Petrochemical reaction process intelligent regulation and control method based on dynamic neural network

Abstract

The invention discloses a petrochemical engineering reaction process intelligent regulation and control method based on a dynamic neural network, and relates to the technical field of petrochemical engineering process control, and the method comprises the following steps: obtaining temperature distribution data of a catalyst bed layer, and fluid flow velocity field data, component concentration gradient data and stirring frequency data of a reactor in real time; and inputting the temperature distribution data and the stirring frequency data into a first branch of a dynamic neural network, and outputting an activity attenuation index of the catalyst bed layer. According to the scheme, by fusing the temperature standard deviation, the stirring dominant frequency component and the by-product weighted value, the dynamic relevance of mass transfer and heat transfer on the surface of the catalyst can be captured, meanwhile, the nonlinear influence of a side reaction path on activity attenuation is considered, the abnormal activity of the catalyst is warned in advance, and a more accurate input basis is provided for regulation and control of the reaction process.

Description

Technical Field

[0001] The present invention relates to the technical field of petrochemical process control, and in particular to an intelligent control method for a petrochemical reaction process based on a dynamic neural network. Background Art

[0002] Petrochemical reaction processes (such as hydrodesulfurization and catalytic cracking) are core links in petroleum refining and chemical production. Their reaction efficiency, catalyst life, and product selectivity directly affect the economic and environmental performance of production. Traditional reaction process control mainly relies on empirical settings or PID control of static process parameters (such as temperature, pressure, feed rate, etc.), which makes it difficult to deal with dynamic problems such as catalyst activity decay, reactor coking, and changes in side reaction paths. Especially in complex reaction processes such as hydrodesulfurization, the activity decay of the catalyst bed often exhibits nonlinear characteristics and is affected by the coupling of multiple factors such as fluid flow state and component concentration distribution. Traditional control methods are difficult to achieve accurate prediction and timely regulation.

[0003] However, in the process of implementing the technical solutions of the embodiments of the present application, the inventors of the present application discovered that the above technology has at least the following technical problems:

[0004] Traditional methods lack the ability to model the dynamic process of catalyst activity decay in real time and cannot accurately predict the spatial changes in active site coverage, resulting in delayed regulation. Summary of the Invention

[0005] The purpose of the present invention is to provide an intelligent control method for petrochemical reaction processes based on dynamic neural networks to solve the problems raised in the above background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for intelligently controlling a petrochemical reaction process based on a dynamic neural network, which is used for intelligently controlling the activity decay index of a catalyst bed in a hydrodesulfurization reactor, wherein the catalyst bed is composed of N active sites at different spatial positions, comprising the following steps:

[0007] Real-time acquisition of catalyst bed temperature distribution data, reactor fluid velocity field data, component concentration gradient data, and stirring frequency data;

[0008] Inputting the temperature distribution data and the stirring frequency data into a first branch of a dynamic neural network, and outputting an activity decay index of the catalyst bed;

[0009] Determining whether the activity decay index is greater than a preset activity decay index threshold, and if so, generating a predicted value of a product position change and a predicted value of a product generation amount deviation of the active site by using the component concentration gradient data through a nonlinear combination simulation;

[0010] The predicted value of the product position change is the difference between the coverage of the product of the next cycle on the active site and the coverage of the product of the current cycle on the active site;

[0011] Determining whether the deviation of the product generation amount from the predicted value is greater than a preset product generation amount deviation threshold, and if so, obtaining the spatial coordinates of the active site, and then adjusting the feed flow rate of the active site;

[0012] Fluid flow state prediction data is calculated based on the fluid flow velocity field data and the product position change prediction value, the dead zone position in the catalyst bed is identified based on the fluid flow state prediction data, and a stirring frequency adjustment instruction is issued to the dead zone position until the activity decay index is lower than the preset activity decay index threshold.

[0013] Compared with the prior art, the present invention has the following beneficial effects:

[0014] 1. This scheme captures the dynamic correlation between mass transfer and heat transfer on the catalyst surface by integrating the temperature standard deviation, the main stirring frequency component, and the by-product weighted value. It also considers the nonlinear effect of the side reaction path on activity decay, provides early warning of abnormal catalyst activity, and provides a more accurate input basis for reaction process control.

[0015] 2. This solution dynamically divides concentration zones and establishes adjustment priorities, enabling differentiated regulation for different reaction states. This avoids the waste of resources caused by full-zone adjustments in traditional methods and improves the accuracy and timeliness of feed flow rate adjustments.

[0016] 3. By integrating fluid dynamics and mass transfer process data, this solution can more accurately predict abnormal areas of flow conditions within the catalyst bed. By further integrating flow state prediction data with temperature distribution data, it can more accurately identify true dead zones. By establishing a compensation model and closed-loop control, it can accurately match the stirring frequency adjustment amount according to different dead zone characteristics, achieve rapid improvement of the dead zone flow state, and effectively suppress fluctuations in the catalyst activity attenuation index caused by dead zone retention. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The disclosure of the present invention is described with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. In the drawings, the same reference numerals are used to refer to the same components. Among them:

[0018] Figure 1 A flowchart of the steps provided by the present invention;

[0019] Figure 2 A schematic diagram of the process of calculating fluid flow state prediction data provided by the present invention;

[0020] Figure 3 A schematic diagram of a process for identifying dead zone positions provided by the present invention;

[0021] Figure 4 A schematic diagram of a flow chart for issuing a stirring frequency adjustment instruction provided by the present invention;

[0022] Figure 5 The present invention provides a scene graph. DETAILED DESCRIPTION

[0023] It is easy to understand that according to the technical solution of the present invention, without changing the essential spirit of the present invention, a person skilled in the art can propose a variety of interchangeable structural modes and implementation modes. Therefore, the following specific embodiments and drawings are only exemplary descriptions of the technical solution of the present invention and should not be regarded as the entire invention or as a limitation or restriction of the technical solution of the present invention.

[0024] Application Overview:

[0025] In the existing technology, the regulation of petrochemical reaction processes has long relied on the empirical setting of static process parameters or traditional control methods. These methods are unable to identify nonlinear attenuation characteristics in the early stages of catalyst activity decay and lack the ability to predict sudden reaction anomalies, resulting in regulatory measures often lagging behind actual operating conditions. For example, during the hydrodesulfurization reaction, the coverage changes of the active sites in the catalyst bed are dynamically coupled with the fluid flow state, and traditional control strategies find it difficult to capture this multi-dimensional coupling effect in a timely manner.

[0026] In order to solve the above problems, it is necessary to establish a dynamic perception system that can synchronously process temperature fields, flow velocity fields and concentration gradient fields; through analysis, it is found that the decay of catalyst activity is essentially the result of the combined effect of temperature distribution changes and fluid flow unevenness, but the dynamic coupling relationship between these two types of parameters has highly nonlinear characteristics; for this reason, it is considered to construct a dynamic neural network model, taking temperature distribution and stirring frequency as input features, and capturing the nonlinear correlation of activity decay through a hierarchical learning mechanism; at the same time, it is necessary to establish a product generation deviation prediction mechanism, combine the component concentration gradient data with the kinetic equation, and simulate the generation of product position change prediction values.

[0027] After introducing the basic concept of the present invention, embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0028] See also Figure 5 The reactor in the application scenario of the present invention is provided with several feed ports and agitators, and a catalyst bed is provided inside. The catalyst bed is composed of N active sites in different spatial positions.

[0029] Example 1:

[0030] See also Figure 1-Figure 4 A petrochemical reaction process intelligent control method based on dynamic neural networks is used to intelligently control the activity decay index of the catalyst bed in a hydrodesulfurization reactor. The catalyst bed is composed of N active sites in different spatial positions, and includes the following steps:

[0031] Real-time acquisition of catalyst bed temperature distribution data, reactor fluid velocity field data, component concentration gradient data, and stirring frequency data;

[0032] Inputting the temperature distribution data and the stirring frequency data into a first branch of a dynamic neural network, and outputting an activity decay index of the catalyst bed;

[0033] Determining whether the activity decay index is greater than a preset activity decay index threshold, and if so, generating a predicted value of a product position change and a predicted value of a product generation amount deviation of the active site by using the component concentration gradient data through a nonlinear combination simulation;

[0034] The predicted value of the product position change is the difference between the coverage of the product of the next cycle on the active site and the coverage of the product of the current cycle on the active site;

[0035] Determining whether the deviation of the product generation amount from the predicted value is greater than a preset product generation amount deviation threshold, and if so, obtaining the spatial coordinates of the active site, and then adjusting the feed flow rate of the active site;

[0036] Fluid flow state prediction data is calculated based on the fluid flow velocity field data and the product position change prediction value, the dead zone position in the catalyst bed is identified based on the fluid flow state prediction data, and a stirring frequency adjustment instruction is issued to the dead zone position until the activity decay index is lower than the preset activity decay index threshold.

[0037] The first branch of the dynamic neural network refers to a neural network structure with multi-level feature extraction capabilities, which can be implemented using a convolutional neural network with long short-term memory units to capture the spatiotemporal correlation characteristics of temperature distribution and stirring frequency.

[0038] The activity decay index is a quantitative indicator reflecting the degree of deactivation of active sites on the catalyst surface. It can be obtained by weighted calculation after normalization, and its numerical change corresponds to the catalyst deactivation rate.

[0039] The preset activity decay index threshold refers to a pre-set critical value used to judge the degree of activity decline of the catalyst or reaction system. It is usually expressed as a numerical value or percentage. The threshold is determined based on reaction kinetics, catalyst life experiments or historical data and is used to guide process adjustments or catalyst regeneration / replacement.

[0040] Nonlinear combination refers to the method of fusing or calculating multiple variables, signals or data through nonlinear relationships;

[0041] The predicted value of product position change represents the trend of change in active site coverage on the catalyst surface. It can be obtained by correlating the kinetic equation with real-time concentration gradient data and used to predict the adsorption equilibrium state of reactants on the catalyst surface.

[0042] The coverage of active sites refers to the change in the proportion of active sites on the catalyst surface occupied by reactants, intermediates or poison molecules with spatial position in catalytic reactors such as fixed beds and fluidized beds.

[0043] Specifically, the temperature sensor array collects the three-dimensional temperature distribution of the catalyst bed in real time, the electromagnetic flowmeter obtains the fluid flow rate field data in real time, and the speed sensor collects the stirring frequency data in real time; the first branch of the dynamic neural network performs multi-scale feature extraction on the temperature distribution data, synchronously calculates the main frequency component of the stirring frequency data, and outputs the activity decay index of the catalyst bed; when the activity decay index output by the neural network exceeds the preset threshold, the current component concentration gradient data is extracted, and the current concentration gradient data is associated with the kinetic equation. By calculating the adsorption equilibrium constant and reaction rate constant of each component at the active site, the product position change prediction value and the product generation amount deviation prediction value are generated; for areas that deviate from the predicted value beyond the limit, the feed control valve opening at the corresponding area coordinate position is adjusted to achieve precise control of the local reaction product concentration; at the same time, based on the convolution fusion of the fluid flow rate field data and the product position change prediction value, the flow stagnation area is identified, the stirring frequency adjustment instruction is triggered, and the flow dead zone is eliminated.

[0044] Compared with existing technologies, traditional methods rely on fixed parameter thresholds for a posteriori control, while this scheme achieves early warning of catalyst activity decay through dynamic neural networks; in existing technologies, stirring frequency adjustment is usually based on empirical formulas, while this scheme combines the dynamic data of the flow velocity field with product distribution predictions to achieve spatially precise positioning of dead zone identification; conventional PID control only performs feedback adjustment on a single variable, while this scheme constructs a multi-physical field collaborative control mechanism of temperature field, flow velocity field and concentration field.

[0045] This technical solution enables early identification of early signs of catalyst activity decline, enabling preventive control measures to be implemented before product distribution anomalies become apparent. Adjusting the feed flow rate at the spatial coordinate level effectively suppresses localized coking. The synergistic effect of a dynamic neural network and a fluid dynamics model enables stirring frequency adjustment to precisely eliminate dead zones and maintain uniform activity distribution across the catalyst bed.

[0046] This application further proposes that real-time acquisition of catalyst bed temperature distribution data includes:

[0047] Acquiring product concentration data, where the product concentration data includes target product concentration data and by-product concentration data;

[0048] Input the target product concentration data and the by-product concentration data into the second branch of the dynamic neural network, and output the predicted value of the activity decay rate of the catalyst bed;

[0049] Determine whether the activity decay rate prediction value is greater than a preset activity decay rate threshold. If so, obtain the spatial position where the target product concentration data and the by-product concentration data are received, and adjust the hydrogen partial pressure data at the spatial position.

[0050] Among them, product concentration data refers to the concentration distribution information of the target product and by-product at different positions of the catalyst bed during the reaction process. Specifically, it can be achieved by using an online gas chromatograph or mass spectrometer to perform real-time sampling and analysis of the reactor outlet gas to characterize the reaction progress and catalyst activity status.

[0051] The second branch of the dynamic neural network refers to a network structure that specifically processes product concentration data. Specifically, it can be constructed by combining an LSTM network based on time series prediction with a convolutional layer. Through historical data training, a mapping relationship between concentration changes and activity decay rate is established to capture nonlinear decay trends.

[0052] The activity decay rate prediction value refers to a quantitative indicator of the rate of decrease in catalyst activity per unit time. It can be obtained by inputting the rate of change of the ratio of the target product to by-product concentrations into a neural network, and is used to evaluate the degree of catalyst performance degradation.

[0053] The preset activity decay rate threshold refers to a pre-set critical value of the activity decay rate, which can be calibrated through experiments according to the catalyst type and process requirements and used to trigger the control mechanism.

[0054] Spatial position refers to the coordinates of the local area where the concentrations of target products and by-products in the catalyst bed are abnormal. Specifically, three-dimensional grid division combined with sensor array positioning can be used to identify activity decay hotspots.

[0055] Hydrogen partial pressure data refers to the partial pressure parameter of hydrogen in the reaction environment. Specifically, it can be dynamically controlled by adjusting the hydrogen flow valve or back pressure controller to suppress side reaction paths.

[0056] Specifically, during the hydrodesulfurization reaction, an online analyzer collects the concentration distribution of the target product and by-products in real time and feeds this information into the second branch of the dynamic neural network. This branch uses a trained model to analyze the concentration gradient and output a predicted value for the activity decay rate. When the predicted value exceeds a preset threshold, the spatial coordinates of the region with abnormal concentration are obtained, and the hydrogen partial pressure is adjusted for that region. For example, if the by-product concentration increases abnormally, the hydrogen partial pressure is dynamically increased to strengthen the main desulfurization reaction pathway and suppress activity decay caused by side reactions.

[0057] Compared with existing technologies, traditional methods rely on fixed thresholds to judge the activity state of the catalyst and are unable to capture the relationship between concentration gradient changes and activity decay rate in real time; this application uses a dynamic neural network to process multi-dimensional concentration data, which can accurately predict the activity decay trend and achieve precise control of local hydrogen partial pressure based on spatial positioning, avoiding the problem of increased energy consumption caused by adjusting the parameters of the entire reactor.

[0058] Through the above technical solution, the present application solves the defect that traditional methods cannot identify nonlinear characteristics in the early stage of activity decay, and can trigger the control mechanism in the early stage of abnormal fluctuations in by-product concentration; by dynamically adjusting the local hydrogen partial pressure, the deactivation of active sites caused by side reaction pathways is effectively suppressed, the service life of the catalyst is extended, and the selectivity of the target product is improved.

[0059] This application further proposes that the hydrogen partial pressure data for adjusting the spatial position include:

[0060] The product deviation value is calculated based on the target product concentration data and the by-product concentration data, combined with the preset expected target product concentration value and the by-product concentration value, and the hydrogen partial pressure adjustment coefficient is generated based on the product deviation value, and the hydrogen partial pressure adjustment coefficient is positively correlated with the product deviation value.

[0061] The product deviation value refers to the comprehensive deviation of the target product concentration and the by-product concentration from the expected value, which can be achieved by using the weighted difference method or the root mean square error method to quantify the degree of deviation between the actual reaction state and the ideal state.

[0062] Among them, the hydrogen partial pressure adjustment coefficient refers to the adjustment parameter generated based on the product deviation value, which can be implemented by a linear proportional function or a segmented mapping algorithm. Its numerical value maintains a positive correlation with the product deviation value and is used to dynamically control the hydrogen partial pressure to correct the reaction path.

[0063] Specifically, when the target product concentration and the by-product concentration deviate from the expected value, the product deviation value is obtained by calculating the weighted difference between the two. For example, when the by-product concentration exceeds the expected value and the target product concentration is lower than the expected value, the product deviation value increases; at this time, a mapping rule between the hydrogen partial pressure adjustment coefficient and the product deviation value is established by looking up the table, and a hydrogen partial pressure adjustment coefficient is generated based on the product deviation value and the mapping rule. This coefficient is used to increase the hydrogen partial pressure at the corresponding spatial position, thereby suppressing side reactions and promoting the main reaction path; by continuously monitoring the concentration changes of the target product and the by-product, the hydrogen partial pressure parameter is corrected in real time to form a closed-loop control.

[0064] Compared with existing technologies, traditional methods usually adjust the hydrogen partial pressure based on a fixed threshold and cannot accurately control it according to dynamic concentration deviations. However, this scheme achieves real-time adaptive optimization of the reaction path by establishing a direct relationship between the product deviation value and the hydrogen partial pressure adjustment coefficient, avoiding the adjustment lag problem caused by static parameter settings.

[0065] Through the above technical solution, the present application can quickly identify abnormal accumulation of by-products in the early stage of catalyst activity decay, and effectively inhibit the progress of side reactions through dynamic compensation of hydrogen partial pressure, thereby maintaining the generation efficiency of the target product and extending the service life of the catalyst.

[0066] The present application further proposes that the activity decay index of the output catalyst bed includes:

[0067] Calculate the standard deviation of the temperature distribution data and the main frequency component of the stirring frequency data, perform weighted calculation on the by-product concentration data, multiply the standard deviation of the temperature distribution data by the main frequency component of the stirring frequency data, sum the product of the standard deviation of the temperature distribution data and the main frequency component of the stirring frequency data with the weighted value of the by-product concentration data, and output the activity decay index after normalizing the sum result.

[0068] The standard deviation of the temperature distribution data refers to the degree of dispersion of the temperature at each point in the catalyst bed from the average temperature. This can be achieved by calculating the standard deviation of the real-time temperature data collected by the temperature sensor array, and is used to characterize the uniformity of the temperature distribution.

[0069] The main frequency component of the stirring frequency data refers to the main component of the frequency distribution during the operation of the stirrer. Specifically, it can be achieved by extracting the main frequency amplitude through fast Fourier transform of the stirring frequency data, which is used to reflect the effect of stirring intensity on the mass transfer of the catalyst surface.

[0070] Among them, the weighted calculation of by-product concentration data refers to assigning different weights according to the contribution of the by-product type to the catalyst activity decay. Specifically, it can be achieved by linearly weighted summing the concentrations of multiple types of by-products using weight coefficients based on the reaction mechanism or experimental calibration, which is used to quantify the impact of side reactions on activity decay.

[0071] Specifically, in the first branch of the dynamic neural network, the product of the standard deviation of the temperature distribution data and the main frequency component of the stirring frequency is used to characterize the coupling effect of temperature fluctuations and stirring intensity on activity decay, and the weighted value of the by-product concentration is used to reflect the cumulative effect of side reaction products; the two are summed and normalized, and the activity decay index is finally output. The introduction of the weighted value of the by-product concentration quantifies the degree of adsorption and accumulation of poisoning substances such as sulfide at the active sites. Therefore, the activity decay index can comprehensively consider the dynamic characteristics of the temperature distribution, the stirring state and the changes in the by-product concentration to achieve a quantitative assessment of the degree of catalyst activity decay.

[0072] Compared with existing technologies, traditional methods usually calculate catalyst activity decay based on only single temperature data or fixed process parameters. However, this scheme can capture the dynamic correlation between catalyst surface mass transfer and heat transfer by integrating temperature standard deviation, stirring main frequency component and by-product weighted value, while considering the nonlinear influence of side reaction pathways on activity decay, thereby improving the accuracy and real-time performance of activity decay prediction.

[0073] Through the above technical solution, the present application can solve the problem that traditional methods cannot identify nonlinear characteristics in the early stage of activity decay. Through the collaborative analysis of multi-dimensional dynamic parameters, it can provide early warning of abnormal catalyst activity and provide more accurate input basis for reaction process control.

[0074] This application further proposes that the predicted values ​​of product position changes and product production deviations from the predicted values ​​of the simulated active sites include:

[0075] Determine whether the activity decay index is greater than the preset activity decay index threshold. If so, extract the current component concentration gradient data, associate the current component concentration gradient data with the kinetic equation, and calculate the adsorption equilibrium constant and reaction rate constant of each component at the active site to generate the product position change prediction value and the product production deviation prediction value.

[0076] The kinetic equation refers to the kinetic model used to describe the reaction of multiple components after adsorption on the catalyst surface. It can be implemented by numerical fitting or experimental calibration of the adsorption equilibrium constant and reaction rate constant, and is used to correlate component concentration gradient data with changes in the reaction path.

[0077] The adsorption equilibrium constant refers to the concentration ratio of each component when the adsorption equilibrium is reached on the catalyst surface. It can be achieved by online monitoring of the catalyst surface coverage or offline calibration of the adsorption isotherm, and is used to quantify the competitive adsorption relationship between different components.

[0078] Among them, the reaction rate constant refers to the amount of reactant conversion per unit active site per unit time. It can be dynamically corrected based on temperature distribution data and catalyst activity decay index to reflect the dynamic changes of active sites during the reaction process.

[0079] Specifically, when the activity decay index exceeds the threshold, the adsorption equilibrium constant of each component is calculated by dynamically correlating the current component concentration gradient data with the kinetic equation to determine the degree of active site coverage. At the same time, the reaction rate constant is combined to predict the failure trend of active sites, which can more accurately capture the correlation between the change in active site coverage and the deviation in product generation. For example, in the hydrodesulfurization reaction, the adsorption competition relationship between sulfide and hydrogen can be characterized by the adsorption equilibrium constant, while the reaction rate constant reflects the change in sulfide removal efficiency with activity decay. Therefore, based on the coupled calculation of adsorption equilibrium and reaction rate, the predicted value of product position change is generated, and the predicted value of product position change and the predicted value of product generation deviation are generated, providing a quantitative basis for subsequent regulation.

[0080] Compared with existing technologies, which usually make predictions based on fixed adsorption parameters or empirical models and cannot dynamically correlate component concentration gradients with kinetics, this solution introduces nonlinear combination simulations of kinetic equations and combines them with real-time component concentration gradient data to more accurately capture the correlation between changes in active site coverage and product generation deviations. For example, traditional PID control only adjusts parameters based on the current product concentration, while this solution can predict activity decay trends within future time windows through dynamic calculation of adsorption equilibrium constants and reaction rate constants.

[0081] Through the above technical solution, the present application can accurately predict the changes in product position and deviations in production at the early stage of catalyst activity decay, and identify anomalies caused by adsorption imbalance or reaction path deviation in advance. For example, when sulfide is excessively adsorbed at the active site, the active site that is about to fail can be quickly located through the abnormal change of the adsorption equilibrium constant, thereby adjusting the feed distribution before the product production deviation actually occurs, avoiding local coking in the reactor or aggravation of side reactions.

[0082] The present application further proposes adjusting the feed flow rate of the active site including:

[0083] generating a flow rate adjustment priority instruction based on the deviation of the product generation amount from the predicted value and the activity decay index;

[0084] Calculate the mean concentration of the reactor components based on the component concentration gradient data;

[0085] Determine whether the component concentration gradient data is greater than the component concentration mean value, and if so, mark it as a high concentration area, otherwise mark it as a low concentration area;

[0086] When the deviation of the product generation amount from the predicted value is greater than the preset product generation amount deviation threshold:

[0087] Determine whether the active site is located in a high concentration area, and if so, reduce the feed flow rate of the active site.

[0088] Among them, the flow rate adjustment priority instruction refers to the flow rate control sequence decision instruction dynamically generated based on the deviation of product generation from the predicted value and the activity decay index. It can be achieved by setting a weight coefficient to perform weighted calculation on the degree of deviation and the degree of activity decay, which is used to determine the urgency of flow rate adjustment in different areas.

[0089] The component concentration mean refers to the arithmetic mean of the component concentration gradient data at all spatial positions in the reactor. This can be achieved by collecting data through a distributed sensor network and performing real-time calculations to establish a baseline reference value for the concentration distribution.

[0090] Among them, the high-concentration area refers to the set of spatial locations where the component concentration gradient data exceeds the calculated mean. Specifically, the area type can be automatically divided by setting the concentration threshold to identify the side reaction risk area caused by excessive aggregation of reactants.

[0091] The low-concentration area refers to the set of spatial locations where the component concentration gradient data is lower than the mean component concentration in the reactor. This can be achieved by reverse screening using the same calculation logic as the high-concentration area. This classification method helps to optimize material supply in a targeted manner.

[0092] Among them, feed flow rate adjustment refers to the dynamic adjustment of the material inlet flow rate corresponding to a specific spatial coordinate. Specifically, open-loop or closed-loop control can be achieved through an electromagnetic flow valve or a variable frequency pump to balance the distribution of reactants and the reaction rate.

[0093] Specifically, when it is detected that the product generation amount deviates from the predicted value by more than a preset threshold, the system automatically associates the abnormal active site with the concentration distribution state in the reactor; by comparing the concentration gradient value of the area where the coordinate is located with the overall mean, if it belongs to a high concentration area, the flow rate reduction operation is triggered; this operation gives priority to high-concentration areas to avoid the aggravation of by-product generation caused by local excess of reactants, and at the same time ensures that flow rate changes in key areas take effect in a timely manner by adjusting the priority; when the threshold is not triggered in the low-concentration area, the system maintains the original control strategy to reduce unnecessary operational disturbances.

[0094] Compared with existing technologies, traditional methods only adjust the homogenized flow rate of the entire area based on a fixed threshold, and are unable to distinguish the impact of concentration differences on the reaction process; this solution dynamically divides concentration areas and establishes adjustment priorities, which can implement differentiated regulation for different reaction states.

[0095] Through the above technical solution, the present application achieves a rapid response to active site anomalies and effectively suppresses the expansion of side reaction paths in high-concentration areas; through the priority control strategy, it avoids the waste of resources caused by full-area adjustments in traditional methods and improves the accuracy and timeliness of feed flow rate adjustment.

[0096] This application further proposes that the flow rate adjustment priority instruction includes:

[0097] Determine whether the deviation of the product generation amount from the predicted value is greater than a preset product generation amount deviation threshold; otherwise, obtain the spatial coordinates of the active site, and then adjust the feed flow rate of the active site:

[0098] Determining whether the active site is located in a low concentration region, and if so, increasing the feed flow rate of the active site;

[0099] When it is necessary to reduce the feed flow rate in the high-concentration area and increase the feed flow rate in the low-concentration area at the same time, the feed flow rate in the high-concentration area should be reduced first. After the feed flow rate adjustment in the high-concentration area is completed, the flow rate increase in the low-concentration area should be recalculated.

[0100] Specifically, when the system detects that there is a need for flow rate adjustment in both high-concentration areas and low-concentration areas, it first performs a flow rate reduction operation on the high-concentration area. For example, by adjusting the opening of the feed valve at the corresponding position to reduce the material input in this area, after the adjustment is completed and the concentration distribution data is updated, the flow rate increase in the low-concentration area is recalculated based on the new concentration gradient mean. This process uses a phased adjustment strategy to ensure that the overload risk in the high-concentration area is controlled first, while avoiding secondary imbalance caused by excessive supply in the low-concentration area.

[0101] Compared with existing technologies, traditional methods usually adopt parallel or random sequence adjustments when encountering adjustment needs in multiple areas, which can easily cause local overload or offset the adjustment effects. However, this solution establishes a priority adjustment mechanism to enable immediate response to the overload risk in high-concentration areas, and recalculate the needs of low-concentration areas based on the adjusted data, forming a dynamic closed-loop optimization path.

[0102] Through the above technical solution, this application effectively solves the resource conflict problem of traditional control methods in multi-region adjustment scenarios. Through a phased optimization strategy, the reaction imbalance state in the high-concentration area is preferentially eliminated, and then the low-concentration area is accurately supplemented based on real-time data, thereby improving the stability and efficiency of the overall control process.

[0103] This application further proposes that the calculated fluid flow state prediction data include:

[0104] Establishing fluid velocity vector field matrix data according to fluid velocity field data;

[0105] Convert the predicted value of product position change into mass transfer coefficient matrix data;

[0106] Fusion of vector field matrix data and transfer coefficient matrix data through convolutional neural network;

[0107] The output includes fluid flow state prediction data including vortex intensity and flow uniformity.

[0108] The fluid velocity vector field matrix data refers to the fluid velocity direction and magnitude data recorded in a two-dimensional or three-dimensional grid format. Specifically, this can be achieved by collecting velocity information through a sensor array and then mapping it into a matrix format according to spatial coordinates.

[0109] Mass transfer coefficient matrix data refers to the spatial distribution data reflecting the efficiency of material transfer, which can be calculated by combining the change in active site coverage in the predicted value of product position change with the mass transfer coefficient equation.

[0110] Convolutional neural network fusion refers to the use of convolutional layers to extract the spatial correlation characteristics of vector fields and mass transfer coefficients. Specifically, it can be achieved by using an architecture design of multi-layer convolution kernels superimposed on pooling layers.

[0111] Eddy intensity refers to the energy quantification index of fluid rotational motion, and flow uniformity refers to the standard deviation index of flow velocity distribution. The combination of the two is used to characterize the dynamic characteristics of fluid flow state.

[0112] Specifically, the fluid velocity field data is converted into vector field matrix data to capture the spatial distribution pattern of the flow velocity in the catalyst bed; the predicted value of the product position change is converted into mass transfer coefficient matrix data to reflect the local changes in the contact efficiency between the reactants and the catalyst surface; the two types of matrix data are subjected to feature extraction and fusion by convolutional neural networks, which can identify the nonlinear coupling relationship between the flow velocity distribution and the mass transfer efficiency, and thus output flow state prediction data including eddy intensity and flow uniformity. For example, the convolution layer can extract the edge features of the flow velocity vector field, perform superposition analysis with the high gradient area of ​​the mass transfer coefficient, and then predict the degree to which the local flow state deviates from the normal range.

[0113] Compared with existing technologies, traditional methods typically estimate flow states based solely on single velocity measurement data or empirical formulas, failing to effectively correlate product position changes with fluid dynamics. Existing technologies also lack the collaborative analysis of multi-source data through matrix transformation and neural network fusion, resulting in lags and errors in identifying dead zone locations. This solution achieves refined modeling and prediction of flow states by integrating fluid dynamics and mass transfer process data.

[0114] Through the above technical solution, the present application can more accurately predict abnormal areas of flow conditions in the catalyst bed, such as identifying the decrease in mass transfer efficiency caused by local circulation flow through eddy current intensity, and judging the risk of activity attenuation caused by uneven flow velocity distribution through flow uniformity; this predictive ability provides a reliable basis for timely adjustment of the stirring frequency, thereby reducing the formation of dead zones and maintaining the stability of catalyst activity.

[0115] The present application further proposes that the dead zone position in the catalyst bed be identified as follows:

[0116] calculating an average value of the fluid flow state prediction data based on the fluid flow state prediction data;

[0117] Calculate the deviation between the predicted data of fluid flow state in each region and the average value. The specific calculation formula is as follows:

[0118]

[0119] Where, represents the actual flow velocity field, represents the average velocity field, Indicates the deviation between the actual velocity field and the average velocity field, represents the deviation norm;

[0120] Mark areas with a deviation greater than 30% as candidate dead zones;

[0121] Acquire temperature distribution data of the candidate dead zone, and verify the authenticity of the candidate dead zone based on the temperature distribution data of the candidate dead zone;

[0122] According to the verification results, the real dead zone is recorded, the position of the real dead zone is obtained, and the dead zone feature data is generated based on the deviation of the real dead zone.

[0123] The deviation refers to the degree of difference between the predicted data of the fluid flow state in each region and the average value, which can be achieved by calculating the percentage deviation or the standard deviation multiple, and is used to quantify the degree of abnormality of the local flow state.

[0124] Among them, the candidate dead zone refers to the area where the deviation exceeds the preset threshold. Specifically, a dynamic threshold or a fixed threshold can be used for screening. For example, the area with a deviation exceeding 30% is preliminarily marked as a candidate area.

[0125] Among them, verifying the authenticity of the candidate dead zone means combining the temperature distribution data to determine whether flow stagnation actually exists in the candidate area. This can be achieved by detecting whether the temperature gradient of the candidate dead zone is lower than the preset range. For example, when the temperature gradient is lower than 5°C / m, the existence of the dead zone is confirmed.

[0126] The dead zone feature data refers to a quantitative index including the dead zone position coordinates and deviation, and specifically, a three-dimensional coordinate system can be used to mark the spatial position.

[0127] Specifically, after calculating the average value of the fluid flow state prediction data, the flow state data of each area will be compared with the average value one by one to calculate the corresponding deviation; when the deviation of a certain area exceeds 30%, the area will be marked as a candidate dead zone; then, by retrieving the temperature distribution data of the area, verifying whether its temperature change is consistent with the flow stagnation phenomenon, the real dead zone can be identified more accurately. For example, if the temperature gradient of the candidate dead zone is lower than the normal range, it is confirmed as a real dead zone and its location is recorded. Finally, dead zone feature data including location and deviation is generated.

[0128] In some specific embodiments, the deviation threshold of the candidate dead zone can be dynamically adjusted according to the reactor type. For example, the threshold is set to 25% in a fixed bed reactor and to 35% in a fluidized bed reactor.

[0129] Compared with existing technologies, existing methods usually only judge dead zones based on single flow rate data or temperature data, which is prone to misjudgment due to sensor errors or local interference; however, this method can more accurately identify the true dead zone by fusing flow state prediction data with temperature distribution data, avoiding false markings caused by data noise.

[0130] Through the above technical solution, the present application can solve the problem that traditional methods cannot accurately identify the location and severity of the dead zone of the catalyst bed, reduce the misjudgment rate through a multi-dimensional data verification mechanism, and provide precise spatial coordinates for subsequent stirring frequency adjustment, thereby improving the uniformity of fluid distribution in the reactor and slowing down the rate of catalyst activity decay.

[0131] The present application further proposes that issuing a stirring frequency adjustment instruction includes:

[0132] The optimal stirring frequency compensation value is calculated based on the dead zone characteristic data. The specific calculation formula is as follows:

[0133]

[0134] Where, Indicates the optimal stirring frequency compensation value, Indicates the maximum compensation amplitude, Indicates the deviation response sensitivity, Indicates the dead zone impact weight;

[0135] Converting the compensation value into stirring control signal data of the reactor; adjusting the stirring frequency according to the stirring control signal data;

[0136] Obtaining the temperature distribution data of the current catalyst bed and calculating the real-time activity decay index in combination with the adjusted stirring frequency data;

[0137] It is determined whether the real-time activity decay index is greater than a preset activity decay index threshold, and if so, the compensation value is iteratively optimized.

[0138] The optimal stirring frequency compensation value refers to the stirring frequency increment or decrement required to adjust the dead zone position. Specifically, a compensation model can be established based on the spatial coordinates and severity parameters in the dead zone characteristic data, for example, by using linear regression or genetic algorithm to solve the optimal frequency adjustment amount.

[0139] The stirring control signal data refers to the compensation value converted into a control instruction recognizable by the reactor stirring device. Specifically, a digital signal conversion module can be used to map the frequency value into a pulse width modulation signal.

[0140] Among them, iterative optimization of the compensation value refers to recalculating the compensation value when the activity decay index does not meet expectations. Specifically, a closed-loop feedback mechanism can be used to adjust the compensation model parameters according to the deviation between the flow state improvement data and the expected target.

[0141] Specifically, based on the dead zone characteristic data, the compensation model is used to calculate the stirring frequency adjustment amount at the corresponding position, and a pulse width modulation signal is generated to adjust the stirring frequency; after the adjustment, the flow state data is collected in real time through the sensor. If the activity decay index does not drop below the threshold, the compensation value is recalculated and the stirring frequency is updated to form a dynamic adjustment loop. For example, when the activity decay index does not drop below the threshold after adjustment, the compensation model will output a higher stirring frequency increment and adjust the stirring frequency through the signal conversion module; if the flow uniformity in the area after adjustment does not meet expectations, the weight coefficient in the compensation model is recalibrated to generate a new stirring frequency compensation value.

[0142] Compared with existing technologies, traditional methods only adjust the stirring frequency according to fixed rules, cannot dynamically optimize the adjustment amount based on the dead zone spatial location and severity, and lack real-time feedback and iteration mechanisms; this solution establishes a compensation model and closed-loop control to accurately match the stirring frequency adjustment amount according to different dead zone characteristics, and ensure the adjustment effect through continuous monitoring and iterative optimization.

[0143] Through the above technical solution, the present application solves the problem that the traditional method cannot dynamically adapt to the dead zone characteristics, resulting in low stirring adjustment efficiency, achieves rapid improvement of the dead zone flow state, and effectively suppresses the fluctuation of the catalyst activity attenuation index caused by dead zone retention.

[0144] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0145] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A dynamic neural network-based intelligent control method for petrochemical reaction processes, which is used for intelligently controlling the activity decay index of a catalyst bed in a hydrodesulfurization reactor. The catalyst bed is composed of N active sites at different spatial locations, and is characterized by: The following steps are involved: Real-time acquisition of catalyst bed temperature distribution data, reactor fluid velocity field data, component concentration gradient data, and stirring frequency data; Inputting the temperature distribution data and the stirring frequency data into a first branch of a dynamic neural network, and outputting an activity decay index of the catalyst bed; Determining whether the activity decay index is greater than a preset activity decay index threshold, and if so, generating a predicted value of a product position change and a predicted value of a product generation amount deviation of the active site by using the component concentration gradient data through a nonlinear combination simulation; The predicted value of the product position change is the difference between the coverage of the product of the next cycle on the active site and the coverage of the product of the current cycle on the active site; Determining whether the deviation of the product generation amount from the predicted value is greater than a preset product generation amount deviation threshold, and if so, obtaining the spatial coordinates of the active site, and then adjusting the feed flow rate of the active site; Fluid flow state prediction data is calculated based on the fluid flow velocity field data and the product position change prediction value, the dead zone position in the catalyst bed is identified based on the fluid flow state prediction data, and a stirring frequency adjustment instruction is issued to the dead zone position until the activity decay index is lower than the preset activity decay index threshold.

2. The method for intelligent control of petrochemical reaction processes based on dynamic neural networks according to claim 1, characterized in that: The real-time acquisition of temperature distribution data of the catalyst bed comprises: Acquiring product concentration data, wherein the product concentration data includes target product concentration data and by-product concentration data; Inputting target product concentration data and by-product concentration data into a second branch of a dynamic neural network, and outputting a predicted value of the activity decay rate of the catalyst bed; Determine whether the activity decay rate prediction value is greater than a preset activity decay rate threshold. If so, obtain the spatial position where the target product concentration data and the by-product concentration data are received, and adjust the hydrogen partial pressure data at the spatial position.

3. The method for intelligent control of petrochemical reaction processes based on dynamic neural networks according to claim 2, characterized in that: The hydrogen partial pressure data for adjusting the spatial position includes: A product deviation value is calculated based on the target product concentration data and the by-product concentration data, combined with the preset expected target product concentration value and the by-product concentration value, and a hydrogen partial pressure adjustment coefficient is generated based on the product deviation value, and the hydrogen partial pressure adjustment coefficient is positively correlated with the product deviation value.

4. The method for intelligent control of petrochemical reaction processes based on dynamic neural networks according to claim 2, characterized in that: The activity decay index of the catalyst bed output includes: Calculate the standard deviation of the temperature distribution data and the main frequency component of the stirring frequency data, perform weighted calculation on the by-product concentration data, multiply the standard deviation of the temperature distribution data by the main frequency component of the stirring frequency data, sum the product of the standard deviation of the temperature distribution data and the main frequency component of the stirring frequency data with the weighted value of the by-product concentration data, and output the activity decay index after normalizing the summation result.

5. The method for intelligent control of petrochemical reaction processes based on dynamic neural networks according to claim 1, characterized in that: The simulation of generating the predicted value of the product position change and the predicted value of the product production amount deviation of the active site includes: When the activity decay index is greater than the preset activity decay index threshold, the current component concentration gradient data is extracted, and the current component concentration gradient data is associated with the kinetic equation. By calculating the adsorption equilibrium constant and reaction rate constant of each component at the active site, the product position change prediction value and the product production amount deviation prediction value are generated.

6. The method for intelligent control of petrochemical reaction processes based on dynamic neural networks according to claim 1, characterized in that: The adjusting of the feed flow rate of the active site comprises: generating a flow rate adjustment priority instruction based on the deviation of the product generation amount from the predicted value and the activity decay index; Calculate the mean concentration of the reactor components based on the component concentration gradient data; Determine whether the component concentration gradient data is greater than the component concentration mean value, and if so, mark it as a high concentration area, otherwise mark it as a low concentration area; When the deviation of the product generation amount from the predicted value is greater than the preset product generation amount deviation threshold: Determine whether the active site is located in a high concentration area, and if so, reduce the feed flow rate of the active site.

7. The method for intelligent control of petrochemical reaction processes based on dynamic neural networks according to claim 6, characterized in that: The flow rate adjustment priority instruction includes: Determine whether the deviation of the product generation amount from the predicted value is greater than a preset product generation amount deviation threshold; otherwise, obtain the spatial coordinates of the active site, and then adjust the feed flow rate of the active site: Determining whether the active site is located in a low concentration region, and if so, increasing the feed flow rate of the active site; When it is necessary to reduce the feed flow rate in the high-concentration area and increase the feed flow rate in the low-concentration area at the same time, the feed flow rate in the high-concentration area should be reduced first. After the feed flow rate adjustment in the high-concentration area is completed, the flow rate increase in the low-concentration area should be recalculated.

8. The method for intelligent control of petrochemical reaction processes based on dynamic neural networks according to claim 1, characterized in that: The fluid flow state prediction data obtained by calculation includes: Establishing fluid velocity vector field matrix data according to the fluid velocity field data; Converting the predicted value of the product position change into mass transfer coefficient matrix data; fusing the vector field matrix data and the transfer coefficient matrix data through a convolutional neural network; The output includes fluid flow state prediction data including vortex intensity and flow uniformity.

9. The method for intelligent control of petrochemical reaction processes based on dynamic neural networks according to claim 1, characterized in that: The identifying of the dead zone position in the catalyst bed comprises: Calculating an average value of the fluid flow state prediction data based on the fluid flow state prediction data; Calculate the deviation between the predicted data of fluid flow state in each area and the average value; Marking the area with a deviation greater than 30% as a candidate dead zone; Acquiring temperature distribution data of the candidate dead zone, and verifying the authenticity of the candidate dead zone according to the temperature distribution data of the candidate dead zone; The real dead zone is recorded according to the verification result, the position of the real dead zone is obtained, and the dead zone feature data is generated in combination with the deviation of the real dead zone.

10. The method for intelligent control of petrochemical reaction processes based on dynamic neural networks according to claim 9, characterized in that: The issuing of the stirring frequency adjustment instruction comprises: Calculating an optimal stirring frequency compensation value according to the dead zone characteristic data; Converting the compensation value into stirring control signal data of the reactor; adjusting the stirring frequency according to the stirring control signal data; Obtaining the temperature distribution data of the current catalyst bed and calculating the real-time activity decay index in combination with the adjusted stirring frequency data; It is determined whether the real-time activity decay index is greater than a preset activity decay index threshold, and if so, the compensation value is iteratively optimized.

Images