Quality evaluation method applied to tank type batch finished gasoline blending formula

Abstract

The invention discloses a quality evaluation method applied to a tank-type batch finished gasoline blending formula, and the method comprises the following steps: obtaining quality index data in a plurality of tank-type batch finished gasoline blending formulas, carrying out the preprocessing, and determining the output variable of the quality evaluation; an IPRFCM algorithm is adopted to calculate a membership matrix, and offline batch data division is carried out; constructing sub-models adapted to each batch of data by using an NGBoost algorithm, and performing dynamic adaptive fusion on all the sub-models in combination with the membership matrix to obtain an IPRFCM-NGBoost adaptive integrated model; and inputting a to-be-processed tank-type batch finished gasoline blending formula into the IPRFCM-NGBoost adaptive integrated model, and performing evaluation by using national VIB and enterprise production standards to obtain a corresponding quality evaluation result. According to the method, through efficient data preprocessing, accurate batch division and clustering, high-performance sub-model construction and fusion and strong cause analysis capability, the accuracy and interpretability of finished gasoline blending formula quality prediction are remarkably improved.

Description

Technical Field

[0001] The invention relates to the field of tank-type batch finished gasoline blending formula, and more specifically to a quality evaluation method applied to tank-type batch finished gasoline blending formula. Background Art

[0002] In the petroleum refining industry, finished gasoline blending is the process of mixing different types of petroleum components and additives to produce high-quality gasoline with specific performance standards and in line with market demand. With the continuous improvement of the country's gasoline quality standards and the limitations of measurement and control technology upgrades, many refining companies are facing many challenges.

[0003] Among them, although the tank-type online blending method can guide the addition flow of components through ratio closed-loop control, when the online analyzer fails, the system will switch to open-loop control, which directly leads to the quality of the blending formula becoming a key factor affecting the quality of finished oil. In addition, the traditional blending formula is still generated based on linear empirical ratios, which is inconsistent with the nonlinear characteristics of the blending process itself. Coupled with the changes in the quality of primary product components caused by differences in raw material properties, the blending process becomes extremely complicated. Therefore, companies often use more conservative formulas to ensure product quality, which not only causes excess product quality, but also increases production costs.

[0004] Currently, in the quality assessment process of tank batch finished gasoline blending formula, the purity of the blending formula data set has become a major challenge. Unqualified samples mixed in the data, such as blending failure cases and records that do not meet national standards, will undoubtedly erode the performance foundation of the model if they are not properly cleaned up before being put into model training. In addition, the dimensional differences between data features are like obstacles on the model optimization path, and an effective normalization processing strategy is urgently needed to smooth them out to ensure the smoothness and efficiency of the training process.

[0005] Furthermore, the current method is unable to accurately identify key quality indicators and fails to accurately identify factors that play a decisive role in the quality of finished gasoline. This omission not only slows down the model's prediction speed, but also limits its ability to capture core blending variables. At the same time, the fuzzy boundaries between blending component batches are ignored, resulting in insufficient batch classification accuracy, which indirectly weakens the accuracy of model predictions.

[0006] At the level of model building and integration, the existing fusion strategy is too rigid and lacks the flexibility to dynamically adjust the fusion weight according to batch characteristics, which to some extent limits the further improvement of model prediction accuracy. In addition, the processing of prediction results is often limited to the pursuit of a single accuracy, while ignoring the inherent uncertainty in the prediction of finished gasoline quality, which may lead to increased decision-making risks in practical applications.

[0007] Finally, the explanatory shortcomings of existing technical solutions cannot be ignored. The lack of an analytical framework that covers both the overall and local aspects makes it difficult for technicians to fully understand the key factors affecting gasoline quality, which in turn limits the pertinence and effectiveness of subsequent optimization measures.

[0008] Therefore, how to design a quality evaluation method for tank batch finished gasoline blending formula, comprehensively optimize the data preprocessing process, accurately screen key quality indicators, build a flexible and efficient model integration system, improve the performance and robustness of the prediction model, and enhance the model's explanatory ability to provide solid support for the continuous optimization of gasoline blending formula is an urgent problem that technical personnel in this field need to solve. Summary of the invention

[0009] In view of this, the present invention provides a quality evaluation method applied to tank batch finished gasoline blending formula, which not only needs to be able to accurately predict the quality indicators of finished gasoline, but also needs to have good interpretability so that technicians can understand the logic behind the prediction results and make more reasonable decisions.

[0010] In order to achieve the above object, the present invention adopts the following technical solution:

[0011] A quality evaluation method for tank batch finished gasoline blending formula includes the following steps:

[0012] S1. Obtaining multiple tank-type batch finished gasoline blending formulas, preprocessing the gasoline blending formulas, and determining output variables for quality evaluation;

[0013] S2. Based on the pre-processed gasoline blending formula, the membership matrix is ​​calculated using the IPRFCM algorithm to perform offline batch data division;

[0014] S3. Use the NGBoost algorithm to build a sub-model that adapts to each batch of data, and dynamically and adaptively fuse all sub-models in combination with the membership matrix to obtain the IPRFCM-NGBoost adaptive integration model;

[0015] S4. Input the blending formula of the tank batch finished gasoline to be processed into the IPRFCM-NGBoost adaptive integrated model, and evaluate it using the National VI B and enterprise production standards to obtain the corresponding quality evaluation results.

[0016] S5. Based on the quality evaluation results, the SHAP-LIME algorithm is used to perform cause analysis.

[0017] Preferably, in S5, the SHAP-LIME algorithm is used to perform cause analysis, including:

[0018] SHAP global interpretation: Analyze the characteristic influence relationship and characteristic influence degree, and identify the important component characteristic variables that affect the quality of finished gasoline;

[0019] LIME local interpretation: Calculate the instance feature contribution interval under each tank batch finished gasoline blending formula, and clarify the component-mass mapping interval that affects the quality of the tank batch finished gasoline.

[0020] Preferably, in S1, the gasoline blending formula is pretreated, including:

[0021] S111, remove unqualified data in the gasoline blending formula; the unqualified data include: data that is not successfully blended, does not meet the National VI B standard for gasoline, does not meet the 3σ principle, does not conform to the actual production process, and is non-numerical type data;

[0022] S112, using min-max normalization to adjust the scale of the gasoline blending formula after removing unqualified data; the min-max normalization is expressed as:

[0023]

[0024] Among them, X normalized is the normalized data, X is the original data point, and X min and X max are the minimum and maximum values ​​in the data, respectively.

[0025] Preferably, in S1, determining the output variables of the quality evaluation includes:

[0026] S121, calculating the maximum mutual information between each quality indicator and the target variable in the gasoline blending formula; and setting a mutual information threshold, and taking the quality indicator whose mutual information value is greater than the mutual information threshold as an important quality indicator;

[0027] S122 uses the PCA algorithm to perform principal component analysis on each quality indicator in the gasoline blending formula and performs dimensionality reduction processing; and sets a cumulative variance contribution rate threshold, selects the minimum number of principal components whose cumulative variance contribution rate is greater than the cumulative variance contribution rate threshold, and determines the key quality indicator;

[0028] S123. Based on the important quality indicators and key quality indicators, feature importance is sorted to determine output variables for quality evaluation.

[0029] Preferably, in S2, the IPRFCM algorithm is used to calculate the membership matrix and perform offline batch data partitioning, including:

[0030] S21, screening regression variables and selecting regression orders, and setting relevant parameters;

[0031] S22, Initialize regression coefficients And the membership matrix U (0) ;

[0032] S23, based on the initialized membership matrix U (0) , calculate the covariance matrix Σ for each batch i ;

[0033] S24. For each batch g, through the membership matrix U (g-1) , calculate the corresponding regression coefficient

[0034] S25. Using the covariance matrix E i and the regression coefficient Update the membership matrix U of the current batch (g) ;

[0035] S26, repeating the above steps S22 to S25, processing all batches of data until the preset conditions are met, and obtaining the membership matrix U;

[0036] S27. Cluster the offline batch data using the membership matrix U, and assign a fuzzy membership to each batch.

[0037] Preferably, the objective function of the IPRFCM algorithm is expressed as:

[0038]

[0039] Where C is the number of clusters, N is the number of data points, and T is the transpose of the matrix. is the membership degree of the kth data point to the i-th cluster, x jk is the j-th attribute value of the k-th data point, is the polynomial regression estimate for the kth data point, is the polynomial regression coefficient of the ith cluster, is the inverse matrix of the covariance matrix of the ith cluster.

[0040] Preferably, in S3, the NGBoost algorithm traverses the optimal increment d based on the constraint conditions to maximize the reduction of the loss function, including: a base learner f, a probability distribution p of a parameter, and a scoring rule s;

[0041] The NGBoost algorithm satisfies:

[0042] E y～Q [S(Q,y)]≤E y～Q [S(P,y)]

[0043] The constraints are expressed as:

[0044] Among them, E y～Q [S(Q,y)] and E y～Q [S(P,y)] is the expected loss under the two distributions Q and P respectively, S(Q,y) and S(P,y) are the scoring rules; g is the gradient, which is used to calculate the norm of d; ε is a constant used to limit the size of the increment d.

[0045] Preferably, in the NGBoost algorithm, the extreme value of the objective function S(θ+d) under the constraint condition is solved by the Lagrange multiplier method; the Lagrange multiplier method is expressed as:

[0046]

[0047] Among them, θ is the current model parameter and λ is the Lagrange multiplier.

[0048] Preferably, in S3, all sub-models are dynamically and adaptively fused in combination with the membership matrix, including: dynamically determining weights according to the membership of each data point in different clusters, using the weights as fusion weights of each sub-model, and dynamically and adaptively fusion of the sub-models;

[0049] The method for determining the fusion weight includes:

[0050] S31, calculating the membership degree of each data point in different clusters;

[0051] S32. For each data point, multiply the prediction results of each batch sub-model by its corresponding membership degree;

[0052] S33. Add all weighted prediction results to obtain the final prediction value.

[0053] It can be seen from the above technical solution that compared with the prior art, the technical solution of the present invention has the following advantages:

[0054] Beneficial effects:

[0055] 1. By eliminating unqualified data and adopting min-max normalization processing, the stability and convergence efficiency of model training are effectively improved; the maximum mutual information and principal component analysis methods are used to screen key quality indicators, which significantly improves the efficiency and effect of the model.

[0056] 2. The improved IPRFCM algorithm is used to reduce the computational cost and improve the accuracy of batch division through regression variable screening and regression order selection; the dynamic adaptive fusion strategy using fuzzy membership as weight can flexibly adjust the fusion weight of the model according to the characteristics of different batches, thereby improving the overall prediction accuracy of the integrated model.

[0057] 3. The NGBoost algorithm not only has excellent prediction performance, but also has the ability to predict the entire probability distribution, which is very suitable for dealing with uncertainty problems in gasoline quality prediction; by optimizing model parameters through the Lagrange multiplier method, it can effectively deal with optimization problems under constraints and improve the model's prediction and generalization capabilities.

[0058] 4. The SHAP-LIME algorithm provides in-depth explanations at both global and local levels, which helps to identify key factors affecting gasoline quality and propose targeted improvement measures. By clarifying the key component variables that affect the quality of finished gasoline and their correction ranges, it provides a solid scientific basis for the optimization and correction of defective formulas. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0060] Figure 1 A flow chart of a quality evaluation method for tank-type batch finished gasoline blending formula provided by an embodiment of the present invention;

[0061] Figure 2 A schematic diagram of a pretreatment process for the gasoline blending formula provided in an embodiment of the present invention;

[0062] Figure 3 A schematic diagram of a process for determining output variables for quality evaluation provided by an embodiment of the present invention;

[0063] Figure 4 A heat map generated based on the maximum mutual information coefficient provided in an embodiment of the present invention;

[0064] Figure 5 A schematic diagram of an offline batch data partitioning process provided by an embodiment of the present invention;

[0065] Figure 6 A schematic diagram of a process for determining the number of cluster divisions provided in an embodiment of the present invention;

[0066] Figure 7 A schematic diagram of a process for determining a fusion weight provided in an embodiment of the present invention;

[0067] Figure 8 A schematic diagram showing the comparison of the octane number prediction results before and after the improvement of the IPRFCM-SAE adaptive integrated model provided in an embodiment of the present invention;

[0068] Fig. 9A schematic diagram showing the comparison of the octane number prediction results before and after the improvement of the IPRFCM-RF adaptive integrated model provided in an embodiment of the present invention;

[0069] Fig.10 A schematic diagram showing the comparison of the octane number prediction results before and after the improvement of the IPRFCM-NGBoost adaptive integrated model provided in an embodiment of the present invention;

[0070] Fig.11 A schematic diagram of a causal analysis process using the SHAP-LIME algorithm provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0071] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0072] like Figure 1 As shown, this embodiment provides a quality evaluation method for tank batch finished gasoline blending formula, comprising the following steps:

[0073] S1. Obtaining multiple tank-type batch finished gasoline blending formulas, preprocessing the gasoline blending formulas, and determining output variables for quality evaluation;

[0074] S2. Based on the pre-processed gasoline blending formula, the membership matrix is ​​calculated using the IPRFCM algorithm to perform offline batch data division;

[0075] S3. Use the NGBoost algorithm to build a sub-model that adapts to each batch of data, and dynamically and adaptively fuse all sub-models in combination with the membership matrix to obtain the IPRFCM-NGBoost adaptive integration model;

[0076] S4. Input the blending formula of the tank batch finished gasoline to be processed into the IPRFCM-NGBoost adaptive integrated model, and evaluate it using the National VI B and enterprise production standards to obtain the corresponding quality evaluation results;

[0077] S5. Based on the quality evaluation results, the SHAP-LIME algorithm is used to perform cause analysis.

[0078] This method screens key quality indicators through maximum mutual information and principal component analysis, significantly improving model efficiency and prediction accuracy. Combining the improved IPRFCM algorithm with the NGBoost algorithm, it achieves accurate division of batch data and dynamic adaptive fusion of the model, improving the overall performance of the prediction model. The SHAP-LIME technology is used to provide model interpretability, which helps to identify and solve the key factors that lead to excess quality of finished gasoline.

[0079] The above steps and related technical features are further described in detail below:

[0080] In this embodiment S1, Figure 2 As shown, the gasoline blending formula is pretreated, including:

[0081] S111, remove unqualified data in the gasoline blending formula; the unqualified data include: data that is not successfully blended, does not meet the National VI B standard for gasoline, does not meet the 3σ principle, does not conform to the actual production process, and is non-numerical type data;

[0082] S112, using min-max normalization to adjust the scale of the gasoline blending formula after removing unqualified data; the min-max normalization is expressed as:

[0083]

[0084] Among them, X normalized is the normalized data, X is the original data point, and X min and X max are the minimum and maximum values ​​in the data, respectively.

[0085] Further, such as Figure 3 As shown, the output variables of quality evaluation are determined, including:

[0086] S121, calculating the maximum mutual information between each quality indicator and the target variable in the gasoline blending formula; and setting a mutual information threshold, and taking the quality indicator whose mutual information value is greater than the mutual information threshold as an important quality indicator;

[0087] S122 uses the PCA algorithm to perform principal component analysis on each quality indicator in the gasoline blending formula and performs dimensionality reduction processing; and sets a cumulative variance contribution rate threshold, selects the minimum number of principal components whose cumulative variance contribution rate is greater than the cumulative variance contribution rate threshold, and determines the key quality indicator;

[0088] S123. Based on the important quality indicators and key quality indicators, feature importance is sorted to determine output variables for quality evaluation.

[0089] Specifically, the output variables of the quality evaluation are determined from 26 original quality indicators, and the 26 original quality indicators are shown in Table 1:

[0090] Table 1

[0091]

[0092]

[0093] In order to scientifically select variables for predicting the quality of blending formulas, the process flow was deeply analyzed, and the two most critical performance indicators, octane number and anti-knock value, were first determined. Subsequently, 9 non-numerical indicators and indicators with serious data missing, such as water-soluble acids or alkalis, were eliminated. Furthermore, the feature selection method of variance filtering was adopted to further eliminate 4 indicators with a numerical variance of 0, such as unwashed colloid content, as well as 4 indicators with too high variance, such as 50% evaporation temperature, conductivity, density, and too low variance, such as terminal distillation point.

[0094] Furthermore, the maximum mutual information method was used to explore the correlation between the six blending components and the remaining seven quality indicators (including 10% evaporation temperature, vapor pressure, sulfur content, benzene content, olefin content, aromatic content, and oxygen content). The maximum mutual information coefficient between each blending component and the quality index was calculated, and a thermodynamic map was generated based on these coefficients.

[0095] like Figure 4 As shown in the figure, the indicators with low correlation, such as 10% evaporation temperature, vapor pressure and oxygen content, are excluded. The quality indicators with closer correlation with the blending components are selected as important quality indicators.

[0096] Next, PCA was used to reduce the dimensionality of the 26 original quality indicators. When the cumulative variance contribution rate was set to 90%, only the first three principal components were needed to reach this threshold. By analyzing the contribution of these three principal components to the original quality indicators, it was found that the five quality indicators of octane number, anti-knock value, benzene content, olefin content, and aromatic content had the highest contribution to these three principal components and were selected as key quality indicators.

[0097] Finally, based on the important quality indicators and key quality indicators, the feature importance is ranked to determine the output variables of the quality evaluation.

[0098] In this embodiment S2, Figure 5 As shown, the IPRFCM algorithm is used to calculate the membership matrix and perform offline batch data partitioning, including:

[0099] S21, screening regression variables and selecting regression orders, and setting relevant parameters;

[0100] S22, Initialize regression coefficients And the membership matrix U (0) ;

[0101] S23, based on the initialized membership matrix U (0) , calculate the covariance matrix Σ for each batch i ;

[0102] S24. For each batch g, through the membership matrix U (g-1) , calculate the corresponding regression coefficient It can more accurately capture the complex relationships inherent in the data set.

[0103] S25. Using the covariance matrix E i and the regression coefficient Update the membership matrix U of the current batch (g) ;

[0104] S26, repeat the above steps S22 to S25, process all batches of data until the preset conditions are met, and obtain the membership matrix U; the preset conditions here may refer to the change of the membership matrix being less than a preset threshold, or reaching a predetermined maximum number of iterations. This means that after each iteration of updating the membership matrix, if the change of the membership matrix is ​​very small (i.e., the change between two adjacent iterations is less than the set threshold), or the maximum allowed number of iterations has been reached, then the algorithm will stop iterating and consider that it has converged, and the membership matrix U obtained at this time is the final result.

[0105] S27. Cluster the offline batch data using the membership matrix U, and assign a fuzzy membership to each batch.

[0106] In this step, the IPRFCM algorithm is an improvement on the PRFCM algorithm. The PRFCM algorithm analyzes data through polynomial regression, where one attribute is considered as the dependent variable and the other attributes are used as independent variables. It introduces a polynomial regression model for each cluster in the traditional FCM clustering algorithm, replacing the concept of standard cluster centers, thereby enhancing the accuracy of clustering specific data types.

[0107] The improved IPRFCM algorithm reduces computational cost and improves classification accuracy by preselecting key regressors and regression orders and updating similarity metrics, and then uses these improvements to accurately partition offline batches.

[0108] At the same time, if Figure 6As shown in the figure, in order to determine the number of cluster divisions, the elbow method and the silhouette coefficient method are used to comprehensively determine the number of clusters. Specifically, it includes: using different numbers of clusters of the IPRFCM algorithm to calculate WSS and find the "elbow" position. For each number of clusters, calculate the silhouette coefficient and find the maximum value of the silhouette coefficient. If the number of clusters corresponding to the "elbow" position and the maximum value of the silhouette coefficient are the same, then the number of clusters is determined to be the optimal number. If the two are inconsistent, the number of clusters near the maximum value of the silhouette coefficient can be selected to balance the compactness and clarity of the clusters.

[0109] The objective function of the improved IPRFCM algorithm is expressed as:

[0110]

[0111] Where C is the number of clusters, N is the number of data points, and T is the transpose of the matrix. is the membership degree of the kth data point to the i-th cluster, x jk is the j-th attribute value of the k-th data point, is the polynomial regression estimate for the kth data point, is the polynomial regression coefficient of the ith cluster, is the inverse matrix of the covariance matrix of the ith cluster.

[0112] In this algorithm, the inverse of the covariance matrix is ​​used to measure the difference between each data point and the regression estimate of the cluster to which it belongs. The smaller this difference is, the better the association between the data point and the cluster.

[0113] By minimizing the objective function J IPRFCM The algorithm continuously adjusts the regression coefficient of each cluster so that the difference between the data point and the regression estimate of the cluster to which it belongs is as small as possible. The data points within each cluster are clustered more closely together while maintaining a certain distance from other clusters. The improved IPRFCM algorithm improves the accuracy of clustering and better adapts to data sets with complex relationships by introducing the concepts of multivariate regression model and covariance matrix, as well as optimizing the membership function.

[0114] In this embodiment S3, the NGBoost algorithm traverses the optimal increment d based on the constraint conditions to maximize the reduction of the loss function, including: a base learner f, a probability distribution p of a parameter, and a scoring rule s;

[0115] The NGBoost algorithm satisfies:

[0116] E y～Q [S(Q,y)]≤E y～Q [S(P,y)]

[0117] The constraints are expressed as:

[0118] Among them, E y～Q [S(Q,y)] and E y～Q [S(P,y)] is the expected loss under the two distributions Q and P respectively, S(Q,y) and S(P,y) are the scoring rules; g is the gradient, which is used to calculate the norm of d; ε is a constant used to limit the size of the increment d.

[0119] Furthermore, the extreme value of the objective function S(θ+d) under the constraint condition is solved by the Lagrange multiplier method; the Lagrange multiplier method is expressed as:

[0120]

[0121] Among them, θ is the current model parameter and λ is the Lagrange multiplier.

[0122] Furthermore, all sub-models are dynamically and adaptively fused in combination with the membership matrix, including: dynamically determining weights according to the membership of each data point in different clusters, using the weights as fusion weights of each sub-model, and dynamically and adaptively fusing the sub-models;

[0123] like Figure 7 As shown, the determination of fusion weights includes:

[0124] S31, calculating the membership degree of each data point in different clusters;

[0125] S32. For each data point, multiply the prediction results of each batch sub-model by its corresponding membership degree;

[0126] S33. Add all weighted prediction results to obtain the final prediction value.

[0127] In this step, the NGBoost algorithm maximizes the reduction of the loss function by traversing the optimal increment d based on the constraints. This process involves the base learner f, the probability distribution of the parameters p, and the scoring rule s. The Lagrange multiplier method is used to solve the extreme value of the objective function S(θ+d) under the constraints to optimize the model parameters θ, thereby improving the model's predictive and generalization capabilities.

[0128] Furthermore, all sub-models are dynamically and adaptively fused in combination with the membership matrix, including calculating the membership of each data point in different clusters, multiplying the prediction results of each batch of sub-models by their corresponding membership, and finally adding up all weighted prediction results to obtain the final prediction value. This not only improves the accuracy and adaptability of model prediction, but also can flexibly adjust the fusion weights of the model according to the characteristics of each data point in different batches, thereby significantly improving the overall prediction accuracy of the integrated model.

[0129] Through the optimization strategy of the NGBoost algorithm and the dynamic adaptive fusion strategy based on membership, the performance of the model in predicting the quality of finished gasoline blending formula is significantly improved, while maintaining good generalization ability and adaptability.

[0130] In this embodiment S4, the blending formula of the tank batch finished gasoline to be processed is input into the IPRFCM-NGBoost adaptive integrated model to obtain the corresponding quality evaluation result.

[0131] The blending formula of the tank batch finished gasoline to be processed contains a series of component ratios and other relevant information, such as octane number, anti-knock value, benzene content, olefin content, aromatic content and other key quality indicators.

[0132] The formula is preprocessed as necessary, such as checking whether it complies with the National VI B standard and whether it meets the 3σ principle. The data is scaled using the min-max normalization method defined above to ensure that the data format is consistent with that used when training the model.

[0133] The preprocessed recipe data is input into the trained IPRFCM-NGBoost adaptive ensemble model. The model will automatically identify the batch type to which the recipe belongs based on the previously obtained membership matrix, and dynamically adjust the fusion weights of the sub-models based on the fuzzy membership of the batch.

[0134] The model outputs the predicted quality indicators of the finished gasoline batch, including but not limited to octane number, anti-knock value, benzene content, olefin content, aromatic content, etc. The predicted results are evaluated according to the National VI B standard and the company's production standards to determine whether the finished gasoline meets the quality requirements.

[0135] Specific:

[0136] The key quality indicators are as follows: octane number: 93.5, anti-knock value: 90.2, benzene content: 0.8%, olefin content: 12.5% ​​and aromatic content: 35%.

[0137] After preprocessing, the data was input into the IPRFCM-NGBoost adaptive ensemble model. Based on previous training, the model identified the batch type that the recipe was most likely to belong to, and dynamically adjusted the fusion weights of the sub-models based on the fuzzy membership of the batch. Finally, the quality indicators of the model output prediction are as follows:

[0138] Predicted octane number: 93.2, predicted antiknock value: 90.0, predicted benzene content: 0.82%, predicted olefin content: 12.6% and predicted aromatic content: 34.8%.

[0139] According to the National VI B standard and the production standard of the enterprise, if the prediction results are within the prescribed range, the quality of the batch of finished gasoline is considered to be qualified. If there is an unqualified situation, go to step S5.

[0140] Further, such as Figure 8 , Fig. 9 and Fig.10 As shown in the figure, when SAE is fused as a sub-model, the prediction results of only a small number of samples of the IPRFCM-SAE adaptive integration model completely overlap. Although the performance after improvement has improved to a certain extent compared with that before improvement, the overall prediction effect is still poor. In contrast, the IPRFCM-RF adaptive integration model with RF as a sub-model has a higher overall fit degree than SAE in terms of prediction effect, and the prediction effect after improvement has also improved to a certain extent.

[0141] The integrated model with NGBoost as a sub-model for fusion shows a higher overall fit compared to the previous two. In the improved IPRFCM-NGBoost adaptive integrated model, only some samples do not completely overlap, and the overall fitting trend is more obvious, which once again verifies the superiority of the IPRFCM-NGBoost adaptive integrated model in uncertainty estimation. At the same time, the prediction performance of the three models after improvement has improved compared to before improvement, which shows that when the calculation of membership is more accurate, the integrated model based on this fusion can obtain better prediction accuracy.

[0142] In this embodiment S5, Fig.11 As shown, the SHAP-LIME algorithm is used for cause analysis, including:

[0143] SHAP global interpretation: Analyze the characteristic influence relationship and characteristic influence degree, and identify the important component characteristic variables that affect the quality of finished gasoline;

[0144] LIME local interpretation: Calculate the instance feature contribution interval under each tank batch finished gasoline blending formula, and clarify the component-mass mapping interval that affects the quality of the tank batch finished gasoline.

[0145] Specific:

[0146] Local interpretation focuses on the individual interpretation of the blending formula of a specific tank batch of finished gasoline. First, an instance x to be explained needs to be selected and explained using the LIME local interpreter. Next, samples similar to the instance to be explained are selected in the neighborhood sample set as references, and then the template standard information of instance x is calculated, that is, its similarity with other neighborhood samples. Next, the weight of instance x is calculated based on the template standard information and the information of other neighborhood samples. Finally, based on the sample weight, the instance feature contribution interval is calculated, which represents the specific impact of the feature on the quality of finished gasoline.

[0147] Based on the results of cause analysis, the main ingredient hydrogenated gasoline has basically the same influence on different quality indicators. Although the influence of other components is slightly different, they show the same influence trend within their proportion range. Then, for the correction of defective formula, by adjusting the content of the component to be corrected to a specific range, the positive influence of quality indicators can be achieved, thereby improving the quality of the formula.

[0148] By combining the technical means of SHAP global interpretation and LIME local interpretation, it achieves in-depth analysis of the factors affecting the quality of finished gasoline and their mechanisms from both macro and micro levels. SHAP helps identify the overall influence of key characteristic variables, while LIME is refined to a single tank batch, accurately quantifying the contribution of each component to the quality of finished oil, thereby providing data support for optimizing the blending formula and improving the accuracy and efficiency of product quality control.

[0149] This embodiment proposes a quality evaluation method for tank batch finished gasoline blending formula, which significantly improves the accuracy and interpretability of finished gasoline quality prediction through efficient data preprocessing (elimination of unqualified data, normalization processing), key quality indicator screening (maximum mutual information and principal component analysis), precise batch data division and dynamic adaptive fusion of models (IPRFCM and NGBoost algorithms), and causal analysis using SHAP-LIME technology, thereby optimizing the blending process and improving product quality.

[0150] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0151] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A quality evaluation method for tank batch finished gasoline blending formula, characterized in that: The following steps are involved: S1. Obtaining multiple tank-type batch finished gasoline blending formulas, preprocessing the gasoline blending formulas, and determining output variables for quality evaluation; S2. Based on the pre-processed gasoline blending formula, the membership matrix is ​​calculated using the IPRFCM algorithm to perform offline batch data division; S3. Use the NGBoost algorithm to build a sub-model that adapts to each batch of data, and dynamically and adaptively fuse all sub-models in combination with the membership matrix to obtain the IPRFCM-NGBoost adaptive integration model; S4. Input the blending formula of the tank batch finished gasoline to be processed into the IPRFCM-NGBoost adaptive integrated model, and evaluate it using the National VI B and enterprise production standards to obtain the corresponding quality evaluation results.

2. A quality evaluation method for tank batch finished gasoline blending formula according to claim 1, characterized in that: Also includes: S5. Based on the quality evaluation results, the SHAP-LIME algorithm is used to perform cause analysis.

3. A quality evaluation method for tank batch finished gasoline blending formula according to claim 2, characterized in that: In S5, the SHAP-LIME algorithm is used to perform cause analysis, including: SHAP global interpretation: Analyze the characteristic influence relationship and characteristic influence degree, and identify the important component characteristic variables that affect the quality of finished gasoline; LIME local interpretation: Calculate the instance feature contribution interval under each tank batch finished gasoline blending formula, and clarify the component-mass mapping interval that affects the quality of the tank batch finished gasoline.

4. The quality evaluation method for tank batch finished gasoline blending formula according to claim 1 is characterized in that: In S1, the gasoline blending formula is pre-processed, including: S111, remove unqualified data in the gasoline blending formula; the unqualified data include: data that is not successfully blended, does not meet the National VI B standard for gasoline, does not meet the 3σ principle, does not conform to the actual production process, and is non-numerical type data; S112, using min-max normalization to adjust the scale of the gasoline blending formula after removing unqualified data; the min-max normalization is expressed as: Among them, X normalized is the normalized data, X is the original data point, and X min and X max are the minimum and maximum values ​​in the data, respectively.

5. The quality evaluation method for tank batch finished gasoline blending formula according to claim 1 is characterized in that: In S1, the output variables of the quality evaluation are determined, including: S121, calculating the maximum mutual information between each quality indicator and the target variable in the gasoline blending formula; and setting a mutual information threshold, and taking the quality indicator whose mutual information value is greater than the mutual information threshold as an important quality indicator; S122 uses the PCA algorithm to perform principal component analysis on each quality indicator in the gasoline blending formula and performs dimensionality reduction processing; and sets a cumulative variance contribution rate threshold, selects the minimum number of principal components whose cumulative variance contribution rate is greater than the cumulative variance contribution rate threshold, and determines the key quality indicator; S123. Based on the important quality indicators and key quality indicators, feature importance is sorted to determine output variables for quality evaluation.

6. The quality evaluation method for tank batch finished gasoline blending formula according to claim 1 is characterized in that: In S2, the IPRFCM algorithm is used to calculate the membership matrix and perform offline batch data division, including: S21, screening regression variables and selecting regression orders, and setting relevant parameters; S22, Initialize regression coefficients And the membership matrix U (0) ; S23, based on the initialized membership matrix U (0) , calculate the covariance matrix Σ for each batch i ; S24. For each batch g, through the membership matrix U (g-1) , calculate the corresponding regression coefficient S25. Using the covariance matrix E i and the regression coefficient Update the membership matrix U of the current batch (g) ; S26, repeating the above steps S22 to S25, processing all batches of data until the preset conditions are met, and obtaining the membership matrix U; S27. Cluster the offline batch data using the membership matrix U, and assign a fuzzy membership to each batch.

7. The quality evaluation method for tank batch finished gasoline blending formula according to claim 1 is characterized in that: The objective function of the IPRFCM algorithm is expressed as: Where C is the number of clusters, N is the number of data points, and T is the transpose of the matrix. is the membership degree of the kth data point to the i-th cluster, x jk is the j-th attribute value of the k-th data point, is the polynomial regression estimate for the kth data point, is the polynomial regression coefficient of the ith cluster, is the inverse matrix of the covariance matrix of the ith cluster.

8. The quality evaluation method for tank batch finished gasoline blending formula according to claim 1 is characterized in that: In S3, the NGBoost algorithm traverses the optimal increment d based on the constraint conditions to maximize the reduction of the loss function, including: base learner f, probability distribution p of parameters and scoring rule s; The NGBoost algorithm satisfies: E y～Q [S(Q,y)]≤E y～Q [S(P,y)] The constraints are expressed as: Among them, E y～Q [S(Q,y)] and E y～Q [S(P,y)] is the expected loss under the two distributions Q and P respectively, S(Q,y) and S(P,y) are the scoring rules; g is the gradient, which is used to calculate the norm of d; ε is a constant used to limit the size of the increment d.

9. A quality evaluation method for tank batch finished gasoline blending formula according to claim 8, characterized in that: In the NGBoost algorithm, the Lagrange multiplier method is used to solve the extreme value of the objective function S(θ+d) under the constraint conditions; the Lagrange multiplier method is expressed as: Among them, θ is the current model parameter and λ is the Lagrange multiplier.

10. The quality evaluation method for tank batch finished gasoline blending formula according to claim 1, characterized in that: In S3, all sub-models are dynamically and adaptively fused in combination with the membership matrix, including: dynamically determining weights according to the membership of each data point in different clusters, using the weights as fusion weights of each sub-model, and dynamically and adaptively fusion of the sub-models; The method for determining the fusion weight includes: S31, calculating the membership degree of each data point in different clusters; S32. For each data point, multiply the prediction results of each batch sub-model by its corresponding membership degree; S33. Add all weighted prediction results to obtain the final prediction value.

Images