Bidirectional adversarial training catalytic cracking regeneration process parameter optimization method based on GAN framework

Abstract

The invention discloses a catalytic cracking regeneration process parameter optimization method based on two-way adversarial training of a GAN framework, and the method comprises the steps: constructing a data-driven prediction model and a mechanism-driven prediction model, executing the two-way adversarial training under a generative adversarial network framework, enabling the two models to take the roles of a generator and a discriminator in turn, and achieving the mutual game and collaborative optimization, the data driving model learns to generate parameters conforming to physical laws through physical constraint verification of the mechanism model, the mechanism driving model is fused into actual production optimal practice through historical experience verification of the data model, and a fusion optimization model is generated and applied to process parameter control of the regenerator of the catalytic cracking device. Efficient recovery of catalyst activity and effective reduction of regeneration energy consumption are achieved, and the technical problems that a traditional method depends on artificial experience, optimization precision is not high, a data model lacks physical constraints, and a mechanism model is difficult to adapt to complex working conditions are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of petroleum refining and artificial intelligence technology, and particularly relates to a method for optimizing catalytic cracking regeneration process parameters based on a GAN framework bidirectional adversarial training. BACKGROUND

[0002] Catalytic cracking is the most important secondary processing technology in the petroleum refining industry, and is widely used in the conversion of heavy distillate oil into gasoline, diesel and liquefied gas and other light oil products. In the catalytic cracking process, the catalyst contacts with the raw oil in the reactor to occur cracking reaction, and at the same time, the coke is inevitably formed on the surface of the catalyst, which leads to the rapid decrease of the catalyst activity. In order to maintain the catalyst activity, the coke on the surface of the catalyst must be removed by burning in the regenerator to restore the catalyst activity. The catalyst is continuously circulated between the reactor and the regenerator to form the core operation mode of the catalytic cracking unit.

[0003] The core function of the regenerator is to burn and remove the coke on the surface of the deactivated catalyst under a high-temperature oxidation atmosphere. The key process parameters of the regeneration process include the regenerator temperature, the coke-burning air volume and the regeneration time. The regenerator temperature directly affects the coke burning rate and the thermal stability of the catalyst. If the temperature is too low, the coke burning will be incomplete, and if the temperature is too high, the catalyst may be hydrothermally deactivated and the equipment safety risk may be caused. The coke-burning air volume determines the oxygen supply amount and affects the completeness of the coke burning and the regeneration efficiency. The regeneration time is related to the material balance and the production load of the unit. There is a complex coupling relationship among the three parameters, and they are affected by various factors such as the amount of coke on the catalyst, the type of catalyst and the properties of raw materials.

[0004] The traditional optimization of catalytic cracking regeneration process parameters mainly relies on the experience judgment of process engineers, and the regeneration parameters are manually adjusted by observing the regenerator temperature curve, flue gas composition and other indicators and relying on years of accumulated experience. This method has obvious shortcomings: first, the optimization precision is not high, and it is difficult to find the truly optimal parameter combination; second, the response is lagging, and it is not possible to adjust in time when the working conditions change; third, the energy consumption control is not good, and a conservative high-temperature long-time regeneration strategy is often used, resulting in energy waste; and fourth, the consistency is poor, and there are differences in the judgment of different operators.

[0005] The generative adversarial network is a deep learning framework that has achieved great success in image generation, data enhancement and other fields in recent years. The core idea of GAN is to train the generator and the discriminator in an adversarial manner, so that the generator can generate realistic data. In the traditional GAN, the roles of the generator and the discriminator are fixed, and the generator is responsible for generating data, and the discriminator is responsible for judging the authenticity. However, in the application of industrial process optimization, both the data-driven model and the mechanism-driven model have their own generation and discrimination abilities, and forcing them to be fixed in a single role cannot fully exert their respective advantages.

[0006] Therefore, it is urgent to solve the technical problems of existing catalytic cracking regeneration process parameter optimization, such as relying on artificial experience, low optimization accuracy, lack of physical constraints in data-driven models, difficulty of mechanism-driven models to adapt to actual complex working conditions, and simple model fusion methods. SUMMARY

[0007] In view of the technical problems of existing catalytic cracking regeneration process parameter optimization, such as relying on artificial experience, low optimization accuracy, lack of physical constraints in data-driven models, difficulty of mechanism-driven models to adapt to actual complex working conditions, and simple model fusion methods, the application provides a catalytic cracking regeneration process parameter optimization method based on GAN framework bidirectional adversarial training.

[0008] The application realizes a bidirectional adversarial training mechanism under the generation adversarial network framework by constructing a data-driven prediction model and a mechanism-driven prediction model, so that the two models take turns to play the roles of generator and discriminator, the data-driven model generates regeneration parameters as the generator and accepts the physical constraint verification of the mechanism-driven model as the discriminator in the first adversarial training stage, the mechanism-driven model generates regeneration parameters as the generator and accepts the historical experience verification of the data-driven model as the discriminator in the second adversarial training stage, the depth game and collaborative optimization between the models are realized through the two-stage alternating adversarial training, a fusion optimization model is generated and applied to the actual parameter control of the regenerator, and the technical problems in the traditional method are solved.

[0009] To achieve the above object, the application adopts the following technical scheme:

[0010] The catalytic cracking regeneration process parameter optimization method based on GAN framework bidirectional adversarial training comprises the following steps:

[0011] Obtain catalyst regeneration historical data, wherein the catalyst regeneration historical data comprises catalyst coke deposition amount, regenerator temperature, oxygen concentration, regeneration time, coke burning air volume and post-regeneration catalyst activity;

[0012] Construct a data-driven prediction model and a mechanism-driven prediction model, wherein the data-driven prediction model learns the nonlinear mapping relationship between regeneration parameters and catalyst activity recovery in the catalyst regeneration historical data based on a deep neural network, and the mechanism-driven prediction model establishes the physical and chemical constraints of the regeneration process based on catalyst coke combustion kinetics equation and catalyst deactivation-regeneration cycle mechanism;

[0013] The bidirectional adversarial training is performed under a GAN framework, including: in a first adversarial training stage, setting the data-driven prediction model as a generator, setting the mechanism-driven prediction model as a discriminator, generating a first optimized regeneration parameter combination based on the catalyst coke amount, the first optimized regeneration parameter combination including a first regenerator temperature set value and a first coke burning air volume set value, inputting the first optimized regeneration parameter combination into the discriminator, and evaluating whether the first optimized regeneration parameter combination meets the physical and chemical conditions of complete coke combustion based on a catalyst coke combustion kinetics equation by the discriminator, if the evaluation result is not met, generating a first adversarial loss signal, and adjusting the neural network weight parameters of the data-driven prediction model according to the first adversarial loss signal; in a second adversarial training stage, setting the mechanism-driven prediction model as a generator, setting the data-driven prediction model as a discriminator, generating a second optimized regeneration parameter combination based on the catalyst coke amount, the second optimized regeneration parameter combination including a second regenerator temperature set value and a second coke burning air volume set value, inputting the second optimized regeneration parameter combination into the discriminator, and evaluating the deviation degree of the second optimized regeneration parameter combination from the actual production optimal parameters based on the catalyst regeneration historical data by the discriminator, if the deviation degree exceeds a preset threshold, generating a second adversarial loss signal, and adjusting the kinetic parameters of the mechanism-driven prediction model according to the second adversarial loss signal; alternately performing the first adversarial training stage and the second adversarial training stage until the first adversarial loss signal and the second adversarial loss signal are converged to a convergence threshold.

[0014] A fusion optimization model is generated, which integrates the prediction results of the data-driven prediction model and the mechanism-driven prediction model, and determines target regeneration parameters according to the current catalyst coke amount, the target regeneration parameters including a target regenerator temperature, a target coke burning air volume and a target regeneration time.

[0015] The target regeneration parameters are applied to a regenerator of a catalytic cracking device, the regenerator performs catalyst regeneration operation according to the target regenerator temperature and the target coke burning air volume, and efficient recovery of catalyst activity is realized.

[0016] Compared with the prior art, the present application has the following beneficial effects:

[0017] The application innovatively introduces the GAN framework into the catalytic cracking regeneration process parameter optimization, and realizes the deep fusion of the data-driven model and the mechanism-driven model through the bidirectional adversarial training mechanism. Compared with the traditional simple weighted fusion method, the bidirectional adversarial training makes the two models respectively strengthen the advantages and make up for the deficiencies in the mutual game, the data-driven model learns to generate parameters conforming to the physical law through the physical constraint verification of the mechanism model, and the mechanism-driven model learns to be close to the optimal practice of actual production through the historical experience verification of the data model, and finally the generated fusion optimization model has theoretical rigor and practical effectiveness.

[0018] The application sets up a physical rationality verification mechanism in the first adversarial training stage, ensures that the regeneration parameters output by the data-driven model meet the basic physical and chemical conditions of coke combustion, and avoids the unreasonable prediction that may occur in the pure data-driven method, such as violating the conservation law and exceeding the safety constraints. Through the historical experience verification mechanism set in the second adversarial training stage, the mechanism-driven model integrates the successful experience of actual production on the basis of theoretical calculation, and avoids the problem that the pure mechanism model may have a theoretically feasible but poor actual effect.

[0019] The application directly applies the trained fusion optimization model to the actual control of the catalytic cracking unit regenerator, realizes the complete closed loop from data acquisition, model training to process parameter optimization control. Through real-time monitoring of the regeneration effect and feedback to the model, continuous learning and optimization are realized, and the model performance is continuously improved with running time. Industrial application shows that after using the method of the application, the catalyst activity recovery rate is increased by 3% to 5%, the regeneration energy consumption is reduced by 8% to 12%, the regenerator temperature fluctuation is reduced by 40%, and the device running stability is significantly improved.

[0020] The bidirectional adversarial training mechanism proposed in the application has good universality, and is not only suitable for catalytic cracking regeneration process, but also can be applied to parameter optimization of other petroleum refining processes such as continuous reforming, hydrocracking, delayed coking, etc., and provides a new technical path for data-mechanism fusion modeling of industrial processes. BRIEF DESCRIPTION OF DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the application or the related art, the following will briefly introduce the drawings needed to be used in the embodiment or related art description. Obviously, the drawings in the following description are only some embodiments of the application, and those skilled in the art can also obtain other drawings according to these drawings without creating any inventive labor.

[0022] Figure 1 is the overall process schematic diagram of the GAN framework bidirectional adversarial training catalytic cracking regeneration process parameter optimization method of the application.

[0023] Figure 2This is a schematic diagram illustrating the construction principle of the mechanism-driven prediction model of this invention.

[0024] Figure 3 This is a schematic diagram illustrating the generation and application of the fusion optimization model of this invention. Detailed Implementation

[0025] The following is in conjunction with the appendix Figures 1-3 The present invention will be further described in detail with reference to specific embodiments.

[0026] Reference Figure 1 This invention provides a method for optimizing process parameters of catalytic cracking regeneration based on bidirectional adversarial training using a GAN framework. The method includes five main stages: data acquisition and preprocessing, model building, bidirectional adversarial training, generation of fusion optimization model, and application control of regeneration parameters.

[0027] I. Data Acquisition and Preprocessing Stage:

[0028] First, raw regeneration data is collected from the distributed control system (DCS) of the catalytic cracking unit. Catalytic cracking units are typically equipped with a comprehensive DCS system that records key parameters in real time, such as regenerator temperature, main air flow rate, regenerated flue gas composition, and catalyst circulation rate. In a typical catalytic cracking unit, the sampling frequency for the regenerator dense-phase bed temperature is 1 time / second, the sampling frequency for the main air flow rate is 1 time / second, and the sampling frequency for flue gas oxygen and carbon dioxide content is 1 time / 10 seconds.

[0029] The raw regeneration data has several quality issues requiring preprocessing. During the regenerator start-up and shutdown phases, the system is in a non-steady-state state, resulting in drastic fluctuations in temperature and flow data; these abnormal data points need to be removed. The criteria are as follows: when the regenerator temperature is below 550℃ or above 800℃, it is considered a start-up or shutdown phase; when the main air flow rate is below 20,000 Nm³ / h, it is considered abnormal operation. Furthermore, due to instrument malfunctions or communication interruptions, data may be missing for certain periods. For short-term data gaps (less than 5 minutes), linear interpolation is used to fill in the gaps; for long-term data gaps (more than 5 minutes), the data for that period is not included in subsequent analysis.

[0030] High-frequency noise exists in the temperature and flow data, requiring smoothing filtering. A moving average filtering method is used, applying a 60-second sliding window to the temperature data and a 30-second sliding window to the flow data, effectively removing high-frequency noise while preserving the data trend characteristics.

[0031] After preprocessing, characteristic parameters for each regeneration cycle need to be extracted from the continuous time-series data to generate catalyst regeneration history data. The regeneration process of the catalytic cracking unit is periodic, with a typical regeneration cycle of 2–4 hours. By identifying the peak and trough values ​​of flue gas carbon dioxide content, the start and end times of a single regeneration cycle can be determined. For each regeneration cycle, the following characteristic parameters are extracted:

[0032] Catalyst coke content: Calculated based on the difference in carbon content of the catalyst before and after regeneration and the catalyst circulation rate. In actual industrial plants, the catalyst carbon content is usually measured by an online carbon content analyzer. The catalyst carbon content before regeneration is generally 0.8% to 1.5%, and after regeneration it decreases to 0.03% to 0.08%.

[0033] Regenerator temperature: The average value of the dense-phase bed temperature during the regeneration cycle, typically ranging from 650 to 750°C. The dense-phase bed temperature reflects the intensity of coke combustion and is the most important control parameter in the regeneration process.

[0034] Oxygen concentration: The average oxygen content in the regenerated flue gas is taken, with a typical range of 1% to 5%. Too low an oxygen concentration indicates insufficient oxygen supply, and the carbon deposits may not burn completely; too high an oxygen concentration indicates excessive oxygen, resulting in energy waste.

[0035] Regeneration time: The duration from the start of carbon deposit combustion (when the carbon dioxide content in the flue gas begins to rise) to the end of combustion (when the carbon dioxide content in the flue gas drops to a stable low value), typically ranging from 2 to 4 hours.

[0036] Coking air volume: This is the average value of the main air flow rate during the regeneration cycle, typically ranging from 30,000 to 50,000 standard cubic meters per hour. The coking air volume determines the oxygen supply and is a key parameter for controlling the combustion rate.

[0037] Catalyst activity after regeneration: Determined through microreactor activity evaluation experiments. The microreactor activity index reflects the catalyst's cracking capacity, typically ranging from 65 to 75. The activity recovery rate can be calculated as the ratio of the activity after regeneration to the activity of the fresh catalyst.

[0038] After a year of data accumulation, a typical catalytic cracking unit can obtain approximately 2,000 to 3,000 complete regeneration cycle data. These data are divided into training, validation, and test sets in a 7:2:1 ratio for subsequent model training, validation, and performance evaluation.

[0039] II. Model Building Phase:

[0040] This invention constructs two core models: a data-driven prediction model and a mechanism-driven prediction model.

[0041] The data-driven prediction model employs a multi-layer fully connected neural network architecture. The network's input layer receives four input features: catalyst coking amount, catalyst cycle count, feedstock char content (Kondratiev residue), and feedstock density. Catalyst coking amount is the most important input feature, directly determining the amount of carbon that needs to be removed by combustion. Catalyst cycle count reflects the degree of catalyst aging; aged catalysts are generally more difficult to regenerate. Feedstock char content reflects the degree of feedstock heavyness; feedstocks with high char content tend to generate more coking. Feedstock density is also related to coking tendency; high-density feedstocks typically contain more polycyclic aromatic hydrocarbons (PAHs).

[0042] The network consists of three hidden layers: the first hidden layer contains 256 neurons, the second contains 128 neurons, and the third contains 64 neurons. This progressively decreasing layer design aligns with the feature extraction principle in deep learning: shallow layers extract low-level features, and deep layers extract high-level abstract features. All hidden layers employ the ReLU activation function, which is defined as... , where x is the input of the neuron. The ReLU function can introduce nonlinearity while avoiding the vanishing gradient problem, thus accelerating network training.

[0043] The output layer contains three neurons, outputting the predicted optimal regenerator temperature, optimal scorch airflow, and optimal regeneration time, respectively. The output layer uses a sigmoid activation function, which is defined as follows: , where x is the input to the output layer neuron and e is the base of the natural logarithm. The Sigmoid function normalizes the output value to the 0-1 interval, facilitating subsequent inverse normalization to obtain the actual physical parameter values.

[0044] The normalization of different regeneration parameters and their specified intervals are explained separately below.

[0045] For the regenerator temperature, the normalized formula is: ,in The normalized temperature value is T, where T is the actual temperature in °C, and 650°C and 750°C are the lower and upper limits of the temperature, respectively (i.e., the specified range of the regenerator temperature is the 0-1 interval of (650°C, 750°C)). The inverse normalization formula is... .

[0046] For the scorching air volume, the normalized formula is: ,in The normalized air volume value is given by F, where F is the actual air volume in standard cubic meters per hour (STM). 30000 and 50000 represent the lower and upper limits of the air volume, respectively, meaning the specified range for the charring air volume is the 0-1 interval (30000, 50000). The inverse normalization formula is as follows: .

[0047] For the regeneration time, the normalization formula is: ,in Here, t represents the normalized time value, where t is the actual regeneration time in hours. 2 hours and 4 hours are the lower and upper limits of the regeneration time, respectively. Therefore, the specified range for the regeneration time is the 0-1 interval within the range of (2,4) hours. The inverse normalization formula is... .

[0048] The data-driven prediction model was trained using the Adam optimization algorithm, which combines the advantages of momentum and adaptive learning rate adjustment. The initial learning rate was set to 0.001, the batch size to 32, and the number of training epochs to 200. The loss function used was the mean squared error, defined as... ,in The loss of the data model is N, where N is the number of samples. For the true optimal parameters of the i-th sample, These are the parameters predicted by the model.

[0049] Reference Figure 2 The mechanism-driven prediction model is constructed based on the kinetic equations of catalyst coke combustion. Catalyst coke combustion is a complex heterogeneous catalytic oxidation reaction involving elementary steps such as oxygen adsorption on the catalyst surface, the reaction of carbon atoms with adsorbed oxygen, and product desorption. Under actual industrial conditions, this process can be described by an extended form of the Voorhies coke combustion kinetic equation.

[0050] The equation for the rate of carbon deposit combustion is: ,in, The value represents the carbon deposition combustion rate, expressed in kg / (kg catalyst·h). It is a frequency factor and its unit is related to the reaction order. The apparent activation energy is given in kJ / mol, R is the universal gas constant with a value of 8.314 J / (mol·K), and T is the regenerator temperature in K. This refers to the oxygen concentration, expressed in mol / L. The carbon content on the catalyst surface is expressed as a percentage. The reaction order of oxygen is given. Let be the order of the coking reaction, and e be the base of the natural logarithm.

[0051] For typical fluidized catalytic cracking catalysts, the apparent activation energy The value ranges from 100-150 kJ / mol, and 125 kJ / mol is preferred in this embodiment. Frequency factor The value of is closely related to the catalyst type. For Y-type molecular sieve-based catalysts, the preferred value is . The oxygen reaction order is typically 0.5-1.0, reflecting the effect of oxygen concentration on the combustion rate; in this embodiment, 0.8 is preferred. The coking reaction order is typically 1.0-1.5, reflecting the effect of coking content on the combustion rate; in this embodiment, 1.2 is preferred.

[0052] The time required for complete combustion of carbon deposits can be obtained by integrating the above rate equation:

[0053] ,

[0054] in, The time for burning carbon deposits is expressed in hours. The amount of catalyst coke before regeneration is expressed as a percentage. The target residual carbon content is typically required to be below 0.05%.

[0055] Catalyst deactivation-regeneration cycle constraints include three aspects. The constraints on complete combustion of coke deposits require calculated... The regeneration time should not exceed the actual usable time, and the amount of residual carbon after regeneration should be... Below 0.05%. If this constraint is not met, it indicates insufficient oxygen supply or excessively low temperature, preventing the carbon deposits from burning completely within a limited time.

[0056] The thermal stability of the catalyst limits the regenerator temperature to no more than 750°C. Molecular sieve catalysts undergo hydrothermal deactivation at high temperatures, resulting in the destruction of their crystal structure and a permanent decrease in activity. Industrial experience shows that the hydrothermal deactivation rate of the catalyst accelerates significantly when the regenerator temperature exceeds 750°C; therefore, strict control of the upper temperature limit is necessary.

[0057] Economic constraints related to regenerative energy consumption limit the coking air volume to be minimized while still meeting the combustion requirements of the coke deposits. While excessive oxygen supply can accelerate the combustion rate, it leads to energy waste, and excessively high flue gas volume increases the load on subsequent flue gas treatment. The optimization objective is to make the coking air volume as close as possible to the theoretical requirement, while ensuring complete combustion of the coke deposits.

[0058] The mechanism-driven prediction model also incorporates an adaptive adjustment mechanism for kinetic parameters. Different catalyst types (such as rare earth Y-type, ultrastable Y-type, ZSM-5 modified, etc.) and different feedstock properties (such as vacuum gas oil, catalytic diesel, coal tar, etc.) affect the properties of coke and the ease of combustion. By establishing the correlation between catalyst type, feedstock properties, and kinetic parameters, adaptive adjustment of parameters can be achieved. For example, for feedstocks containing a high proportion of coal tar, the generated coke contains more refractory polycyclic aromatic hydrocarbon structures, and the apparent activation energy increases by 10-20 kJ / mol, requiring a corresponding increase in regeneration temperature.

[0059] III. Two-way confrontation training phase:

[0060] This invention innovatively implements a bidirectional adversarial training mechanism within the GAN framework, which is the core innovation that distinguishes it from traditional model fusion methods.

[0061] In traditional GANs, the roles of generator and discriminator are fixed: the generator generates data, and the discriminator distinguishes between true and false data. However, in the scenario of optimizing regeneration parameters in catalytic cracking, both data-driven and mechanism-driven models possess generation and discrimination capabilities. Data-driven models can generate predicted optimal parameters based on coke deposition and determine whether given parameters conform to historical best practices; mechanism-driven models can generate theoretically optimal parameters based on kinetic calculations and determine whether given parameters satisfy physicochemical constraints. Therefore, this invention designs a bidirectional adversarial training mechanism, allowing the two models to take turns playing the roles of generator and discriminator, achieving a deeper level of mutual promotion.

[0062] The two-way combat training is conducted in two alternating phases.

[0063] In the first adversarial training phase, the data-driven prediction model acts as the generator, while the mechanism-driven prediction model acts as the discriminator. The specific process is as follows:

[0064] A batch of samples is randomly selected from the training set, and their catalyst coking amount is used as input. The coking amount is input into the data to drive the prediction model, and the model outputs the first optimized regeneration parameter combination, including the first regenerator temperature setpoint and the first coking air volume setpoint.

[0065] The first optimized regeneration parameter combination is input into the mechanism-driven prediction model, which acts as a discriminator. The discriminator performs a physical feasibility test based on the catalyst coking combustion kinetics equation. The test process involves inputting the first regenerator temperature setpoint... and the first burnt air volume setting value Substitute the values ​​into the carbon deposit combustion rate equation to calculate the combustion rate at that temperature and oxygen concentration. Oxygen concentration The theoretical combustion time can be calculated based on the coking air volume and the oxygen content in the air (21%). The integral combustion rate equation yields the theoretical combustion time. and final carbon residue .

[0066] The physical rationality judgment criterion is: if More than 4 hours (the upper limit of actual usable regeneration time), or Exceeding 0.05% (residual carbon content requirement), or Below 600℃ (the minimum ignition temperature for coke combustion), or the calculated oxygen consumption exceeds... If the total amount of oxygen that can be provided, or the calculated heat release from combustion, results in a temperature rise exceeding 200°C, it is deemed physically unreasonable.

[0067] If the physical condition is deemed unreasonable, the discriminator outputs a discrimination score close to 0. For example, 0.1; if the violation is deemed physically reasonable, the discriminator outputs a discrimination score close to 1, such as 0.9. The specific value of the discrimination score can be adjusted according to the severity of the violation; the more severe the violation, the closer the score is to 0.

[0068] The first adversarial loss is calculated based on the discriminant score. The goal of the first adversarial loss is to make the parameters generated by the data-driven model able to deceive the mechanism-driven discriminator, i.e., to make the discriminator output a score close to 1. The loss function is defined as:

[0069] ,

[0070] in, As the first instance of combat losses, The physical plausibility score output by the mechanism-driven discriminator is used to determine the mechanism. Represents the natural logarithm. When When approaching 0, Approaching infinity generates a strong punitive signal; when When it approaches 1, A value approaching 0 indicates that the generated parameters are sufficient to deceive the discriminator.

[0071] Using the backpropagation algorithm, the first adversarial loss Calculate the gradient of the neural network weight parameters in the data-driven prediction model, and update the parameters using gradient descent. The update formula is:

[0072] ,

[0073] in, The parameters represent the data-driven model. The learning rate is preferably 0.0001. The gradient of the loss with respect to the parameters.

[0074] Through multiple iterations, the data-driven model gradually learns to generate regeneration parameters that satisfy physical constraints, and its output no longer violates the basic laws of carbon deposit combustion.

[0075] In the second adversarial training phase, the roles are reversed: the mechanism-driven prediction model acts as the generator, while the data-driven prediction model acts as the discriminator. The specific process is as follows:

[0076] Similarly, the amount of carbon deposits of a batch of samples randomly selected from the training set is used as input. The amount of carbon deposits is input into the mechanism-driven prediction model, and the model calculates and outputs the second optimized combination of regeneration parameters based on the kinetic equations, including the second regenerator temperature setpoint and the second coking air volume setpoint.

[0077] The second optimized regeneration parameter combination is used as the input data to drive the prediction model, which acts as a discriminator. The discriminator is validated based on historical catalyst regeneration data. The validation process involves retrieving cases from historical data with similar coke deposit amounts to the current input (similar defined as a difference in coke deposit amount less than 0.1%), and then selecting typical cases with excellent regeneration effects. The criteria for excellent regeneration effects are: catalyst activity recovery rate greater than 95% after regeneration, and coking air volume lower than the average of all cases with that coke deposit amount.

[0078] Calculate the similarity between the second optimized regeneration parameter combination and the typical case. The similarity calculation formula is:

[0079] ,

[0080] Where S represents the similarity score and its value ranges from 0 to 1. The temperature setpoint for the second regenerator. The temperature of a typical historical case and These are the upper and lower limits of the temperature, namely 750℃ and 650℃. Set the second burnout airflow value. The air volume for typical historical cases. and These are the upper and lower limits of air volume, namely 50,000 and 30,000 standard cubic meters per hour, respectively.

[0081] If the similarity S is lower than the preset similarity threshold (preferably 0.7), the discriminator outputs a discrimination score close to 0. For example, 0.1; if the similarity S is higher than the threshold, the discriminator outputs a discrimination score close to 1, such as 0.9.

[0082] The second adversarial loss is calculated based on the discriminant score. The goal of the second adversarial loss is to enable the parameters generated by the mechanism-driven model to deceive the data-driven discriminator, i.e., to make the discriminator believe that the parameters are highly similar to historical best practices. The loss function is defined as:

[0083] ,

[0084] in, For the second confrontation loss, The historical experience conformity score output by the data-driven discriminator.

[0085] The second adversarial loss is used to adjust the dynamic parameters of the mechanism-driven prediction model. While the framework of the mechanism model is based on physical equations, the dynamic parameters within it (such as activation energy) are adjusted accordingly. Frequency factor The reaction order (n and m) has a certain adjustment space. By adjusting these parameters using gradient descent, the mechanistic model's predictions can be made closer to optimal practices in actual production while maintaining physical meaning. The update formula is:

[0086] ,

[0087] in, These represent the dynamic parameters of the mechanism-driven model.

[0088] The two adversarial training phases alternate, with each phase consisting of 5 iterations before switching to the other. The entire bidirectional adversarial training process continues until both adversarial loss signals converge to the convergence threshold. The convergence criterion is: within 10 consecutive iterations, and The variation range is less than 0.001. In actual training, convergence is usually achieved after 100-200 rounds of alternating training.

[0089] The innovative value of two-way adversarial training lies in the following: the data-driven model learns physical constraints through the first stage of adversarial training, avoiding unreasonable predictions from pure data methods; the mechanism-driven model incorporates practical experience through the second stage of adversarial training, avoiding the disconnect between pure theoretical methods and practice; the two models achieve complementary advantages and deep integration in mutual game, and finally reach a Nash equilibrium state, at which point the prediction results of the two models are both in line with physical laws and close to optimal practice.

[0090] IV. Generation stage of fusion optimization model:

[0091] Reference Figure 3 After bidirectional adversarial training, the data-driven prediction model and the mechanism-driven prediction model are optimized respectively, and the two need to be merged to generate the final fusion optimized model.

[0092] The fusion strategy employs a dynamic weighted fusion method. Unlike traditional fixed weights or simple averaging, this invention dynamically calculates the confidence scores of the two models based on the current operating conditions and determines the fusion weights based on these confidence scores.

[0093] Prediction confidence of data-driven prediction models The calculation is based on the coverage of the current input data in the training data. The calculation method is as follows: find the k samples in the training dataset that are closest to the current carbon deposition amount (preferred). ), calculate the average difference between the amount of carbon deposits in these k samples and the current amount of carbon deposits. .like A very small value indicates that the current operating conditions are adequately covered in the training data, and the data model's prediction reliability is high; if A large value indicates that the current operating conditions are outside the training range, and the extrapolation prediction reliability of the data model is low.

[0094] The confidence level is calculated using the following formula:

[0095] ,

[0096] in, The confidence level of the data-driven model is 0-1. The attenuation coefficient, preferably 5, is used to control the rate at which the confidence level decreases as the difference increases, and e is the base of the natural logarithm. Average coke deposit variation, expressed as a percentage. That is, the current operating conditions are exactly the same as those of the training samples. ;along with Increase Exponential decay.

[0097] Prediction confidence of mechanism-driven prediction models The calculations are based on the similarity between the current catalyst type and feedstock properties and the standard operating conditions at the time the model was established. The kinetic parameters of the mechanistic model are determined based on a specific catalyst type (such as rare earth Y-type) and a specific feedstock (such as vacuum gas oil). When the actual catalyst or feedstock differs significantly from the standard operating conditions, the applicability of the model will decrease.

[0098] Similarity can be calculated using key properties of the catalyst (such as specific surface area, pore volume, and acidity) and key properties of the feedstock (such as density, Causeway residue, and aromatic content). Euclidean distance is used to measure similarity.

[0099] ,

[0100] Where D is the Euclidean distance, and P is the number of property parameters, preferably 6 (3 key parameters each for catalyst and feedstock). This is the normalized value of the current i-th property parameter. This is the normalized value of the i-th property parameter under standard operating conditions.

[0101] The confidence level is calculated using the following formula: ,in, The confidence level of the mechanism-driven model is 0-1. This is the attenuation coefficient, preferably 2, used to control the rate at which the confidence level decreases with increasing distance. When D=0, the current operating condition is exactly the same as the standard operating condition. As D increases, Exponential decay.

[0102] After obtaining the confidence scores of the two models, the Softmax normalization method is used to determine the fusion weights.

[0103]

[0104] in, and These are the fusion weights for the data-driven model and the mechanism-driven model, respectively. Softmax normalization ensures that the sum of the two weights is 1, and the model with higher confidence receives more weights.

[0105] The final target regeneration parameters are obtained through weighted fusion:

[0106] ,

[0107]

[0108] in, These are the target regenerator temperature, the target charring air volume, and the target regeneration time, respectively. The predicted values ​​are from the data-driven model. These are the predicted values ​​from the mechanism-driven model.

[0109] The advantages of the dynamic weighted fusion strategy are as follows: When the operating condition is within the training data coverage and close to the standard operating condition, the confidence of both models is high, and the fusion weight is close to 0.5, making full use of the information from both models; when the operating condition exceeds the training range but is still close to the standard operating condition, the confidence of the data model decreases while the confidence of the mechanism model remains high, and the fusion weight automatically tilts towards the mechanism model, leveraging the extrapolation capability of the mechanism model; when the operating condition is within the training range but deviates from the standard operating condition, the confidence of the mechanism model decreases while the confidence of the data model remains high, and the fusion weight automatically tilts towards the data model, utilizing the data model's adaptability to actual complex operating conditions.

[0110] V. Regeneration Parameter Application and Control Stage:

[0111] The target regeneration parameters generated by the fusion optimization model need to be applied to the actual control of the catalytic cracking unit to complete the closed loop from model prediction to process optimization.

[0112] The regenerator control system of a catalytic cracking unit typically employs a distributed control system (DCS), possessing comprehensive automatic control functions. Target regenerator temperature. This is converted into a dense phase bed temperature control command for the regenerator. The dense phase bed temperature is the most important controlled variable in the regenerator, and temperature control is achieved by adjusting the main airflow. The control logic uses a PID controller, with the PID parameters being: proportional gain... Integral time Second, differential time Second.

[0113] The governing equations are: ,in, The output of the PID controller is the main airflow regulation command. Temperature deviation is the difference between the target temperature and the actual measured temperature. For time, It is the integral variable.

[0114] Target charring air volume This is converted to the main fan speed setpoint. The main fan typically uses variable frequency speed control, and the speed is positively correlated with the flow rate. After setting the speed, the main air flow rate will be adjusted accordingly. However, due to the pressure drop in the regenerator bed, the flow rate and speed are not strictly linearly related. In practical applications, closed-loop flow control is used. The actual flow rate is fed back through a flow measurement instrument, and the main fan speed is fine-tuned to ensure that the actual flow rate tracks the target coking air volume. The flow control accuracy is better than ±500 Nm³ / h.

[0115] During the regeneration process, the regeneration effect needs to be monitored in real time. Key monitoring indicators include the oxygen content and carbon dioxide content in the flue gas at the regenerator outlet. The oxygen content reflects the sufficiency of the oxygen supply, while the carbon dioxide content reflects the intensity of coke combustion.

[0116] The combustion process of carbon deposits can be divided into three stages. In the initial stage of combustion, as the catalyst enters the regenerator and oxygen is supplied, the carbon deposits are gradually ignited, and the carbon dioxide content in the flue gas rises rapidly while the oxygen content remains at a low level. In the middle stage of combustion, a large amount of carbon deposits burns violently, the carbon dioxide content reaches its peak, and the oxygen content remains low; at this point, the regenerator temperature rises to its highest point. In the later stage of combustion, as the carbon deposits are gradually depleted, the carbon dioxide content in the flue gas begins to decrease, while the oxygen content gradually increases.

[0117] When the oxygen content in the flue gas rises to 3%–5% and the carbon dioxide content drops to a stable low value (usually below 1%), the combustion of the carbon deposits is considered basically complete. The time from the start to the completion of combustion is recorded as the actual regeneration time. .

[0118] Comparison with actual regeneration time With target regeneration time Calculate the relative deviation:

[0119] ,

[0120] in, The relative deviation of regeneration time is expressed as a percentage. If the deviation exceeds 10%, it indicates that there is a significant bias in the model's prediction, and the bias information needs to be fed back to the fusion optimization model.

[0121] The bias feedback mechanism enables online learning and continuous optimization of the model. Actual data from the current regeneration (coke deposition, optimal parameters, and regeneration effect) are added as new samples to the catalyst regeneration history data. When the accumulated number of new samples reaches a specified number (which can be preset and adjusted, for example, 100 regeneration cycles can be optimized, corresponding to approximately 1-2 months of running time), the model retraining process is triggered.

[0122] Model retraining uses updated historical data to re-execute bidirectional adversarial training within the GAN framework, generating a new fusion-optimized model. The retraining process can be performed offline, after which the new model is deployed to the online control system, replacing the old model. Through this periodic retraining, the model can continuously learn the latest operating characteristics of the device, adapt to long-term trends such as catalyst aging and changes in feedstock properties, and maintain prediction accuracy and optimization performance.

[0123] The method of this invention has been tested in an industrial application trial on a 1.5 million tons / year catalytic cracking unit of a refinery, and has achieved significant results.

[0124] The unit employs a riser reactor with a designed processing capacity of 1.5 million tons per year. The main feedstock is vacuum-pressed wax oil, and the catalyst is a rare-earth Y-type molecular sieve-based catalyst. The regenerator uses a fully regeneration mode, with a designed dense-phase bed temperature of 680-720℃.

[0125] Before applying the method of this invention, the operation of the regenerator in this device mainly relied on the experience of the shift operators. A typical operating procedure was as follows: based on the catalyst return rate and the estimated amount of coke, a target temperature for the dense-phase bed was set, usually around 700°C; the main air flow rate was adjusted to maintain the bed temperature at the target value; the oxygen content in the flue gas was observed, and combustion was considered complete when the oxygen content reached 4%–5%. This operating method has significant shortcomings: the temperature setting is conservative, typically using a higher temperature to ensure complete coke combustion, but high temperatures accelerate catalyst deactivation; the main air flow rate adjustment relies mainly on manual operation, resulting in a delayed response and large fluctuations; and there is a lack of precise differentiation for different amounts of coke, often leading to a one-size-fits-all approach.

[0126] After applying the method of this invention, a bidirectional adversarial training optimization system based on the GAN framework was established. First, regeneration history data from the device over the past two years, totaling approximately 2800 complete regeneration cycles, were collected. The data includes catalyst coke deposition, regenerator temperature curves, main air flow rate curves, flue gas composition curves, and catalyst activity data after regeneration for each regeneration cycle.

[0127] Both data-driven and mechanism-driven prediction models were constructed. The data model employed a three-layer fully connected network of 256-128-64, while the mechanism model was based on the Voorhies coke combustion kinetic equation, considering the actual catalyst type and feedstock properties of the device. The kinetic parameters were calibrated as follows: activation energy 126 kJ / mol, frequency factor... The oxygen reaction order is 0.75, and the coking reaction order is 1.15.

[0128] Bidirectional adversarial training was performed within the GAN framework, with a total of 150 rounds of alternating training. The first adversarial loss converged from the initial 2.3 to 0.08, and the second adversarial loss converged from the initial 1.9 to 0.06, achieving a good balance between the two models.

[0129] The trained fusion optimization model was deployed to the unit's DCS system, enabling automated optimization of regeneration parameters. The system automatically collects the current catalyst coke content every hour, calls the fusion optimization model to calculate the target regeneration parameters, and sends the parameters to the regenerator control system for execution.

[0130] Comparative analysis of data from three consecutive months of operation shows that:

[0131] The catalyst activity recovery effect was significantly improved. Before application, the average micro-reaction activity index of the regenerated catalyst was 68.5, which increased to 70.8 after application, and the activity recovery rate increased from 92.3% to 95.6%, an increase of 3.3 percentage points. The reason for the improved activity is that the optimized regeneration parameters can be precisely adjusted according to the amount of coke deposits, ensuring complete combustion of the coke deposits while avoiding hydrothermal deactivation caused by excessively high temperatures.

[0132] Regeneration energy consumption has been significantly reduced. Before application, the average coking air volume was 42,000 Nm³ / h, which decreased to 38,500 Nm³ / h after application, a reduction of 8.3%. The main fan power consumption decreased from an average of 320 kW to 290 kW, a reduction of 9.4%. The reason for the reduced energy consumption is that the optimized model can calculate the theoretical oxygen demand based on the actual amount of coke deposits, avoiding excessive air supply in traditional operations. Based on 8,000 hours of operation per year, the annual electricity cost savings are approximately 180,000 yuan.

[0133] The regenerator's operational stability has been significantly improved. Before application, the standard deviation of the dense phase bed temperature was 12.8℃, which decreased to 7.5℃ after application, a reduction of 41.4% in fluctuation. The reason for the reduced temperature fluctuation is that the optimized system has achieved automated control, avoiding the lag and inconsistency of manual operation.

[0134] The accuracy of regeneration time prediction has been improved. After application, the average relative deviation between the actual regeneration time and the model prediction time was 6.2%, while the deviation estimated by operators before application was usually 15% to 20%. The improved prediction accuracy helps to optimize production planning and accurately control material balance.

[0135] In addition, operators reported that the system significantly reduced labor intensity, freeing them from frequent manual adjustments and allowing them to devote more energy to equipment inspection and handling of abnormal operating conditions.

[0136] The technical solution of the present invention has good versatility and scalability, and can be modified in various ways according to actual needs.

[0137] Regarding catalyst types, this invention is applicable not only to traditional rare-earth Y-type molecular sieve catalysts, but also to other types of fluidized catalytic cracking catalysts such as ultrastable Y-type, ZSM-5 modified, and Beta molecular sieves. The coke combustion kinetic parameters of different catalysts differ and can be adapted through an adaptive kinetic parameter adjustment mechanism. For multi-metal catalysts (such as nickel-vanadium catalysts), the catalytic effect of metals on coke combustion also needs to be considered, requiring the addition of corresponding correction terms to the mechanism-driven model.

[0138] Regarding feedstock types, this invention is applicable not only to traditional vacuum gas oil but also to other heavy distillate oils such as catalytic diesel, coal tar, slurry oil, and deasphalted oil. The properties of the coke deposits generated from different feedstocks vary significantly. For example, coal tar tends to produce recalcitrant polycyclic aromatic hydrocarbon deposits, requiring higher regeneration temperatures. Model parameters can be adjusted by modifying feedstock property parameters (such as Causeway residue, aromatic content, and hydrogen-to-carbon ratio) to adapt to different feedstocks.

[0139] Regarding the regeneration mode, this invention mainly targets the full regeneration mode (residual carbon after regeneration is less than 0.05%), but it can also be extended to the partial regeneration mode (residual carbon 0.1% to 0.3%). In the partial regeneration mode, residual carbon control constraints need to be added to the objective function to allow a certain amount of residual carbon while ensuring catalyst activity, so as to save regeneration energy consumption.

[0140] In terms of model structure, data-driven prediction models are not limited to fully connected neural networks; they can also employ time-series models such as recurrent neural networks (RNNs) and long short-term memory networks (LSTMs) to better capture the dynamic characteristics of the regeneration process. Mechanism-driven prediction models are not limited to the Voorhies equation; they can also employ more complex multi-step reaction kinetics models or computational fluid dynamics models to improve the accuracy of the physical description.

[0141] Regarding fusion strategies, in addition to confidence-based dynamic weighted fusion, more advanced methods such as Bayesian fusion and Kalman filter fusion can also be used. For cases where prediction results differ significantly, an expert system can be introduced for arbitration, combining the experience and judgment of process engineers to ensure the reliability of the optimization results.

[0142] In terms of control execution, in addition to PID control of temperature and flow, advanced control strategies such as model predictive control (MPC) and adaptive control can be introduced to achieve multivariate constraint optimization. At the same time, multiple control objectives such as regenerator temperature, flue gas oxygen content, and catalyst circulation rate are considered to seek the global optimum under multiple constraints.

[0143] In terms of application scope, this invention is not only applicable to catalytic cracking regeneration processes, but can also be extended to other petroleum refining and petrochemical processes. For example, catalyst regeneration in continuous reforming, catalyst sulfidation in hydrocracking, and coke gasification in delayed coking all involve complex physicochemical processes and multi-parameter optimization problems. The bidirectional adversarial training method of the GAN framework of this invention can be used to construct a fusion optimization system of data-driven and mechanism-driven models, achieving intelligent optimization of process parameters and precise control of the production process.

[0144] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for optimizing catalytic cracking regeneration process parameters based on bidirectional adversarial training using a GAN framework, characterized in that... include: Acquire historical data on catalyst regeneration, including catalyst coke deposit amount, regenerator temperature, oxygen concentration, regeneration time, coke burner air volume, and catalyst activity after regeneration; Constructing data-driven and mechanism-driven prediction models; Performing bidirectional adversarial training within the GAN framework includes: In the first adversarial training phase, the data-driven prediction model is set as the generator, and the mechanism-driven prediction model is set as the discriminator. A first optimized regeneration parameter combination is generated based on the catalyst coke deposition amount. The first optimized regeneration parameter combination includes a first regenerator temperature setpoint and a first coking airflow setpoint. The first optimized regeneration parameter combination is input into the discriminator, which evaluates whether the first optimized regeneration parameter combination meets the physicochemical conditions for complete coke combustion based on the catalyst coke combustion kinetic equation. If the evaluation result is not met, a first adversarial loss signal is generated. The neural network weight parameters of the data-driven prediction model are adjusted according to the first adversarial loss signal. In the second adversarial training phase, the mechanism-driven... The prediction model is set as a generator, and the data-driven prediction model is set as a discriminator. A second optimized regeneration parameter combination is generated based on the catalyst coke deposition. The second optimized regeneration parameter combination includes a second regenerator temperature setpoint and a second coking air volume setpoint. The second optimized regeneration parameter combination is input into the discriminator, which evaluates the deviation between the second optimized regeneration parameter combination and the actual optimal parameters based on the catalyst regeneration historical data. If the deviation exceeds a preset threshold, a second adversarial loss signal is generated. The kinetic parameters of the mechanism-driven prediction model are adjusted according to the second adversarial loss signal. The first adversarial training phase and the second adversarial training phase are executed alternately until both the first adversarial loss signal and the second adversarial loss signal converge to the convergence threshold. A fusion optimization model is generated, which integrates the prediction results of the data-driven prediction model and the mechanism-driven prediction model. The target regeneration parameters are determined based on the current catalyst coke content. The target regeneration parameters include the target regenerator temperature, the target coke burning air volume, and the target regeneration time. The target regeneration parameters are applied to the regenerator of the catalytic cracking unit, and the regenerator is controlled to perform catalyst regeneration operation according to the target regenerator temperature and the target coke burning air volume, so as to achieve efficient recovery of catalyst activity.

2. The method according to claim 1, characterized in that, The data-driven prediction model is based on deep neural networks to learn the nonlinear mapping relationship between regeneration parameters and catalyst activity recovery in the historical data of catalyst regeneration. The mechanism-driven prediction model is based on the catalyst coking combustion kinetic equation and the catalyst deactivation-regeneration cycle mechanism to establish physicochemical constraints on the regeneration process. Before obtaining the catalyst regeneration history data, the process also includes: Raw regeneration data is collected from the distributed control system of the catalytic cracking unit. The raw regeneration data includes the regenerator temperature curve, the main air flow curve, the regeneration flue gas oxygen content curve, and the catalyst circulation flow data. The raw regeneration data is preprocessed, including: removing abnormal data during the regenerator start-up and shutdown phases, filling in data gaps caused by instrument malfunctions, and performing smoothing filtering on temperature and flow data; Based on the regenerator temperature curve and the main air flow curve, the combustion rate and cumulative combustion amount of catalyst coke in each regeneration cycle are calculated to generate the catalyst regeneration history data.

3. The method according to claim 2, characterized in that, The construction of the data-driven prediction model specifically includes: A multi-layer fully connected neural network is constructed, comprising an input layer, three hidden layers, and an output layer. The input layer receives the catalyst coke amount, the number of catalyst cycles, and the raw material property parameters, while the output layer outputs the predicted optimal regeneration parameters. The three hidden layers contain 256, 128, and 64 neurons respectively, and each hidden layer uses the ReLU activation function to introduce nonlinear mapping capability. The output layer employs a Sigmoid activation function to normalize the output regeneration parameters to a specified range, where the regenerator temperature is normalized to the 0-1 range corresponding to 650-750℃, and the coking air volume is normalized to the 0-1 range corresponding to 30000-50000 standard cubic meters per hour.

4. The method according to claim 2, characterized in that, Constructing mechanism-driven prediction models specifically includes: A catalyst coking combustion model based on the Voorhies coking kinetics equation is established, which describes the quantitative relationship between the coking combustion rate and the regenerator temperature, oxygen concentration and catalyst coking amount. The catalyst deactivation-regeneration cycle constraints are set, including: complete combustion condition constraints for coke deposits, catalyst thermal stability constraints, and regeneration energy consumption economic constraints. The complete combustion condition constraints for coke deposits require that the amount of residual coke in the catalyst after regeneration be less than 0.05%. The catalyst thermal stability constraints limit the regenerator temperature to no more than 750°C to prevent hydrothermal deactivation of the catalyst. The regeneration energy consumption economic constraints limit the coking air volume to be minimized while meeting the coke deposit combustion requirements. An adaptive adjustment mechanism for kinetic parameters is introduced to dynamically adjust the reaction activation energy and frequency factor in the catalyst coking combustion model according to different catalyst types and raw material properties.

5. The method according to claim 1, characterized in that, The specific steps of performing bidirectional adversarial training within the GAN framework include: In the first adversarial training phase, the physical constraint violation degree of the first optimized regeneration parameter combination is calculated. The physical constraint violation degree is determined by substituting the first regenerator temperature setpoint and the first coking air volume setpoint into the catalyst coking combustion kinetic equation to calculate the theoretical coking combustion time. If the theoretical coking combustion time exceeds the actual usable regeneration time or the calculated regenerated residual carbon content exceeds 0.05%, then it is determined that there is a physical constraint violation. In the second adversarial training phase, the historical data deviation of the second optimized regeneration parameter combination is calculated. The historical data deviation is determined by retrieving historical regeneration cases with similar coke deposits from the catalyst regeneration historical data, calculating the difference between the second regenerator temperature setpoint and the historical optimal temperature, and the difference between the second coking air volume setpoint and the historical optimal air volume, and summing the absolute values ​​of the two differences to obtain the historical data deviation. A bidirectional adversarial loss function is constructed, which includes a generator loss term and a discriminator loss term. The generator loss term measures the ability of the generated regenerated parameters to deceive the discriminator, and the discriminator loss term measures the ability of the discriminator to correctly distinguish between the true optimal parameters and the generated parameters.

6. The method according to claim 1, characterized in that, The generated fusion optimization model specifically includes: The prediction confidence of the data-driven prediction model is calculated. The prediction confidence is determined based on the coverage of the current catalyst coke amount in the training data. If the current catalyst coke amount falls in a dense region of the training data, the confidence is high; if it falls in a sparse region, the confidence is low. The prediction confidence of the mechanism-driven prediction model is calculated, and the prediction confidence is determined based on the similarity between the current catalyst type and feedstock properties and the standard operating conditions when the mechanism-driven prediction model was established. Based on the prediction confidence of the data-driven prediction model and the prediction confidence of the mechanism-driven prediction model, a dynamic weighted fusion strategy is adopted to determine the fusion weights of the data-driven prediction model and the mechanism-driven prediction model respectively. The regeneration parameters output by the two models are weighted and averaged to generate the target regeneration parameters.

7. The method according to claim 1, characterized in that, The application of the target regeneration parameters to the regenerator of the catalytic cracking unit specifically includes: The target regenerator temperature is converted into a regenerator dense phase bed temperature control command, and the regenerator dense phase bed temperature is controlled to track the target regenerator temperature by adjusting the main air flow rate. The target coking air volume is converted into a main fan speed set value, and the main air flow is controlled by adjusting the main fan speed to achieve the target coking air volume. Real-time monitoring of oxygen and carbon dioxide content in the flue gas at the regenerator outlet. When the oxygen content in the flue gas rises to 3% to 5% and the carbon dioxide content drops to a stable low value, the carbon deposit combustion is determined to be complete, and the actual regeneration time is recorded. If the actual regeneration time is compared with the target regeneration time, and the deviation exceeds 10%, the deviation information is fed back to the fusion optimization model for parameter optimization in the next regeneration cycle.

8. The method according to claim 1, characterized in that, The method further includes: Collect activity evaluation data of the regenerated catalyst, including the catalyst micro-reaction activity index, gasoline yield and dry gas yield; The activity evaluation data is correlated with the target regeneration parameters and stored in the catalyst regeneration history data for continuous learning and optimization of the data-driven prediction model and the mechanism-driven prediction model. When the cumulative number of newly added regeneration cases reaches a specified number, the model retraining process is triggered, and bidirectional adversarial training is re-executed under the GAN framework to update the fusion optimization model.

9. The method according to claim 1, characterized in that, The first adversarial training phase also includes: A physical rationality verification mechanism is set up. When the first optimized regeneration parameter combination generated by the data-driven prediction model meets one of the following conditions, it is judged to be physically unreasonable: the first regenerator temperature setpoint is lower than the minimum ignition temperature of coke combustion; the calculated oxygen consumption exceeds the total amount of oxygen that the first coke burning air volume can provide; the calculated combustion heat release causes the regenerator temperature rise to exceed the threshold. When a physical inconsistency is determined, the discriminator outputs a discrimination score close to 0, generating a high-intensity first adversarial loss signal, which forces the data-driven prediction model to significantly adjust its parameters to generate physically reasonable regeneration parameters.

10. The method according to claim 1, characterized in that, The second adversarial training phase also includes: A historical experience verification mechanism is set up to screen typical cases with excellent regeneration effect from the historical data of catalyst regeneration. The typical cases with excellent regeneration effect meet the requirements that the catalyst activity recovery rate after regeneration is greater than 95% and the regeneration energy consumption is lower than the average level. The similarity between the second optimized regeneration parameter combination and the typical case is calculated. If the similarity is lower than the preset similarity threshold, the discriminator outputs a discrimination score close to 0 and generates a high-intensity second adversarial loss signal. Through the historical experience verification mechanism, the mechanism-driven prediction model incorporates successful experiences from actual production on the basis of theoretical calculations, avoiding the output of regeneration parameters that are theoretically feasible but have poor practical effects.

Images