Machine learning method and system for assisting in predicting hydrogen yield from pyrolysis of recalcitrant waste

Abstract

The application discloses a machine learning method and system for assisting in predicting hydrogen yield of pyrolysis of difficult biodegradable waste, wherein the preface part comprises the steps of data set construction, data preprocessing, data set division, model training optimization, feature influence analysis and yield prediction; and the machine learning method comprises the following working steps: step S1, multi-source experimental data set directional construction: searching the Web of Science core database to screen published literatures on hydrogen production by pyrolysis of difficult biodegradable waste, extracting 269 groups of standardized experimental records to construct a hydrogen yield special data set; setting 12-dimensional input features and 1-dimensional output labels, wherein the input features include 12 features such as element components C%, H%, N%, O%, S% and the like. The application can quickly predict the pyrolysis hydrogen yield based on the element composition, industrial analysis parameters and operation condition parameters of the sample, without the need of a large number of pyrolysis experiments, so that the labor and material costs are significantly reduced, and the application has the technical advantages of high efficiency, low consumption and generalization.

Description

Technical Field

[0001] This invention relates to the field of recalcitrant waste technology, and in particular to a machine learning method and system for assisting in predicting the yield of hydrogen gas from the pyrolysis of recalcitrant waste. Background Technology

[0002] Recalcitrant wastes, due to their stable structure and difficult conversion, have become a key challenge in the resource utilization of solid waste. Pyrolysis can convert recalcitrant wastes into gaseous, liquid, and solid products under anaerobic conditions, providing a highly efficient and environmentally friendly thermochemical conversion method. Among these methods, high-value utilization pathways, such as hydrogen production from pyrolysis, are considered an important direction for achieving clean energy substitution. However, the composition of complex recalcitrant wastes varies significantly, and their pyrolysis process is influenced by multiple factors, including operating conditions and industrial analytical parameters, resulting in complex product formation patterns. In particular, hydrogen yield is significantly affected by the nonlinear effects of multiple parameters. Traditional mechanistic models and experimental optimization methods are insufficient to effectively reveal the complex coupling relationships between variables and to achieve rapid response and quantitative prediction under process conditions. Currently, this field still lacks general methods and tools capable of accurately modeling and efficiently optimizing the prediction of hydrogen yield from pyrolysis under multi-factor coupling conditions.

[0003] Machine learning (ML) provides a new technical approach for modeling and optimizing pyrolysis systems. This method can leverage multi-source data learning and nonlinear regression algorithms to uncover the implicit correlations between elemental composition, industrial analysis parameters, operating conditions, and pyrolysis products, achieving efficient prediction and optimization of hydrogen yield. However, different machine learning models exhibit significant differences in prediction accuracy and stability, with performance influenced by factors such as dataset quality, feature dimensionality, and hyperparameter settings. To address this issue, this invention conducted statistical and correlation analyses on the pyrolysis dataset, optimized feature selection and dataset partitioning ratios, and compared the performance of various models and optimization algorithms. Considering prediction accuracy, generalization ability, and interpretability, the Bayesian Optimization-Gradient Boosting (BGGB) model was ultimately selected as the optimal solution. This model accurately captures the nonlinear relationship between elemental composition, industrial analysis parameters, and operating conditions on hydrogen yield, and achieves interpretable predictions through methods such as SHAP and Partial Dependency Graph (PDP), thereby significantly improving the accuracy and reliability of pyrolysis hydrogen prediction. Therefore, there is a great need for a machine learning method and system to assist in predicting the yield of hydrogen from the pyrolysis of recalcitrant waste, so as to achieve high-precision prediction and collaborative optimization of the yield of hydrogen from pyrolysis in complex systems and improve the yield and selectivity of hydrogen.

[0004] Based on the above, we have designed a machine learning method and system to assist in predicting the hydrogen yield from the pyrolysis of recalcitrant waste to address the aforementioned problems. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a machine learning method and system for assisting in the prediction of hydrogen yield from the pyrolysis of recalcitrant waste.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A machine learning method and system for assisting in predicting the hydrogen yield from the pyrolysis of recalcitrant waste includes the following preliminary steps: dataset construction, data preprocessing, dataset partitioning, model training and optimization, feature influence analysis, and yield prediction. The system comprises the following steps: Step S1: Targeted Construction of Multi-Source Experimental Dataset: Search the Web of Science core database to screen publicly available literature on hydrogen production from the pyrolysis of recalcitrant waste, extract 269 sets of standardized experimental records to construct a dedicated hydrogen yield dataset; set 12-dimensional input features and 1-dimensional output labels. The input features include elemental composition (C%, H%, N%, O%, S), industrial analysis indicators (volatile matter, fixed carbon, moisture, ash), and pyrolysis operation parameters (pyrolysis temperature, heating rate, pyrolysis time); the output label is hydrogen yield. Step S2, Multi-method coupled data preprocessing: Pearson correlation matrix is ​​used to complete quantitative analysis of feature correlation; dual imputation mechanism is used to handle missing values: mean imputation is used to fill mildly missing samples, and random forest regression model is used to complete highly missing features; outlier identification and removal are completed based on the 3σ criterion to eliminate the interference of extreme experimental data on model fitting; Step S3: Multi-ratio cross-validation to select the optimal dataset partitioning scheme: Set the training set / test set partitioning ratio to 0.6 / 0.4, 0.7 / 0.3, 0.8 / 0.2, and 0.9 / 0.1, and perform 5 random repetitions for each ratio; train all 12 types of machine learning regression models on each ratio dataset respectively, using R... 2 The optimal division ratio was selected based on a comprehensive evaluation of three indicators: RMSE, MAE, and RMSE. The 12 types of regression models include linear regression, ridge regression, Lasso regression, SVM, decision tree, random forest (RF), gradient boosting (GB), XGBoost, LightGBM, AdaBoost, KNN, and artificial neural network (ANN). Step S4: Triple Hyperparameter Optimization Comparison to Identify the Optimal Bayesian Optimization-Gradient Boosting (BO-GB) Prediction Model: The dataset is split into an 8:2 ratio, and five-fold cross-validation is used. RF and GB models are trained using three hyperparameter tuning strategies: grid search (GS), random search (RS), and Bayesian optimization (BO). The training and test sets (R and Gradient) are compared.2 RMSE and MAE were used to select the BO-GB combination with the lowest generalization error and the best anti-overfitting ability as the core prediction model; the performance index R of the BO-GB model on the test set was... 2 =0.780, RMSE=6.96, MAE=4.67; Step S5, Multi-level interpretability coupling analysis of feature contribution: Simultaneously perform global feature importance ranking, SHAP global contribution quantification, partial dependency graph PDP single feature response analysis, and SHAP interaction effect multi-feature coupling analysis on the BO-GB model; quantitatively rank the weights of key features affecting hydrogen yield, with the order of importance being pyrolysis temperature > heating rate > H / C hydrogen-carbon ratio > pyrolysis time > moisture > ash content, clarifying the law of multi-parameter nonlinear coupling effect; Step S6, Real-time Yield Prediction Output: Input the 12 characteristic parameters of the recalcitrant waste to be tested into the trained BO-GB model, and directly output the predicted value of pyrolysis hydrogen yield under the corresponding working conditions.

[0007] Preferably, in step S2, missing value repair adopts a hierarchical judgment logic: if the missing rate of a single feature is less than 10%, mean imputation is used; if the missing rate of a single feature is ≥10%, random forest regression is used to fit and fill in the missing values ​​based on the remaining complete features, so as to avoid data distortion caused by a single imputation method.

[0008] Preferably, the selection rule for determining the optimal partitioning ratio in step S3 is: prioritize selecting the test set R. 2 The optimal split ratio is the one that minimizes the difference in RMSE between the training and test sets and avoids significant overfitting, while excluding splitting schemes that have high training set accuracy but significantly increase test error.

[0009] Preferably, in step S4, the number of Bayesian optimization iterations for BO is set to 50. A Gaussian process surrogate model and an expected improvement acquisition function are used, which significantly reduces the number of hyperparameter traversals compared to grid search and random search, and achieves global optimal hyperparameter convergence under limited computing resources.

[0010] Preferably, in step S5, the SHAP interaction effect value is used to quantify the gain / suppression effect of the synergistic effect of the two sets of features on the hydrogen yield. Combined with the PDP curve, it can intuitively distinguish the linear and nonlinear control ranges of a single feature, providing a quantitative basis for the optimization of pyrolysis conditions.

[0011] Preferably, the recalcitrant waste includes at least one of cross-linked resin, waste plastic, fiber-reinforced composite material, recalcitrant organic sludge, and polymeric solid waste.

[0012] A machine learning system for assisting in predicting the yield of hydrogen gas from the pyrolysis of recalcitrant waste includes six functional units that communicate with each other in sequence: data acquisition unit, preprocessing unit, dataset partitioning unit, model optimization and evaluation unit, contribution evaluation unit, and hydrogen gas prediction unit. The data acquisition unit is used to import multiple sets of raw pyrolysis experimental data extracted from literature, complete the data format standardization, and output the raw dataset with 12-dimensional features and hydrogen yield labels to the preprocessing unit. The preprocessing unit has a built-in Pearson correlation calculation module, a 3σ outlier removal module, and a double missing value repair module, and outputs a standardized dataset that has been cleaned and feature correlation filtered to the dataset partitioning unit. The dataset partitioning unit is configured with a multi-ratio dataset random partitioning module, a multi-model batch training module, and a multi-index performance comparison module, and outputs the 8:2 optimal partitioned training set and test set to the model optimization and evaluation unit. The model optimization and evaluation unit integrates three types of hyperparameter tuning sub-modules: grid search, random search, and Bayesian optimization. It batch trains RF and GB models and compares the R-values ​​of multiple tuning schemes. 2 RMSE, MAE, and output the best BO-GB model after training to the contribution evaluation unit; The contribution assessment unit has built-in sub-modules for global feature importance calculation, SHAP value solving, PDP plotting, and SHAP interaction effect analysis, and outputs feature weight ranking and multi-parameter coupling influence law to the pyrolysis hydrogen prediction unit. The pyrolysis hydrogen prediction unit receives the waste elements to be tested, industrial analysis, and pyrolysis operating parameters, calls the BO-GB model to complete forward inference, and outputs the predicted hydrogen yield value in a visual manner; the units work together in a pipeline manner to realize a fully automated closed-loop process from raw data input to hydrogen yield prediction.

[0013] Preferably, the dual missing value repair module in the preprocessing unit runs the mean imputation algorithm and the random forest regression imputation algorithm in parallel, and automatically switches the repair strategy according to the missing value threshold of each feature.

[0014] Preferably, the model optimization and evaluation unit has a built-in five-fold cross-validation logic, which performs cross-validation performance verification every time it is optimized, thus avoiding random deviations in model performance caused by a single partition.

[0015] Preferably, the contribution evaluation unit is configured with a visualization output interface, which can export feature importance bar charts, SHAP summary charts, single feature partial dependency curves, and dual feature SHAP interactive heatmaps.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention provides a machine learning method for predicting the yield of hydrogen from pyrolysis. A Bayesian optimization-gradient boosting model is constructed; the model requires only 12 input parameters, including elemental composition (C%, H%, N%, O%, S%), industrial analysis parameters (volatile matter, fixed carbon, moisture, ash content), and operating condition parameters (pyrolysis temperature, heating rate, pyrolysis time). The performance index of the prediction model has been verified to be R0. 2 =0.780, RMSE=6.96, MAE=4.67. The method of this invention can quickly predict the yield of pyrolysis hydrogen based on the elemental composition of the sample, industrial analysis parameters, and operating condition parameters, without the need for a large number of pyrolysis experiments, significantly reducing manpower and material costs, and has the technical advantages of high efficiency, low consumption, and scalability.

[0017] 2. The model constructed in this invention can accurately identify the key factors affecting the generation of hydrogen from pyrolysis. Feature importance analysis results show that the yield of hydrogen from pyrolysis is mainly affected by factors such as pyrolysis temperature, heating rate, hydrogen-to-carbon ratio (H / C), pyrolysis time, moisture content, and ash content. Therefore, the model of this invention not only has high predictive accuracy but also reveals the dominant roles of various factors in the generation of different pyrolysis products, providing a scientific basis for optimizing pyrolysis operating conditions and improving the efficiency of hydrogen production from pyrolysis.

[0018] 3. The method for predicting hydrogen gas from the pyrolysis of recalcitrant waste described in this invention overcomes the bottlenecks of traditional mechanistic models, which struggle to reveal the coupling and nonlinear effects of multiple factors, suffer from low prediction accuracy, and lack sufficient generalization ability. The developed method for predicting hydrogen gas from the pyrolysis of recalcitrant waste provides new insights for the application of machine learning in modeling the pyrolysis process of recalcitrant waste and predicting product yields. Attached Figure Description

[0019] Figure 1 This is a flowchart of a machine learning method for assisting in predicting the yield of hydrogen gas from the pyrolysis of recalcitrant waste, as disclosed in this invention. Figure 2 This is a structural diagram of a machine learning system disclosed in this invention that assists in predicting the yield of hydrogen gas from the pyrolysis of recalcitrant waste. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0021] This invention relates to a machine learning method and system for assisting in the prediction of hydrogen yield from the pyrolysis of recalcitrant waste. By integrating big data analysis and machine learning algorithms, a complex relationship is constructed between elemental composition, industrial analysis parameters, operating condition parameters, and hydrogen yield from pyrolysis. The novel model of this invention can improve the prediction efficiency and accuracy of hydrogen yield from pyrolysis and can be extended to the product prediction and optimization applications of other types of solid waste pyrolysis systems.

[0022] Reference Figure 1 As shown in the figure, this invention discloses a machine learning method to assist in predicting the yield of hydrogen gas from the pyrolysis of recalcitrant waste, comprising the following steps: Pyrolysis experimental data of various recalcitrant wastes were collected to construct a pyrolysis parameter dataset for predicting hydrogen yield. Perform statistical analysis, correlation analysis, and feature engineering on the dataset to construct the model input features; Hydrogen prediction datasets with different partition ratios were used to train various machine learning regression models, and the model performance was compared to select the best dataset partition ratio. The divided hydrogen prediction dataset was input into various machine learning regression models, and grid search, random search and Bayesian optimization methods were combined for training and performance evaluation in order to select the optimal prediction model. A quantitative evaluation of the hydrogen prediction model was conducted to analyze the effects of elemental composition, industrial analytical parameters, and operating conditions on the yield of pyrolysis hydrogen and to identify the main influencing factors. Input the elemental composition, industrial analysis parameters, and operating condition parameters into the optimization model to obtain the predicted results of pyrolysis hydrogen yield.

[0023] Furthermore, pyrolysis parameters of recalcitrant waste are collected, including: The dataset input features include: C%, H%, N%, O%, S%, volatile matter, fixed carbon, moisture, ash, pyrolysis temperature, heating rate and pyrolysis time, and one output variable: hydrogen yield.

[0024] Specifically, all data used in this invention are derived from a systematic literature search of the Web of Science core database. During the search process, keyword combinations such as "pyrolysis of recalcitrant waste," "hydrogen production," "pyrolysis oil," and "pyrolysis gas" were selected to screen and compare publicly available literature. After rigorous quality control and data cleaning, 269 high-quality, publicly available experimental data records were ultimately obtained to construct a hydrogen prediction dataset.

[0025] Furthermore, the correlation of input features is analyzed using the Pearson correlation matrix; preprocessing methods include outlier identification and removal, mean imputation, and random forest regression to repair missing feature variables.

[0026] Furthermore, machine learning regression algorithm models include: Machine learning regression algorithms include: Linear Regression, Ridge Regression, Lasso Regression, Support Vector Machine (SVM), Decision Tree (DT), Random Forest, Gradient Boosting, Extreme Gradient Boosting (XGBoost), Lightweight Gradient Boosting (LightGBM Regressor), Adaptive Boosting (AdaBoost), K-Nearest Neighbors (KNN), and Artificial Neural Network (ANN) models.

[0027] Furthermore, datasets with different partition ratios were used to train linear regression, ridge regression, Lasso regression, support vector machine (SVM), decision tree (DT), random forest, gradient boosting, extreme gradient boosting (XGBoost), lightweight gradient boosting (LightGBM Regressor), adaptive boosting (AdaBoost), K-nearest neighbor (KNN) algorithm, and artificial neural network (ANN) models. R0 was then calculated. 2 , MAE and RMSE, to filter the dataset with the best split ratio.

[0028] Specifically, three models—random forest, gradient boosting regression, and extreme gradient boosting—were used on the dataset, and their performance was compared with different training / test set split ratios (0.6 / 0.4, 0.7 / 0.3, 0.8 / 0.2, 0.9 / 0.1). To reduce the impact of randomness, each ratio underwent five repeated random split experiments, and the results were measured using R... 2 RMSE and MAE were used as evaluation metrics. The results show that Boosting-type models outperform other methods overall, and the 80 / 20 split ratio performs best in terms of model accuracy and robustness. Therefore, 80% of the data was used as the training set and 20% as the test set for subsequent model construction.

[0029] Furthermore, for the dataset with the optimal split ratio, grid search, random search, and Bayesian optimization methods are combined with the selected gradient boosting (GB) and random forest (RF) prediction models for optimization. The R-values ​​of the training and testing datasets are then calculated. 2 The optimal hyperparameter optimization-prediction model combination with the best prediction performance is obtained by filtering the MAE and RMSE values. Refer to the table below for a comparison of the prediction performance of GB and RF under different hyperparameter optimization strategies.

[0031]

[0032] Specifically, for a hydrogen prediction dataset with an 8:2 training / test split, the predictive performance of gradient boosting regression, random forest, extreme gradient boosting, adaptive boosting, and K-nearest neighbor models were evaluated. Five-fold cross-validation (CV) combined with random search was used to optimize the model's hyperparameters, and the coefficient of determination (R²) was used as the benchmark. 2 The prediction performance of the training set and the test set is comprehensively compared using three metrics: root mean square error (RMSE) and mean absolute error (MAE).

[0033] Experimental results show that, compared to other models, Gradient Boosting Regression (GB) exhibits high consistency on both the training and test sets, with R0... 2 With an accuracy close to 0.9 and both RMSE and MAE at low levels, the model effectively captures complex nonlinear relationships and is suitable for current data structures with multivariate coupling features. Random Forest (RF) also exhibits good stability; its test set performance is slightly lower than GB, but it is more robust in error control, demonstrating strong noise resistance. Therefore, both GB and RF, while ensuring high prediction accuracy, possess good generalization performance and are the preferred models in this invention. Referring to Table 1, in the comparison of three hyperparameter optimization methods—random search, grid search, and Bayesian optimization (BO)—the BO-optimized GB model achieves better generalization ability at the expense of some training accuracy. This balance is crucial for preventing overfitting and improving model robustness. Therefore, this invention ultimately selects the Bayesian-optimized gradient boosting (BO-GB) model as the core model for subsequent mechanistic explanation and application analysis.

[0034] Furthermore, global importance analysis, the SHAP algorithm, partial dependency graph (PDP), and SHAP interaction effect values ​​were used to evaluate the characteristic contributions of the selected Bayesian optimization-gradient boosting models to the elemental composition, industrial analysis parameters, and operating condition parameters of recalcitrant waste. Through quantitative calculation and visualization analysis of the contributions of each parameter, the main influencing factors on the pyrolysis hydrogen yield and their degree of influence on hydrogen generation were determined.

[0035] Specifically, an interpretability analysis was conducted on the Bayesian Optimization-Gradient Boosting (BO-GB) model for hydrogen prediction. Global feature importance analysis, the SHAP algorithm, partial dependency graphs (PDP), and SHAP interaction effect analysis were employed to systematically evaluate the contributions of elemental composition of recalcitrant organic matter, industrial analytical parameters, and operating conditions to the yield of pyrolysis products. Through PDP and SHAP interaction effect analysis, the influence of each feature parameter on the generation of pyrolysis hydrogen and oil and gas was revealed. The results of global importance ranking and SHAP value analysis showed that in pyrolysis hydrogen prediction, the top six key features were: pyrolysis temperature > heating rate > H / C > pyrolysis time > moisture > ash content. Furthermore, when the input parameters for recalcitrant waste are: C%=10, N%=10, O%=10, H%=10, S%=10, heating rate=0 ℃ / min, pyrolysis time=30 min, pyrolysis temperature=600 ℃, ash content=10, fixed carbon content=10, moisture content=10, and volatile matter content=10, the model output prediction result is: hydrogen yield 18.74%.

[0036] Reference Figure 2 As shown, the present invention also provides a machine learning system for assisting in the prediction of hydrogen yield from the pyrolysis of recalcitrant waste. The system is used to implement any of the aforementioned machine learning methods for assisting in the prediction of hydrogen yield from the pyrolysis of recalcitrant waste, and includes, in sequence, a data acquisition unit, a preprocessing unit, a dataset partitioning unit, a model optimization and evaluation unit, a contribution evaluation unit, and a pyrolysis product prediction unit. The data acquisition unit is used to receive various externally input pyrolysis experimental data of non-biodegradable wastes, format the data, and output the pyrolysis parameter dataset to the preprocessing unit. The preprocessing unit receives the pyrolysis parameter dataset output by the data acquisition unit, analyzes the correlation of the input features through the Pearson correlation matrix, performs outlier identification and removal, and uses a combination of mean imputation and random forest regression to repair missing feature variables; after preprocessing, it outputs a dataset with improved features to the dataset partitioning unit. The dataset partitioning unit receives the cleaned dataset output from the preprocessing unit and partitions the dataset according to different training and test set ratios. The partitioning results are then input into various machine learning regression models for training and comparison to select the optimal partitioning ratio. Finally, the partitioned dataset and preliminary model performance results are output to the model optimization and evaluation unit. The model optimization and evaluation unit receives the training and test set data output by the dataset partitioning unit, and combines grid search, random search, and Bayesian optimization methods to optimize the parameters of various models; based on the prediction performance metric R... 2 The accuracy and stability of the model are evaluated using RMSE and MAE, and the optimal hydrogen prediction model is output to the contribution evaluation unit. The contribution assessment unit receives the optimal model output by the model optimization assessment unit, and analyzes the influence of elemental composition, industrial analysis parameters and operating condition parameters on the pyrolysis hydrogen yield through global importance analysis, SHAP value, partial dependency graph (PDP), and SHAP interaction effect value. It then outputs the key features and contribution results to the pyrolysis hydrogen prediction unit. The pyrolysis hydrogen prediction unit receives the key features and optimal model parameters output by the contribution assessment unit, inputs the elemental composition of the recalcitrant waste, industrial analysis parameters and operating condition parameters, and calculates and outputs the predicted results of pyrolysis hydrogen yield through the integrated prediction model. The data outputs and inputs of each unit are connected sequentially to form an automated processing flow from data acquisition to hydrogen prediction.

[0037] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0038] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A machine learning method for assisting in predicting the hydrogen yield from the pyrolysis of recalcitrant waste, comprising the following pre-processing steps: dataset construction, data preprocessing, dataset partitioning, model training and optimization, feature influence analysis, and yield prediction; characterized in that, The following steps are included: Step S1: Targeted Construction of Multi-Source Experimental Dataset: Search the Web of Science core database to screen publicly available literature on hydrogen production from the pyrolysis of recalcitrant waste, extract 269 sets of standardized experimental records to construct a dedicated hydrogen yield dataset; set 12-dimensional input features and 1-dimensional output labels. The input features include elemental composition (C%, H%, N%, O%, S), industrial analysis indicators (volatile matter, fixed carbon, moisture, ash), and pyrolysis operation parameters (pyrolysis temperature, heating rate, pyrolysis time); the output label is hydrogen yield. Step S2, Multi-method coupled data preprocessing: Pearson correlation matrix is ​​used to complete quantitative analysis of feature correlation; dual imputation mechanism is used to handle missing values: mean imputation is used to fill mildly missing samples, and random forest regression model is used to complete highly missing features; outlier identification and removal are completed based on the 3σ criterion to eliminate the interference of extreme experimental data on model fitting; Step S3: Multi-ratio cross-validation to select the optimal dataset partitioning scheme: Set the training set / test set partitioning ratio to 0.6 / 0.4, 0.7 / 0.3, 0.8 / 0.2, and 0.9 / 0.1, and perform 5 random repetitions for each ratio; train all 12 types of machine learning regression models on each ratio dataset respectively, using R... 2 The optimal division ratio was selected based on a comprehensive evaluation of three indicators: RMSE, MAE, and RMSE. The 12 types of regression models include linear regression, ridge regression, Lasso regression, SVM, decision tree, random forest (RF), gradient boosting (GB), XGBoost, LightGBM, AdaBoost, KNN, and artificial neural network (ANN). Step S4: Triple Hyperparameter Optimization Comparison to Identify the Optimal Bayesian Optimization-Gradient Boosting (BO-GB) Prediction Model: The dataset is split into an 8:2 ratio, and five-fold cross-validation is used. RF and GB models are trained using three hyperparameter tuning strategies: grid search (GS), random search (RS), and Bayesian optimization (BO). The training and test sets (R and Gradient) are compared. 2 RMSE and MAE were used to select the BO-GB combination with the lowest generalization error and the best anti-overfitting ability as the core prediction model; the performance index R of the BO-GB model on the test set was... 2 =0.780, RMSE=6.96, MAE=4.67; Step S5, Multi-level interpretability coupling analysis of feature contribution: Simultaneously perform global feature importance ranking, SHAP global contribution quantification, partial dependency graph PDP single feature response analysis, and SHAP interaction effect multi-feature coupling analysis on the BO-GB model; quantitatively rank the weights of key features affecting hydrogen yield, with the order of importance being pyrolysis temperature > heating rate > H / C hydrogen-carbon ratio > pyrolysis time > moisture > ash content, clarifying the law of multi-parameter nonlinear coupling effect; Step S6, Real-time Yield Prediction Output: Input the 12 characteristic parameters of the recalcitrant waste to be tested into the trained BO-GB model, and directly output the predicted value of pyrolysis hydrogen yield under the corresponding working conditions.

2. The machine learning method for assisting in predicting the hydrogen yield from the pyrolysis of recalcitrant waste according to claim 1, characterized in that, In step S2, missing value imputation adopts a hierarchical judgment logic: if the missing rate of a single feature is less than 10%, mean imputation is used; if the missing rate of a single feature is ≥10%, random forest regression is used to fit and fill in the missing values ​​based on the remaining complete features, so as to avoid data distortion caused by a single imputation method.

3. The machine learning method for assisting in predicting the hydrogen yield from the pyrolysis of recalcitrant waste according to claim 1, characterized in that, The selection rule for determining the optimal partition ratio in step S3 is: prioritize the selection of the test set R. 2 The optimal split ratio is the one that minimizes the difference in RMSE between the training and test sets and avoids significant overfitting, while excluding splitting schemes that have high training set accuracy but significantly increase test error.

4. The machine learning method for assisting in predicting the hydrogen yield from the pyrolysis of recalcitrant waste according to claim 1, characterized in that, In step S4, the number of Bayesian optimization iterations for BO is set to 50. A Gaussian process surrogate model and an expected improvement acquisition function are adopted. Compared with grid search and random search, this significantly reduces the number of hyperparameter traversals and achieves global optimal hyperparameter convergence under limited computing resources.

5. The machine learning method for assisting in predicting the hydrogen yield from the pyrolysis of recalcitrant waste according to claim 1, characterized in that, In step S5, the SHAP interaction effect value is used to quantify the gain / suppression effect of the synergistic effect of the two sets of features on the hydrogen yield. Combined with the PDP curve, it can intuitively distinguish the linear and nonlinear control ranges of a single feature, providing a quantitative basis for the optimization of pyrolysis conditions.

6. The machine learning method for assisting in predicting the hydrogen yield from the pyrolysis of recalcitrant waste according to claim 1, characterized in that, The recalcitrant waste includes at least one of cross-linked resin, waste plastics, fiber-reinforced composite materials, recalcitrant organic sludge, and polymeric solid waste.

7. A machine learning system for assisting in predicting the hydrogen yield from the pyrolysis of recalcitrant waste, characterized in that, The machine learning method described in any one of claims 1 to 6 comprises six functional units that are sequentially interconnected: a data acquisition unit, a preprocessing unit, a dataset partitioning unit, a model optimization and evaluation unit, a contribution evaluation unit, and a pyrolysis hydrogen prediction unit. The data acquisition unit is used to import multiple sets of raw pyrolysis experimental data extracted from literature, complete the data format standardization, and output the raw dataset with 12-dimensional features and hydrogen yield labels to the preprocessing unit. The preprocessing unit has a built-in Pearson correlation calculation module, a 3σ outlier removal module, and a double missing value repair module, and outputs a standardized dataset that has been cleaned and feature correlation filtered to the dataset partitioning unit. The dataset partitioning unit is configured with a multi-ratio dataset random partitioning module, a multi-model batch training module, and a multi-index performance comparison module, and outputs the 8:2 optimal partitioned training set and test set to the model optimization and evaluation unit. The model optimization and evaluation unit integrates three types of hyperparameter tuning sub-modules: grid search, random search, and Bayesian optimization. It batch trains RF and GB models and compares the R-values ​​of multiple tuning schemes. 2 RMSE, MAE, and output the best BO-GB model after training to the contribution evaluation unit; The contribution assessment unit has built-in sub-modules for global feature importance calculation, SHAP value solving, PDP plotting, and SHAP interaction effect analysis, and outputs feature weight ranking and multi-parameter coupling influence law to the pyrolysis hydrogen prediction unit. The pyrolysis hydrogen prediction unit receives the waste elements to be tested, industrial analysis, and pyrolysis operating parameters, calls the BO-GB model to complete forward inference, and outputs the predicted hydrogen yield value in a visual manner; the units work together in a pipeline manner to realize a fully automated closed-loop process from raw data input to hydrogen yield prediction.

8. The machine learning system for assisting in predicting the hydrogen yield from the pyrolysis of recalcitrant waste according to claim 7, characterized in that, The preprocessing unit contains a dual missing value repair module that runs the mean imputation algorithm and the random forest regression imputation algorithm in parallel, and automatically switches the repair strategy according to the missing value threshold of each feature.

9. The machine learning system for assisting in predicting the yield of hydrogen gas from the pyrolysis of recalcitrant waste according to claim 7, characterized in that, The model optimization and evaluation unit incorporates a five-fold cross-validation logic, which performs cross-validation performance verification every time it is optimized, thus avoiding random deviations in model performance caused by a single partition.

10. The machine learning system for assisting in predicting the hydrogen yield from the pyrolysis of recalcitrant waste according to claim 7, characterized in that, The contribution evaluation unit is configured with a visualization output interface, which can export feature importance bar charts, SHAP summary charts, single-feature partial dependency curves, and dual-feature SHAP interactive heatmaps.

Images