Catalytic cracking unit raw material property soft measurement modeling method based on integrated learning algorithm

Abstract

The invention discloses a catalytic cracking unit raw material property soft measurement modeling method based on an integrated learning algorithm, and relates to the technical field of flow industrial process soft measurement. On the basis of a Stacking ensemble learning algorithm, a distillation curve model and a data driving model are connected in series, the soft measurement modeling method for the properties of the catalytic cracking raw oil is provided, and the method has the advantages that the distillation range point, the specific gravity, the carbon residue value, the PONA value and other properties of raw materials of a catalytic cracking device can be comprehensively and accurately predicted in real time; the defect that an existing soft measurement model is single in raw material property prediction result is overcome, and the requirement for industrial optimization control can be better met. The embodiment shows that compared with a traditional machine learning method, the soft measurement model has the advantages that the prediction result is more accurate, the average absolute percentage error of each distillation range point is smaller than 2%, and the soft measurement model can accurately and comprehensively predict the properties of the raw materials of the catalytic cracking device in real time.

Description

Technical Field

[0001] The present invention relates to the technical field of soft sensor technology for process industry processes, and specifically relates to a soft sensor modeling method for the properties of feedstocks in a fluid catalytic cracking unit based on an ensemble learning algorithm, which is used to predict in real time the properties of feedstocks in a fluid catalytic cracking unit, such as distillation range points, specific gravity, carbon residue value, and PONA value. Background Art

[0002] In recent years, soft sensor technology has been widely developed and applied. This technology constructs the relationship between difficult-to-measure variables (dominant variables) and related easy-to-measure variables (auxiliary variables) in the research object through mathematical means to achieve real-time measurement of difficult-to-measure variables. Fluid catalytic cracking (FCC) is a typical secondary crude oil processing technology with difficult-to-measure variables. This technology converts low-value wax oil and residue into high-value light products such as gasoline, diesel, and liquefied gas. The fluctuations in the properties of the feedstock of the FCC unit directly affect the catalytic cracking reaction and cause changes in the product distribution. Careless operation will seriously reduce its economic benefits. The properties of the feedstock mainly include various distillation range points, specific gravity, carbon residue value, and PONA value (contents of paraffin, olefin, naphthene, and aromatic hydrocarbon), etc. These properties are obtained through sampling and analysis, so there is a certain time lag. The untimely measurement information of the feedstock properties causes the actual operation trajectory of the FCC unit to deviate from its optimal operation trajectory. Therefore, the realization of real-time measurement of the feedstock properties is of great significance for improving the efficiency of the unit.

[0003] The soft sensor model of feedstock properties is mainly based on data-driven methods. Current research mainly focuses on single feedstock properties. For example, in "Soft sensor modeling for fraction yield of crude oil based on ensemble deep learning" (Chemometrics and Intelligent Laboratory Systems, 2020, 204:104087), a soft sensor model is constructed by combining CNN, NNR, and RVFL networks to predict the mass fraction of crude oil fractions. However, important properties such as density and sulfur content, which are correlated with the mass fraction of fractions and jointly affect the crude oil processing process, are ignored. This makes the model unable to comprehensively reflect the comprehensive characteristics of crude oil and difficult to meet the demand for a comprehensive understanding of feedstock properties in the crude oil processing technology, limiting its application value in industrial production. "Soft sensor for distillation column feeds" (2009 Chinese Control and Decision Conference, CCDC2009) mainly focuses on the research of feed composition, introduces the sequential quadratic programming method (SQP) and the continuous tabu search algorithm (CTS), and uses the temperature distribution to predict the change of feed composition. However, this method is also limited to the prediction of single properties and cannot fully present the comprehensive situation of the feed, making it difficult to provide a comprehensive basis for the optimal control of the distillation column. "Soft Sensing of Feed Pulp Concentration in Thickener Based on Cross Validation LSTM Hyper-parameters Optimization" (2009 Chinese Control and Decision Conference, CCDC2021) uses the method of cross-validation combined with the long short-term memory network (LSTM) to predict the feed concentration. Although LSTM has certain advantages in dealing with dynamic characteristics and nonlinear problems due to its gating mechanism, important feed properties such as chemical composition and particle size distribution are ignored, making it unable to fully reflect the overall characteristics of the feed and difficult to meet the demand for comprehensive monitoring of feed properties in industrial optimal control. In the actual operation of the FCC unit, different indicators of feedstock properties are correlated with each other and jointly affect the cracking reaction. The research on single properties is difficult to comprehensively reflect the feedstock and cannot meet the demand for the optimal control of the unit. Therefore, the present invention proposes a soft measurement method for feedstock properties based on an ensemble learning algorithm, using a probability distribution to replace the discrete distillation range points in describing the feedstock properties to describe its distillation curve in order to obtain more comprehensive feedstock property data. Summary of the Invention

[0004] The object of the present invention is to solve the problems of insufficient comprehensiveness and accuracy of feedstock properties in the soft measurement prediction results, and to provide a soft measurement modeling method for feedstock properties of a fluid catalytic cracking unit based on an ensemble learning algorithm, which can predict properties such as the distillation range points, specific gravity, carbon residue value, and PONA value of the feedstock in real time, accurately, and comprehensively. The design of the soft measurement model for feedstock properties uses certain real-time measurable variables in the FCC unit as inputs, and the outputs are the distillation range points, specific gravity, carbon residue value, PONA value, etc. of the feedstock, and the output variables reflect the comprehensive characteristics of the feedstock. Since the influence of feedstock properties on the operation process is very complex and its mechanism model is difficult to obtain, a data-driven method can be used to achieve it. To facilitate modeling, a probability distribution function is introduced to describe the discrete distillation range point data; in addition, different contents of light and heavy components in the feedstock lead to different reaction processes, which will be manifested as different working conditions. Therefore, the feedstock oil is divided into three categories: light, medium, and heavy, and three corresponding sub-models of the ensemble learning algorithm are constructed to achieve accurate prediction of different types of feedstock.

[0005] To achieve the above object of the invention, the present invention provides the following technical solutions.

[0006] A soft measurement modeling method for feedstock properties of a fluid catalytic cracking unit based on an ensemble learning algorithm, which is used to realize the prediction of feedstock properties in the fluid catalytic cracking process. The method includes the following steps:

[0007] 1) Obtain data samples of various feedstock oils and different working conditions in the fluid catalytic cracking unit;

[0008] 2) Divide the data samples into three sub-datasets: light oil, medium oil, and heavy oil, perform normalization processing on each sub-dataset, and independently allocate training sets and test sets;

[0009] 3) Screen the input variables of the soft measurement model through the Pearson correlation coefficient method and the SHAP analysis method;

[0010] 4) Establish a feedstock oil classification model based on the K-Nearest Neighbors Classification (KNNC), optimize the hyperparameters through grid search, and use the classification result as the basis for selecting the data-driven sub-model;

[0011] 5) Fit the true boiling point distillation curve of the feedstock oil with the Weibull distribution, and use the fitting parameters (shape parameter) and (scale parameter) as the output variables of the data-driven model, and the parameter (location parameter) is used to calculate the distillation temperature of a certain distillation range point. (Here, the shape of the curve is determined by the scale parameter and the shape parameter, and after the curve shape is determined, the input location parameter is used to calculate the function value of a point inside the curve. For example, when input , it is calculated that which is the distillation temperature at the 10% distillation range point).

[0012] 6) The Stacking ensemble learning algorithm is adopted as the method for data-driven modeling, and four basic learners, namely K-Nearest Neighbors Regression (KNNR), Gradient Boosting Regression (GBR), Random Forest (RFR), and Linear Regression (LR), are selected. The cross-validation method is combined with the grid search strategy for parameter optimization, and the model performance is evaluated by indicators such as Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE). Finally, the real-time prediction of the distillation range points (0%, 5%, 10%, 30%, 50%, 70%), specific gravity, carbon residue value, and PONA values (contents of paraffins, olefins, naphthenes, and aromatics) of the feedstock oil in the fluid catalytic cracking unit is achieved.

[0013] In step 1), the data samples of various feedstock oils and different operating conditions of the fluid catalytic cracking unit can be actual industrial production data or process simulation data; the actual industrial production data needs to cover the property data of different feedstock oils processed and the production process parameters corresponding to different feedstock oils, such as temperature, pressure, flow rate, etc.; the process simulation needs to include the atmospheric and vacuum distillation unit (CDU) and the fluid catalytic cracking unit (FCCU). Through simulation, a series of data such as atmospheric residue and vacuum gas oil obtained by CDU processing feedstock oils with different weights can be obtained. Then, the products from CDU are mixed according to different wax-residue ratios and used as the input of FCCU, and the operating parameters of the FCCU model under different feedstock oils and operating conditions and the corresponding feedstock oil property data are collected;

[0014] In step 2), the data samples are divided into three sub-datasets of light oil, middle oil, and heavy oil. The data samples are divided into three sub-datasets of light oil, middle oil, and heavy oil according to the mass ratio of the feedstock oil in each distillation range interval (<343 °C, 343 - 482 °C, >482 °C); the training set and the test set are independently allocated, and the training set and the test set are independently divided according to a ratio within the range of 7:3 - 8:2.

[0015] In step 4), the input of the feedstock oil classification model is processed by Robustscale standardization, the hyperparameters of the model are determined by the grid search method, and the Kappa consistency coefficient between the classification result and the feedstock oil category is ≥0.85, and the classification accuracy is >90%.

[0016] In step 5), the true boiling point distillation curve of the feedstock oil is fitted using the Weibull distribution. By comparing the fitting accuracies of various probability distributions for the true boiling point distillation data of the feedstock oil, the Weibull distribution is selected to fit the true boiling point distillation curve of the feedstock oil. The root mean square error and mean absolute percentage error are used to measure the fitting accuracy. The fitting accuracies of the Weibull and Kumaraswamy distributions are higher than those of the Gamma, Beta, Lognorm, and Johnson SB distributions. Further, based on the error comparison at the indicative distillation range points (10%, 30%, 50%, and 70%), the Weibull distribution model is used to describe the distillation curve of the feedstock. The temperature values corresponding to each distillation range point can be calculated through the quantile function of the Weibull probability distribution:

[0017] (1)

[0018] where is the distillation temperature (°C); is the scale parameter of the Weibull distribution, which affects the extent of the distribution; is the mass percentage of the distillate; is the shape parameter of the Weibull distribution, which determines the specific shape characteristics of the distribution.

[0019] In step 6), the Stacking ensemble learning algorithm is used as a data-driven modeling method. The cross-validation method is combined with the grid search strategy for parameter optimization, and the root mean square error (RMSE) and mean absolute percentage error (MAPE) metrics are used to evaluate the model performance, realizing the real-time prediction of the distillation range points (0%, 5%, 10%, 30%, 50%, 70%), specific gravity, carbon residue value, and PONA value of the feedstock oil for the fluid catalytic cracking unit.

[0020] After screening by the Pearson correlation coefficient method and SHAP analysis method, the input variables are determined to be the feed flow rate, atomizing steam temperature, riser outlet temperature, mid-riser temperature, reactor pressure, regenerator pressure, dense-phase bed temperature of the regenerator, bulk density of the dense-phase bed of the regenerator, main air flow rate, flue gas temperature, etc.

[0021] In step 6), the selection of the base learners in the process of constructing the Stacking ensemble learning algorithm model is crucial. Each base learner should be complementary to capture different characteristics of the FCC process. Four base learners, namely K-NearestNeighbors Regression (KNNR), Gradient Boosting Regression (GBR), Random Forest (RFR), and Linear Regression (LR), are selected. Each base learner performs 5-10 fold cross-validation respectively, adjusts the hyperparameters of the base learner and generates meta-features for further training and integration. The models are trained on three sub-datasets of feedstock respectively, and the best model parameters of the Stacking ensemble learning algorithm are selected according to the prediction performance.

[0022] The present invention classifies feedstocks with different weights, uses the Weibull distribution to replace the discrete distillation range points on the distillation curve, and identifies the characteristics of the FCC unit through an ensemble learning algorithm to establish a soft sensor model for the feedstock properties of the FCC unit, which can realize real-time, accurate and comprehensive prediction of the feedstock properties. Compared with the prior art, the prominent technical effects and advantages of the present invention are as follows:

[0023] 1. The present invention proposes a light / medium / heavy three-classification strategy based on the mass ratio in the distillation range interval and establishes independent sub-models. The KNNC classification model is used to realize the dynamic identification of the feedstock category (classification accuracy > 90%, Kappa coefficient ≥ 0.85), and the prediction logic is optimized specifically to solve the problem of the differential influence of different feedstocks on the reaction process.

[0024] 2. The present invention introduces the Weibull probability distribution function into the feedstock distillation curve modeling, and characterizes the continuous distillation range characteristics through the shape parameter and the scale parameter. Compared with the discrete point modeling, 78% of the data redundancy is reduced, and the RMSE is reduced to 2.52 (compared with 3.26 of the Kumaraswamy distribution), realizing the refined description of the distillation curve. The quantile function supports the rapid calculation of any distillation range point, breaking through the lag of traditional offline analysis.

[0025] 3. The present invention constructs a heterogeneous base learner combination including KNNR, GBR, RFR, and LR, generates meta-features through 5-fold cross-validation, and captures local / global non-linear relationships and linear trends in the FCC process. The meta-learner integrates the outputs of the base models to achieve error correction and feature complementarity, and the RMSE is reduced by more than 20% compared with a single model (the percentage result here is calculated from the 50% distillation range point of the minimum reduction effect). The present invention adopts an independent sub-dataset training strategy to avoid prediction interference between different raw material categories and improve the generalization ability of the model.

[0026] 4. The present invention adopts a variable co-optimization strategy, combines the Pearson correlation coefficient and SHAP analysis to screen out key input variables, and reduces 23% of redundant variables. The introduction of Robustscale standardization processing reduces the influence of outliers on the KNNC classification model, and the classification accuracy reaches 90.9%. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a schematic structural diagram of the soft sensor model for raw material properties;

[0028] Figure 2 It is a schematic diagram of the CDU and FCCU process models; among them, the left figure is the process model of the atmospheric and vacuum distillation unit, and the right figure is the process model of the fluid catalytic cracking unit;

[0029] Figure 3 It is a Pearson correlation coefficient matrix diagram;

[0030] Figure 4 It is a diagram of the evaluation results of operating parameters by the SHAP method;

[0031] Figure 5 It is the error distribution of the Weibull distribution and the Kumaraswamy distribution at the 10%, 30%, 50%, and 70% distillation range points; among them, a is the 10% distillation range point, b is the 30% distillation range point, c is the 50% distillation range point, and d is the 70% distillation range point;

[0032] Figure 6 It is a comparison diagram of the distillation curve fitted by the Weibull distribution and the actual data;

[0033] Figure 7 It is a structural diagram of the Stacking ensemble learning algorithm;

[0034] Figure 8 It is a flow chart of the soft sensor modeling steps for raw material properties;

[0035] Figure 9 It is a comparison diagram of the prediction results at the 50% distillation range point;

[0036] Figure 10 It is a comparison chart of the predicted results of the specific gravity of the feedstock oil. Specific implementation manners

[0037] To further understand the present invention, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0038] The structural schematic diagram of the soft sensor model for the properties of the feedstock is as Figure 1 shown. It is designed that its input variables adopt some real-time measurable variables in the Fluid Catalytic Cracking Unit (FCCU), and the output variables cover the properties of each distillation range point, specific gravity, carbon residue value, PONA value, etc. of the feedstock oil. The specific operation process is as follows: the input variables first pass through the feedstock oil classification model to realize the dynamic identification of the light and heavy characteristics of the oil products, and then enter the corresponding Stacking data-driven sub-models (light, medium, and heavy) to predict the distillation curve parameters and other output variables. The data of each discrete distillation range point is calculated through the distillation curve model. This structure can realize the accurate prediction of the properties of different types of feedstocks.

[0039] Data acquisition: Based on the production data of a domestic device, process models of its Crude Distillate Unit & Vacuum Distillate Unit (CDU&VDU) and fluid catalytic cracking unit are established using HYSYS software as Figure 2 shown. The left figure is the crude distillation unit. The crude oil first enters the atmospheric column, and through distillation, light components such as top gas, light naphtha, water, heavy naphtha, light wax oil, and kerosene are separated. At the same time, atmospheric residue is produced at the bottom of the column. The atmospheric residue further enters the vacuum distillation column and continues to be distilled under reduced pressure to separate products such as non-condensable gas, vacuum gas oil, light vacuum wax oil, and heavy vacuum wax oil, and vacuum residue is produced at the bottom of the column, realizing the preliminary fractionation of the crude oil.

[0040] Figure 2 The right figure in [[ ]] is the fluid catalytic cracking unit. The atmospheric residue and vacuum wax oil after being processed by the crude distillation unit are mixed and used as the feedstock oil of the FCCU reaction unit, and cracking reactions occur under the action of the catalyst to generate a product mixture. Products such as hydrogen sulfide, dry gas, and liquefied petroleum gas are cut out through the first separation, and then the bottom oil of the column is cut into products such as gasoline, diesel, and slurry through the second separation, realizing the conversion of heavy oil into light fuel oil and chemical raw materials with high added value and improving the utilization value of crude oil.

[0041] The established process model above is used to simulate the operating parameters and feedstock property data under different feedstocks and steady-state conditions. Four types of crude oils with different properties in the HYSYS database: Arabian crude oil (light, high sulfur), Daqing crude oil (medium, low sulfur), Italian crude oil (heavy, low sulfur), and Venezuelan crude oil (heavy, high sulfur) are input into the CDU&VDU process model. Then, the atmospheric residue and vacuum gas oil output from the model are mixed at different wax residue ratios (0.5:0.5 - 0.7:0.3) and used as multiple feedstocks for the FCCU process model. An orthogonal experimental design is carried out in combination with the operating parameter change range shown in Table 1. While ensuring the quality of FCC products, 880 sets of data on operating parameters and feedstock properties under the steady-state conditions of the FCC unit are finally obtained. The above simulation process can ensure a wide range of feedstock sources and diverse unit operating conditions.

[0042] Table 1 Operating parameter change range

[0043]

[0044] The dataset is divided into three sub-datasets of light oil, middle oil, and heavy oil according to the mass ratio of the feedstock in each distillation range (<343 °C, 343 - 482 °C, >482 °C), with 276, 366, and 238 groups of data respectively. Each sub-dataset is normalized and the training set and test set are independently allocated according to the ratio of 8:2.

[0045] Combined with prior knowledge and on-site experience, 13 key operating variables are selected as candidate input variables from three parts: the feed, the riser, and the regenerator, including feed flow rate (FMF), atomizing steam temperature (ST), atomizing steam flow rate (SMF), atomizing steam pressure (SP), riser outlet temperature (RiOT), riser middle temperature (RiMT), cyclone separator outlet temperature (CSOT), reactor pressure (RIP), regenerator dense bed temperature (ReBT), regenerator pressure (ReP), regenerator dense bed bulk density (ReBPD), regenerator main air flow rate (RArMF), flue gas temperature (FGT), etc. Calculate the Pearson correlation coefficient between variables to evaluate their linear correlation degree and obtain the Figure 3 shown correlation coefficient matrix diagram. Then, the matrix diagram is used for variable screening. Blue indicates negative correlation and red indicates positive correlation. If there is a phenomenon that the linear correlation degree between two variables is close to ±1, then one of the variables is a redundant variable. The results show that there is a high linear correlation between the cyclone separator outlet temperature (CSOT) and the riser outlet temperature (RiOT). Since the riser is the core of the reactor, the variable of its outlet temperature is retained. Finally, the SHAP (SHapley Additive exPlanations) method is used to evaluate the importance of each variable as Figure 4As shown, each point in the figure represents the actual contribution value of a variable sample. According to the SHAP value analysis, the average comprehensive contribution degrees of the atomizing steam pressure (SP) and flow rate (SMF) are lower than 0.15. These variables can be deleted according to the evaluation results. The input variables of the soft sensor model finally selected are shown in Table 2.

[0046] Table 2 Key operating variables related to the properties of the feedstock

[0047]

[0048] The input variables of the feedstock classification model based on the K-nearest neighbor classification algorithm are processed by Robustscale standardization and scaled using the median and interquartile range (IQR) of the dataset, thus effectively reducing the impact of outliers on the model performance. The hyperparameters of the K-nearest neighbor classification model are determined by the grid search method, including: the number of neighbors (n_neighbors: 1 - 50), the weight function (weights: ['uniform', 'distance']), the distance metric (metric: ['euclidean','manhattan', 'cosine']), the algorithm type (algorithm: ['auto', 'ball_tree', 'kd_tree', 'brute']), and the leaf size (leaf_size: 10 - 100). In this embodiment, it is finally determined that n_neighbors = 1, weights = 'uniform', metric ='manhattan', algorithm = 'ball_tree', and leaf_size = 80. The Kappa consistency coefficient between the optimized feedstock classification model result and the feedstock category is ≥0.85, and the classification accuracy is >90%.

[0049] When the catalytic cracking feedstock is evaluated by true boiling point distillation, a series of discrete data points are obtained. Due to the limited information of individual data points, these data points can be correlated to describe the physical properties of the feedstock in detail. Polynomial regression or probability distribution function fitting can be used to extract the data characteristics. Gamma, Beta, Weibull, Kumaraswamy, Lognorm and Johnson SB distribution functions can better describe the distillation data of catalytic cracking feedstock. In order to select the probability distribution function that best suits the characteristics of catalytic cracking feedstock, different distribution functions are first applied to the research object and their fitting accuracies are compared. The results are shown in Table 3. In Table 3, RMSE and MAPE represent the root mean square error and mean absolute percentage error of the fitting results of different distribution functions at all distillation range points. It can be seen from Table 3 that the Weibull distribution and Kumaraswamy distribution have better fitting effects. To further compare the fitting results of these two distributions in detail, the error distributions at the indicative distillation range points of 10%, 30%, 50% and 70% are selected as shown in Figure 5 shown.

[0050] Table 3 Comparison of fitting errors of different distribution functions for distillation curves

[0051]

[0052] It can be seen from Figure 5 that the Weibull distribution has smaller errors than the Kumaraswamy distribution at the 10%, 30% and 70% distillation range points; the errors of the two distributions are similar at the 50% distillation range point. Therefore, the Weibull distribution model is used in the present invention to describe the distillation curve of the feedstock. The temperature values corresponding to each distillation range point can be calculated by the quantile function of the Weibull probability distribution, as shown in Equation 1:

[0053] (1)

[0054] In the formula, is the distillation temperature (°C); is the scale parameter of the Weibull distribution, which affects the stretching degree of the distribution; is the shape parameter of the Weibull distribution, which determines the specific shape characteristics of the distribution; is the mass percentage of the distillate for subsequent calculation of the distillation temperature at each distillation range point. When data-driven modeling, the parameter is used as the model output variable. Figure 6 is the comparison between the actual temperature and the model result at each distillation range point of a certain group of feedstock in the dataset, indicating that the distillation curve model established based on the Weibull distribution can better fit each distillation range point.

[0055] In the FCC, there are strong couplings and strong non - linearities among various parameters. A single machine - learning algorithm not only has difficulty in comprehensively capturing and interpreting its complex non - linear features but also is prone to underfitting. To overcome this shortcoming, the present invention adopts an ensemble learning algorithm as the method for data - driven modeling. Bagging, Boosting, and Stacking are three main ensemble learning methods. Bagging constructs multiple independent base models by sampling multiple sub - samples with replacement from the training set, and then integrates the prediction results of the models through voting or averaging. However, the methods of voting or averaging are difficult to fuse the features extracted by the base learners. Boosting sequentially constructs a series of weak learners, and each new model focuses on reducing the error of the previous model. This serial training depends on the results of the previous base learner and is prone to error accumulation for objects with strong - coupling complex non - linear relationships. Stacking is a two - stage learning algorithm. In the first stage, multiple different types of base learners are used to make preliminary predictions on the dataset, and base learners with different characteristics extract different features from the data. In the second stage, the preliminary prediction results are used as the input of the second - layer meta - learner to generate the final prediction result. The Stacking algorithm can flexibly combine models with different characteristics, and the models can correct each other to reduce bias and variance, performing well when dealing with complex non - linear relationships and being suitable for objects with strong - coupling relationships such as fluid catalytic cracking units. Figure 7 The Stacking ensemble learning algorithm structure diagram shown can represent sub - models for light, medium, and heavy feedstocks.

[0056] In the process of constructing the ensemble learning algorithm model, the selection of base learners is crucial. Each base learner should be complementary to capture different characteristics of the FCC process. K - Nearest Neighbors Regression (KNNR) focuses on predicting the output value closest to the sample according to the spatial distance to capture local non - linear relationships in the system; Gradient Boosting Regression (GBR) and Random Forest (RFR) both have the ability to capture global non - linear relationships in the data. Among them, GBR iteratively constructs new models to gradually reduce the residuals. The new model focuses on reducing the error of the previous model, thereby reducing the overall bias. RFR reduces the variance by integrating multiple decision trees, improving the stability and generalization ability of the model. The hyper - parameters of each base learner are shown in Table 4.

[0057] Table 4 Hyper - parameters of each base learner

[0058]

[0059] Each base learner performs 5-fold cross-validation separately and generates meta-features for further training and integration. The models are trained on three feedstock sub-datasets respectively, and the best model parameters of the Stacking ensemble learning algorithm are selected according to the prediction performance. The overall modeling process of the soft sensor for online predicting FCC feedstock properties is as Figure 8 shown.

[0060] The prediction accuracy of the model is evaluated by the root mean square error (RMSE) and the mean absolute percentage error (MAPE). The calculations of RMSE and MAPE are shown in Eqs. (2) and (3):

[0061] (2)

[0062] (3)

[0063] wherein, represents the number of samples, represents the true value of the th sample, represents the predicted value of the th sample. The prediction results of the artificial neural network model (ANN model) with hyperparameters optimized by particle swarm optimization and the Stacking-Weibull model (Baseline model) without classifying the feedstock are evaluated by Eqs. (2)-(3) as shown in Table 5. It can be seen from Table 5 that the MAPE of the prediction results of the Stacking-Weibull model for the 0%, 10%, 30%, 50% and 70% distillation range points are all less than 1%; the paired t-test shows that compared with the ANN model, p = 1.062×10 -12 , compared with the Baseline model, p = 0.027 for the Stacking-Weibull model. Both are lower than the significance level α = 0.05. The RMSE of the Stacking-Weibull model is significantly lower than that of the ANN model and the Baseline model; although the prediction accuracy of the 5% distillation range point is slightly lower, it is also significantly better than the ANN model and the Baseline model. It can be seen from Table 6 that the accuracy of the prediction results of the Stacking-Weibull model for various properties of the feedstock is higher than that of the other two control models. Figure 9 and Figure 10They respectively represent the prediction results of different models for the 50% distillation point and the specific gravity of the feedstock oil. It can be clearly seen that the predicted values of the Stacking-Weibull model for the 50% distillation point and the specific gravity of the feedstock oil are better than those of the other two models. The same results also apply to other distillation points, the carbon residue value, and the PONA value. This indicates that the synergistic effect of the classification strategy and the integrated learning of the present invention significantly improves the model's ability to characterize the properties of complex feedstocks.

[0064] Table 5 Comparison of prediction errors for each distillation point

[0065]

[0066] Table 6 Comparison of prediction errors for other feedstock oil properties

[0067]

[0068] The above cases are used to explain the effectiveness and feasibility of the present invention, indicating that the integrated learning algorithm can better identify the characteristics of the FCC unit, the Weibull distribution can well describe the distillation curve of the feedstock oil, and the fusion of the feedstock oil classification sub-models can more effectively improve the model performance. The soft sensor model of the present invention can achieve real-time, comprehensive, and accurate prediction of the properties of FCC feedstocks.

[0069] The above embodiments are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. The protection scope of the present invention shall be determined by the content recorded in the claims, and any equivalent replacement, improvement, or modification based on the concept of the present invention shall be covered within the protection scope of the present invention.

Claims

1. A soft-sensing modeling method for raw material properties of a catalytic cracking unit based on an ensemble learning algorithm, characterized in that The following steps are involved: 1) Obtain data samples of various crude oils and different operating conditions of the catalytic cracking unit; 2) Divide the data samples into three sub-datasets: light oil, medium oil, and heavy oil. Normalize each sub-dataset and independently allocate training and test sets. 3) Screen the input variables of the soft sensor model through the Pearson correlation coefficient method and SHAP analysis method; 4) Establish a crude oil classification model based on the K nearest neighbor classification algorithm, optimize hyperparameters through grid search, and use the classification results as the basis for selecting data-driven sub-models; 5) Use Weibull distribution to fit the actual boiling point distillation curve of crude oil, and set the fitting shape parameter and scale parameter As the output variable of the data-driven model, the location parameter Used to calculate the distillation temperature of a certain distillation point; 6) The Stacking ensemble learning algorithm is used as the data-driven modeling method, and four basic learners are selected: K nearest neighbor regression, gradient boosting regression, random forest regression and linear regression. The cross-validation method is combined with the grid search strategy to optimize the parameters, and the model performance is evaluated by the root mean square error and mean absolute percentage error indicators. Finally, the real-time prediction of the distillation point, specific gravity, residual carbon value and PONA value of the catalytic cracking unit feedstock is achieved.

2. A soft-sensing modeling method for raw material properties of a catalytic cracking unit based on an ensemble learning algorithm as claimed in claim 1, characterized in that In step 1), the source of the data sample includes actual industrial production data or process simulation data; the actual industrial production data must cover the property data of different crude oils processed therein and the production process parameters corresponding to different crude oils, such as temperature, pressure, and flow rate; the process simulation must include a atmospheric and vacuum unit CDU and a catalytic cracking unit FCCU, and the atmospheric and vacuum unit CDU processes crude oils of different weights through simulation to obtain corresponding atmospheric residue oil and vacuum wax oil series data, and then the products from the atmospheric and vacuum unit CDU are mixed according to different wax-residue ratios and used as input to the catalytic cracking unit FCCU, and the operating parameters of the catalytic cracking unit FCCU model under different crude oils and operating conditions and the corresponding crude oil property data are collected.

3. A soft-sensing modeling method for raw material properties of a catalytic cracking unit based on an ensemble learning algorithm as claimed in claim 1, characterized in that In step 2), the division method is to divide the crude oil into three sub-datasets of light oil, medium oil and heavy oil according to the mass proportion of the crude oil in the three distillation ranges <343°C, 343~482°C and >482°C, and the ratio of the training set to the test set is in the range of 7:3~8:

2.

4. A soft-sensing modeling method for raw material properties of a catalytic cracking unit based on an ensemble learning algorithm as claimed in claim 1, characterized in that In step 3), the input variables of the soft sensor model are screened, and the screened variables include: feed flow rate, atomizing steam temperature, riser outlet temperature, riser middle temperature, reactor pressure, regenerator pressure, regenerator dense bed temperature, regenerator dense bed packing density, main air flow rate, flue gas temperature.

5. A soft-sensing modeling method for raw material properties of a catalytic cracking unit based on an ensemble learning algorithm as claimed in claim 1, characterized in that In step 4), the input of the crude oil classification model is processed by Robustscale standardization, the hyperparameters of the model are determined by grid search method, the Kappa consistency coefficient of the classification result and the crude oil category is ≥0.85, and the classification accuracy is >90%.

6. A soft-sensing modeling method for raw material properties of a catalytic cracking unit based on an ensemble learning algorithm as claimed in claim 1, characterized in that In step 5), the quantile function expression of the Weibull distribution is: (1) In the formula, is the distillation temperature; is the scale parameter of the Weibull distribution, which affects the stretch of the distribution; is the mass percentage of distillate; It is the shape parameter of the Weibull distribution, which determines the specific shape characteristics of the distribution.

7. A soft-sensing modeling method for feedstock properties of a catalytic cracking unit based on an ensemble learning algorithm as claimed in claim 1, characterized in that In step 6), the selection of the basic learners of the Stacking ensemble learning algorithm requires that the basic learners should be complementary to capture the different characteristics of the catalytic cracking process; 5-10 fold cross validation is performed on each basic learner, the hyperparameters of the basic learner are adjusted and meta-features are generated for further training and integration, the models are trained on three raw oil sub-datasets, and the optimal model parameters of the Stacking ensemble learning algorithm are selected according to the prediction performance.

Images