阅读说明：本技术 一种基于分子级机理模型与大数据技术的催化裂化装置模拟预测方法 (Catalytic cracking unit simulation prediction method based on molecular level mechanism model and big data technology ) 是由 元梦琪 涂文辉 何恺源 于 2021-09-16 设计创作，主要内容包括：本发明涉及一种基于分子级机理模型与大数据技术的催化裂化装置模拟预测方法,本发明能够对催化裂化过程的产物收率和产品性质进行预测；本发明建立催化裂化过程的分子级机理模型,该分子级机理模型不仅可以提高预测精度,还可以适用于不同的装置,具有良好的外延性；另外,基于大数据技术对由实际装置运行状态造成机理模型预测偏差进行校正,不仅抓住催化裂化反应的本质,还能反映不同催化装置的特点,精准地预测产物收率以及关键产品性质,可实现工业级装置的准确过程模拟。(The invention relates to a catalytic cracking unit simulation prediction method based on a molecular-level mechanism model and a big data technology, which can predict the product yield and the product property in the catalytic cracking process; the invention establishes a molecular-level mechanism model of the catalytic cracking process, and the molecular-level mechanism model not only can improve the prediction precision, but also can be suitable for different devices and has good extensibility; in addition, the mechanism model prediction deviation caused by the actual device running state is corrected based on a big data technology, so that the essence of catalytic cracking reaction is grasped, the characteristics of different catalytic devices can be reflected, the product yield and the key product property can be accurately predicted, and the accurate process simulation of an industrial device can be realized.)

1. A catalytic cracking unit simulation prediction method based on a molecular level mechanism model and a big data technology is characterized by comprising the following steps:

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

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

(3) establishing a deviation compensation prediction model based on a machine learning algorithm based on the processing result of the step (2); the input of the deviation compensation prediction model is the actual operation parameter of the device, the output is the deviation between the mechanism prediction value and the actual output value of the historical working condition, and the prediction precision of the model is improved through the deviation compensation of the actual working condition;

(4) and (3) establishing a communication mechanism with real-time data of the catalytic cracking unit in the production process through the step (2) and the step (3), reading the data of the catalytic cracking unit in real time, and updating the model correction and deviation compensation prediction model to realize automatic updating of the prediction model.

2. The catalytic cracking unit simulation prediction method based on molecular-scale mechanism model and big data technology as claimed in claim 1, characterized in that: the method for establishing the raw material molecule analytical model in the step (1) specifically comprises the following steps:

(1.1) constructing a raw material molecule library: performing experimental analysis and characterization on the raw material, determining the core structure of a raw material molecule, and adding a side chain, a branched chain and a methyl group on the basis of the core molecular structure according to a certain strategy to obtain a raw material molecule library;

(1.2) analysis of molecular concentration of raw Material: setting the initial value of the molecular concentration according to the probability distribution according to the composition characteristics of the raw material, and then adjusting the distribution parameters and the molecular concentration through a global optimization algorithm to ensure that the final molecular concentration distribution can meet various macroscopic properties of the raw material and can be analyzed into detailed molecular compositions according to the macroscopic physical properties of the feeding material, such as density, carbon residue, sulfur content, nitrogen content, group compositions and distillation range; and performing inversion analysis on the molecular composition of the catalytic cracking raw material according to various macroscopic properties which can be obtained through experimental analysis by using a molecular concentration composition construction technology.

3. The catalytic cracking unit simulation prediction method based on molecular-scale mechanism model and big data technology as claimed in claim 1, characterized in that: the method of the molecular-level kinetic model in the step (1) is specifically as follows:

(1.3) compiling a reaction rule, and constructing a reaction network: according to the carbonium ion mechanism of catalytic cracking reaction, reactant selection rules and product generation rules are respectively established for different types of reactions, a rule function is respectively compiled for each type of reaction, a large class of reaction rules containing five main reactions of cracking, ring opening, isomerism, hydrogen transfer and condensation are compiled, and a reaction network is automatically generated by applying the reaction rules to raw material molecules by using a computer-assisted technology; wherein the number of reaction rules is preferably 10 to 50.

4. The catalytic cracking unit simulation prediction method based on molecular-scale mechanism model and big data technology as claimed in claim 1, characterized in that: the method of the riser reactor model in step (1) is specifically as follows:

(1.4) establishing a reactor model, solving the model, and calculating the molecular concentration distribution of a product; the reactor model comprises common catalytic cracking processes such as a single riser model, a MIP double riser model, a DCC main and auxiliary riser parallel model and the like; and combining the reaction network, the stoichiometry, the reaction rate equation and the kinetic parameters with the reactor model to obtain a complete catalytic cracking reactor model.

5. The catalytic cracking unit simulation prediction method based on molecular-scale mechanism model and big data technology as claimed in claim 1, characterized in that: the method for cutting the model by the product in the step (1) is specifically as follows:

(1.5) product cleavage model: cutting and separating the oil-gas mixture molecules of the product from the reactor into dry gas, liquefied gas, gasoline, diesel oil, oil slurry and coke flow products according to the quality requirements of the products; wherein, the product cutting model can adopt a simple cutting model based on boiling point cutting, and simultaneously considers the influence of overlapping factors;

the specific method of the physical property model in the step (1) is as follows:

(1.6) physical Properties calculation: and calculating the properties of the product, including various physical properties of gasoline and diesel oil, by using the molecular concentration of each product through a physical property calculation model by adopting a group contribution method and an empirical correlation method.

6. The catalytic cracking unit simulation prediction method based on molecular-scale mechanism model and big data technology as claimed in claim 1, wherein the specific method for parameter correction in step (2) is: correcting the parameters of the model through actual industrial data to finish the molecular level mechanism model; the industrial data includes reactor structural dimension parameters, catalyst parameters, feed and discharge property measurement data (LIMS data), and plant operating parameters (DCS data).

7. The catalytic cracking unit simulation prediction method based on molecular-scale mechanism model and big data technology as claimed in claim 1, characterized in that: the specific method of the step (3) is as follows:

(3.1) data acquisition and arrangement: reading DCS and LIMS historical production data of the device, establishing a database, standardizing a format and establishing an index rule, and facilitating later-stage query and calling;

(3.2) data preprocessing: extracting data from a database to perform processing work, including missing interpolation processing, abnormal value processing, data smoothing and noise reduction and data normalization processing;

(3.3) variable association analysis: calculating the correlation among all variables through a correlation algorithm, and selecting the most relevant variable for modeling by combining expert experience analysis; specifically, the variable correlation analysis method selects any one of Pearson correlation analysis, transfer entropy and Glangel causal analysis, combines expert experience analysis, and selects variables with strong correlation from a plurality of variables for modeling;

(3.4) Steady-State analysis: carrying out steady state detection on the system through the established steady state analysis rule, extracting corresponding working conditions under each steady state, and establishing a database by using the obtained steady state working conditions, so that subsequent updating and use are facilitated;

(3.5) establishing a deviation compensation prediction model: inputting input variables corresponding to historical working conditions into the established mechanism model to obtain a predicted value of the mechanism model, then calculating the deviation between the predicted value of the mechanism and an actual output value of the historical working conditions, reducing the input dimensionality of the input variables through principal component analysis, establishing a deviation compensation prediction model through a machine learning algorithm, and adding deviation compensation to the predicted value of the mechanism model to obtain a final prediction result.

8. The catalytic cracking unit simulation prediction method based on molecular-scale mechanism model and big data technology as claimed in claim 1, characterized in that: the missing interpolation processing method in the step (3.2) selects any one of linear interpolation, cubic spline interpolation, mean value interpolation and Lagrange interpolation; selecting any one of a 3 sigma criterion method, a box plot method and a Grabbs test method by the abnormal value identification method; the data noise reduction smoothing method selects a steady quadratic regression method, eliminates high-frequency noise signals and keeps the low-frequency data trend.

9. The catalytic cracking unit simulation prediction method based on molecular-scale mechanism model and big data technology as claimed in claim 1, characterized in that: the steady state analysis method in the step (3.4) is divided into univariate steady state analysis and system steady state analysis:

(i) univariate steady state analysis: trend extraction is performed on the data set by adopting a wavelet transform method, and signals are decomposed into noise in a high frequency band and a low frequency band representing the signal trendCarrying out finite continuous approximation on the measured data to obtain an approximate function f (t) of the process variable, establishing a steady-state index beta (beta is more than or equal to 0 and less than or equal to 1) to represent the stability of the working condition state of the variable, when the beta is 0, the process variable is in an unsteady state, when the beta is 1, the process variable is in a steady state, and 0<β<1, indicates 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 derived from the first derivative f' (t) of the data trend₀) And the second derivative f ″ (t)₀) Jointly determined according to the following criteria:

θ(t)＝|f'(t)|+γ|f"(r)|

in the formula:

wherein T is_s、T_w、T_uThe threshold value is determined by the following method: selecting a section of data with a steady process in a historical database as a reference datum, extracting the variation trend of the process variable through wavelet transformation to obtain a first derivative sequence and a second derivative sequence of the process variable at a sampling point, and respectively solving the percentile of the first derivative sequence and the second derivative sequence, so that: t is_sIs a percentile of the first derivative, T_wIs a percentile of the second derivative, and can be selected from 90% quantile or 95% quantile

T_u＝αT_s

Alpha is an adjustable parameter and is generally an integer between [2 and 5 ]; by the above formula, a steady state judgment threshold value of the process variable can be obtained;

(ii) and (3) analyzing the system steady state: obtaining steady-state index beta of each variable by adopting a univariate steady-state detection rule_iThe steady state condition of the whole system is determined by weighting the steady state indexes of all variables according to the following formula:

wherein p is the number of key characteristic variables of the system, beta_i(t) is the steady state index of the ith variable, u_iIs the weight of the ith variable.

10. The catalytic cracking unit simulation prediction method based on molecular-scale mechanism model and big data technology as claimed in claim 1, characterized in that: the machine learning algorithm in the step (3.5) selects any one of a feedforward neural network, a cyclic neural network, a support vector machine, a least square method, a least square support vector machine and an extreme learning machine regression algorithm.

Technical Field

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

Background

The catalytic cracking is an important oil refining process, the total processing capacity of the catalytic cracking is listed as the top of various oil refining processes, and the technical complexity of the catalytic cracking is also at the top of various oil refining processes, so the catalytic cracking plays a significant role in the oil refining industry. The traditional catalytic cracking process simulation is mostly established on a lumped dynamics method, the lumped dynamics method divides complex components in catalytic cracking into a plurality of lumped components according to the dynamics characteristics, and each lumped component is considered as a virtual single component in the dynamics simulation. Thus, conventional lumped kinetic models are typically only predictive of product production, not product properties, and do not reflect changes in feedstock composition because of the large differences that can occur between the compositions of oils of the same properties.

With the increasingly strict requirements on the quality of the finished oil, a method for predicting the product yield and accurately predicting the product property is urgently needed. The advent of molecular-scale kinetic models provides the possibility to solve this problem. By analyzing the molecular composition of the raw materials, a molecular-level reaction kinetic network is established, and the conversion rule of reactant and product molecules in a reactor is calculated, so that the molecular composition and the product properties are accurately predicted. Compared with the traditional lumped method, the method is more accurate in prediction and wider in model adaptability. However, the molecular-level mechanism model has not been widely applied in the industry because of the complexity, many parameters and large calculation amount of the model.

The defects of the existing method mainly comprise the following aspects: (1) the traditional lumped model is poor in prediction accuracy and extensibility, control over the actual operation condition of the device is lacked, and deviation between a prediction result and actual data is difficult to avoid. (2) The catalytic cracking process is complex, the influence variables are many, the construction difficulty of a complete molecular-level mechanism model is high, the model is difficult to be suitable for different devices, the practical applicability is poor, and the actual operation condition of the device is lack of control. (3) Independent big data model, because do not consider the nature of reaction, the causal relationship between the data does not correspond, and the ductility of model is poor. In addition, the aspects of variable causal relevance, causal response time delay and the like of the big data model are not considered sufficiently, and the accuracy of the model is seriously influenced by the quality of data preprocessing.

Disclosure of Invention

The invention aims to overcome the defects and provide a catalytic cracking device simulation prediction method based on a molecular mechanism model and a big data technology, and the method can predict the product yield and the product property in the catalytic cracking process; the invention establishes a molecular-level mechanism model of the catalytic cracking process, and the molecular-level mechanism model not only can improve the prediction precision, but also can be suitable for different devices and has good extensibility; in addition, the mechanism model prediction deviation caused by the actual device running state is corrected based on a big data technology, so that the essence of catalytic cracking reaction is grasped, the characteristics of different catalytic devices can be reflected, the product yield and the key product property can be accurately predicted, and the accurate process simulation of an industrial device can be realized.

The invention achieves the aim through the following technical scheme: a catalytic cracking unit simulation prediction method based on a molecular level mechanism model and a big data technology comprises the following steps:

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

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

(3) establishing a deviation compensation prediction model based on a machine learning algorithm based on the processing result of the step (2); the input of the deviation compensation prediction model is the actual operation parameter of the device, the output is the deviation between the mechanism prediction value and the actual output value of the historical working condition, and the prediction precision of the model is improved through the deviation compensation of the actual working condition;

(4) and (3) establishing a communication mechanism with real-time data of the catalytic cracking unit in the production process through the step (2) and the step (3), reading the data of the catalytic cracking unit in real time, and updating the model correction and deviation compensation prediction model to realize automatic updating of the prediction model.

Preferably, the method for establishing the raw material molecule analysis model in step (1) specifically comprises the following steps:

(1.1) constructing a raw material molecule library: performing experimental analysis and characterization on the raw material, determining the core structure of a raw material molecule, and adding a side chain, a branched chain and a methyl group on the basis of the core molecular structure according to a certain strategy to obtain a raw material molecule library;

(1.2) analysis of molecular concentration of raw Material: setting the initial value of the molecular concentration according to the probability distribution according to the composition characteristics of the raw material, and then adjusting the distribution parameters and the molecular concentration through a global optimization algorithm to ensure that the final molecular concentration distribution can meet various macroscopic properties of the raw material and can be analyzed into detailed molecular compositions according to the macroscopic physical properties of the feeding material, such as density, carbon residue, sulfur content, nitrogen content, group compositions and distillation range; and performing inversion analysis on the molecular composition of the catalytic cracking raw material according to various macroscopic properties which can be obtained through experimental analysis by using a molecular concentration composition construction technology.

Preferably, the method of the molecular-scale kinetic model in step (1) is specifically as follows:

(1.3) compiling a reaction rule, and constructing a reaction network: according to the carbonium ion mechanism of catalytic cracking reaction, reactant selection rules and product generation rules are respectively established for different types of reactions, a rule function is respectively compiled for each type of reaction, a large class of reaction rules containing five main reactions of cracking, ring opening, isomerism, hydrogen transfer and condensation are compiled, and a reaction network is automatically generated by applying the reaction rules to raw material molecules by using a computer-assisted technology; wherein the number of reaction rules is preferably 10 to 50.

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

(1.4) establishing a reactor model, solving the model, and calculating the molecular concentration distribution of a product; the reactor model comprises catalytic cracking processes such as a single riser model, a MIP double riser model, a DCC main and auxiliary riser parallel model and the like; and combining the reaction network, the stoichiometry, the reaction rate equation and the kinetic parameters with the reactor model to obtain a complete catalytic cracking reactor model.

Preferably, the method for cutting the model by the product in the step (1) specifically comprises the following steps:

(1.5) product cleavage model: cutting and separating the oil-gas mixture molecules of the product from the reactor into dry gas, liquefied gas, gasoline, diesel oil, oil slurry and coke flow products according to the quality requirements of the products; wherein, the product cutting model can adopt a simple cutting model based on boiling point cutting, and simultaneously considers the influence of overlapping factors;

the specific method of the physical property model in the step (1) is as follows:

(1.6) physical Properties calculation: and calculating the properties of the product, including various physical properties of gasoline and diesel oil, by using the molecular concentration of each product through a physical property calculation model by adopting a group contribution method and an empirical correlation method.

Preferably, the specific method for correcting the parameters in step (2) is as follows: correcting the parameters of the model through actual industrial data to finish the molecular level mechanism model; the industrial data includes reactor structural dimension parameters, catalyst parameters, feed and discharge property measurement data (LIMS data), and plant operating parameters (DCS data).

Preferably, the specific method of step (3) is as follows:

(3.1) data acquisition and arrangement: reading DCS and LIMS historical production data of the device, establishing a database, standardizing a format and establishing an index rule, and facilitating later-stage query and calling;

(3.2) data preprocessing: extracting data from a database to perform processing work, wherein the processing work comprises missing value interpolation processing, abnormal value processing, data smoothing noise reduction and data normalization processing;

(3.3) variable association analysis: calculating the correlation among all variables through a correlation algorithm, and selecting the most relevant variable for modeling by combining expert experience analysis; specifically, the variable association analysis method selects one or more of Pearson correlation analysis, transfer entropy and Glangel causal analysis, and combines expert experience analysis to select variables with strong association from a plurality of variables for modeling;

(3.4) Steady-State analysis: carrying out steady state detection on the system through the established steady state analysis rule, extracting corresponding working conditions under each steady state, and establishing a database by using the obtained steady state working conditions, so that subsequent updating and use are facilitated;

(3.5) establishing a deviation compensation prediction model: inputting input variables corresponding to historical working conditions into the established mechanism model to obtain a predicted value of the mechanism model, then calculating the deviation between the predicted value of the mechanism and an actual output value of the historical working conditions, reducing the input dimensionality of the input variables through principal component analysis, establishing a deviation compensation prediction model through a machine learning algorithm, and adding deviation compensation to the predicted value of the mechanism model to obtain a final prediction result.

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; selecting any one of a 3 sigma criterion method, a box plot method and a Grabbs test method by the abnormal value identification method; the data noise reduction smoothing method selects a steady quadratic regression method, eliminates high-frequency noise signals and keeps the low-frequency data trend.

Preferably, the steady-state analysis method in the step (3.4) is divided into univariate steady-state analysis and system steady-state analysis:

(i) univariate steady state analysis: the method comprises the steps of extracting trends of a data set by adopting a wavelet transform method, carrying out finite continuous approximation on measured data by decomposing signals into high-frequency noise and low-frequency bands representing signal trends to obtain an approximation function f (t) of a process variable, establishing a steady-state index beta (beta is more than or equal to 0 and less than or equal to 1) to represent the stability degree of the working condition state of the variable, and when the beta is 0, the process variable is in an unsteady state, and when the beta is 1, the process variable is in a steady state, and 0 & ltWhen beta < 1, it means that the process variable is in a transition state, and the closer beta is to 1, the more stable the state of the process variable is; the steady state index β (t) is derived from the first derivative f' (t) of the data trend₀) And the second derivative f ″ (t)₀) Jointly determined according to the following criteria:

θ(t)＝|f′(t)|+γ|f″(t)|

in the formula:

wherein T is_s、T_w、T_uThe threshold value is determined by the following method: selecting a section of data with a steady process in a historical database as a reference datum, extracting the variation trend of the process variable through wavelet transformation to obtain a first derivative sequence and a second derivative sequence of the process variable at a sampling point, and respectively solving the percentile of the first derivative sequence and the second derivative sequence, so that: t is_sIs a percentile of the first derivative, T_wIs percentile of second derivative, generally 90% quantile or 95% quantile can be selected, wherein T_u＝αT_s

Alpha is an adjustable parameter and is generally an integer between [2 and 5 ]; by the above formula, a steady state judgment threshold value of the process variable can be obtained;

(ii) and (3) analyzing the system steady state: obtaining steady-state index beta of each variable by adopting a univariate steady-state detection rule_iThe steady state condition of the whole system is determined by weighting the steady state indexes of all variables according to the following formula:

wherein p is the number of key characteristic variables of the system, beta_i(t) is the steady state index of the ith variable, u_iIs the weight of the ith variable.

Preferably, the machine learning algorithm in the step (3.5) selects any one of a feedforward neural network, a recurrent neural network, a support vector machine, a least square method, a least square support vector machine, and an extreme learning machine regression algorithm.

The invention has the beneficial effects that: (1) the invention adopts a form of combining a molecular-level mechanism model and a big data technology, the molecular-level mechanism model describes the essential rule of the catalytic cracking process, the big data model fully utilizes the production historical data, can reflect the characteristics of different devices and the actual running state of the devices, and can efficiently process a data set with long time span and huge data amount; (2) according to the invention, the yield and the property of the product are calculated through the mechanism model, the deviation between the actual result and the mechanism prediction result is calculated by utilizing the big data model, and the deviation between the actual result calculated by the data model and the mechanism prediction result is increased by the mechanism model, so that the accurate prediction of the product yield and the product property is realized, and meanwhile, when the unknown working condition is faced, the model has good extrapolation performance and can be used for exploring the unknown working condition; (3) the invention adopts the molecular-level mechanism model based on structure-oriented aggregation, and compared with the traditional aggregation model, the molecular-level mechanism model has more accurate prediction precision and wider prediction range.

Drawings

FIG. 1 is a schematic flow diagram of the present invention;

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

FIG. 3 is a schematic view of various properties that can be calculated by the property model of the present invention;

FIG. 4 is a graph showing the results of a univariate steady state analysis in an embodiment of the present invention;

FIG. 5 is a diagram illustrating the steady state analysis of the system in an embodiment of the present invention;

FIG. 6 is a schematic diagram of a multi-layer neural network architecture of the present invention;

FIG. 7 is a graph illustrating the product yield prediction error of the dual-core driving model of the present invention;

FIG. 8 is a schematic diagram of product property prediction errors of the dual-core driving model according to the present invention.

Detailed Description

The invention will be further described with reference to specific examples, but the scope of the invention is not limited thereto:

example (b): the present invention will be described in further detail below with reference to a specific embodiment of a catalytic cracking unit simulation in a refinery. As shown in fig. 1, a catalytic cracking unit simulation prediction method based on a molecular-scale mechanism model and a big data technology includes the following steps:

(1) and constructing a raw material molecule library. The number of molecules that make up the catalytic cracking feedstock is quite large, while the number of homologue core structures that make up these molecules is much smaller. Adding side chains, branched chains and methyl groups according to a certain strategy on the basis of a core structure to obtain a series of subsets, calculating properties such as distillation range density of each molecule through physical property retrieval and a molecular physical property calculation model, further screening the subsets by utilizing carbon number and distillation range constraints, deleting unreasonable molecules to obtain a final raw material molecule library, and constructing the raw material molecule library containing 48 core 2473 molecules.

(2) And (4) analyzing the concentration of the raw material molecules. The molecular level reaction mechanism model needs to input molecular concentration, certain fragment information of molecular composition is obtained through different experimental analysis methods, complete molecular information of the raw material is indirectly deduced through splicing a plurality of fragments, an initial value of the molecular concentration is set according to the composition characteristics of the raw material and certain probability distribution, and then distribution parameters and the molecular concentration are adjusted through a specific global optimization algorithm, so that the final molecular concentration distribution can meet various macroscopic properties of the raw material. The accuracy of the concentration of the molecules of the raw materials has great influence on the calculation result of the mechanism model, while the molecular analysis algorithm is only a tool for obtaining the concentration of the molecules, and the accuracy of the result depends on the accuracy and completeness of experimental analysis data.

(3) Writing a reaction rule to construct a reaction network. The catalytic cracking reaction mechanism is very complex, the widely accepted mechanism is a carbonium ion mechanism, under the action of an acid center of a molecular sieve catalyst, hydrocarbon molecules capture protons to form carbonium ions, the carbonium ions are subjected to beta fracture to generate olefin and new carbonium ions, and the carbonium ions can also undergo isomerization reaction, so that the main reaction of the catalytic cracking system is formed. Aiming at a catalytic cracking chemical reaction mechanism, reaction rules are sequentially established for different types of reactions, wherein the reaction rules comprise a reactant selection rule and a product production rule, a rule function is compiled for each type of reaction, 25 types of reaction rules comprising main reactions such as cracking, ring opening, isomerism, hydrogen transfer, condensation and the like are compiled, then a computer-assisted technology is utilized to apply the reaction rules to raw material molecules, a reaction network is automatically generated, and a reaction network comprising 5216 reactions is generated (a reaction network schematic diagram is shown in figure 2). The type of reaction carried out on different catalysts will vary somewhat, and therefore the reaction rules will also need to be adapted appropriately according to experimental data.

(4) And establishing a reactor model. According to the quasi-plug flow reactor model, a mass balance equation and a heat balance equation are established, a reaction network, chemometrics, a reaction rate equation and kinetic parameters are combined with the reactor model to obtain a complete catalytic cracking reactor model which is expressed as a huge differential equation set, and the concentration conversion rule of each molecule in the reactor can be calculated by solving the differential equation set.

(5) And cutting and separating the product. A simple cut model based on boiling point cuts was used according to the typical distillation range of each product, while taking into account the effect of the overlap factor.

(6) And (4) a physical property calculation model. The invention calculates the properties of the products by using the molecular concentrations of the products and calculates the physical properties of the catalytic gasoline and the catalytic diesel oil by using a group contribution method and an empirical correlation method based on a related physical property calculation model, and the calculable physical properties are shown in figure 3.

(7) And (6) correcting model parameters. And acquiring actual production data of a factory, including the property analysis data of feeding materials and products and the operating parameters of the device, taking the minimum sum of squares of errors between a predicted value and an actual output value as a target, taking an empirical value or an experimental value as an initial value, and starting a global optimization algorithm to carry out optimization solution on the model parameters.

(8) And (6) data acquisition and arrangement. In the embodiment, complete historical production data of a catalytic cracking unit DCS system and LIMS detection of a certain refinery about one year are collected, wherein operation parameters and state parameters of the DCS system are automatically recorded once every 10 minutes, feeding material property detection data of the system is analyzed once every 3 days, and product discharging material properties of the system are analyzed once every 24 hours. And the collected DCS and LIMS data are sorted, and a basic database convenient for subsequent reading and retrieval is established.

(9) And (6) data processing. And carrying out descriptive statistical analysis on the collected DCS operation parameters and state parameters and the material detection properties of the LIMS, and analyzing the distribution rule and statistical information of the data. And processing the missing value and the outlier of the data, and preferably, judging the outlier of the data set by adopting a percentile method. And filling missing values and outliers by using an interpolation fitting method. And (3) performing noise reduction processing on the data by using a quadratic regression function, and filtering high-frequency noise signals to obtain a data set which is more in line with the data change trend.

(10) And (5) performing variable association analysis. 542 DCS operation and state variables are acquired, 340 LIMS property detection variables are acquired, and in order to simplify the input variables and retain effective information as much as possible, a group of variables most relevant to a prediction model need to be screened out. According to the actual situation of a catalytic cracking production device, combining the experience and professional knowledge of experts and matching with corresponding correlation analysis algorithms, analyzing the variables of three systems of reaction, fractionation and absorption stability, exploring the correlation relationship between each operation parameter and each state parameter, and selecting the operation and state parameters which have large influence on the yield and the property of the product as input variables of subsequent modeling, wherein the correlation analysis algorithms adopt transfer entropy algorithms.

(11) And (4) steady state analysis. In the face of a complex nonlinear system of catalytic cracking, a wavelet decomposition method based on trend extraction is adopted to perform steady-state analysis on the system. The wavelet transform has good time-frequency domain positioning and multi-resolution analysis capability, can analyze different frequency characteristic information contained in a historical data center, and carries out finite continuous approximation on measured data by decomposing signals into high-frequency-band noise and low-frequency bands representing signal trends, so as to obtain an approximate function of a process variable. Through the established univariate steady-state criterion and the system steady-state judgment criterion, all variables are subjected to steady-state analysis independently, and then the system is subjected to steady-state judgment, and the univariate steady-state analysis and the system state judgment are respectively shown in fig. 4 and 5. After the steady state of the system is determined, the system data in the steady state is extracted. And taking the average value of each section of steady-state data as variable data representing the steady-state section, and establishing a database according to the variable data, so that the subsequent analysis, extraction and use are facilitated.

(12) And establishing a deviation compensation model based on machine learning. After variable association analysis and steady-state detection analysis, input variables and corresponding data sets are determined, the data sets of the input variables are input into the established mechanism model, prediction deviations of the mechanism model can be obtained, and the deviations are used as output variables of the neural network. In order to further distinguish the influence of different variables on the product yield and the product property and improve the prediction accuracy of the model, a two-layer neural network model with double-input is adopted as a learning algorithm, and the network structure is shown in fig. 6. Inputting the variables of the reaction system and the raw oil property variables which have large influence on the product yield and distribution from a first hidden layer after the principal component analysis dimensionality reduction, and inputting the variables of the fractionation system and the absorption stabilizing system which have small influence on the product yield and distribution from a second hidden layer after the principal component analysis dimensionality reduction; two layers of neural networks respectively containing 5 to 15 neurons are constructed, sample data are randomly divided into a training set, a verification set and a test set, and a gradient descent and error reverse transfer method is adopted for parameter learning. And after training, obtaining an input and output prediction model. The error of the model is shown in fig. 7 and 8.

(13) And the model is automatically updated. And (7) the steps (7) to (12) can be combined with real-time data in the production process, and the real-time data of the device is used for carrying out parameter correction and automatic updating of the deviation compensation model, so that the automatic updating of the model is realized.

While the invention has been described in connection with specific embodiments thereof, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims.