A fixed bed reactor diffusion behavior analysis method based on partition SHAP analysis

Abstract

The application discloses a fixed bed reactor diffusion behavior analysis method based on partition SHAP analysis, comprising the following steps: step S1, determining a feature set affecting the performance of a reactor, and constructing a structured data set; step S2, constructing a candidate proxy model set covering at least two types of regression models, and selecting a proxy model with the optimal comprehensive evaluation index; step S3, global Shapley value attribution; step S4, external diffusion control interval identification; step S5, internal diffusion influence strength analysis; and step S6, diffusion control interval output. Through two-stage sliding window partition SHAP analysis along the reactor operation parameter axis related to external diffusion control and the catalyst particle structure parameter axis related to internal diffusion control, the method realizes data-driven identification of the external diffusion control interval and internal diffusion influence strength analysis under multi-dimensional parameter coupling.

Description

Technical Field

[0001] This invention relates to the field of intelligent analysis technology for fixed-bed reactors, and in particular to a method for analyzing the diffusion behavior of fixed-bed reactors based on sliding window partitioning (SHAP) analysis. Background Technology

[0002] Fixed-bed reactors are among the most widely used reactor types in chemical production, with large-scale industrial applications in major chemical processes such as propane dehydrogenation, methane dry reforming, acetylene hydrogenation, and methanol-to-olefins. During the transition from ideal kinetic testing to industrial operation, the reactor gradually evolves from a stage largely unconstrained by internal and external diffusion to a stage dominated by both. Accurately identifying the critical point where the influence of external diffusion constraints becomes negligible and analyzing the impact of catalyst particle structure parameters on reactor operating parameters under different degrees of diffusion restriction is crucial for bridging reaction kinetics research and industrial reactor design.

[0003] Existing methods for identifying external diffusion control zones have three main shortcomings. First, traditional empirical criteria such as the Mears number can only provide qualitative conclusions about the diffusion zone at a single operating point, making it difficult to characterize the continuous change of the diffusion control zone under multivariate coupled operating conditions. Second, traditional empirical criteria rely on manual estimation of parameters such as reaction order and effective diffusion coefficient at the particle size, and the given critical points are highly sensitive to deviations in parameter estimation. Third, although recent studies have used machine learning methods and global SHAP analysis to reveal the mapping relationship between input features and reactor performance, global SHAP can only reflect the average contribution of each feature across the entire dataset, failing to capture the local evolution of parameter contributions with changing operating conditions, and also unable to directly identify the critical point where the reactor enters a negligible diffusion control zone.

[0004] Therefore, there is an urgent need for a data-driven identification method that can simultaneously utilize the efficient predictive capabilities and interpretable analytical capabilities of machine learning proxy models to quantitatively characterize the continuous evolution of the reactor diffusion control range under multidimensional parameter coupling. Summary of the Invention

[0005] The purpose of this invention is to overcome the technical deficiencies in the prior art and provide a method for analyzing the diffusion behavior of a fixed-bed reactor based on sliding window partitioning (SHAP) analysis.

[0006] The technical solution adopted to achieve the purpose of this invention is: A method for analyzing diffusion behavior in a fixed-bed reactor based on sliding window partitioning (SHAP) analysis includes the following steps: Step S1: Determine the set of features that affect reactor performance. The features include at least one reactor operating parameter related to external diffusion control and one catalyst particle structure parameter related to internal diffusion control. Sample the feature and reactor performance parameter data from the reactor numerical model or reactor physical experiment to construct a structured dataset. Then, perform feature standardization and divide the dataset into training and test sets. Step S2: Construct a set of candidate surrogate models covering at least two types of regression models. For each candidate model, use a cross-validation strategy to optimize hyperparameters. For each reactor performance index, select the surrogate model with the best comprehensive evaluation index as the optimal surrogate model. Step S3: Calculate the Shapley contribution value of each feature of each sample on the training set or test set for the optimal surrogate model after training, obtain the global contribution ratio of each feature on the entire dataset to the reactor performance index, and identify the reactor operation parameter features and catalyst particle structure parameter features that play a dominant role in the reactor performance index. Step S4, External diffusion control zone identification: Divide the dataset into several continuous sliding windows along the reactor operating parameter axis that plays a dominant role. Calculate the local Shapley contribution ratio of each feature independently within each window. Based on the trend of the local contribution ratio of the reactor operating parameter features decreasing from dominant to near zero as the window progresses, identify the external diffusion critical point where the reactor enters the external diffusion negligible zone. Perform independent cross-validation between the critical point identification results and traditional physical criteria. Step S5, Internal diffusion influence intensity analysis: Based on the external diffusion critical point obtained in step S4, a high convection subset where external diffusion control is negligible is selected from the dataset. Within the high convection subset, several continuous sliding windows are divided along the dominant catalyst particle structure parameter axis. The evolution of the local Shapley contribution ratio of each feature to the reactor performance index along the catalyst particle structure parameter axis is independently calculated. The influence intensity of internal diffusion on reactor performance under different particle structure conditions is analyzed. Step S6: Summarize and output the external diffusion critical point and the local contribution ratio of each feature under different particle structure conditions.

[0007] In the above technical solution, in step S1, the reactor operating parameters related to external diffusion control include mass hourly space velocity, volume hourly space velocity, apparent gas velocity, or flow rate; the catalyst particle structure parameters related to internal diffusion control include catalyst particle size, porosity, pore size, shape factor, or pore tortuosity factor. In the external diffusion control zone identification stage of step S4, the reactor performance index is a reaction rate-like physical quantity that reflects the speed of particle-scale reaction, including the target product generation rate or reactant consumption rate. In the internal diffusion influence intensity analysis stage of step S5, the reactor performance index is the target product generation rate or the internal diffusion effectiveness factor reflecting the utilization rate of catalyst particles.

[0008] In the above technical solution, in step S1, the sampling method of the structured dataset includes one or more of orthogonal design, mixed horizontal orthogonal design, Latin hypercube design, or optimal space-filling design; the sampling method should ensure the balanced coverage of the feature space by the dataset and sufficient interactive combinations between different features.

[0009] In the above technical solution, in step S2, the regression model categories of the candidate surrogate model set include linear models, nearest neighbor models, kernel method models, ensemble tree models, probabilistic regression models, and neural network models; the candidate surrogate models include elastic networks, k-nearest neighbor regression, support vector regression, random forests, gradient boosting regression, extreme random trees, Gaussian process regression, or multilayer perceptrons.

[0010] In the above technical solution, in step S2, the hyperparameter optimization uses the cross-validation set determination coefficient as the evaluation index, and the hyperparameter configuration with the largest value of the index is taken as the optimal hyperparameter of the candidate model; when selecting the optimal surrogate model for each reactor performance index, one or more of the test set determination coefficient, test set root mean square error, and cross-validation set determination coefficient are comprehensively considered as a comprehensive evaluation index.

[0011] In the above technical solution, the determination of the external diffusion control critical point in step S4 includes the following three conditions: (1) The local contribution of reactor operating parameter characteristics related to external diffusion control decreased from the initial dominant level to near zero; (2) The contribution of this feature has decreased to the lowest level among all features or is lower than the median contribution of the remaining features; (3) The proportion of local contributions of the remaining features tends to stabilize along the window advancement direction, and the proportion of overall feature contributions enters a plateau period.

[0012] In the above technical solution, in step S4, the traditional physical criterion includes the Mears criterion.

[0013] In another aspect of the present invention, a system for implementing the method for analyzing diffusion behavior in a fixed-bed reactor based on sliding window partitioning (SHAP) analysis includes: The dataset construction module executes step S1 and outputs a structured feature-label dataset; The proxy model training module executes step S2 and outputs the optimal proxy model corresponding to each performance index. The global Shapley value attribution module executes step S3 and outputs the global contribution percentage of each feature to the performance index. The external diffusion control zone identification module executes step S4 and outputs the reactor's external diffusion control critical point. The internal diffusion influence intensity analysis module executes step S5 and outputs the local contribution ratio of each feature under different particle structure windows; The diffusion control interval output module executes step S6 and outputs the diffusion control interval identification results and optimization suggestions.

[0014] In another aspect of the present invention, a fixed-bed reactor diffusion behavior analysis apparatus is provided, the apparatus comprising a processor and a memory; the memory stores a program or instructions, which are loaded and executed by the processor to implement the fixed-bed reactor diffusion behavior analysis method based on sliding window partitioning (SHAP) analysis.

[0015] In another aspect of the present invention, a computer-readable storage medium stores a program or instructions that, when executed by a processor, implement the method for analyzing diffusion behavior in a fixed-bed reactor based on sliding window partitioning (SHAP) analysis.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This method uses a two-stage sliding window partitioning (SHAP) analysis along the reactor operating parameter axis related to external diffusion control and the catalyst particle structure parameter axis related to internal diffusion control. It decouples the contribution ratio of each feature from the average value on the entire dataset to a local evolution curve that changes continuously along the operating condition axis, thereby realizing data-driven identification of the external diffusion control interval and analysis of the intensity of internal diffusion influence under multi-dimensional parameter coupling. Attached Figure Description

[0017] Figure 1 This is a schematic flowchart of the fixed-bed reactor diffusion behavior analysis method based on partitioned SHAP analysis provided in an embodiment of the present invention.

[0018] Figure 2 This is a schematic diagram of the structure of the fixed-bed reactor diffusion behavior analysis system based on partitioned SHAP analysis provided in an embodiment of the present invention.

[0019] Figure 3 The diagram shows a sliding window partitioning SHAP analysis along the reactor operating parameter axis related to external diffusion control in an embodiment of the present invention (a) and a partitioning SHAP analysis along the catalyst particle structure parameter axis related to internal diffusion control within a high convection subset (b). Detailed Implementation

[0020] The present invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0021] Example 1 like Figure 1 As shown, a method for analyzing diffusion behavior in a fixed-bed reactor based on sliding window partitioning (SHAP) analysis includes the following steps: Step S1, Dataset Construction: A set of features affecting reactor performance indicators is determined. This set includes at least one reactor operating parameter related to external diffusion control and one catalyst particle structure parameter related to internal diffusion control. Other features can be flexibly expanded according to the specific reaction system. The reactor operating parameter related to external diffusion control can be a physical quantity characterizing the convection intensity within the reactor, such as mass hourly space velocity (MHV), volumetric hourly space velocity (VHSV), apparent gas velocity, and flow rate. The catalyst particle structure parameter related to internal diffusion control can be a physical quantity characterizing the transport characteristics within the particle size, such as catalyst particle size, porosity, pore size, shape factor, and pore tortuosity factor. The reactor performance indicators are selected according to different analysis stages: In the external diffusion control zone identification stage, the performance indicators must reflect the reaction speed at the particle size and be directly related to the strength of external diffusion control, including but not limited to intrinsic reaction rate physical quantities such as target product formation rate and reactant consumption rate; in the internal diffusion influence intensity analysis stage, the performance indicators can be the target product formation rate or physical quantities such as the internal diffusion effectiveness factor reflecting catalyst particle utilization. Based on existing reactor numerical models or reactor physical experimental data, the aforementioned feature space is sampled to obtain sample pairs of operating conditions and corresponding performance indicators, forming a structured dataset. The sampling method should ensure the dataset's balanced coverage of the feature space and sufficient interactive combinations between different features; it can be one or more of the following: orthogonal design, mixed-level orthogonal design, Latin hypercube design, and optimal space-filling design. The structured dataset is then subjected to feature standardization and divided into training and test sets according to a preset ratio.

[0022] Step S2, Agent Model Training: Construct a candidate model set covering at least two types of regression models, including but not limited to linear models, nearest neighbor models, kernel method models, ensemble tree models, probabilistic regression models, and neural network models. Candidate models may include, but are not limited to, elastic networks, k-nearest neighbor regression, support vector regression, random forests, gradient boosting regression, extreme random trees, Gaussian process regression, and multilayer perceptrons.

[0023] For each candidate model, a cross-validation strategy is employed to optimize hyperparameters. The cross-validation set determination coefficient is used as the hyperparameter evaluation metric, and the hyperparameter configuration with the highest determination coefficient is selected as the optimal hyperparameter for that candidate model. For each reactor performance index to be predicted, a horizontal comparison is conducted among all candidate models. One or more of the following are considered as a comprehensive evaluation metric: test set determination coefficient, test set root mean square error, and cross-validation set determination coefficient. The surrogate model with the best comprehensive evaluation metric is selected as the optimal surrogate model for that index. The optimal surrogate models are independent of each other across different performance indices and do not require the same algorithm to be used.

[0024] Step S3, Global Shapley Value Attribution: The Shapley contribution value of each feature for each sample on the training or test set is calculated for the optimal surrogate model after training, obtaining the global contribution ratio of each feature to the reactor performance index across the entire dataset. The Shapley value calculation methods include, but are not limited to, Kernel SHAP, Tree SHAP, Deep SHAP, and precise Shapley value calculation. Based on the global contribution ratio, the importance of each feature is ranked, identifying the features that play a dominant role and those that play a secondary role in the reactor performance index, providing a global background reference for subsequent partitioned SHAP analysis.

[0025] Step S4, Identification of the external diffusion control region: The dataset is divided into several continuous sliding windows along the axis of reactor operating parameters related to external diffusion control, with no specific limitations on window width, step size, and overlap rate. Within each sliding window, the local Shapley contribution ratio of each feature is independently calculated for the selected optimal surrogate model. The local contribution ratios within each window are arranged according to window position, yielding the evolution curve of each feature's contribution ratio along the axis of reactor operating parameters related to external diffusion control. The evolution curve of the local contribution ratio of reactor operating parameter features related to external diffusion control is examined along the window advancement direction. When the feature's contribution ratio simultaneously meets the following three conditions, the corresponding window position is taken as the critical point for the reactor to enter the negligible external diffusion range: 1) The feature's contribution ratio decreases from its initial dominant level to near zero, meaning that the feature's local contribution to reactor performance prediction no longer has a significant driving effect; 2) The feature's contribution ratio has decreased to the lowest level among all features or is lower than the median level of the contribution ratios of the remaining features; 3) The local contribution ratios of the remaining features tend to stabilize along the window advancement direction, and the overall feature contribution ratio enters a plateau period. The above critical point identification results are independently cross-validated with traditional empirical criteria, which may be the Mears number or other physical criteria used to determine whether external diffusion control is negligible.

[0026] Step S5, Analysis of the Intensity of Internal Diffusion Influence: Based on the external diffusion critical point obtained in the previous step, a subset of the dataset where external diffusion control is negligible is selected, denoted as the high convection subset. Within the high convection subset, several continuous sliding windows are divided along the axis of catalyst particle structure parameters related to internal diffusion control. Within each window, the local Shapley contribution of each feature to the reactor performance indicators is independently calculated, further analyzing the relative effects of catalyst particle structure parameters and reactor operating parameters under different degrees of internal diffusion constraint. For small sample scenarios, a virtual single-factor scanning curve of continuously changing catalyst particle structure parameters can be constructed based on a trained surrogate model under fixed other features to avoid statistical fluctuations of discrete sample points masking the true variation of the catalyst particle structure parameter effect.

[0027] Step S6, Diffusion control interval output: The external diffusion critical point and the local contribution ratio of each feature under different particle structure conditions are summarized and output as the final result of reactor diffusion control zone identification and internal diffusion influence intensity analysis.

[0028] Example 2 This embodiment, based on Example 1, takes a propane dehydrogenation fixed-bed reactor as an example to analyze its diffusion behavior.

[0029] A method for analyzing the diffusion behavior of a propane dehydrogenation fixed-bed reactor includes the following steps: Step S1, Dataset Construction: In this embodiment, the reactor operating parameters related to external diffusion control are mass hourly space velocity (MHV), reaction temperature, and reaction pressure; the catalyst particle structure parameters related to internal diffusion control are catalyst particle size, porosity, and pore size, forming a six-dimensional input feature space. In the external diffusion control interval identification stage, the reactor performance index is selected based on the propylene formation rate; in the internal diffusion influence intensity analysis stage, both the propylene formation rate and the internal diffusion effectiveness factor are selected. A hybrid horizontal orthogonal design is used to generate a working condition configuration table within the six-dimensional feature space. For each working condition, a validated particulate analytical reactor numerical model is called for batch solutions, outputting the true labels of each reactor performance index, resulting in a structured dataset containing 240 working conditions. The structured dataset is then subjected to feature standardization, and the training and test sets are divided in a 9:1 ratio.

[0030] Step S2, Agent Model Training: A candidate model set covering eight representative regression algorithms across six categories—linear models, nearest neighbor models, kernel method models, ensemble tree models, probabilistic regression models, and neural network models—was constructed. The candidate models included elastic networks, k-nearest neighbor regression, support vector regression, random forests, gradient boosting regression, extreme random trees, Gaussian process regression, and multilayer perceptrons. For each candidate model, a five-fold cross-validation strategy was employed to optimize hyperparameters, using the cross-validation set coefficient of determination as the evaluation metric. For each reactor performance index to be predicted, a horizontal comparison was conducted among all candidate models. The comprehensive evaluation metric considered the test set coefficient of determination, the test set root mean square error, and the cross-validation set coefficient of determination. For the propylene production rate, Gaussian process regression was selected as the optimal surrogate model; similarly, for the internal diffusion efficiency factor, Gaussian process regression was also selected as the optimal surrogate model. Both surrogate models achieved coefficients of determination exceeding 0.99 on the test set.

[0031] Step S3, Global Shapley Value Attribution: For the Gaussian process regression surrogate model, the Kernel SHAP method was used to calculate the Shapley contribution value of each feature for each sample in the training set, obtaining the global contribution ratio of each feature to the reactor performance index across the entire dataset. The analysis results show that for the propylene formation rate, the overall reactor performance index, mass hourly space velocity (MHSV) contributes 56.7% and temperature contributes 13.1%, together dominating the contribution. For the internal diffusion efficiency factor, the relative contributions of the three catalyst particle structure parameters—particle size, porosity, and pore size—are 33.3%, 21.5%, and 16.8%, respectively, totaling 71.6%, while MHSV and temperature contribute only 2.3% and 1.3%, respectively, totaling less than 4%. This confirms that the internal diffusion efficiency factor is a particle-scale intrinsic property dominated by particle geometry and pore structure.

[0032] Step S4, Identification of the external diffusion control region: Along the dominant reactor operating parameter axis (mass hourly space velocity axis in this embodiment), the 240 samples were divided into five consecutive sliding windows, corresponding to five intervals: low hourly space velocity, low-medium hourly space velocity, medium hourly space velocity, medium-high hourly space velocity, and high hourly space velocity. Within each sliding window, the local Shapley contribution ratio of the six-dimensional features was independently calculated for the optimal surrogate model. Figure 3 The analysis results in (a) show that the local contribution of mass space velocity (MSV) decreased from 61.36% in the lowest space velocity window to 2.25% in the highest space velocity window, exhibiting a trend of changing from a dominant feature to a peripheral feature; at the same time, the contribution of temperature increased from 13.83% to 28.36% and stabilized as the dominant factor, and the total contribution of catalyst particle structure parameters also increased significantly from the lower level in the low space velocity window. When the space velocity enters window 5 (i.e., exceeding 150 h), -1After this, the contribution of mass space velocity (MSV) decreased to 2.25%, while the combined contributions of temperature and catalyst particle structure parameters stabilized, and the contribution ratios of each characteristic entered a plateau period. Based on the combined criteria of MSV contribution ratio decreasing to near zero, reaching its minimum, and overall characteristic contribution ratio stabilizing, the critical space velocity for the reactor to enter the negligible external diffusion range was identified as approximately 150 h⁻¹. -1 The Mears criterion was calculated under various operating conditions based on the particle analytical numerical model. The critical space velocity for external diffusion given by the Mears criterion was compared with the critical space velocity identified by the data-driven method. The two showed a high degree of agreement, which mutually confirmed the reliability of the identification results from two independent paths: data-driven and physical mechanism.

[0033] Step S5, Analysis of the Intensity of Internal Diffusion Influence: Based on the external diffusion critical space velocity obtained in the previous step, the mass space velocity is higher than 150 h⁻¹. -1 The samples were screened to form a high-convection subset. Within this subset, three continuous sliding windows—small, medium, and large particle sizes—were divided along the dominant catalyst particle structure parameter axis (particle size axis in this example). Partitioned SHAP analysis was conducted with propylene formation rate and internal diffusion efficiency as objective functions, respectively. The analysis results with propylene formation rate as the objective function showed that... Figure 3 In (b), the combined contribution of porosity and pore size monotonically increased from 30.3% in the small particle size window to 44.5% in the large particle size window, while the contribution of temperature decreased from 32.4% to 26.9%. The influence of catalyst particle structure parameters on the reaction rate significantly increased with increasing particle size. Analysis targeting the internal diffusion efficiency factor showed that the combined contribution of catalyst particle structure parameters remained above 50% across the three particle size windows. Furthermore, based on a Gaussian process regression surrogate model, virtual single-factor scans of particle size were conducted at four different temperature levels. The average reaction rate decreased with increasing particle size, with a decrease of approximately 42% at 873 K. The virtual scan curves of the internal diffusion efficiency factor almost completely overlapped at the four temperature levels, and the internal diffusion efficiency factor monotonically decreased from 0.93 to 0.40 with increasing particle size, indicating that the effect of temperature change was negligible.

[0034] Results Analysis. The critical space velocity of 150 h⁻¹ obtained in this embodiment during the external diffusion control zone identification phase. -1 The results are consistent with the traditional Mears criterion, confirming the physical reliability of the data-driven identification method based on partitioned SHAP. The evolution of the contribution ratio of catalyst particle structure parameters obtained in the internal diffusion influence intensity analysis stage along the particle size axis reveals the gradual emergence of the catalyst pore structure regulation space within the significant range of internal diffusion influence, providing a quantitative basis for the pore structure optimization of large-particle-size industrial catalysts.

[0035] As can be seen from the above embodiments, the method of the present invention, through two-stage sliding window partitioning (SHAP) analysis along the reactor operating parameter axis related to external diffusion control and the catalyst particle structure parameter axis related to internal diffusion control, decouples the contribution ratio of each feature from the average value on the entire dataset to a local evolution curve that continuously varies along the operating condition axis. This can quantitatively characterize the continuous evolution process of the diffusion control interval with the operating condition, making up for the deficiency of traditional empirical criteria in the ability to distinguish under multivariate coupled operating conditions. Once the surrogate model of this method is trained, the prediction time for a single operating condition can be compressed from several hours in particle analytical numerical simulation to seconds, which can significantly reduce the computational cost of high-dimensional parameter space scanning. It can be extended to a variety of strongly exothermic or strongly endothermic heterogeneous catalytic reaction systems and has broad application prospects.

[0036] Example 3 like Figure 2 As shown, this embodiment, based on Embodiment 1, constructs a fixed-bed reactor diffusion behavior analysis system based on partitioned SHAP analysis, including: The dataset construction module 110 executes step S1 and outputs a structured feature-label dataset; The agent model training module 120 executes step S2 and outputs the optimal agent model corresponding to each performance index. The global Shapley value attribution module 130 executes step S3 and outputs the global contribution ratio of each feature to the performance index. The external diffusion control zone identification module 140 executes step S4 and outputs the reactor external diffusion control critical point. The internal diffusion influence intensity analysis module 150 executes step S5 and outputs the local contribution ratio of each feature under different particle structure windows; The diffusion control interval output module 160 executes step S6 and outputs the diffusion control interval identification result and optimization suggestions.

[0037] The above description is only a preferred embodiment of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for analyzing diffusion behavior in a fixed-bed reactor based on sliding window partitioning (SHAP) analysis, characterized in that, Includes the following steps: Step S1: Determine the set of features that affect reactor performance. The features include at least one reactor operating parameter related to external diffusion control and one catalyst particle structure parameter related to internal diffusion control. Sample the feature and reactor performance parameter data from the reactor numerical model or reactor physical experiment to construct a structured dataset. Then, perform feature standardization and divide the dataset into training and test sets. Step S2: Construct a set of candidate surrogate models covering at least two types of regression models. For each candidate model, use a cross-validation strategy to optimize hyperparameters. For each reactor performance index, select the surrogate model with the best comprehensive evaluation index as the optimal surrogate model. Step S3: Calculate the Shapley contribution value of each feature of each sample on the training set or test set for the optimal surrogate model after training, obtain the global contribution ratio of each feature on the entire dataset to the reactor performance index, and identify the reactor operation parameter features and catalyst particle structure parameter features that play a dominant role in the reactor performance index. Step S4, External diffusion control zone identification: Divide the dataset into several continuous sliding windows along the reactor operating parameter axis that plays a dominant role. Calculate the local Shapley contribution ratio of each feature independently within each window. Based on the trend of the local contribution ratio of the reactor operating parameter features decreasing from dominant to near zero as the window progresses, identify the external diffusion critical point where the reactor enters the external diffusion negligible zone. Perform independent cross-validation between the critical point identification results and traditional physical criteria. Step S5, Internal diffusion influence intensity analysis: Based on the external diffusion critical point obtained in step S4, a high convection subset where external diffusion control is negligible is selected from the dataset. Within the high convection subset, several continuous sliding windows are divided along the dominant catalyst particle structure parameter axis. The evolution of the local Shapley contribution ratio of each feature to the reactor performance index along the catalyst particle structure parameter axis is independently calculated. The influence intensity of internal diffusion on reactor performance under different particle structure conditions is analyzed. Step S6: Summarize and output the external diffusion critical point and the local contribution ratio of each feature under different particle structure conditions.

2. The method for analyzing diffusion behavior in a fixed-bed reactor based on sliding window partitioning (SHAP) analysis as described in claim 1, characterized in that, In step S1, the reactor operating parameters related to external diffusion control include mass hourly space velocity, volume hourly space velocity, apparent gas velocity, or flow rate; the catalyst particle structure parameters related to internal diffusion control include catalyst particle size, porosity, pore size, shape factor, or pore tortuosity factor. In the external diffusion control zone identification stage of step S4, the reactor performance index is a reaction rate-like physical quantity that reflects the speed of particle-scale reaction, including the target product generation rate or reactant consumption rate. In the internal diffusion influence intensity analysis stage of step S5, the reactor performance index is the target product generation rate or the internal diffusion effectiveness factor reflecting the utilization rate of catalyst particles.

3. The method for analyzing diffusion behavior in a fixed-bed reactor based on sliding window partitioning (SHAP) analysis as described in claim 1, characterized in that, In step S1, the sampling method of the structured dataset includes one or more of the following: orthogonal design, mixed horizontal orthogonal design, Latin hypercube design, or optimal space-filling design; the sampling method should ensure that the dataset has balanced coverage of the feature space and that there are sufficient interactive combinations between different features.

4. The method for analyzing diffusion behavior in a fixed-bed reactor based on sliding window partitioning (SHAP) analysis as described in claim 1, characterized in that, In step S2, the regression model categories of the candidate surrogate model set include linear models, nearest neighbor models, kernel method models, ensemble tree models, probabilistic regression models, and neural network models; the candidate surrogate models include elastic networks, k-nearest neighbor regression, support vector regression, random forests, gradient boosting regression, extreme random trees, Gaussian process regression, or multilayer perceptrons.

5. The method for analyzing diffusion behavior in a fixed-bed reactor based on sliding window partitioning (SHAP) analysis as described in claim 1, characterized in that, In step S2, the hyperparameter optimization uses the cross-validation set determination coefficient as the evaluation index, and the hyperparameter configuration with the largest value of the index is taken as the optimal hyperparameter of the candidate model. When selecting the optimal surrogate model for each reactor performance index, one or more of the test set determination coefficient, test set root mean square error, and cross-validation set determination coefficient are considered as comprehensive evaluation indexes.

6. The method for analyzing diffusion behavior in a fixed-bed reactor based on sliding window partitioning (SHAP) analysis as described in claim 1, characterized in that, In step S4, the determination of the external diffusion control critical point includes the following three conditions: (1) The local contribution of reactor operating parameter characteristics related to external diffusion control decreased from the initial dominant level to near zero; (2) The contribution of this feature has decreased to the lowest level among all features or is lower than the median contribution of the remaining features; (3) The proportion of local contributions of the remaining features tends to stabilize along the window advancement direction, and the proportion of overall feature contributions enters a plateau period.

7. The method for analyzing diffusion behavior in a fixed-bed reactor based on sliding window partitioning (SHAP) analysis as described in claim 1, characterized in that, In step S4, the traditional physical criteria include the Mears criterion.

8. A system for implementing the fixed-bed reactor diffusion behavior analysis method based on sliding window partitioning SHAP analysis as described in any one of claims 1 to 7, characterized in that, include: The dataset construction module executes step S1 and outputs a structured feature-label dataset; The proxy model training module executes step S2 and outputs the optimal proxy model corresponding to each performance index. The global Shapley value attribution module executes step S3 and outputs the global contribution percentage of each feature to the performance index. The external diffusion control zone identification module executes step S4 and outputs the reactor's external diffusion control critical point. The internal diffusion influence intensity analysis module executes step S5 and outputs the local contribution ratio of each feature under different particle structure windows; The diffusion control interval output module executes step S6 and outputs the diffusion control interval identification results and optimization suggestions.

9. A device for analyzing the diffusion behavior of a fixed-bed reactor, characterized in that, The device includes a processor and a memory; the memory stores a program or instructions, which are loaded and executed by the processor to implement the fixed-bed reactor diffusion behavior analysis method based on sliding window partitioning SHAP analysis as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the fixed-bed reactor diffusion behavior analysis method based on sliding window partitioning SHAP analysis as described in any one of claims 1 to 7.

Images