Page 47 - 《橡塑技术与装备》英文版2026年2期

SPECIAL AND COMPREHENSIVE REVIEW

on current data and model predictions. The online learning process is a continuous iterative process that requires close
algorithm enables the model to learn from data updates after collaboration among interdisciplinary teams as well as tracking
each pyrolysis process, eliminating the need to train the and application of the latest technologies. The AI model can
entire model from scratch. The model can quickly adjust its learn optimal operating conditions through historical data and
parameters to adapt to changes. In the actual pyrolysis process, predict outputs under different parameter combinations, thereby
environmental conditions, raw material composition, and achieving automation and refined control of the process.
other factors may change. The machine learning model can 1.8 Development strategy for basic theories
adapt to these changes through continuous learning and self- of prediction and modeling
adjustment, maintaining high performance and stability. This By collecting relevant data on the pyrolysis process
technology combines the adaptive learning capabilities of of waste plastics, statistical analysis or machine learning
artificial intelligence with the principles of pyrolysis processes methods (such as recursive feature elimination, feature
in chemical engineering, aiming to continuously optimize importance evaluation) are used to select the features that
pyrolysis process parameters through real-time data feedback, have the greatest impact on the pyrolysis process. Based on
maximize resource utilization, and minimize environmental the nature of the problem, an appropriate machine learning
impact. model or deep learning model is selected. The performance
1.6 Development strategy for basic theories of the model on the validation set and test set is evaluated,
of ensemble learning and multi-model fusion and appropriate metrics are used to judge the generalization
By developing a diverse set of models, conducting ability of the model. The trained model is applied to the actual
sampling with replacement on the training data, and weighted pyrolysis process of waste plastics to predict key parameters
averaging the prediction results of all models, we can evaluate (such as yield, product type, energy consumption, etc.), and
the importance of each model to the features. Through careful operational parameters are adjusted based on the prediction
design and optimization, ensemble learning and multi-model results to optimize the pyrolysis process. Prediction models are
fusion can play a powerful role in prediction and optimization developed to predict key parameters and product characteristics
in complex scenarios such as waste plastic pyrolysis, while during the pyrolysis process, such as oil yield, gas generation,
improving the robustness and efficiency of decision-making. and carbon black production. These models can be based on
By combining the advantages of multiple machine learning statistical learning, deep learning, or hybrid learning methods,
models and utilizing ensemble learning methods, we can and predict results under different conditions by training a large
enhance prediction accuracy and robustness, better handling amount of experimental data.
the complex and variable pyrolysis process. 1.9 Development strategy for basic theory of
1.7 Development strategy for basic theories by-product analysis and classification
of process optimization and control By collecting all relevant information about the pyrolysis
By collecting key parameters during the pyrolysis process of waste plastics, including but not limited to the
process of waste plastics, we utilize statistical analysis and chemical composition and physical properties of the pyrolysis
machine learning algorithms to screen the features that have products, pyrolysis conditions (such as temperature, pressure,
the greatest impact on the pyrolysis process and create new time, etc.), and other parameters that may affect the product
features. We select appropriate machine learning models characteristics. Data preprocessing includes cleaning data,
based on the complexity of the problem, develop models to removing outliers, standardizing numerical ranges, etc., to
accommodate real-time data streams, continuously update facilitate subsequent analysis and classification. Select an
model parameters, optimize the pyrolysis process, dynamically appropriate machine learning or deep learning model to
adjust pyrolysis conditions based on real-time data and the classify the by-products. Evaluate model performance using
results of predictive models, and ensure that the optimization methods such as cross-validation to select the best model.
and control development of the AI waste plastics pyrolysis Evaluation metrics typically include accuracy, recall, F1

Vol.52，2026 ·3·