Method and system for optimizing biomass catalytic pyrolysis process

Abstract

The invention provides an optimization method and system for a biomass catalytic pyrolysis process, belongs to the crossing field of chemical engineering and artificial intelligence, and aims to solve the problem of collaborative optimization of the yield of multiple competitive aromatic hydrocarbon products. The method comprises the following steps: acquiring a data set containing input parameters and corresponding to multiple target aromatic hydrocarbon yields; training a machine learning prediction model based on the data set; integrating the model with a multi-objective optimization algorithm; the model is called through an algorithm to conduct iterative optimization, a Pareto optimal solution set is obtained with the aim of optimizing the yield of multiple aromatic hydrocarbons at the same time, and the solution set comprises multiple sets of optimized input parameter combinations. According to the method, efficient global optimization can be realized, multi-scheme decision support is provided, and the research and development cost is reduced.

Description

Technical Field

[0001] This application relates to the interdisciplinary field of chemical engineering and artificial intelligence, and in particular to an optimization method for biomass catalytic pyrolysis process based on machine learning and multi-objective optimization. Background Technology

[0002] Biomass catalytic pyrolysis technology is an important pathway for converting renewable biomass resources into high-value aromatics, particularly chemicals such as benzene, toluene, and xylene. The product distribution and yield of this process are influenced by a complex array of factors, including feedstock characteristics such as elemental composition, process parameters such as pyrolysis temperature and catalyst dosage, and catalyst properties such as metal type and acidity. These multidimensional variables exhibit nonlinear coupling relationships, making process optimization extremely challenging.

[0003] In existing technologies, researchers have begun applying machine learning methods, such as random forests, gradient boosting trees, support vector machines, and neural networks, to build predictive models between input parameters and product yields. Specifically, some technical solutions disclose acquiring datasets containing input parameters such as biomass elemental composition and pyrolysis conditions, along with various target products, and training machine learning models to establish a mapping relationship between inputs and outputs, thereby enabling the prediction of biomass pyrolysis products. These methods can learn from historical experimental data and predict the yield of pyrolysis products under specific process conditions, providing a data-driven perspective for understanding this complex process.

[0004] However, existing technologies still have significant shortcomings. On the one hand, most current machine learning applications focus on "prediction," that is, predicting product yields based on given input parameters, but fail to achieve "backward optimization" of input parameters, and cannot proactively find the optimal combination of process parameters to achieve specific production goals. On the other hand, these methods often analyze a single objective (such as total aromatics yield), ignoring the inherent competitive relationships between multiple target products such as benzene, toluene, and xylene, and thus cannot achieve targeted selective control of product distribution. Furthermore, traditional experimental trial-and-error methods or single-parameter optimization methods are inefficient and prone to getting trapped in local optima, making it difficult to perform global optimization in a broad multidimensional parameter space, and therefore difficult to discover breakthrough process formulation combinations. Summary of the Invention

[0005] The purpose of this application is to provide a method for efficient and global optimization of biomass catalytic pyrolysis processes, in order to solve the problem of how to synergistically optimize the yield of multiple competing target aromatic products (such as benzene, toluene, and xylene) within a multidimensional process parameter and catalyst formulation space, and to achieve precise and flexible control of product distribution, thereby overcoming the shortcomings of traditional optimization methods such as low efficiency, easy to get trapped in local optima, and inability to effectively balance multiple objectives.

[0006] To achieve the above objectives, this application provides an optimization method for a biomass catalytic pyrolysis process. The method includes the following steps: Step 1, obtaining a training dataset, which includes multiple sets of input parameters related to the biomass catalytic pyrolysis process, and the yields of various target aromatic products corresponding to the multiple sets of input parameters. The input parameters include at least one of feedstock characteristics, process parameters, and catalyst properties. Step 2, training a machine learning prediction model based on the training dataset to establish a mapping relationship between the input parameters and the yields of the various target aromatic products. Step 3, integrating the machine learning prediction model with a multi-objective optimization algorithm, wherein the machine learning prediction model is used to respond to the call of the multi-objective optimization algorithm to predict the yields of the various target aromatic products under given input parameters. Step 4, based on a preset range of values ​​for the input parameters, iteratively optimizing the machine learning prediction model through the multi-objective optimization algorithm to simultaneously optimize the yields of the various target aromatic products, thereby obtaining a Pareto optimal solution set. The Pareto optimal solution set includes multiple optimized combinations of input parameters.

[0007] Optionally, the raw material characteristics in the input parameters include at least one of the carbon, hydrogen, oxygen, and nitrogen content of biomass; the process parameters include at least one of the pyrolysis temperature, catalytic temperature, and mass ratio of catalyst to biomass; and the catalyst properties include at least one of the type of metal loaded on the catalyst, metal loading, specific surface area, and acidity.

[0008] Optionally, before training the machine learning prediction model, a preprocessing step of the training dataset is further included; the preprocessing step includes at least one of the following steps: normalizing or standardizing the numerical features in the training dataset; and label encoding the categorical variables.

[0009] Furthermore, the preprocessing step also includes: calculating derived features based on domain knowledge, and merging the derived features with the original features as a new input feature set for model training; the derived features include at least one of the hydrogen-to-carbon atomic ratio and the oxygen-to-carbon atomic ratio of the raw materials.

[0010] Optionally, the machine learning prediction model is one of the following: random forest model, gradient boosting tree model, support vector regression model, or neural network model.

[0011] Optionally, the steps of training the machine learning prediction model specifically include: dividing the training dataset into a training set, a validation set, and a test set; training the machine learning prediction model using the training set; and monitoring the model performance during training using the validation set to prevent overfitting.

[0012] Optionally, the multi-objective optimization algorithm is the NSGA-II algorithm.

[0013] Optionally, the optimization objective further includes process cost-related parameters; the process cost-related parameters are the pyrolysis temperatures that need to be minimized.

[0014] This application also provides a biomass catalytic pyrolysis process optimization system, comprising: a data acquisition module for acquiring a training dataset, the training dataset including multiple sets of input parameters related to the biomass catalytic pyrolysis process, and the yields of various target aromatic products corresponding to the multiple sets of input parameters, the input parameters including at least one of feedstock characteristics, process parameters, and catalyst properties; a model training module for training a machine learning prediction model based on the training dataset to establish a mapping relationship between the input parameters and the yields of the various target aromatic products; and an optimization execution module for integrating the machine learning prediction model with a multi-objective optimization algorithm, and iteratively optimizing the machine learning prediction model by calling the multi-objective optimization algorithm based on a preset range of values ​​for the input parameters, with the goal of simultaneously optimizing the yields of the various target aromatic products, thereby obtaining a Pareto optimal solution set; the Pareto optimal solution set including multiple optimized input parameter combinations.

[0015] Optionally, the optimization execution module is specifically configured to use the NSGA-II algorithm as the multi-objective optimization algorithm.

[0016] Compared with the prior art, this application has the following beneficial effects: 1. Achieving coordinated and precise control and global optimization: By combining machine learning prediction models with multi-objective optimization algorithms, it can efficiently and globally search and provide a set of Pareto optimal solutions containing various optimized processes and catalyst formulations. This enables decision-makers to make scientific trade-offs and flexible choices between the yields and selectivity of various competing aromatic products, overcoming the shortcomings of traditional optimization methods such as low efficiency, easy to get trapped in local optima, and inability to flexibly trade off multiple objectives.

[0017] 2. Reduced R&D costs and time: This method can adapt to different biomass raw materials and reaction conditions, and discover the optimal process parameters through data-driven methods, reducing the need for a large number of repeated experiments due to changes in operating conditions, thus significantly reducing R&D costs and time.

[0018] 3. Lowering the technical threshold: This method does not require the construction and solution of complex reaction kinetic equations. It directly learns the inherent laws of the process from the data, which greatly reduces the technical threshold and modeling difficulty for optimizing complex chemical systems.

[0019] 4. Provides flexible decision support. The output Pareto optimal solution set provides decision-makers with a variety of high-performance alternatives, rather than a single optimal solution, supporting flexible and scientific adjustments to production strategies based on external factors. Attached Figure Description

[0020] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0021] Figure 1 This is a schematic flowchart of an optimized method for a biomass catalytic pyrolysis process provided in one embodiment of this application.

[0022] Figure 2 This is a schematic diagram of the structure of a biomass catalytic pyrolysis process optimization system provided in one embodiment of this application.

[0023] Figure 3 This is a schematic diagram of an electronic device structure provided in one embodiment of this application. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] First, let me explain some of the terms used in this application: Co-pyrolysis of biomass and metal-modified catalysts: Essentially, it is a process in which the thermal cracking reaction of biomass macromolecules and the catalytic reconstruction reaction of metal-modified HZSM-5 catalyst proceed in synergy. The core lies in the directional regulation of pyrolysis intermediates by metal active sites. The liquid products of pyrolysis are mainly aromatic hydrocarbons such as benzene, toluene, and xylene, which have high industrial value.

[0026] Data set: mainly includes data on pyrolysis feedstock, process, catalyst parameters, aromatic yield, etc., aiming to study the laws of biomass catalytic pyrolysis to produce aromatics (benzene, toluene, xylene) under different conditions.

[0027] Training set: Contains input data and the corresponding target output, and is the foundation for model learning. The model learns the mapping relationship between input and output through the training set data.

[0028] Validation set: As an intermediate evaluation metric for model performance, it helps us determine the optimal model parameters and structure. During training, we typically use the validation set to evaluate the model and adjust it based on the evaluation results.

[0029] Test set: After the model is trained, the test set is used to evaluate the model's generalization ability. The test set data is unknown to the model and is used to test the model's performance on unseen data.

[0030] Data normalization: Normalization is to linearly scale the data to a specific range, such as [0, 1]. The purpose is to eliminate the scale effect between different features and prevent some features from being ignored during training due to excessive differences in feature value ranges.

[0031] Feature engineering: The types of metals added to the catalyst in the dataset (such as Zn, Fe, Ni, etc.) are labeled using tag encoding, so that the text information is converted into character information.

[0032] NSGA-II optimization algorithm: The liquid products after pyrolysis contain three aromatics: benzene, toluene, and xylene. Since all three cannot be obtained simultaneously, an optimization algorithm is needed to select the target products under the corresponding constraints to maximize the yield of the corresponding products.

[0033] This application provides an overview of the proposed solution: In today's rapidly developing green chemical and renewable energy technologies, biomass catalytic pyrolysis, as a key pathway for converting renewable biomass into high-value aromatics (such as benzene, toluene, and xylene), is receiving increasing attention. However, this process is subject to complex coupling effects from multiple dimensions, including feedstock characteristics, process parameters, and catalyst properties. Traditional experimental trial-and-error methods or single-objective optimization methods are inefficient, struggle to achieve synergistic optimization among competing aromatic products, and are prone to getting trapped in local optima, thus hindering further improvements in process efficiency and economics.

[0034] To address the aforementioned issues, this application provides a method and system for optimizing biomass catalytic pyrolysis processes based on machine learning and multi-objective optimization. The scheme uses feedstock characteristics (e.g., carbon, hydrogen, oxygen, and nitrogen content), process parameters (e.g., pyrolysis temperature, catalytic temperature, catalyst / biomass ratio), and catalyst properties (e.g., metal type, loading, specific surface area, and acidity) as system inputs. The core of this approach lies in first establishing a precise nonlinear mapping relationship between the aforementioned input variables and aromatic yields (benzene, toluene, and xylene) using a machine learning model; then, integrating the trained prediction model as a surrogate model with multi-objective optimization algorithms such as NSGA-II to automatically search for and output the optimal solution that can simultaneously control the selectivity of benzene, toluene, and xylene, ultimately achieving precise and global optimization of aromatic product distribution and yield.

[0035] 1. Dataset creation and data processing: Key data were collected and organized by reading literature, and a dataset of aromatic hydrocarbon production from biomass and metal-modified HZSM-5 catalyst pyrolysis was created.

[0036] 2. Dataset Splitting and Preprocessing: The collected complete experimental dataset was randomly divided into a training set (70%), a validation set (15%), and a test set (15%). Subsequently, key preprocessing steps were performed: numerical features were standardized to eliminate the influence of units; categorical variables such as "metal type" were labeled and encoded; and feature engineering based on domain knowledge was conducted, such as calculating more chemically meaningful derived features like the H / C and O / C atomic ratios, to improve model learning efficiency and prediction performance.

[0037] 3. Model Training and Optimization: Various supervised learning algorithms are used to build models on the training set. A validation set is used to monitor model performance during training and prevent overfitting through techniques such as early stopping. The final evaluation of model performance is based on an independent test set, using root mean square error (RMSE) and coefficient of determination (R²) as the main evaluation metrics to ensure the accuracy and reliability of the model.

[0038] 4. Multi-objective optimization: A trained and validated machine learning prediction model is used as a high-efficiency surrogate model and integrated with the NSGA-II multi-objective optimization algorithm. A physically feasible parameter space boundary is defined for each input variable. The optimization algorithm performs a global search within this space to maximize the predicted yields of benzene, toluene, and xylene, ultimately outputting a set of Pareto optimal solutions, providing multiple high-performance process and catalyst formulation options for decision-making.

[0039] 5. Results Verification and Analysis: The effectiveness of the method of this invention was quantitatively evaluated by comparing the optimized recommended scheme with the original experimental data or the results of traditional optimization methods. The analysis focused on: (1) the improvement in total aromatic yield and monomer selectivity; (2) the screening of metal-modified catalysts with optimal yield; and (3) the robustness of this method under various raw material and reaction conditions. The results showed that this method can stably and significantly improve process performance, which is superior to traditional single-parameter optimization or empirical trial-and-error methods.

[0040] This invention, through a global search algorithm, can discover novel and superior combinations of process conditions and catalyst formulations that exceed the scope of empirical knowledge, thereby achieving a breakthrough improvement in aromatics yield and selectivity. It overcomes the shortcomings of traditional methods that are prone to getting trapped in local optima, and can systematically search the entire feasible parameter space, ensuring that it finds a point that is globally optimal or close to it. It explicitly acknowledges and effectively handles the competing relationships between multiple optimization objectives, providing a series of optimal trade-offs and offering decision-makers unprecedented flexibility and scientific rigor. Compared to the "input-output" black box of traditional optimization methods, this invention, combined with model interpretability techniques (such as SHAP analysis), can quantify the contribution of each input variable to the output, not only providing "how" but also revealing "why," providing strong data support for mechanistic studies.

[0041] After introducing the basic principles of this application, various non-limiting embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0042] Figure 1 This is a schematic flow diagram of an optimized method for a biomass catalytic pyrolysis process provided in one embodiment of this application. Figure 1 As shown, the method includes the following:

[0043] Step S110: Obtain a training dataset, which includes multiple sets of input parameters related to the biomass catalytic pyrolysis process, and the yields of various target aromatic products corresponding to the multiple sets of input parameters. The input parameters include at least one of feedstock characteristics, process parameters, and catalyst properties. This step aims to lay the data foundation for data-driven modeling and optimization. Specifically, a structured training dataset is constructed by systematically collecting and organizing publicly available academic literature, patent documents, and internal experimental data. This dataset contains multiple data records, each corresponding to a complete biomass catalytic pyrolysis experiment.

[0044] The data record consists of two main components: Input parameters: These are the independent variables that affect the pyrolysis process and product distribution, and they cover at least one of the following categories of parameters: Raw material characteristics: such as the mass percentage of carbon, hydrogen, oxygen, and nitrogen in biomass; Process parameters: such as pyrolysis temperature, catalytic temperature, and the mass ratio of catalyst to biomass; Catalyst properties: such as the type of metal supported on the catalyst (e.g., Zn, Fe, Ni), metal loading, specific surface area, and acidity.

[0045] Output parameters: These are the dependent variables measured in each experiment, representing the yields of various target aromatic products to be optimized by this method, typically including the yields of benzene, toluene, and xylene (BTX).

[0046] Using the above method, a tabular dataset containing hundreds or even thousands of data points was constructed. Each row represents an experimental record, containing all the input parameters and the yields of the three target aromatic hydrocarbons.

[0047] Step S120: Based on the training dataset, train a machine learning prediction model to establish a mapping relationship between the input parameters and the yields of the various target aromatic products. After obtaining the training dataset, the core of this step is to use machine learning techniques to build a proxy model that can accurately simulate complex pyrolysis processes.

[0048] First, the dataset can be preprocessed, such as standardizing numerical features, encoding categorical variables, and calculating derived features (e.g., the hydrogen-to-carbon ratio H / C) based on domain knowledge. Then, the dataset is divided into training, validation, and test sets.

[0049] Next, one or more supervised learning algorithms (such as random forest, gradient boosting tree, support vector regression, or neural network) are selected, using the input parameters as features and the various target aromatic yields as prediction targets, to train the model on the training set. During training, the model performance is monitored using a validation set, and overfitting is prevented through strategies such as early stopping. Finally, the model's prediction accuracy and generalization ability are evaluated using an independent test set. Thus, the trained machine learning prediction model can establish a high-precision nonlinear mapping relationship between the input parameters and the yields of various aromatics.

[0050] Step S130: Integrate the machine learning prediction model with the multi-objective optimization algorithm, wherein the machine learning prediction model is used to respond to the call of the multi-objective optimization algorithm to predict the yield of each of the multiple target aromatic products under given input parameters. This step is crucial for achieving automation and global optimization. We integrate the high-precision predictive machine learning model trained in step S120 with a multi-objective optimization algorithm (preferably the non-dominated sorting genetic algorithm NSGA-II).

[0051] In this integrated architecture, the machine learning prediction model is encapsulated as a highly efficient fitness evaluation function. When the multi-objective optimization algorithm generates a set of candidate input parameter combinations during subsequent iterative optimization, it can directly call the prediction model without conducting real experiments, thereby obtaining predicted values ​​for the yields of various target aromatic hydrocarbons within milliseconds. This integrated approach replaces time-consuming physical experiments or complex simulations with rapid model predictions, providing feasibility for large-scale, global parameter optimization.

[0052] Step S140: Based on the preset range of the input parameters, the machine learning prediction model is called through the multi-objective optimization algorithm to perform iterative optimization, with the goal of simultaneously optimizing the yield of each of the multiple target aromatic products, thereby obtaining a Pareto optimal solution set; the Pareto optimal solution set includes multiple sets of optimized input parameter combinations.

[0053] In this step, a multi-objective optimization algorithm integrating a proxy model is activated to perform automated global optimization.

[0054] First, a physically feasible range of values ​​is defined for all input parameters to be optimized, forming the algorithm's search space. Simultaneously, the optimization objective is clearly defined as simultaneously optimizing the yields of the various target aromatic products (e.g., maximizing the yields of benzene, toluene, and xylene).

[0055] Subsequently, an optimization algorithm (such as NSGA-II) initializes a population within the search space and begins iterative evolution. In each generation, the algorithm uses an ensemble predictive model to evaluate the fitness (i.e., predicted yield) of each individual in the population, then performs non-dominated sorting according to Pareto dominance and calculates crowding to maintain population diversity. New generations of the population are generated through genetic operations such as selection, crossover, and mutation, and this process is repeated iteratively until a termination condition is met (e.g., reaching the maximum number of generations).

[0056] At the end of the optimization process, the algorithm outputs a Pareto optimal solution set consisting of all non-dominated solutions in the current population. This solution set contains multiple sets of optimized input parameter combinations, each representing an optimal trade-off among multiple competing target aromatic yields. This provides process decision-makers with a rich and scientific set of optimization options, enabling precise and flexible control over the distribution of biomass catalytic pyrolysis products.

[0057] In some embodiments, the raw material characteristics in the input parameters include at least one of the carbon, hydrogen, oxygen, and nitrogen content of biomass; the process parameters include at least one of the pyrolysis temperature, catalytic temperature, and mass ratio of catalyst to biomass; and the catalyst properties include at least one of the type of metal loaded on the catalyst, metal loading, specific surface area, and acidity.

[0058] The input parameters are independent variables that affect the pyrolysis process and product distribution. These input parameters can be further subdivided into several categories with clear physical meanings to more accurately describe the complex influencing factors of the biomass catalytic pyrolysis process.

[0059] The raw material characteristics include the content of at least one of the four key elements—carbon, hydrogen, oxygen, and nitrogen—in biomass. These elemental compositions directly determine the chemical structure, calorific value, reaction pathway, and precursor formation during pyrolysis.

[0060] The process parameters include at least one of pyrolysis temperature, catalytic temperature, and the mass ratio of catalyst to biomass. These are key process variables that are operable and controllable. The pyrolysis temperature mainly affects the primary pyrolysis of biomass, the catalytic temperature affects the secondary reactions of intermediate products on the catalyst (such as aromatization), and the mass ratio of catalyst to biomass directly determines the contact probability and reaction depth between the reactants and the active sites.

[0061] The catalyst properties include at least one of the following: the type of metal supported on the catalyst (e.g., zinc, iron, nickel, etc.), the metal loading (usually expressed as a percentage by mass), the specific surface area of ​​the catalyst (typically measured in square meters per gram), and the acidity of the catalyst (e.g., the total acid content or the distribution of different acid strengths). These properties collectively determine the catalyst's activity, selectivity, and stability, and are the core factors controlling the final distribution of aromatic products.

[0062] The output parameters are the experimentally measured yields of the target products, i.e., the dependent variables. In this embodiment, the optimization objective is to simultaneously maximize the yields of the three key aromatic products; therefore, the specific output parameters are: the yield of benzene, the yield of toluene, and the yield of xylene (including the sum of the ortho, meta, and para isomers). It should be noted that yield is usually defined as the percentage of the mass of a specific aromatic hydrocarbon produced relative to the initial dry biomass mass.

[0063] In some embodiments, a step of preprocessing the training dataset is further included before training the machine learning prediction model; The preprocessing steps include at least one of the following steps: Normalize or standardize the numerical features in the training dataset; Label the categorical variables.

[0064] First, the numerical features in the training dataset are normalized or standardized. Since the numerical features in the input parameters (e.g., pyrolysis temperature, element content, mass ratio, etc.) typically have different dimensions and value ranges (e.g., temperature values ​​are in the hundreds, while element content values ​​are usually less than one hundred), this can lead to difficulties in model convergence or the assignment of inappropriate weights to certain features. Normalization (e.g., scaling the data to the [0, 1] interval) or standardization (e.g., converting the data to a distribution with a mean of 0 and a standard deviation of 1 using the Z-score method) can eliminate these dimensional differences, accelerate model convergence, and improve its performance.

[0065] Secondly, categorical variables are labeled and encoded. Among the input parameters are non-numerical categorical variables such as "the type of metal loaded on the catalyst" (e.g., zinc, iron, nickel). Machine learning models cannot directly process this type of textual information. Therefore, it is necessary to convert it into numerical form through encoding techniques. As a basic and effective implementation method, this embodiment uses label encoding, that is, assigning a unique integer value identifier to each different category (e.g., mapping "zinc" to 0, "iron" to 1, and "nickel" to 2), thereby converting categorical features into numerical features that the model can learn and process.

[0066] In some embodiments, the preprocessing step further includes: calculating derived features based on domain knowledge, and merging the derived features with the original features as a new input feature set for model training; the derived features include at least one of the hydrogen-to-carbon atomic ratio and the oxygen-to-carbon atomic ratio of the raw materials.

[0067] It should be noted that the data preprocessing steps also include deep operations such as feature engineering based on domain knowledge. That is, by calculating derived features with clear physicochemical significance, the representational power of the input features is enhanced, and these derived features are merged with the original features that have undergone basic preprocessing to form a new, more informative set of input features for use in subsequent machine learning model training.

[0068] The derived characteristics preferably include at least one of the hydrogen-to-carbon ratio (H / C) and oxygen-to-carbon ratio (O / C) of the raw material. The calculation of these derived characteristics can be based on expertise in the field of chemical engineering. The hydrogen-to-carbon atomic ratio (H / C) is calculated using the following formula: H / C = (mass percentage of hydrogen / atomic weight of hydrogen) / (mass percentage of carbon / atomic weight of carbon). This ratio is a key indicator for assessing the hydrogen saturation and aromatization potential of biomass feedstocks.

[0069] The oxygen-to-carbon ratio (O / C) is calculated using the following formula: O / C = (mass percentage of oxygen / atomic weight of oxygen) / (mass percentage of carbon / atomic weight of carbon). This ratio reflects the oxygen content of the raw material and directly affects the formation trend of oxygen-containing byproducts (such as acids and ketones) and coke during pyrolysis.

[0070] By introducing domain-knowledge-driven feature engineering, simple numerical processing is elevated to a level that integrates mechanistic understanding. Through the introduction of derived features such as H / C and O / C, machine learning models no longer merely perform "black box" data fitting, but are able to learn more fundamental chemical principles. This results in the dual technical benefits of improved model prediction accuracy and generalization ability, as well as enhanced model interpretability.

[0071] In some embodiments, the machine learning prediction model is one of a random forest model, a gradient boosting tree model, a support vector regression model, or a neural network model. Preferably, the machine learning prediction model is one of a random forest model, a gradient boosting tree model, a support vector regression model, or a neural network model. This choice is based on the inherent high dimensionality, strong nonlinearity, and complex variable coupling characteristics of the biomass catalytic pyrolysis process.

[0072] Specifically, all of the above models have been shown to have significant advantages in handling such complex regression problems: Random forest models, as an ensemble learning algorithm, effectively reduce the risk of overfitting that may arise from a single decision tree by constructing multiple decision trees and combining their outputs. They also exhibit good robustness to noise and imbalance in the data. Furthermore, they offer high training efficiency and provide feature importance assessments, enhancing the model's interpretability.

[0073] Gradient boosting tree models (such as XGBoost and LightGBM) are another powerful ensemble approach. They iteratively train a series of weak learners (decision trees), with each tree working to correct the prediction error of the previous tree. This mechanism typically enables them to achieve extremely high prediction accuracy, and is particularly effective in accurately capturing the subtle nonlinear relationships between process, catalyst, and performance.

[0074] Support vector regression (SVR) models are particularly suitable for situations where the number of training samples is not extremely large. Its core idea is to find a regression hyperplane that maximizes the tolerance boundary of the sample points. By selecting an appropriate kernel function (such as a radial basis function kernel), it can elegantly handle nonlinear problems and typically exhibits good generalization ability due to its principle of minimizing structural risk.

[0075] As a network structure composed of multiple interconnected nodes (neurons), neural network models possess the most powerful representation learning capabilities, capable of approximating any complex nonlinear function with arbitrary precision. For the deep and complex mapping relationships between multidimensional input parameters and multiple output yields involved in this application, deep neural networks exhibit unparalleled fitting potential.

[0076] Using any of the above models, a high-precision surrogate model can be effectively learned from historical data, which is a fundamental prerequisite for achieving subsequent efficient, global multi-objective optimization. The improved model prediction accuracy directly ensures the reliability and practical value of the Pareto optimal solution set searched by the optimization algorithm. Furthermore, these four models provide flexible options for implementing this invention, allowing users to choose the most suitable modeling tool based on the scale, quality, and computational resources of their data, thereby achieving the technical effects of this invention in different application scenarios.

[0077] In some embodiments, the step of training the machine learning prediction model specifically includes: The training dataset is divided into a training set, a validation set, and a test set; The machine learning prediction model is trained using the training set, and the model performance is monitored during training using the validation set to prevent overfitting.

[0078] First, the training dataset is divided into training, validation, and test sets. This is a crucial step designed to scientifically evaluate and ensure model performance. Specifically, the preprocessed complete dataset is randomly divided into three disjoint subsets according to a predetermined ratio (e.g., 70%:15%:15%). Training set: Used to directly learn and fit the parameters of machine learning prediction models, and is the basis for the model to learn the mapping relationship between input and output.

[0079] Validation set: This set does not participate in the direct training of model parameters. Its core role is to indirectly monitor the model's performance during training. By periodically evaluating the model's performance on the validation set (e.g., calculating the root mean square error between predicted and true values), it's possible to determine if the model is overfitting (i.e., the model has excessively memorized the training data and lost the ability to predict new data). Based on the performance feedback from the validation set, strategies such as early stopping (terminating training prematurely when validation set performance no longer improves) can be implemented, and it can be used for hyperparameter tuning to find the optimal model configuration.

[0080] Test set: After the model training and tuning process is fully completed, the test set is used to perform a one-time, independent evaluation of the model's final generalization performance. The test set simulates "unknown" data that the model will encounter in future applications, and its evaluation results (e.g., root mean square error and coefficient of determination R²) are considered reliable indicators of the model's ability to be applied in real-world situations.

[0081] Secondly, based on the partitioned dataset, the machine learning prediction model is trained using the training set, and the model performance is monitored during training using the validation set to prevent overfitting. This operation constitutes the core loop of model training. In practice, this typically means calculating the model's performance metrics on the validation set after each training epoch or every few epochs. Once it is detected that the performance on the validation set begins to deteriorate continuously while the performance on the training set continues to improve, this is a typical signal of overfitting. At this point, an early stopping mechanism will be triggered, retaining the model version with the best performance on the validation set, thereby ensuring that the final model not only fits the known data well but also has strong predictive capabilities for unknown data.

[0082] The preferred multi-objective optimization algorithm is the NSGA-II algorithm, i.e., a non-dominated sorting genetic algorithm with an elitist strategy. This choice is based on the core requirement in the optimization of biomass catalytic pyrolysis process that multiple objectives compete with each other and a set of balanced solutions is needed.

[0083] The NSGA-II algorithm is a classic and efficient multi-objective evolutionary algorithm, which perfectly meets the technical requirements of this invention through the following core mechanisms: Fast non-dominated sorting: This algorithm first sorts the individuals in the population according to their Pareto dominance. This allows the algorithm to prioritize solutions that perform better on multiple objective functions (such as the yields of benzene, toluene, and xylene), thereby guiding the search process closer to the true Pareto front.

[0084] Crowding Calculation and Comparison: Within the same non-dominated hierarchy, NSGA-II evaluates the distribution density of solutions by calculating the crowding distance of each individual to its surroundings. It tends to retain individuals with higher crowding, i.e., solutions located in sparse regions. This mechanism greatly ensures the diversity of the solution set, guaranteeing that the final Pareto optimal solution set can broadly cover the entire objective space, providing decision-makers with a range of optimization schemes with different focuses, rather than clustering in a local area.

[0085] Elite preservation strategy: During the evolutionary process, the NSGA-II algorithm always merges superior individuals from the parent generation with their offspring before making a selection. This strategy prevents the loss of superior genes, accelerates the convergence process of the algorithm, and improves the globality and reliability of the optimization results.

[0086] Relevance of Technical Effects: Using the NSGA-II algorithm as the optimization engine of this invention directly brings the following key beneficial effects: Collaborative optimization of competitive objectives: It fundamentally acknowledges and effectively addresses the inherent competitive relationship between the yields of benzene, toluene, and xylene, and can automatically find the optimal balance point among these objectives.

[0087] Provides decision support: Its output Pareto optimal solution set is not a single solution, but a series of high-performance solutions with different trade-offs among multiple objectives. This provides a direct basis for process decision-makers to make flexible and scientific decisions based on external factors such as market demand and cost.

[0088] Global search capability: Combining the global search characteristics of genetic algorithms, NSGA-II can escape local optima and conduct efficient exploration in a broad multidimensional parameter space, thus potentially discovering groundbreaking combinations of process and catalyst formulations that transcend traditional experience.

[0089] In summary, the choice of the NSGA-II algorithm is not a simple, common-sense decision, but rather a result of its unique technical mechanism being highly compatible with the core technical problem of "multi-objective collaborative optimization" that this invention aims to solve. It is one of the key technical means to achieve the expected technical effects of this invention.

[0090] In some embodiments, and in more advanced ones, the optimization objective is not limited to maximizing aromatics yield, but may further incorporate parameters related to process costs, thereby achieving comprehensive optimization of the economics and sustainability of the production process. That is, the optimization objective also includes process cost-related parameters; these process cost-related parameters are the pyrolysis temperatures that need to be minimized.

[0091] The parameter related to process cost is preferably the pyrolysis temperature that needs to be minimized. In industrial practice, pyrolysis temperature is one of the main sources of energy consumption, directly determining the operating cost of the reaction unit. Higher pyrolysis temperatures usually mean higher energy consumption and equipment requirements. Therefore, minimizing the pyrolysis temperature as an optimization objective allows the optimization algorithm to simultaneously consider energy consumption costs when searching for high-performance formulations, thereby automatically selecting more economical process schemes that can still achieve high product yields under relatively mild conditions.

[0092] In terms of technical implementation, the multi-objective optimization algorithm (such as NSGA-II) can naturally handle complex optimization problems that simultaneously involve maximizing objectives (such as toluene and xylene yields) and minimizing objectives (such as pyrolysis temperature and benzene yield). The algorithm automatically finds the optimal set of trade-off solutions among these competing or even conflicting objectives through non-dominated sorting.

[0093] Ultimately, the Pareto optimal solution set output by the system will become a multi-dimensional decision space. Decision-makers can clearly see the trade-offs between "economic benefits" (yield of high-value products) and "operating costs" (pyrolysis temperature) for different solutions, and make the most scientific and comprehensive strategic choices based on current market demand and energy prices. This marks an upgrade of the optimization method of this invention from simple technical performance optimization to a comprehensive decision-making tool supporting sustainable development.

[0094] Suppose that in a specific production scenario, decision-makers are not only concerned with the yield of high-value products (such as toluene and xylene), but also highly focused on the energy costs of the production process. This embodiment will demonstrate how to configure an optimization system to meet this complex optimization requirement.

[0095] At this point, in step S140, a new set of composite optimization objectives is defined for the multi-objective optimization algorithm: 1. Objective 1 (Economic Benefit): Maximize toluene yield. 2. Objective 2 (Economic Benefit): Maximize xylene yield. 3. Objective 3 (Process Cost): Minimize pyrolysis temperature. 4. Objective 2 (Economic Benefit): Maximize benzene yield.

[0096] At this point, the multi-objective optimization algorithm needs to handle four conflicting and competing objectives simultaneously. For example, increasing the pyrolysis temperature usually promotes the aromatization reaction and may simultaneously increase the yield of all aromatics (including the undesirable benzene), which conflicts with objectives three and four.

[0097] It should be noted that the internal mechanisms of multi-objective optimization algorithms (such as non-dominated sorting) can naturally handle such complex situations involving both maximization and minimization objectives. In terms of technical implementation, the minimization objective function is typically used... The problem is transformed into maximizing its opposite. The problem is thus unified into a maximization problem for processing. When evaluating each candidate solution, the algorithm considers its performance on all four objectives simultaneously and looks for "non-dominated" solutions that cannot improve any one objective without harming any of the other objectives.

[0098] After sufficient algebraic iterations, the output Pareto optimal solution set will be a Pareto front in the four-dimensional objective space. Each solution in this set represents a unique balance between economic benefits and operating costs.

[0099] For example, the final Pareto optimal solution set may contain schemes of the following types: The "ultimate economy" solution may correspond to very high yields of benzene, toluene, and xylene, but requires higher pyrolysis temperatures. The yields of benzene, toluene, and xylene need to be determined based on the user's actual needs. Generally, the demand for production is primarily for benzene, which has the highest yield.

[0100] "Low-cost operation" scheme: It has the lowest pyrolysis temperature among all schemes, which saves energy significantly, but its product yield may not be the highest.

[0101] The "Comprehensive Balanced" approach performs well across all four objectives. While it doesn't excel in any single area, it may be the most commercially attractive option overall.

[0102] This embodiment fully demonstrates the powerful flexibility of the method in this application. It provides a general optimization framework, allowing users to freely define and combine optimization objectives according to specific industrial application scenarios and business goals, thereby obtaining highly customized process optimization solutions with practical guiding value.

[0103] The solution provided in this application will be described below with reference to various preferred embodiments: The core idea of ​​this solution is to construct an automated and intelligent optimization closed loop encompassing "data acquisition → model building → integrated optimization → decision output." It replaces complex physicochemical mechanism models or time-consuming experimental trial and error with machine learning models, enabling rapid and accurate prediction of process outcomes. Furthermore, by utilizing multi-objective optimization algorithms, it globally searches within a broad space of process parameters and catalyst formulations to find the optimal solution set that simultaneously considers multiple competing objectives. Ultimately, this provides process designers and decision-makers with a scientific, flexible, and efficient optimization tool.

[0104] Specifically, the steps include the following: Data infrastructure construction and preprocessing: The starting point of the solution is to establish a high-quality, structured training dataset. This dataset is derived from historical experiments, literature, and patent data, and each record contains: Input parameters: comprehensively cover the three major categories of variables affecting the pyrolysis process: Raw material characteristics: such as the C, H, O, and N element content of biomass.

[0105] Process parameters: such as pyrolysis temperature, catalytic temperature, and catalyst / biomass mass ratio.

[0106] Catalyst properties: such as the type of supported metal, metal loading, specific surface area, and acidity.

[0107] Output parameters: These are the yields of various target aromatic hydrocarbons to be optimized, typically the yields of benzene, toluene, and xylene (BTX).

[0108] To improve model performance, the data needs to undergo rigorous preprocessing, including: Numerical standardization: Eliminating the differences in the dimensions of different parameters.

[0109] Categorical variable encoding: converting textual information such as "metal type" into numerical values.

[0110] Derivative feature engineering: Based on domain knowledge, calculate more physicochemically significant features such as the hydrogen-to-carbon ratio (H / C) and the oxygen-to-carbon ratio (O / C) to enhance the model's ability to learn intrinsic laws.

[0111] Establishment of a high-precision proxy model: Machine learning algorithms are used to learn complex input-output mappings from preprocessed data. This approach offers high model flexibility and can utilize: Random Forest: Strong resistance to overfitting and high training efficiency.

[0112] Gradient boosting trees (such as XGBoost): typically have extremely high prediction accuracy.

[0113] Support vector regression: suitable for small sample data, with good generalization ability.

[0114] Neural networks have the strongest ability to fit extremely complex nonlinear relationships.

[0115] To ensure model quality, a training set, validation set, and test set partitioning strategy is adopted. The training set is used for model fitting, the validation set is used to monitor the training process and implement early stopping to prevent overfitting, and finally, the generalization ability of the model is evaluated through an independent test set, thereby ensuring that the final surrogate prediction model has both high accuracy and high reliability.

[0116] Integrated multi-objective global optimization: Integrating the trained high-precision surrogate model with the NSGA-II multi-objective optimization algorithm is key to achieving automated global optimization. In this architecture: The surrogate model acts as a fast fitness evaluation function, predicting the BTX yield for any given combination of input parameters within milliseconds.

[0117] The NSGA-II algorithm, within a predefined feasible parameter space, drives iterative evolution of the population through mechanisms such as fast non-dominated sorting, crowding calculation, and elite retention strategies. Its optimization objective is set to simultaneously maximize (or minimize, depending on demand) the yield of multiple target aromatics.

[0118] A flexible and scalable optimization target system: The significant advantage of this approach lies in the flexibility and comprehensiveness of its optimization objectives. In addition to the core aromatics yield target, economic considerations can be seamlessly incorporated. Process cost objectives: For example, pyrolysis temperature is taken as an objective that needs to be minimized in order to directly correlate with and optimize energy consumption costs.

[0119] Results output and decision support: The final output of the optimization process is a Pareto optimal solution set. This solution set is not a single answer, but a collection of solutions that make different optimal trade-offs among multiple competing objectives. For example, the solution set may simultaneously include a "high benzene yield solution," an "equilibrium yield solution," and a "high toluene yield solution," as well as a comprehensive solution that considers low energy consumption. Decision-makers can then use this set, combined with real-time market demand and raw material prices, to select the most suitable production strategy.

[0120] Collaborative optimization and global search: It fundamentally solves the problem of simultaneously optimizing multiple competing objectives and can break out of local optima to achieve global optimization.

[0121] Significant cost reduction and efficiency improvement: Replacing a large number of physical experiments with "virtual experiments" greatly shortens the R&D cycle and reduces R&D costs.

[0122] Mechanism-independent and low-barrier: No need to construct complex kinetic equations, reducing the technical barriers to optimizing complex chemical processes.

[0123] Scientific and flexible decision-making: Provides multiple optimal trade-off solutions to support dynamic and flexible production decisions.

[0124] In summary, this application provides a complete, efficient, and intelligent method for optimizing biomass catalytic pyrolysis processes, offering a powerful technical tool for achieving precise control and maximizing the benefits of high-value utilization of biomass resources.

[0125] In some embodiments, this application also provides a biomass catalytic pyrolysis process optimization system, which can be used to execute the method embodiments of this application. For details not disclosed in the system embodiments of this application, please refer to the method embodiments of this application.

[0126] Specifically, refer to Figure 2 The biomass catalytic pyrolysis process optimization system provided in this application includes: Data acquisition module 21 is used to acquire training dataset, which includes multiple sets of input parameters related to the biomass catalytic pyrolysis process, and the yields of various target aromatic products corresponding to the multiple sets of input parameters. The input parameters include at least one of feedstock characteristics, process parameters and catalyst properties. The model training module 22 is used to train a machine learning prediction model based on the training dataset to establish a mapping relationship between the input parameters and the yields of the various target aromatic products. The optimization execution module 23 is used to integrate the machine learning prediction model with the multi-objective optimization algorithm, and based on the preset range of the input parameters, to call the machine learning prediction model through the multi-objective optimization algorithm to perform iterative optimization, with the goal of simultaneously optimizing the yield of each of the multiple target aromatic products, thereby obtaining a Pareto optimal solution set; the Pareto optimal solution set includes multiple sets of optimized input parameter combinations.

[0127] The optimization execution module is specifically configured to use the NSGA-II algorithm as the multi-objective optimization algorithm.

[0128] Below, for reference Figure 3 This describes an electronic device according to embodiments of the present application. Figure 3 A block diagram of an electronic device according to an embodiment of this application is illustrated.

[0129] like Figure 3 As shown, the electronic device 300 includes one or more processors 310 and memory 320.

[0130] The processor 310 may be a central processing unit (CPU) or other form of processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device 300 to perform desired functions.

[0131] The memory 320 may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory. The non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor 310 may execute the program instructions to implement the optimization methods of the biomass catalytic pyrolysis process of the various embodiments of this application described above, and / or other desired functions. Various contents, such as category correspondences, may also be stored in the computer-readable storage medium.

[0132] In one example, the electronic device 300 may also include an input device 330 and an output device 340, which are interconnected via a bus system and / or other forms of connection mechanism (not shown).

[0133] In addition, the input device 330 may also include, for example, a keyboard, mouse, interface, etc. The output device 340 can output various information to the outside, including analysis results, etc. The output device 340 may include, for example, a display, speaker, printer, and communication network and its connected remote output devices, etc.

[0134] Of course, for the sake of simplicity, Figure 3 Only some of the components of the electronic device relevant to this application are shown in this illustration; components such as buses, input / output interfaces, etc., are omitted. In addition, the electronic device may include any other suitable components depending on the specific application.

[0135] In addition to the methods and apparatus described above, embodiments of this application may also be computer program products, comprising computer program instructions that, when executed by a processor, cause the processor to perform the steps in the optimized methods of the biomass catalytic pyrolysis process according to various embodiments of this application as described in the "Exemplary Methods" section of this specification.

[0136] The computer program product can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of this application. The programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0137] Furthermore, embodiments of this application may also be computer-readable storage media storing computer program instructions thereon, which, when executed by a processor, cause the processor to perform the steps in the optimized method of the biomass catalytic pyrolysis process according to various embodiments of this application as described in the "Exemplary Methods" section above.

[0138] The computer-readable storage medium may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0139] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this application to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations thereof.

Claims

1. An optimization method for a biomass catalytic pyrolysis process, characterized in that, Includes the following steps: Obtain a training dataset, which includes multiple sets of input parameters related to the biomass catalytic pyrolysis process, and the yields of various target aromatic products corresponding to the multiple sets of input parameters. The input parameters include at least one of feedstock characteristics, process parameters, and catalyst properties. Based on the training dataset, a machine learning prediction model is trained to establish a mapping relationship between the input parameters and the yields of the various target aromatic products. The machine learning prediction model is integrated with a multi-objective optimization algorithm, wherein the machine learning prediction model is used to respond to the call of the multi-objective optimization algorithm to predict the yield of each of the multiple target aromatic products under given input parameters; Based on the preset range of the input parameters, the machine learning prediction model is called through the multi-objective optimization algorithm to perform iterative optimization, with the goal of simultaneously optimizing the yield of each of the multiple target aromatic products, thereby obtaining a Pareto optimal solution set; the Pareto optimal solution set includes multiple sets of optimized input parameter combinations.

2. The method according to claim 1, characterized in that, The raw material characteristics in the input parameters include at least one of the carbon, hydrogen, oxygen, and nitrogen content of biomass; the process parameters include at least one of the pyrolysis temperature, catalytic temperature, and mass ratio of catalyst to biomass; the catalyst properties include at least one of the type of metal loaded on the catalyst, metal loading, specific surface area, and acidity.

3. The method according to claim 1, characterized in that, Before training the machine learning prediction model, the method also includes a step of preprocessing the training dataset; The preprocessing steps include at least one of the following steps: Normalize or standardize the numerical features in the training dataset; Label the categorical variables.

4. The method according to claim 3, characterized in that, The preprocessing step further includes: calculating derived features based on domain knowledge, and merging the derived features with the original features as a new input feature set for model training; The derived characteristics include at least one of the hydrogen-to-carbon atomic ratio and the oxygen-to-carbon atomic ratio of the raw material.

5. The method according to claim 1, characterized in that, The machine learning prediction model is one of the following: random forest model, gradient boosting tree model, support vector regression model, or neural network model.

6. The method according to claim 1, characterized in that, The specific steps for training the machine learning prediction model include: The training dataset is divided into a training set, a validation set, and a test set; The machine learning prediction model is trained using the training set, and the model performance is monitored during training using the validation set to prevent overfitting.

7. The method according to claim 1, characterized in that, The multi-objective optimization algorithm is the NSGA-II algorithm.

8. The method according to claim 1, characterized in that, in, The optimization objectives of the multi-objective optimization algorithm also include parameters related to process cost; The parameter related to the process cost is the pyrolysis temperature that needs to be minimized.

9. A biomass catalytic pyrolysis process optimization system, characterized in that, include: The data acquisition module is used to acquire a training dataset, which includes multiple sets of input parameters related to the biomass catalytic pyrolysis process, and the yields of various target aromatic products corresponding to the multiple sets of input parameters. The input parameters include at least one of feedstock characteristics, process parameters, and catalyst properties. The model training module is used to train a machine learning prediction model based on the training dataset to establish a mapping relationship between the input parameters and the yields of the various target aromatic products. The optimization execution module is used to integrate the machine learning prediction model with a multi-objective optimization algorithm, and based on the preset range of the input parameters, to call the machine learning prediction model through the multi-objective optimization algorithm to perform iterative optimization, with the goal of simultaneously optimizing the yield of each of the multiple target aromatic products, thereby obtaining a Pareto optimal solution set; the Pareto optimal solution set includes multiple sets of optimized input parameter combinations.

10. The system according to claim 9, characterized in that, The optimization execution module is specifically configured to use the NSGA-II algorithm as the multi-objective optimization algorithm.

Images