cross-validation for training and hyperparameter optimization to improve the model's generalization ability.

8. The method as described in claim 1, characterized in that, In step S3, the evaluation index is evaluated by the coefficient of determination R 2 , the root mean square error RMSE, and the mean absolute error MAE, and the optimal model is determined by the comprehensive evaluation result.

9. The method as described in claim 1, characterized in that, In step S4, the interpretability analysis includes feature importance analysis and cross-validation in conjunction with SHAP analysis.

10. The method as described in claim 1, characterized in that... In step S5, the heuristic optimization algorithm is the particle swarm optimization algorithm (PSO).

Images