WO2023040512A1 - Catalytic cracking unit simulation and prediction method based on molecular-level mechanism model and big data technology - Google Patents

Abstract

Provided is a catalytic cracking unit simulation and prediction method based on a molecular-level mechanism model and big data technology. The product yield and product properties of a catalytic cracking process are predicted; a molecular-level mechanism model for the catalytic cracking process is established; and furthermore, on the basis of big data technology, a mechanism model prediction deviation caused by an actual unit running state is corrected, thereby realizing the process simulation of an industrial-grade unit.

Description

A simulation and prediction method for catalytic cracking unit based on molecular mechanism model and big data technology technical field

The invention relates to the technical fields of petroleum refining and petrochemical production, in particular to a method for simulating and predicting a catalytic cracking unit based on a molecular-level mechanism model and big data technology.

Background technique

Catalytic cracking is an important oil refining process. Its total processing capacity ranks first among various petroleum processing processes, and its technical complexity also ranks first among all kinds of oil refining processes. Therefore, catalytic cracking occupies a pivotal position in the oil refining industry. The simulation of catalytic cracking process by traditional technology is mostly based on the method of lumped kinetics. The lumped kinetics method divides the complex components in catalytic cracking into several lumped components according to the kinetic characteristics. In , each lump is considered as a virtual single component. Therefore, the traditional lumped kinetic model can usually only predict the yield of the product, but cannot predict the properties of the product, and the lumped kinetic model cannot reflect the change of the composition of the raw material, because there may be great differences between the composition of the oil products with the same properties. big difference.

With the increasingly stringent requirements on the quality of refined oil, it is urgent to find a method that can not only predict the product yield, but also accurately predict the product properties. The emergence of molecular-level dynamics models provides the possibility to solve this problem. By analyzing the molecular composition of raw materials, a molecular-level reaction kinetic network is established, and the conversion laws of reactant and product molecules in the reactor are calculated, and then the molecular composition and product properties are accurately predicted. This method predicts more accurately than the traditional lumped method, and the model adaptability is more extensive. However, due to the complexity of the model, many parameters, and a large amount of calculation, the molecular-level mechanism model has not been widely used in the industry.

The shortcomings of existing methods mainly include the following aspects: (1) The prediction accuracy and extension of traditional lumped models are not good, and the lack of control over the actual operation of the device makes it difficult to avoid deviations between prediction results and actual data. (2) The catalytic cracking process is complex, with many influencing variables, and it is difficult to construct a complete molecular-level mechanism model, which is difficult to apply to different devices, has poor practical applicability, and lacks control over the actual operation of the device. (3) The independent big data model does not consider the nature of the response, the causal relationship between the data does not correspond, and the extension of the model is poor. In addition, the causal correlation of variables and the time delay of causal response in the big data model are insufficiently considered, and the quality of data preprocessing seriously affects the accuracy of the model.

Contents of the invention

In order to overcome the above shortcomings, the present invention aims to provide a method for simulating and predicting a catalytic cracking unit based on a molecular-level mechanism model and big data technology. The present invention can predict the product yield and product properties of the catalytic cracking process; the present invention Invented and established a molecular-level mechanism model of the catalytic cracking process. The molecular-level mechanism model can not only improve the prediction accuracy, but also be applicable to different devices and has good extension; in addition, based on big data technology, the mechanism caused by the actual device operation Correction of model prediction deviations not only captures the essence of catalytic cracking reactions, but also reflects the characteristics of different catalytic devices, accurately predicts product yields and key product properties, and enables accurate process simulation of industrial-level devices.

The present invention achieves the above object through the following technical solutions: a method for simulating and predicting a catalytic cracking unit based on a molecular-level mechanism model and big data technology, comprising the following steps:

(1) Establish a catalytic cracking unit model, wherein the catalytic cracking unit model includes a raw material molecular analysis model, a molecular dynamics model, a riser reactor model, a product cutting model, and a physical property model;

(2) based on actual industrial data, the parameters of the model built in step (1) are corrected;

(3) Based on the processing results of step (2), a deviation compensation prediction model based on a machine learning algorithm is established; wherein, the input of the deviation compensation prediction model is the actual operating parameters of the device, and the output is the difference between the predicted value of the mechanism and the actual output value of the historical working condition The deviation of the actual working conditions is compensated to improve the prediction accuracy of the model;

(4) Through steps (2) and (3), a communication mechanism can be established with the real-time data of the catalytic cracking unit in the production process, and the device data can be read in real time and the model correction and deviation compensation prediction model can be updated to realize the prediction model Automatic updates.

As preferably, the method of the raw material molecular analytical model that described step (1) establishes is specifically as follows:

(1.1) Raw material molecular library construction: carry out experimental analysis and characterization of the raw material, determine the core structure of the raw material molecule, add side chains, branch chains and methyl groups on the basis of the core molecular structure according to a certain strategy, and obtain the raw material molecular library;

(1.2) Raw material molecular concentration analysis: the initial value of the molecular concentration is set according to the composition characteristics of the raw material according to the probability distribution, and then the distribution parameters and molecular concentration are adjusted through the global optimization algorithm, so that the final molecular concentration distribution can meet the requirements of the raw material. Macroscopic properties can be analyzed into detailed molecular composition according to the macroscopic physical properties of the feed such as density, carbon residue, sulfur content, nitrogen content, group composition, and distillation range; through the molecular concentration composition construction technology, according to various macroscopic properties that can be obtained through experimental analysis Properties Inversion analysis of molecular composition of FCC feedstock.

As preferably, the method of the molecular-level dynamics model in the described step (1) is specifically as follows:

(1.3) Write reaction rules and construct reaction network: According to the carbenium ion mechanism of catalytic cracking reaction, reactant selection rules and product generation rules were established for different types of reactions, and a class of rule functions was written for each type of reaction , compiled a large class of reaction rules including cracking, ring-opening, isomerization, hydrogen transfer, and condensation, and used computer-aided technology to apply reaction rules to raw material molecules to automatically generate reaction networks; among them, the types of reaction rules are preferably within 10 pcs - 50 pcs.

As preferably, the method of the riser reactor model in the described step (1) is specifically as follows:

(1.4) set up a reactor model, solve the model, and calculate the product molecular concentration distribution; the built reactor model includes catalytic cracking processes such as single riser model, MIP double riser model, DCC main and auxiliary riser parallel model; The network, stoichiometry, reaction rate equation and kinetic parameters are combined with the above reactor model to obtain a complete FCC reactor model.

As preferably, the method of the product cutting model in described step (1) is specifically as follows:

(1.5) Product cutting model: For the product oil-gas mixture molecules coming out of the reactor, the mixed oil-gas is cut and separated into stream products of dry gas, liquefied gas, gasoline, diesel oil, oil slurry, and coke according to various product quality requirements; , the product cutting model can adopt a simple cutting model based on boiling point cutting, while considering the influence of overlapping factors;

The concrete method of the physical property model in described step (1) is as follows:

(1.6) Calculation of physical properties: Through the physical property calculation model, the group contribution method and the empirical correlation method are used to calculate the properties of the products with the molecular concentration of each product, including the physical properties of gasoline and diesel.

As preferably, the specific method of parameter correction in the step (2) is: correct the parameters of the model through actual industrial data, and complete the molecular-level mechanism model; the industrial data include reactor structural size parameters, catalyst parameters, feed-in and discharge parameters, etc. Property detection data (LIMS data), device operating parameters (DCS data).

As preferably, the concrete method of described step (3) is as follows:

(3.1) Data collection and arrangement: read the DCS and LIMS historical production data of the device, establish a database, standardize the format and establish indexing rules to facilitate later query and call;

(3.2) Data preprocessing: Extract data from the database for processing, including missing value interpolation processing, outlier processing, data smoothing and noise reduction, and data normalization processing;

(3.3) Variable correlation analysis: Calculate the correlation between variables through the correlation algorithm, and combine expert experience analysis to select the most relevant variables for modeling; specifically, the method of variable correlation analysis chooses Pearson correlation analysis One or more of , transfer entropy, Granger causality analysis, combined with expert experience analysis, select variables with strong correlation from many variables for modeling;

(3.4) Steady-state analysis: The steady-state detection of the system is carried out through the established steady-state analysis rules, and the corresponding working conditions in each steady state are extracted, and a database is established using the obtained steady-state working conditions to facilitate subsequent updates and use;

(3.5) Establish a deviation compensation prediction model: input the input variables corresponding to the historical working conditions into the established mechanism model to obtain the predicted value of the mechanism model, and then calculate the deviation between the predicted value of the mechanism and the actual output value of the historical working condition, through The principal component analysis reduces the input dimension of the input variables, establishes the deviation compensation prediction model through the machine learning algorithm, and adds the deviation compensation to the predicted value of the mechanism model to obtain the final prediction result.

As preferably, the missing value interpolation processing method in the step (3.2) selects any one of linear interpolation, cubic spline interpolation, mean value interpolation, and Lagrange interpolation; the outlier identification method selects 3σ criterion method, box line Any one of the graph method and Grubbs test method; the data noise reduction smoothing method selects the robust quadratic regression method to eliminate high-frequency noise signals and retain low-frequency data trends.

As preferably, the method for the steady-state analysis in described step (3.4) is divided into univariate steady-state analysis and system steady-state analysis:

(i) Univariate steady-state analysis: use the wavelet transform method to extract the trend of the data set, and decompose the signal into the noise of the high frequency band and the low frequency band representing the trend of the signal to perform finite continuous approximation on the measured data, and obtain the process The approximate function f(t) of the variable establishes a steady-state index β (0≤β≤1) to represent the stability of the variable working condition. When β=0, the process variable is in an unsteady state. When β=1, The process variable is in a steady state. When 0<β<1, it means that the process variable is in a transition state, and the closer β is to 1, the more stable the state of the process variable is; the steady state index β(t) is determined by the first derivative of the data trend f '(t ₀ ) and the second derivative f”(t ₀ ) are jointly determined according to the following criteria:

θ(t)=|f'(t)|+γ|f″(t)|

In the formula:

Among them, T _s , T _w , and _Tu are the thresholds, and the determination method is as follows: select a section of data in the historical database that is in a steady state as a reference, and extract the change trend of the process variable through wavelet transform to obtain the process variable at the sampling point. The first-order derivative sequence and the second-order derivative sequence of , by calculating their percentiles respectively, then: T _s is the percentile of the first-order derivative, T _w is the percentile of the second-order derivative, generally 90 % quantile or 95% quantile, where T _u = αT _s

α is an adjustable parameter, which generally takes an integer between [2, 5]; through the above formula, the steady-state judgment threshold of the process variable can be obtained;

(ii) System steady-state analysis: The steady-state index β _i of each variable is obtained by using the single-variable steady-state detection rule, and the steady-state condition of the entire system is determined by the weighting of the steady-state index of each variable according to the following formula:

Among them, p is the number of key characteristic variables of the system, β _i (t) is the steady-state index of the i-th variable, and u _i is the weight of the i-th variable.

As preferably, the machine learning algorithm in the step (3.5) is selected from any one of a feedforward neural network, a recurrent neural network, a support vector machine, a least squares method, a least squares support vector machine, and an extreme learning machine regression algorithm.

The beneficial effects of the present invention are: (1) The present invention adopts the form of combining molecular-level mechanism model and big data technology, the molecular-level mechanism model describes the essential law of the catalytic cracking process, and the big data model makes full use of production history data, It can reflect the characteristics of different devices and the actual operating status of the device, and can efficiently process data sets with long time spans and huge data volumes; (2) the present invention calculates the yield and properties of products through the mechanism model, and uses the big data model to calculate the actual results The deviation from the mechanism prediction result, the mechanism model increases the data model to calculate the deviation between the actual result and the mechanism prediction result, and realizes the accurate prediction of product yield and product properties. At the same time, when facing unknown working conditions, the model has a good appearance (3) The present invention uses a molecular-level mechanism model based on structure-oriented lumping. Compared with the traditional lumped model, the molecular-level mechanism model has more accurate prediction accuracy and is more Wide forecast range.

Description of drawings

Fig. 1 is a schematic flow sheet of the present invention;

Fig. 2 is a schematic diagram of the reaction network of the present invention;

Fig. 3 is a schematic diagram of various physical properties that can be calculated by the physical property model of the present invention;

Fig. 4 is the schematic diagram of the univariate steady-state analysis result in the embodiment of the present invention;

Fig. 5 is the schematic diagram of the system steady-state analysis result in the implementation case of the present invention;

Fig. 6 is a schematic diagram of a multilayer neural network structure of the present invention;

Fig. 7 is a schematic diagram of the product yield prediction error of the dual-core drive model of the present invention;

Fig. 8 is a schematic diagram of the product property prediction error of the dual-core driver model of the present invention.

Detailed ways

The present invention is further described below in conjunction with specific embodiment, but protection scope of the present invention is not limited thereto:

Embodiment: The present invention will be described in further detail below through the specific implementation of the simulation of a catalytic cracking unit in a refining and chemical enterprise. As shown in Figure 1, a method for simulating and predicting a catalytic cracking unit based on a molecular-level mechanism model and big data technology includes the following steps:

(1) Construct the raw material molecular library. The number of molecules that make up FCC feedstock is enormous, and the number of homologue core structures that make up those molecules is much smaller. On the basis of the core structure, side chains, branch chains and methyl groups are added according to a certain strategy to obtain a series of molecular sets, and the properties such as the distillation range density of each molecule are calculated through the physical property search and molecular physical property calculation model, and the carbon number and distillation range constraints are used. The molecular set is further screened, unreasonable molecules are deleted, and the final raw material molecular library is obtained. Currently, a raw material molecular library containing 48 cores and 2473 molecules has been constructed.

(2) Raw material molecular concentration analysis. The molecular-level reaction mechanism model needs to input the molecular concentration. The information of a certain fragment of the molecular composition is obtained through different experimental analysis methods. Through the splicing of multiple fragments, the complete molecular information of the raw material can be inferred indirectly. The characteristics are set according to a certain probability distribution, and then the distribution parameters and molecular concentration are adjusted through a specific global optimization algorithm, so that the final molecular concentration distribution can meet the macroscopic properties of the raw materials. The accuracy of the molecular concentration of raw materials has a great influence on the calculation results of the mechanism model, and the molecular analysis algorithm is only a tool to obtain the molecular concentration, and the accuracy of the results depends more on the accuracy and completeness of the experimental analysis data.

(3) Write the reaction rules to construct the reaction network. The reaction mechanism of catalytic cracking is very complicated, and the widely accepted mechanism is the carbanion mechanism. Under the action of the acidic center of the molecular sieve catalyst, the hydrocarbon molecules capture protons to form carbonium ions, and the carbonium ions undergo β-cleavage to generate olefins and new carbonium ions. Carbanions also undergo an isomerization reaction, which constitutes the main reaction in the catalytic cracking system. According to the chemical reaction mechanism of catalytic cracking, the reaction rules are established for different types of reactions in turn. The reaction rules include reactant selection rules and product production rules. A class of rule functions is written for each type of reaction. 25 types of reaction rules for the main reactions such as structure, hydrogen transfer, and condensation, and then use computer-aided technology to apply the reaction rules to the raw material molecules to automatically generate a reaction network. A reaction network containing 5216 reactions was generated (the reaction network diagram is shown in Figure 2 shown). The types of reactions carried out on different catalysts will have some changes, so the reaction rules also need to be adjusted appropriately according to the experimental data.

(4) Build the reactor model. According to the pseudo-planar plug-flow reactor model, the mass balance equation and heat balance equation are established, and the reaction network, stoichiometry, reaction rate equation and kinetic parameters are combined with the above reactor model to obtain a complete catalytic cracking reactor model , which is expressed as a huge differential equation system, and the concentration conversion law of each molecule in the reactor can be calculated by obtaining the differential equation system.

(5) The product is cut and separated. According to the typical distillation range range of each product, a simple cut model based on boiling point cut was adopted, and the influence of overlap factor was considered.

(6) Physical property calculation model. The present invention is based on the relevant physical property calculation model, adopts the group contribution method and the empirical correlation method, uses the molecular concentration of each product to calculate the properties of the product, and calculates the various physical properties of catalytic gasoline and catalytic diesel oil. The calculable physical properties are shown in the figure 3.

(7) Model parameter correction. Collect the actual production data of the factory, including the property analysis data of the feed material and the product and the operating parameters of the device. The goal is to minimize the sum of squared errors between the predicted value and the actual output value, and the empirical value or experimental value is the initial value. Enable the global The optimization algorithm optimizes and solves the model parameters.

(8) Data collection and collation. In this example, the complete historical production data of the DCS system and LIMS of a catalytic cracking unit in a refinery were collected for about one year. The operating parameters and state parameters of the DCS system were automatically recorded every 10 minutes, and the detection data of the feed material properties of the device system were recorded every 10 minutes. Analyze once every 3 days, and analyze the properties of the output material of the device system every 24 hours. Organize the collected DCS and LIMS data, and establish a basic database for subsequent reading and retrieval.

(9) Data processing. Descriptive statistical analysis is carried out on the collected DCS operating parameters and state parameters and the material detection properties of LIMS, and the distribution rules and statistical information of the data are analyzed. The missing values and outliers of the data are processed, and as a preference, the percentile method is used to judge the outliers of the data set. Missing and outlier values are imputed using an interpolation fitting method. Use the quadratic regression function to denoise the data, filter high-frequency noise signals, and obtain a data set that is more in line with the trend of data changes.

(10) Variable correlation analysis. 542 DCS operation and state variables and 340 LIMS property detection variables were collected. In order to simplify the input variables and retain as much valid information as possible, it is necessary to filter out a group of variables most relevant to the prediction model. According to the actual situation of the catalytic cracking production unit, combined with the experience and professional knowledge of experts, and with the corresponding correlation analysis algorithm, the variables of the three major systems of reaction, fractionation and absorption stability are analyzed, and the relationship between each operating parameter and state parameter is explored. Correlation relationship, the operation and state parameters that have a greater impact on product yield and properties are selected as input variables for subsequent modeling, and the correlation analysis algorithm uses the transfer entropy algorithm.

(11) Steady-state analysis. Facing the complex nonlinear system of catalytic cracking, the steady-state analysis of the system is carried out by wavelet decomposition method based on trend extraction. Wavelet transform has good time-frequency domain positioning and multi-resolution analysis capabilities. It can analyze the different frequency characteristic information contained in the historical data center, and analyze the measurement data by decomposing the signal into high-frequency noise and low-frequency band representing the signal trend. Finite continuous approximation to obtain an approximate function of the process variable. Through the established univariate steady-state criterion and system steady-state judgment criterion, the steady-state analysis is performed on all variables separately, and then the steady-state judgment is performed on the system. Figure 4 and Figure 5 show the univariate steady-state analysis and the state of the system respectively judge. After judging the steady state of the system, extract the system data in the steady state. The average value of each segment of steady-state data is taken as the variable data representing the steady-state segment, and a database is established based on these variable data to facilitate subsequent analysis, extraction and use.

(12) Establish a deviation compensation model based on machine learning. After variable correlation analysis and steady-state detection analysis, the input variables and corresponding data sets are determined, and the data sets of input variables are input into the established mechanism model to obtain the prediction deviation of the mechanism model, which will be used as the neural network. output variable. In order to further distinguish the influence of different variables on product yield and product properties, and improve the prediction accuracy of the model, a two-layer neural network model with dual input is used as the learning algorithm. The network structure is shown in Figure 6. The reaction system variables and raw oil property variables that have a greater impact on product yield and distribution are input from the first hidden layer after dimensionality reduction through principal component analysis, while the fractionation system and absorption system that have less impact on product yield and distribution The variables of the stable system are input from the second hidden layer after dimensionality reduction by principal component analysis; a two-layer neural network containing 5 to 15 neurons is constructed, and the sample data is randomly divided into training set, verification set and test set, Parameter learning adopts the method of gradient descent and error backpropagation. After training, a predictive model of input and output is obtained. The error of the model is shown in Figure 7 and Figure 8.

(13) The model is automatically updated. Steps (7)-(12) can be combined with real-time data in the production process, and the real-time data of the device is used to perform parameter correction and automatic update of the deviation compensation model, thereby realizing automatic update of the model.

The above descriptions are the specific embodiments of the present invention and the technical principles used. If the changes made according to the conception of the present invention do not exceed the spirit covered by the description and accompanying drawings, they should still be Belong to the protection scope of the present invention.

Claims (10)

1. A method for simulating and predicting a catalytic cracking unit based on a molecular-level mechanism model and big data technology, characterized in that it includes the following steps:

   (1) Establish a catalytic cracking unit model, wherein the catalytic cracking unit model includes a raw material molecular analysis model, a molecular dynamics model, a riser reactor model, a product cutting model, and a physical property model;

   (2) based on actual industrial data, the parameters of the model built in step (1) are corrected;

   (3) Based on the processing results of step (2), a deviation compensation prediction model based on a machine learning algorithm is established; wherein, the input of the deviation compensation prediction model is the actual operating parameters of the device, and the output is the difference between the predicted value of the mechanism and the actual output value of the historical working condition The deviation of the actual working conditions is compensated to improve the prediction accuracy of the model;

   (4) Through steps (2) and (3), a communication mechanism can be established with the real-time data of the catalytic cracking unit in the production process, and the device data can be read in real time and the model correction and deviation compensation prediction model can be updated to realize the prediction model Automatic updates.
2. A method for simulating and predicting a catalytic cracking unit based on a molecular-level mechanism model and big data technology according to claim 1, characterized in that: the method of the raw material molecular analysis model established in the step (1) is specifically as follows:

   (1.1) Raw material molecular library construction: carry out experimental analysis and characterization of the raw material, determine the core structure of the raw material molecule, add side chains, branch chains and methyl groups on the basis of the core molecular structure according to a certain strategy, and obtain the raw material molecular library;

   (1.2) Raw material molecular concentration analysis: the initial value of the molecular concentration is set according to the composition characteristics of the raw material according to the probability distribution, and then the distribution parameters and molecular concentration are adjusted through the global optimization algorithm, so that the final molecular concentration distribution can meet the requirements of the raw material. Macroscopic properties can be analyzed into detailed molecular composition according to the macroscopic physical properties of the feed such as density, carbon residue, sulfur content, nitrogen content, group composition, and distillation range; through the molecular concentration composition construction technology, according to various macroscopic properties that can be obtained through experimental analysis Properties Inversion analysis of molecular composition of FCC feedstock.
3. A method for simulating and predicting a catalytic cracking unit based on a molecular-level mechanism model and big data technology according to claim 1, wherein the method of the molecular-level kinetic model in the step (1) is specifically as follows:

   (1.3) Write reaction rules and construct reaction network: According to the carbenium ion mechanism of catalytic cracking reaction, reactant selection rules and product generation rules were established for different types of reactions, and a class of rule functions was written for each type of reaction , compiled a large class of reaction rules including cracking, ring-opening, isomerization, hydrogen transfer, and condensation, and used computer-aided technology to apply reaction rules to raw material molecules to automatically generate reaction networks; among them, the types of reaction rules are preferably within 10 pcs - 50 pcs.
4. A method for simulating and predicting a catalytic cracking unit based on a molecular-level mechanism model and big data technology according to claim 1, wherein the method for the riser reactor model in the step (1) is specifically as follows:

   (1.4) Build a reactor model, solve the model, and calculate the product molecular concentration distribution; the built reactor model includes common catalytic cracking processes such as single riser model, MIP double riser model, DCC main and auxiliary riser parallel model; The reaction network, stoichiometry, reaction rate equation and kinetic parameters are combined with the above reactor model to obtain a complete FCC reactor model.
5. A method for simulating and predicting a catalytic cracking unit based on a molecular-level mechanism model and big data technology according to claim 1, wherein the method of the product cutting model in the step (1) is specifically as follows:

   (1.5) Product cutting model: For the product oil-gas mixture molecules coming out of the reactor, the mixed oil-gas is cut and separated into stream products of dry gas, liquefied gas, gasoline, diesel oil, oil slurry, and coke according to various product quality requirements; , the product cutting model can adopt a simple cutting model based on boiling point cutting, while considering the influence of overlapping factors;

   The concrete method of the physical property model in described step (1) is as follows:

   (1.6) Calculation of physical properties: Through the physical property calculation model, the group contribution method and the empirical correlation method are used to calculate the properties of the products with the molecular concentration of each product, including the physical properties of gasoline and diesel.
6. A method for simulating and predicting a catalytic cracking unit based on a molecular-level mechanism model and big data technology according to claim 1, characterized in that, the specific method of parameter correction in the step (2) is: through actual industrial data to The parameters of the model are corrected to complete the molecular-level mechanism model; industrial data include reactor structure size parameters, catalyst parameters, property detection data of incoming and outgoing materials (LIMS data), and device operating parameters (DCS data).
7. A method for simulating and predicting a catalytic cracking unit based on a molecular-level mechanism model and big data technology according to claim 1, characterized in that: the specific method of the step (3) is as follows:

   (3.1) Data collection and arrangement: read the DCS and LIMS historical production data of the device, establish a database, standardize the format and establish indexing rules to facilitate later query and call;

   (3.2) Data preprocessing: Extract data from the database for processing, including missing interpolation processing, outlier processing, data smoothing and noise reduction, and data normalization processing;

   (3.3) Variable correlation analysis: Calculate the correlation between variables through the correlation algorithm, and combine expert experience analysis to select the most relevant variables for modeling; specifically, the method of variable correlation analysis chooses Pearson correlation analysis , transfer entropy, and Granger causality analysis, combined with expert experience analysis, select highly correlated variables from many variables for modeling;

   (3.4) Steady-state analysis: The steady-state detection of the system is carried out through the established steady-state analysis rules, and the corresponding working conditions in each steady state are extracted, and a database is established using the obtained steady-state working conditions to facilitate subsequent updates and use;

   (3.5) Establish a deviation compensation prediction model: input the input variables corresponding to the historical working conditions into the established mechanism model to obtain the predicted value of the mechanism model, and then calculate the deviation between the predicted value of the mechanism and the actual output value of the historical working condition, through The principal component analysis reduces the input dimension of the input variables, establishes the deviation compensation prediction model through the machine learning algorithm, and adds the deviation compensation to the predicted value of the mechanism model to obtain the final prediction result.
8. A method for simulating and predicting a catalytic cracking unit based on a molecular-level mechanism model and big data technology according to claim 1, characterized in that: the missing interpolation processing method in the step (3.2) selects linear interpolation and cubic spline interpolation Any one of , mean interpolation, and Lagrangian interpolation; the outlier identification method chooses any one of the 3σ criterion method, box plot method, and Grubbs test method; the data noise reduction smoothing method chooses robust quadratic regression method to eliminate high-frequency noise signals and retain low-frequency data trends.
9. A method for simulating and predicting catalytic cracking units based on molecular-level mechanism models and big data technology according to claim 1, characterized in that: the steady-state analysis method in the step (3.4) is divided into univariate steady-state analysis and a steady-state analysis of the system:

   (i) Univariate steady-state analysis: use the wavelet transform method to extract the trend of the data set, and decompose the signal into the noise of the high frequency band and the low frequency band representing the trend of the signal to perform finite continuous approximation on the measured data, and obtain the process The approximate function f(t) of the variable establishes a steady-state index β (0≤β≤1) to represent the stability of the variable working condition. When β=0, the process variable is in an unsteady state. When β=1, The process variable is in a steady state. When 0<β<1, it means that the process variable is in a transition state, and the closer β is to 1, the more stable the state of the process variable is; the steady state index β(t) is determined by the first derivative of the data trend f '(t ₀ ) and the second derivative f”(t ₀ ) are jointly determined according to the following criteria:

   θ(t)=|f'(t)|+γ|f"(r)|

   In the formula:

   

   

   Among them, T _s , T _w , and _Tu are the thresholds, and the determination method is as follows: select a section of data in the historical database that is in a steady state as a reference, and extract the change trend of the process variable through wavelet transform to obtain the process variable at the sampling point. The first-order derivative sequence and the second-order derivative sequence of , by calculating their percentiles respectively, then: T _s is the percentile of the first-order derivative, T _w is the percentile of the second-order derivative, generally 90 % quantile or 95% quantile, where

   T _u =αT _s

   α is an adjustable parameter, which generally takes an integer between [2,5]; through the above formula, the steady-state judgment threshold of the process variable can be obtained;

   (ii) System steady-state analysis: The steady-state index β _i of each variable is obtained by using the single-variable steady-state detection rule, and the steady-state condition of the entire system is determined by the weighting of the steady-state index of each variable according to the following formula:

   

   Among them, p is the number of key characteristic variables of the system, β _i (t) is the steady-state index of the i-th variable, and u _i is the weight of the i-th variable.
10. A method for simulating and predicting catalytic cracking units based on molecular-level mechanism models and big data technology according to claim 1, characterized in that: the machine learning algorithm in the step (3.5) selects feedforward neural network and recurrent neural network , support vector machine, least squares method, least squares support vector machine, extreme learning machine regression algorithm.

Images