Methods, apparatus and systems for predicting the microreaction activity of catalysts in catalytic cracking units

Abstract

This invention provides a method, apparatus, and system for predicting the microreaction activity of catalysts in catalytic cracking units, belonging to both the chemical industry and the field of artificial intelligence. The method includes: preprocessing variable data to obtain a processed sample set; constructing an initial prediction model for the microreaction activity of catalytic cracking catalysts based on the sample set; optimizing the initial prediction model using a fast descent algorithm to obtain a target prediction model for the microreaction activity of catalytic cracking catalysts; and using the target prediction model to predict the preprocessed test data to obtain predicted values ​​for catalyst microreaction activity. Addressing the problem of limited and low-frequency data related to catalyst microreaction activity in existing systems, this invention utilizes a local anomaly factor algorithm based on feature attributes to identify anomalous data and compensate for missing data, and establishes a catalyst microreaction activity prediction model based on a recursive RBF neural network, thus achieving the prediction of catalyst microreaction activity.

Description

Technical Field

[0001] This invention belongs to both the chemical industry and the field of artificial intelligence, and specifically relates to a method for predicting the micro-reaction activity of catalysts in a catalytic cracking unit, a device for predicting the micro-reaction activity of catalysts in a catalytic cracking unit, and a system for predicting the micro-reaction activity of catalysts in a catalytic cracking unit. Background Technology

[0002] Catalysts in catalytic cracking units are the most widely used catalysts in petroleum processing, and my country has a huge amount of scrapped FCC catalysts. Because there are no online instruments to monitor catalyst microreaction activity, real-time microreaction activity values ​​cannot be obtained. Frequent adjustments to the load, feedstock, and production plan of catalytic cracking units result in low frequency and delayed analysis of catalyst reactivity. The replenishment of fresh catalyst and the discharge of spent catalyst during production are quantitatively controlled based on experience, rather than dynamically adjusted according to real-time catalyst microreaction activity values. Consequently, catalyst consumption and balancer activity cannot remain relatively stable. Therefore, research is urgently needed on real-time prediction methods for catalyst microreaction activity in catalytic cracking units to improve the intelligence level of these units.

[0003] To obtain the reactivity of catalysts in catalytic cracking units, Cui Yufeng et al. analyzed the mathematical relationship between catalyst activity and micro-reaction activity, derived the hydrothermal deactivation kinetic equation of the catalyst, and established a mathematical model of the micro-reaction activity of the equilibrium agent in industrial units. This enabled the simulation and prediction of the micro-reaction activity of the equilibrium catalyst in catalytic cracking units, and the unit parameters were guided and determined through the simulation calculations. Deng Mingbo et al. established the mathematical relationship between catalyst activity and micro-reaction activity using mathematical derivation methods, determined the hydrothermal deactivation kinetic parameters through experimental data on catalyst hydrothermal deactivation, and achieved the simulation and prediction of catalyst equilibrium activity. Although the above research methods can obtain the catalyst reactivity, they all rely on mathematical mechanism model calculations to obtain the catalyst reactivity, which are relatively complex and time-consuming.

[0004] To address the challenge of predicting catalyst reactivity, Li Fangsheng et al. proposed a catalyst prediction method for denitrification units. They established a multi-layer SCR catalyst lifetime prediction model based on the SVM algorithm, predicting catalyst lifetime through changes in relative catalyst activity. Tang Shijie et al. designed a catalyst lifetime prediction method for coal-fired power plants. By comparing established curve fitting, grey prediction, BP neural network, and grey neural network prediction models, experimental results showed that the BP neural network modeling accuracy exceeded other methods. Li Debo et al. established a mathematical model for optimizing the replacement cycle and strategy of SCR catalysts in coal-fired power plants, studying catalyst renewal strategies and evaluating their benefits. Liu Xinping et al. proposed a BP network-based catalyst activity prediction method for oil refining processes. By analyzing changes in catalyst activity and related process parameters, they adjusted catalyst dosage to improve production efficiency and ensure stable operation under optimal conditions. These research methods use data-driven methods such as neural networks to predict catalyst reactivity, enabling real-time acquisition of catalyst micro-reaction activity based on raw material and product data. Compared with traditional mathematical mechanism models, data-driven calculation methods improve calculation accuracy and shorten time. However, the calculation errors of these data-driven methods are large, and there is still room for further improvement. Furthermore, the data collection frequency for catalyst microreaction activity and related variables of the reaction regeneration system varies greatly, and the data is difficult to clean, which often makes it more difficult to establish a predictive model.

[0005] Currently, research on methods for predicting the microreaction activity of catalysts in catalytic cracking units is not mature. Furthermore, the large differences in the frequency of data collection for relevant variables, as well as the problems of abnormal and missing data, further increase the difficulty of modeling. The process variables related to the microreaction activity of catalysts in catalytic cracking units cannot be fully grasped. Therefore, how to better process the dataset to obtain accurate relevant variables and how to improve the accuracy of prediction are still difficult issues that need to be addressed. Summary of the Invention

[0006] The purpose of this invention is to provide a method, apparatus, and system for predicting the microreaction activity of catalysts in catalytic cracking units, so as to at least solve the above-mentioned problem of improving prediction accuracy.

[0007] To achieve the above objectives, a first aspect of the present invention provides a method for intelligent prediction of the microreaction activity of a catalyst in a catalytic cracking unit, the method comprising:

[0008] Preprocess the variable data to obtain the processed sample set;

[0009] An initial prediction model for the microreaction activity of catalytic cracking catalysts was constructed based on the sample set.

[0010] The initial prediction model of the catalyst was optimized using a fast descent algorithm to obtain a target prediction model for the microreaction activity of the catalytic cracking catalyst.

[0011] The target prediction model is used to predict the preprocessed test data to obtain the predicted value of catalyst microreaction activity.

[0012] Optionally, the preprocessing of the variable data to obtain the processed sample set includes:

[0013] Get all variable data;

[0014] Determine the upper and lower limit ranges for each variable's data;

[0015] The variable data is segmented using a window segmentation method;

[0016] A feature-based local anomaly factor algorithm is used to identify anomalous data segments and anomalous data within those segments, and then remove the anomalous data.

[0017] Compensate for missing data in variable data after removing outliers;

[0018] The compensated variable data is then normalized to obtain normalized data.

[0019] The normalized data is dimensionality reduced to obtain the input variable data set.

[0020] A preset number of data sets are selected from the input variable data set as a sample set.

[0021] Optionally, determining the upper and lower limit ranges of each variable's data includes:

[0022] Analyze the operating status of the variable data;

[0023] Based on the described operating status, obtain the upper and lower limit range values ​​of each variable data;

[0024] The operating states include: low load operating state, medium load operating state, and full load operating state.

[0025] Optionally, the variable data can be segmented using a window segmentation method, including:

[0026] Let the dataset of the variable data be D = {X1, X2, X3, ..., X...} j}, the j-th vector is X j ={x j1 ,x j2 ,…,x jn}, where n is the total amount of data;

[0027] The variable data is divided using a clustering algorithm to determine the segmentation threshold;

[0028] The window size s is determined based on the segmentation threshold, the amount of variable data, and the running status of the variable data.

[0029] The window slides from the first data point of each vector to obtain the data segment S.

[0030] Optionally, a local anomaly factor algorithm based on feature attributes is used to identify anomalous data segments and anomalous data within those segments, and to remove anomalous data, including:

[0031] The k-th distance between data segment S and data segment O is calculated using the following formula:

[0032]

[0033] In the formula, I is the number of attributes, and f(S) i ) is the i-th dimension attribute value of data segment S, f(O) i ) is the i-th dimension attribute value of data segment O, and the attribute value includes: entropy value, mean value, maximum value and peak interval;

[0034] The expression for the entropy value is:

[0035]

[0036] Where m = {1, 2, ..., M}, M is the number of subspaces, p(u m ) is the uth m The probability distribution function of each subspace is expressed as follows:

[0037]

[0038] k(u m ) is the subspace u m The number of data items included;

[0039] Calculate the local anomaly factor of data segment S. The evaluation function for the local anomaly factor of data segment S is:

[0040]

[0041] The local reachability density of data segment S is:

[0042]

[0043] Among them, lrd k (S) represents the local reachability density of data segment S, lrd k (O) represents the local reachability density of data segment O, L k(S) is the set of all data segments in dataset D whose distance to data segment S does not exceed its k-th distance, LOF k (S) represents the local anomaly factor value of data segment S; ∑reach-distance(S,O) represents the distances of all neighboring data L within data segment S. k The sum of reachable distances of (S);

[0044] Based on the Local Outlier Factor (LOF) k Values ​​are used to classify the data segments into different levels.

[0045] The local anomaly factor algorithm is used to identify anomalous data in data segments of different levels.

[0046] All identified abnormal data are screened and removed.

[0047] Optionally, compensate for missing data in the variable data after removing outliers, including:

[0048] Filter relevant variables for missing data;

[0049] Based on principal component analysis, the top E variables with the largest and smallest contribution rates were selected from the relevant variables and used as input variables for the RBF neural network missing value compensation model.

[0050] An RBF neural network missing data compensation model is established. The hidden layer structure of the RBF neural network missing data compensation model is determined to have E+2 nodes based on the input variables and experimental construction method; the output layer is the compensation output of the missing data.

[0051] The particle swarm optimization algorithm was used to optimize the RBF neural network missing compensation model, and the optimized RBF neural network missing compensation model was obtained.

[0052] The optimized RBF neural network missing data compensation model is used to perform soft measurement of the compensation variable to obtain the compensation value of the missing data.

[0053] Optionally, the dimensionality reduction processing of the normalized data to obtain the input variable data set includes:

[0054] Partial least squares method is used to reduce the dimensionality of normalized data;

[0055] The correlation coefficients and contribution rates of variables with values ​​greater than a threshold are obtained by using principal component analysis to form the input variable data set.

[0056] Optionally, an initial prediction model for the microreaction activity of the catalytic cracking catalyst is constructed based on the sample set, including:

[0057] A basic model of the microreaction activity of catalytic cracking catalysts based on a recursive RBF neural network is constructed. This basic model includes:

[0058] Input layer: This layer contains n input variables u i (t), i = 1, 2, ..., n;

[0059] Hidden layer: The hidden layer contains J neurons, and the output expression of each hidden layer neuron is:

[0060]

[0061] Among them, c j (t) is the center vector of the j-th neuron, σ j (t) is the width of the j-th neuron, and h is the input vector of the j-th hidden layer neuron. j (t) is:

[0062] h j (t)=[u1(t),u2(t),u3(t),u4(t),v j [(t)×y(t-1)];

[0063] Where y(t-1) is the output of the recurrent RBF neural network at time t-1, v j (t) represents the feedback connection weights between the output neuron and the j-th hidden layer neuron at time t;

[0064] Output layer:

[0065]

[0066] Where w(t) is the connection weight vector between the hidden layer and the output layer, θ(t) is the output vector of the hidden layer neuron, and θ j (t) represents the output of the j-th hidden layer neuron, w j y(t) represents the connection weights between the j-th hidden layer neuron and the output neuron, and y(t) represents the output of the recursive RBF neural network at time t.

[0067] Optionally, the method further includes:

[0068] After obtaining the target prediction model for the microreaction activity of the catalytic cracking catalyst, the performance of the target prediction model is evaluated based on the root mean square error and accuracy. The RMSE expression is as follows:

[0069]

[0070] The formula for calculating its prediction accuracy is:

[0071]

[0072] Where z = 1, 2, ..., Z, Z is the number of test samples, and y d y(t) is the expected output of the catalyst microreaction activity, and y(t) is the actual output of the catalyst microreaction activity.

[0073] A second aspect of the present invention provides an intelligent prediction device for the microreaction activity of a catalyst in a catalytic cracking unit, comprising: a controller, used for:

[0074] Preprocess the variable data to obtain the processed sample set;

[0075] An initial prediction model for the microreaction activity of catalytic cracking catalysts was constructed based on the sample set.

[0076] The initial prediction model was optimized using a fast descent algorithm to obtain a target prediction model for the microreaction activity of the catalytic cracking catalyst.

[0077] The target prediction model is used to predict the preprocessed test data to obtain the predicted value of catalyst microreaction activity.

[0078] Optionally, the preprocessing of the variable data to obtain the processed sample set includes:

[0079] Get all variable data;

[0080] Determine the upper and lower limit ranges for each variable's data;

[0081] The variable data is segmented using a window segmentation method;

[0082] A feature-based local anomaly factor algorithm is used to identify anomalous data segments and anomalous data within those segments, and then remove the anomalous data.

[0083] Compensate for missing data in variable data after removing outliers;

[0084] The compensated variable data is then normalized to obtain normalized data.

[0085] The normalized data is dimensionality reduced to obtain the input variable data set.

[0086] A preset number of data sets are selected from the input variable data set as a sample set.

[0087] Optionally, determining the upper and lower limit ranges of each variable's data includes:

[0088] Analyze the operating status of the variable data;

[0089] Based on the described operating status, obtain the upper and lower limit range values ​​of each variable data;

[0090] The operating states include: low load operating state, medium load operating state, and full load operating state.

[0091] Optionally, the variable data can be segmented using a window segmentation method, including:

[0092] Let the dataset of the variable data be D = {X1, X2, X3, ..., X...} j}, the j-th vector is X j ={x j1 ,x j2 ,…,x jn}, where n is the total amount of data;

[0093] The variable data is divided using a clustering algorithm to determine the segmentation threshold;

[0094] The window size s is determined based on the segmentation threshold, the amount of variable data, and the running status of the variable data.

[0095] The window slides from the first data point of each vector to obtain the data segment S.

[0096] Optionally, a local anomaly factor algorithm based on feature attributes is used to identify anomalous data segments and anomalous data within those segments, and to remove anomalous data, including:

[0097] The k-th distance between data segment S and data segment O is calculated using the following formula:

[0098]

[0099] In the formula, I is the number of attributes, and f(S) i ) is the i-th dimension attribute value of data segment S, f(O) i ) is the i-th dimension attribute value of data segment O, and the attribute value includes: entropy value, mean value, maximum value and peak interval;

[0100] The expression for the entropy value is:

[0101]

[0102] Where m = {1, 2, ..., M}, M is the number of subspaces, p(u m ) is the uth m The probability distribution function of each subspace is expressed as follows:

[0103]

[0104] k(u m ) is the subspace u m The number of data items included;

[0105] Calculate the local anomaly factor of data segment S. The evaluation function for the local anomaly factor of data segment S is:

[0106]

[0107] The local reachability density of data segment S is:

[0108]

[0109] Among them, lrd k (S) represents the local reachability density of data segment S, lrd k (O) represents the local reachability density of data segment O, L k (S) is the set of all data segments in dataset D whose distance to data segment S does not exceed its k-th distance, LOF k (S) represents the local anomaly factor value of data segment S; ∑reach-distance(S,O) represents the distances of all neighboring data L within data segment S. k The sum of reachable distances of (S);

[0110] Based on the Local Outlier Factor (LOF) k Values ​​are used to classify the data segments into different levels.

[0111] The local anomaly factor algorithm is used to identify anomalous data in data segments of different levels.

[0112] All identified abnormal data are screened and removed.

[0113] Optionally, compensate for missing data in the variable data after removing outliers, including:

[0114] Filter relevant variables for missing data;

[0115] Based on principal component analysis, the top E variables with the largest and smallest contribution rates were selected from the relevant variables and used as input variables for the RBF neural network missing value compensation model.

[0116] An RBF neural network missing data compensation model is established. The hidden layer structure of the RBF neural network missing data compensation model is determined to have E+2 nodes based on the input variables and experimental construction method; the output layer is the compensation output of the missing data.

[0117] The particle swarm optimization algorithm was used to optimize the RBF neural network missing compensation model, and the optimized RBF neural network missing compensation model was obtained.

[0118] The optimized RBF neural network missing data compensation model is used to perform soft measurement of the compensation variable to obtain the compensation value of the missing data.

[0119] Optionally, the dimensionality reduction processing of the normalized data to obtain the input variable data set includes:

[0120] Partial least squares method is used to reduce the dimensionality of normalized data;

[0121] The correlation coefficients and contribution rates of variables with values ​​greater than a threshold are obtained by using principal component analysis to form the input variable data set.

[0122] Optionally, an initial prediction model for the microreaction activity of the catalytic cracking catalyst is constructed based on the sample set, including:

[0123] A basic model of the microreaction activity of catalytic cracking catalysts based on a recursive RBF neural network is constructed. This basic model includes:

[0124] Input layer: This layer contains n input variables u i (t), i = 1, 2, ..., n;

[0125] Hidden layer: The hidden layer contains J neurons, and the output expression of each hidden layer neuron is:

[0126]

[0127] Among them, c j (t) is the center vector of the j-th neuron, σ j (t) is the width of the j-th neuron, and h is the input vector of the j-th hidden layer neuron. j (t) is:

[0128] h j (t)=[u1(t),u2(t),u3(t),u4(t),v j [(t)×y(t-1)];

[0129] Where y(t-1) is the output of the recurrent RBF neural network at time t-1, v j (t) represents the feedback connection weights between the output neuron and the j-th hidden layer neuron at time t;

[0130] Output layer:

[0131]

[0132] Where w(t) is the connection weight vector between the hidden layer and the output layer, θ(t) is the output vector of the hidden layer neuron, and θ j (t) represents the output of the j-th hidden layer neuron, w j y(t) represents the connection weights between the j-th hidden layer neuron and the output neuron, and y(t) represents the output of the recursive RBF neural network at time t.

[0133] Optionally, the controller is also used for:

[0134] After obtaining the target prediction model for the microreaction activity of the catalytic cracking catalyst, the performance of the target prediction model is evaluated based on the root mean square error and accuracy. The RMSE expression is as follows:

[0135]

[0136] The formula for calculating its prediction accuracy is:

[0137]

[0138] Where z = 1, 2, ..., Z, Z is the number of test samples, and y d y(t) is the expected output of the catalyst microreaction activity, and y(t) is the actual output of the catalyst microreaction activity.

[0139] A third aspect of the present invention provides an intelligent prediction system for the microreaction activity of a catalyst in a catalytic cracking unit, the system comprising:

[0140] The data processing module is used to preprocess the variable data to obtain the processed sample set;

[0141] An initial prediction model construction module is used to construct an initial prediction model of the microreaction activity of the catalytic cracking catalyst based on the sample set.

[0142] The initial prediction model optimization module is used to optimize the initial prediction model using a fast descent algorithm to obtain a target prediction model for the microreaction activity of the catalytic cracking catalyst.

[0143] The data prediction module is used to predict the preprocessed test data using the target prediction model to obtain the predicted value of catalyst microreaction activity.

[0144] On the other hand, the present invention provides a machine-readable storage medium storing instructions for causing a machine to execute the above-described intelligent prediction method for the micro-reaction activity of catalysts in a catalytic cracking unit.

[0145] The above technical solution provides an intelligent prediction method for the micro-reaction activity of catalysts in catalytic cracking units, which solves the shortcomings of the inability to predict the micro-reaction activity of catalysts in real time, the large differences in the frequency of data collection of catalyst micro-reaction activity and related variables of reaction regeneration system, and the difficulty in cleaning the data.

[0146] This invention utilizes a window segmentation method to divide data into segments, identifies abnormal data based on a local anomaly factor algorithm using feature attributes, establishes the correlation between complete data segments and missing data segments with different dimensional variables to compensate for missing data, and establishes a catalyst microreaction activity prediction model based on a recursive RBF neural network to achieve the prediction of catalyst microreaction activity.

[0147] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0148] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings:

[0149] Figure 1 This is a flowchart of an intelligent prediction method for catalyst microreaction activity in a catalytic cracking unit provided by one embodiment of the present invention;

[0150] Figure 2 This is a block diagram of an intelligent prediction system for catalyst microreaction activity in a catalytic cracking unit provided by one embodiment of the present invention. Detailed Implementation

[0151] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0152] Figure 1 This is a flowchart of an intelligent prediction method for the microreaction activity of catalysts in a catalytic cracking unit, provided by one embodiment of the present invention. Figure 1 As shown, the method includes:

[0153] Step 1: Preprocess the variable data to obtain the processed sample set, specifically including:

[0154] 1) Obtain all variable data. In this application, variable data is obtained through online instruments or laboratory analysis, including reactor data, regenerator data, desulfurization and denitrification facility data, product distribution data, and feedstock data of the treatment facilities during the operation of the catalytic cracking unit. Specifically, this includes: regeneration temperature, regeneration pressure, catalyst stockpile, gasoline yield, liquid hydrocarbon yield, feedstock sulfur, nitrogen, and metal content, fresh feed rate, reaction pressure, reaction temperature, feedstock preheating temperature, diesel yield, feedstock nitrogen content, regenerator oxygen content, regenerator dense phase stockpile, regeneration main air volume, riser slurry feed rate, riser top temperature, outlet flue gas temperature, total feed rate, regenerator bottom dense phase temperature, and regenerator dilute phase section pressure.

[0155] 2) Determine the upper and lower limit ranges of each variable data and analyze the operating status to which the variable data belongs; obtain the upper and lower limit ranges of each variable data according to the operating status; the operating status includes: low load operating status (load below 80%), medium load operating status (load 80% to 95%) and full load operating status (load above 95%).

[0156] 3) The variable data is segmented using the window segmentation method. In this embodiment, the segmentation process includes:

[0157] Let the dataset of the variable data be D = {X1, X2, X3, ..., X...} j}, the j-th vector is X j ={x j1 ,x j2 ,…,x jn}, where n is the total amount of data;

[0158] The variable data is divided using a clustering algorithm to determine the segmentation threshold;

[0159] The window size s is determined based on the segmentation threshold, the amount of variable data, and the running status of the variable data.

[0160] The window slides from the first data point of each vector to obtain the data segment S.

[0161] 4) The Local Outlier Factor (LOF) algorithm based on feature attributes is used to identify anomalous data segments and anomalous data within those segments, and then the anomalous data is removed. In this embodiment, removing anomalous data includes the following steps:

[0162] The k-th distance between data segment S and data segment O is calculated using the following formula:

[0163]

[0164] In the formula, I is the number of attributes, and f(S) i )yes The i-th dimension attribute value of data segment S, f(O) i ) is the i-th dimension attribute value of data segment O. The attribute value includes: entropy value, mean, maximum value and peak interval. Since the entropy value can more accurately reflect the distribution state of the data, it is used as one of the key indicators of data attributes.

[0165] The expression for the entropy value is:

[0166]

[0167] Where m = {1, 2, ..., M}, M is the number of subspaces, p(u m ) is the uthm The probability distribution function of each subspace is expressed as follows:

[0168]

[0169] k(u m ) is the subspace u m The number of data items included;

[0170] Calculate the local anomaly factor of data segment S. The evaluation function for the local anomaly factor of data segment S is:

[0171]

[0172] The local reachability density of data segment S is:

[0173]

[0174] Among them, lrd k (S) represents the local reachability density of data segment S, lrd k (O) represents the local reachability density of data segment O, L k (S) is the set of all data segments in dataset D whose distance to data segment S does not exceed its k-th distance, LOF k (S) represents the local anomaly factor value of data segment S; ∑reach-distance(S,O) represents the distances of all neighboring data L within data segment S. k The sum of reachable distances of (S); Local outlier (LOF) of the comparison data. k Value, when LOF k The larger the value, the greater the likelihood that the data is abnormal.

[0175] Based on the Local Outlier Factor (LOF) k Values ​​are used to classify the data segments into different levels.

[0176] The local anomaly factor algorithm is used to identify anomalous data in data segments of different levels.

[0177] All identified abnormal data are screened and removed.

[0178] 5) Compensate for missing data in the variable data after removing outliers. In this embodiment, compensating for missing data includes the following steps:

[0179] In this embodiment, the relevant variables for missing data are obtained by analyzing the information of the missing data.

[0180] Based on principal component analysis, the top E variables with the largest and smallest contribution rates were selected from the relevant variables and used as input variables for the radial basis function (RBF) neural network missing value compensation model.

[0181] An RBF neural network missing data compensation model is established. The hidden layer structure of the RBF neural network missing data compensation model is determined to have E+2 nodes based on the input variables and experimental construction method; the output layer is the compensation output of the missing data.

[0182] The particle swarm optimization algorithm was used to optimize the RBF neural network missing compensation model, and the optimized RBF neural network missing compensation model was obtained.

[0183] The optimized RBF neural network missing data compensation model is used for soft measurement of the compensation variable to obtain the compensation value for the missing data. This compensation process can solve the problem of missing data in data segments caused by low data collection frequency and data removal.

[0184] 6) Normalize the compensated variable data to obtain normalized data. In this embodiment, partial least squares (PLS) is used to reduce the dimensionality of the normalized data, eliminating the impact of differences in dimensions and orders of magnitude on the model training process; principal component analysis is used to analyze the correlation coefficients and contribution rates of variables that are greater than a threshold to form the input variable data set.

[0185] 7) Select a preset number of data sets from the input variable data set as the sample set. This process eliminates the problem of large differences in threshold ranges between different intervals in the variable data set, and also eliminates data gaps or significant differences in data dimensionality between different variables during data collection.

[0186] Step 2: Construct an initial prediction model for the microreaction activity of the catalytic cracking catalyst based on the sample set, including:

[0187] A basic model of the microreaction activity of catalytic cracking catalysts based on a recursive RBF neural network is constructed. This basic model includes:

[0188] Input layer: This layer contains n input variables u i (t), i = 1, 2, ..., n;

[0189] Hidden layer: The hidden layer contains J neurons, and the output expression of each hidden layer neuron is:

[0190]

[0191] Among them, c j (t) is the center vector of the j-th neuron, σ j(t) is the width of the j-th neuron, and h is the input vector of the j-th hidden layer neuron. j (t) is:

[0192] h j (t)=[u1(t),u2(t),u3(t),u4(t),v j [(t)×y(t-1)];

[0193] Where y(t-1) is the output of the recurrent RBF neural network at time t-1, v j (t) represents the feedback connection weights between the output neuron and the j-th hidden layer neuron at time t;

[0194] Output layer:

[0195]

[0196] Where w(t) is the connection weight vector between the hidden layer and the output layer, θ(t) is the output vector of the hidden layer neuron, and θ j (t) represents the output of the j-th hidden layer neuron, w j y(t) represents the connection weights between the j-th hidden layer neuron and the output neuron, and y(t) represents the output of the recursive RBF neural network at time t.

[0197] Step 3: Optimize the initial prediction model using the fast descent algorithm to obtain the target prediction model for the microreaction activity of the catalytic cracking catalyst;

[0198] Step 4: Use the target prediction model to predict the preprocessed test data to obtain the predicted value of catalyst microreaction activity.

[0199] After obtaining the target prediction model for the microreaction activity of the catalytic cracking catalyst, the performance of the target prediction model is evaluated based on the root mean square error and accuracy. The RMSE expression is as follows:

[0200]

[0201] The formula for calculating its prediction accuracy is:

[0202]

[0203] Where z = 1, 2, ..., Z, Z is the number of test samples, and y d y(t) is the expected output of the catalyst microreaction activity, and y(t) is the actual output of the catalyst microreaction activity.

[0204] Example 1

[0205] (1) Data acquisition, organization and cleaning

[0206] In this embodiment, data such as reactor data, regenerator data, desulfurization and denitrification facility data, product distribution data, and raw material data of the treatment facilities during the operation of the catalytic cracking unit are obtained through online instruments or laboratory analysis. These data include the following 50 parameters, each with 500 entries, mainly including: catalyst activity value, regeneration temperature, regeneration pressure, catalyst stockpile, gasoline yield, liquid hydrocarbon yield, feed sulfur, nitrogen, and metal content, fresh feed rate, reaction pressure, reaction temperature, feed preheating temperature, diesel yield, feed nitrogen content, regenerator oxygen content, regenerator dense phase stockpile, regeneration main air volume, riser slurry feed rate, riser upper temperature, outlet flue gas temperature, total feed rate, regenerator bottom dense phase temperature, and regenerator dilute phase section pressure.

[0207] Then, the statistical results of the data collected during the data collection period are analyzed. In this embodiment, the variable data collection period is at full capacity to confirm the upper and lower limit range values ​​of each variable data.

[0208] The variable data is segmented using a window segmentation method, and the dataset of variables for the catalytic cracking reaction system is set as D = {X1, X2, X3, ..., X...} 50}, the j-th vector is X j ={x j1 ,x j2 ,…,x j500}, with the window size set to s, the window slides from the first data of each vector to obtain the data segment S.

[0209] The Local Outlier Factor (LOF) algorithm based on feature attributes is used to identify outlier data segments and outlier data within those segments, and then remove the outlier data.

[0210] To address the issue of missing data in data segments due to infrequent data collection and data removal, an RBF neural network was used to compensate for continuously missing data. First, relevant variables for the missing data were selected as input. Then, principal component analysis was used to select the top 7 variables with the largest contribution rates. The output layer provided the compensation for the missing data. After these data cleaning methods, a dataset of 480 records was finally obtained.

[0211] Then the data is normalized.

[0212] To address the impact of differences in dimensions and orders of magnitude between different data segments in the abnormal dataset on the model training process, the threshold range was eliminated. Partial Least Squares (PLS) was used to reduce the dimensionality of the normalized data. Principal Component Analysis (PCA) was then used to analyze the collected data, yielding correlation coefficients and contribution rates for seven input variables: regenerator oxygen content, riser slurry feed rate, riser top temperature, feed nitrogen content, total feed rate, catalyst reserves, and regeneration main air volume. The output variable to be collected is the catalyst microreaction activity value. 480 data sets were selected and divided into two parts: 250 sets as training samples and 230 sets as test samples.

[0213] (2) Construct an initial prediction model for the microreaction activity of catalytic cracking catalysts based on the sample set.

[0214] A microreaction activity prediction model for catalytic cracking catalysts was designed using a recursive RBF neural network.

[0215] Input layer: This layer contains 7 input variables u i (t), i = 1, 2, ..., 7;

[0216] Hidden layer: The hidden layer contains 4 neurons, and the output expression of each hidden layer neuron is as follows:

[0217]

[0218] Among them, c j (t) is the center vector of the j-th neuron, σ j (t) is the width of the j-th neuron, and h is the input vector of the j-th hidden layer neuron. j (t) is

[0219] h j (t)=[u1(t),u2(t),u3(t),u4(t),v j (t)×y(t-1)]

[0220] y(t-1) is the output of the recurrent RBF neural network at time t-1, v j (t) represents the feedback connection weights between the output neuron and the j-th hidden layer neuron at time t;

[0221] Output layer:

[0222]

[0223] Where w(t) is the connection weight vector between the hidden layer and the output layer, θ(t) is the output vector of the hidden layer neuron, and θ j (t) represents the output of the j-th hidden layer neuron, w jy(t) represents the connection weights between the j-th hidden layer neuron and the output neuron, and y(t) represents the output of the recursive RBF neural network at time t.

[0224] (3) The initial prediction model is optimized using a fast descent algorithm to obtain a target prediction model for the microreaction activity of the catalytic cracking catalyst. The performance of the target prediction model is evaluated based on the root mean square error (RMSE) and accuracy. The RMSE expression is:

[0225]

[0226] The formula for calculating its prediction accuracy is:

[0227]

[0228] Where z = 1, 2, ..., Z, Z is the number of test samples, and y d y(t) is the expected output of the catalyst microreaction activity, and y(t) is the actual output of the catalyst microreaction activity.

[0229] In this embodiment, the root mean square error and accuracy evaluation of the prediction model are shown in Table 1. It can be seen from Table 1 that the error and accuracy of the catalyst micro-activity are within a reasonable range.

[0230] Table 1 Prediction results of different algorithms

[0231]

[0232] Example 2

[0233] (1) Data acquisition, organization and cleaning

[0234] In this embodiment, data such as reactor data, regenerator data, desulfurization and denitrification facility data, product distribution data, and raw material data of the treatment facilities during the operation of the catalytic cracking unit are obtained through online instruments or laboratory analysis. These data include the following 40 parameters, each with 400 entries, mainly including: catalyst activity value, regeneration temperature, regeneration pressure, catalyst stockpile, gasoline yield, liquid hydrocarbon yield, feed sulfur, nitrogen, and metal content, fresh feed rate, reaction pressure, reaction temperature, feed preheating temperature, diesel yield, feed nitrogen content, regenerator oxygen content, regenerator dense phase stockpile, regeneration main air volume, riser slurry feed rate, and riser upper temperature.

[0235] Then, the statistical results of the data collected during the data collection period are analyzed. In this embodiment, the variable data collection period is at full capacity to confirm the upper and lower limit range values ​​of each variable data.

[0236] The variable data is segmented using a window segmentation method, and the dataset of variables for the catalytic cracking reaction system is set as D = {X1, X2, X3, ..., X...}40}, the j-th vector is X j ={x j1 ,x j2 ,…,x j400}, with the window size set to s, the window slides from the first data of each vector to obtain the data segment S.

[0237] The Local Outlier Factor (LOF) algorithm based on feature attributes is used to identify outlier data segments and outlier data within those segments, and then remove the outlier data.

[0238] To address the issue of missing data in data segments due to infrequent data collection and data removal, an RBF neural network was used to compensate for continuously missing data. First, relevant variables for the missing data were selected as input. Then, principal component analysis was used to select the top 7 variables with the largest contribution rates. The output layer provided the compensation for the missing data. After these data cleaning methods, a dataset of 380 records was finally obtained.

[0239] Then the data is normalized.

[0240] To address the impact of differences in dimensions and orders of magnitude between different data segments in the abnormal dataset on the model training process, the threshold range was eliminated. Partial Least Squares (PLS) was used to reduce the dimensionality of the normalized data. Principal Component Analysis (PCA) was then used to analyze the collected data, yielding correlation coefficients and contribution rates for six input variables: regenerator oxygen content, riser slurry feed rate, riser top temperature, feed nitrogen content, total feed rate, catalyst reserves, and regeneration main air volume. The output variable to be collected is the catalyst microreaction activity value. 380 data sets were selected and divided into two parts: 200 sets as training samples and 180 sets as test samples.

[0241] (2) Construct an initial prediction model for the microreaction activity of catalytic cracking catalysts based on the sample set.

[0242] A microreaction activity prediction model for catalytic cracking catalysts was designed using a recursive RBF neural network.

[0243] Input layer: This layer contains 6 input variables u i (t), i = 1, 2, ..., 6;

[0244] Hidden layer: The hidden layer contains 3 neurons, and the output expression of each hidden layer neuron is as follows:

[0245]

[0246] Among them, c j (t) is the center vector of the j-th neuron, σ j(t) is the width of the j-th neuron, and h is the input vector of the j-th hidden layer neuron. j (t) is

[0247] h j (t)=[u1(t),u2(t),u3(t),u4(t),v j (t)×y(t-1)]

[0248] y(t-1) is the output of the recurrent RBF neural network at time t-1, v j (t) represents the feedback connection weights between the output neuron and the j-th hidden layer neuron at time t;

[0249] Output layer:

[0250]

[0251] Where w(t) is the connection weight vector between the hidden layer and the output layer, θ(t) is the output vector of the hidden layer neuron, and θ j (t) represents the output of the j-th hidden layer neuron, w j y(t) represents the connection weights between the j-th hidden layer neuron and the output neuron, and y(t) represents the output of the recursive RBF neural network at time t.

[0252] (3) The initial prediction model is optimized using a fast descent algorithm to obtain a target prediction model for the microreaction activity of the catalytic cracking catalyst. The performance of the target prediction model is evaluated based on the root mean square error (RMSE) and accuracy. The RMSE expression is:

[0253]

[0254] The formula for calculating its prediction accuracy is:

[0255]

[0256] Where z = 1, 2, ..., Z, Z is the number of test samples, and y d y(t) is the expected output of the catalyst microreaction activity, and y(t) is the actual output of the catalyst microreaction activity.

[0257] In this embodiment, the root mean square error and accuracy evaluation of the prediction model are shown in Table 2. It can be seen from Table 2 that the error and accuracy of the catalyst micro-activity are within a reasonable range.

[0258] Table 2 Prediction results of different algorithms

[0259]

[0260] Example 3

[0261] (1) Data acquisition, organization and cleaning

[0262] In this embodiment, data such as reactor data, regenerator data, desulfurization and denitrification facility data, product distribution data, and raw material data of the treatment facilities during the operation of the catalytic cracking unit are obtained through online instruments or laboratory analysis. These data include the following 30 parameters, each with 300 entries, mainly including: catalyst activity value, regeneration temperature, regeneration pressure, catalyst stockpile, gasoline yield, liquid hydrocarbon yield, feedstock sulfur, nitrogen, and metal content, fresh feed rate, reaction pressure, reaction temperature, feedstock preheating temperature, diesel yield, feedstock nitrogen content, regenerator oxygen content, regenerator dense phase stockpile, regeneration main air volume, riser slurry feed rate, and riser upper temperature.

[0263] Then, the statistical results of the data collected during the data collection period are analyzed. In this embodiment, the variable data collection period is at full capacity to confirm the upper and lower limit range values ​​of each variable data.

[0264] The variable data is segmented using a window segmentation method, and the dataset of variables for the catalytic cracking reaction system is set as D = {X1, X2, X3, ..., X...} 30}, the j-th vector is X j ={x j1 ,x j2 ,…,x j300}, with the window size set to s, the window slides from the first data of each vector to obtain the data segment S.

[0265] The Local Outlier Factor (LOF) algorithm based on feature attributes is used to identify outlier data segments and outlier data within those segments, and then remove the outlier data.

[0266] To address the issue of missing data in data segments due to infrequent data collection and data removal, an RBF neural network was used to compensate for continuously missing data. First, relevant variables for the missing data were selected as input. Then, principal component analysis was used to select the top 5 variables with the largest contribution rates. The output layer provided the compensation for the missing data. After these data cleaning methods, a dataset of 290 records was finally obtained.

[0267] Then the data is normalized.

[0268] To address the impact of differences in dimensions and orders of magnitude between different data segments in the abnormal dataset on the model training process, the threshold range was eliminated. Partial Least Squares (PLS) was used to reduce the dimensionality of the normalized data. Principal Component Analysis (PCA) was then used to analyze the collected data, yielding correlation coefficients and contribution rates for six input variables: regenerator oxygen content, riser slurry feed rate, riser top temperature, feed nitrogen content, total feed rate, catalyst reserves, and regeneration main air volume. The output variable to be collected is the catalyst microreaction activity value. 290 data sets were selected and divided into two parts: 150 sets as training samples and 140 sets as test samples.

[0269] (2) Construct an initial prediction model for the microreaction activity of catalytic cracking catalysts based on the sample set.

[0270] A microreaction activity prediction model for catalytic cracking catalysts was designed using a recursive RBF neural network.

[0271] Input layer: This layer contains 5 input variables u i (t), i = 1, 2, ..., 5;

[0272] Hidden layer: The hidden layer contains 3 neurons, and the output expression of each hidden layer neuron is as follows:

[0273]

[0274] Among them, c j (t) is the center vector of the j-th neuron, σ j (t) is the width of the j-th neuron, and h is the input vector of the j-th hidden layer neuron. j (t) is

[0275] h j (t)=[u1(t),u2(t),u3(t),u4(t),v j (t)×y(t-1)]

[0276] y(t-1) is the output of the recurrent RBF neural network at time t-1, v j (t) represents the feedback connection weights between the output neuron and the j-th hidden layer neuron at time t;

[0277] Output layer:

[0278]

[0279] Where w(t) is the connection weight vector between the hidden layer and the output layer, θ(t) is the output vector of the hidden layer neuron, and θ j (t) represents the output of the j-th hidden layer neuron, w jy(t) represents the connection weights between the j-th hidden layer neuron and the output neuron, and y(t) represents the output of the recursive RBF neural network at time t.

[0280] (3) The initial prediction model is optimized using a fast descent algorithm to obtain a target prediction model for the microreaction activity of the catalytic cracking catalyst. The performance of the target prediction model is evaluated based on the root mean square error (RMSE) and accuracy. The RMSE expression is:

[0281]

[0282] The formula for calculating its prediction accuracy is:

[0283]

[0284] Where z = 1, 2, ..., Z, Z is the number of test samples, and y d y(t) is the expected output of the catalyst microreaction activity, and y(t) is the actual output of the catalyst microreaction activity.

[0285] In this embodiment, the root mean square error and accuracy evaluation of the prediction model are shown in Table 3. It can be seen from Table 3 that the error and accuracy of the catalyst micro-activity are within a reasonable range.

[0286] Table 3 Prediction results of different algorithms

[0287]

[0288] A second aspect of the present invention provides an intelligent prediction device for the microreaction activity of a catalyst in a catalytic cracking unit, comprising: a controller, used for:

[0289] Preprocess the variable data to obtain the processed sample set;

[0290] An initial prediction model for the microreaction activity of catalytic cracking catalysts was constructed based on the sample set.

[0291] The initial prediction model was optimized using a fast descent algorithm to obtain a target prediction model for the microreaction activity of the catalytic cracking catalyst.

[0292] The target prediction model is used to predict the preprocessed test data to obtain the predicted value of catalyst microreaction activity.

[0293] In one embodiment, the preprocessing of the variable data to obtain the processed sample set includes:

[0294] All variable data are obtained; in this application, variable data are obtained through online instruments or laboratory analysis of reactor data, regenerator data, desulfurization and denitrification facility data, product distribution data, and raw material data of treatment facilities during the operation of the catalytic cracking unit, specifically including: regeneration temperature, regeneration pressure, catalyst stockpile, gasoline yield, liquid hydrocarbon yield, feed sulfur, nitrogen, and metal content, fresh feed rate, reaction pressure, reaction temperature, feed preheating temperature, diesel yield, feed nitrogen content, regenerator oxygen content, regenerator dense phase stockpile, regeneration main air volume, riser slurry feed rate, riser upper temperature, outlet flue gas temperature, total feed rate, regenerator bottom dense phase temperature, and regenerator dilute phase section pressure, etc.

[0295] Determine the upper and lower limit ranges of each variable data, and analyze the operating state to which the variable data belongs; obtain the upper and lower limit ranges of each variable data based on the operating state; the operating state includes: low load operating state (load below 80%), medium load operating state (load 80% to 95%), and full load operating state (load above 95%).

[0296] The variable data is segmented using a window segmentation method;

[0297] A feature-based local anomaly factor algorithm is used to identify anomalous data segments and anomalous data within those segments, and then remove the anomalous data.

[0298] Compensate for missing data in variable data after removing outliers;

[0299] The compensated variable data is then normalized to obtain normalized data.

[0300] The normalized data is dimensionality reduced to obtain the input variable data set.

[0301] A predetermined number of data sets are selected from the input variable data set as the sample set. This process eliminates the problem of large differences in threshold ranges between different intervals in the variable data set, and also eliminates data gaps or significant differences in data dimensionality between different variables that may have occurred during data collection.

[0302] In one embodiment, determining the upper and lower limit range values ​​of each variable data includes:

[0303] Analyze the operating status of the variable data;

[0304] Based on the described operating status, obtain the upper and lower limit range values ​​of each variable data;

[0305] The operating states include: low load operating state, medium load operating state, and full load operating state.

[0306] In one embodiment, the variable data is segmented using a window segmentation method, including:

[0307] Let the dataset of the variable data be D = {X1, X2, X3, ..., X...} j}, the j-th vector is X j ={x j1 ,x j2 ,…,x jn}, where n is the total amount of data;

[0308] The variable data is divided using a clustering algorithm to determine the segmentation threshold;

[0309] The window size s is determined based on the segmentation threshold, the amount of variable data, and the running status of the variable data.

[0310] The window slides from the first data point of each vector to obtain the data segment S.

[0311] In one embodiment, the Local Outlier Factor (LOF) algorithm based on feature attributes is used to identify and remove anomalous data segments, including:

[0312] The k-th distance between data segment S and data segment O is calculated using the following formula:

[0313]

[0314] In the formula, I is the number of attributes, and f(S) i ) is the i-th dimension attribute value of data segment S, f(O) i ) is the i-th dimension attribute value of data segment O. The attribute value includes: entropy value, mean, maximum value and peak interval. Since the entropy value can more accurately reflect the distribution state of the data, it is used as one of the key indicators of data attributes.

[0315] The expression for the entropy value is:

[0316]

[0317] Where m = {1, 2, ..., M}, M is the number of subspaces, p(u m ) is the uth m The probability distribution function of each subspace is expressed as follows:

[0318]

[0319] k(u m ) is the subspace u m The number of data items included;

[0320] Calculate the local anomaly factor of data segment S. The evaluation function for the local anomaly factor of data segment S is:

[0321]

[0322] The local reachability density of data segment S is:

[0323]

[0324] Among them, lrd k (S) represents the local reachability density of data segment S, lrd k (O) represents the local reachability density of data segment O, L k (S) is the set of all data segments in dataset D whose distance to data segment S does not exceed its k-th distance, LOF k (S) represents the local anomaly factor value of data segment S; ∑reach-distance(S,O) represents the distances of all neighboring data L within data segment S. k The sum of reachable distances of (S); Local outlier (LOF) of the comparison data. k Value, when LOF k The larger the value, the greater the likelihood that the data is abnormal.

[0325] Based on the Local Outlier Factor (LOF) k Values ​​are used to classify the data segments into different levels.

[0326] The local anomaly factor algorithm is used to identify anomalous data in data segments of different levels.

[0327] All identified abnormal data are screened and removed.

[0328] In one embodiment, compensating for missing data in variable data after removing outlier data includes:

[0329] In this embodiment, the relevant variables for missing data are obtained by analyzing the information of the missing data.

[0330] Based on principal component analysis, the top E variables with the largest and smallest contribution rates were selected from the relevant variables and used as input variables for the RBF neural network missing value compensation model.

[0331] An RBF neural network missing data compensation model is established. The hidden layer structure of the RBF neural network missing data compensation model is determined to have E+2 nodes based on the input variables and experimental construction method; the output layer is the compensation output of the missing data.

[0332] The particle swarm optimization algorithm was used to optimize the RBF neural network missing compensation model, and the optimized RBF neural network missing compensation model was obtained.

[0333] The optimized RBF neural network missing data compensation model is used for soft measurement of the compensation variable to obtain the compensation value for the missing data. This compensation process can solve the problem of missing data in data segments caused by low data collection frequency and data removal.

[0334] In one embodiment, the dimensionality reduction processing of the normalized data to obtain the input variable data set includes:

[0335] Partial least squares (PLS) is used to reduce the dimensionality of the normalized data;

[0336] The correlation coefficients and contribution rates of variables with values ​​greater than a threshold are obtained by using principal component analysis to form the input variable data set.

[0337] In one embodiment, constructing an initial prediction model for the microreaction activity of a catalytic cracking catalyst based on the sample set includes:

[0338] A basic model of the microreaction activity of catalytic cracking catalysts based on a recursive RBF neural network is constructed. This basic model includes:

[0339] Input layer: This layer contains n input variables u i (t), i = 1, 2, ..., n;

[0340] Hidden layer: The hidden layer contains J neurons, and the output expression of each hidden layer neuron is:

[0341]

[0342] Among them, c j (t) is the center vector of the j-th neuron, σ j (t) is the width of the j-th neuron, and h is the input vector of the j-th hidden layer neuron. j (t) is:

[0343] h j (t)=[u1(t),u2(t),u3(t),u4(t),v j [(t)×y(t-1)];

[0344] Where y(t-1) is the output of the recurrent RBF neural network at time t-1, v j (t) represents the feedback connection weights between the output neuron and the j-th hidden layer neuron at time t;

[0345] Output layer:

[0346]

[0347] Where w(t) is the connection weight vector between the hidden layer and the output layer, θ(t) is the output vector of the hidden layer neuron, and θ j (t) represents the output of the j-th hidden layer neuron, w j y(t) represents the connection weights between the j-th hidden layer neuron and the output neuron, and y(t) represents the output of the recursive RBF neural network at time t.

[0348] Optionally, the controller is also used for:

[0349] After obtaining the target prediction model for the microreaction activity of the catalytic cracking catalyst, the performance of the target prediction model is evaluated based on the root mean square error and accuracy. The RMSE expression is as follows:

[0350]

[0351] The formula for calculating its prediction accuracy is:

[0352]

[0353] Where z = 1, 2, ..., Z, Z is the number of test samples, and y d y(t) is the expected output of the catalyst microreaction activity, and y(t) is the actual output of the catalyst microreaction activity.

[0354] Figure 2 This is a block diagram of an intelligent prediction system for the microreaction activity of a catalyst in a catalytic cracking unit, provided by one embodiment of the present invention. Figure 2 As shown, the system includes:

[0355] The data processing module is used to preprocess the variable data to obtain the processed sample set;

[0356] An initial prediction model construction module is used to construct an initial prediction model of the microreaction activity of the catalytic cracking catalyst based on the sample set.

[0357] The initial prediction model optimization module is used to optimize the initial prediction model using a fast descent algorithm to obtain a target prediction model for the microreaction activity of the catalytic cracking catalyst.

[0358] The data prediction module is used to predict the preprocessed test data using the target prediction model to obtain the predicted value of catalyst microreaction activity.

[0359] On the other hand, the present invention provides a machine-readable storage medium storing instructions for causing a machine to execute the above-described intelligent prediction method for the micro-reaction activity of catalysts in a catalytic cracking unit.

[0360] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a microcontroller, chip, or processor to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0361] The optional embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details described above. Within the scope of the technical concept of the embodiments of the present invention, various simple modifications can be made to the technical solutions of the embodiments of the present invention, and these simple modifications all fall within the protection scope of the embodiments of the present invention. It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the embodiments of the present invention will not further describe the various possible combinations.

[0362] Furthermore, various different embodiments of the present invention can be combined in any way, as long as they do not violate the spirit of the embodiments of the present invention, they should also be regarded as the content disclosed by the embodiments of the present invention.

Claims

1. A method for intelligent prediction of microreactivity of a catalytic cracking unit catalyst, characterized in that, The method comprises: preprocessing variable data to obtain a processed sample set; constructing an initial prediction model of micro-reaction activity of a catalytic cracking catalyst based on a recursive RBF neural network according to the sample set; optimizing the initial prediction model by using a fast descent algorithm to obtain a target prediction model of micro-reaction activity of the catalytic cracking catalyst; predicting preprocessed to-be-tested data by using the target prediction model to obtain a predicted value of the micro-reaction activity of the catalyst; the preprocessing of the variable data to obtain the processed sample set comprises: obtaining all variable data; determining upper and lower limit interval values of each variable data; segmenting the variable data by using a window segmentation method; identifying abnormal data segments and abnormal data in the abnormal data segments by using a local outlier factor algorithm based on characteristic attributes, and eliminating the abnormal data; compensating for missing data in the variable data after the abnormal data are eliminated; normalizing the compensated variable data to obtain normalized data; performing dimension reduction processing on the normalized data to obtain an input variable data group; selecting a preset number of data groups in the input variable data group as a sample set; the compensation for the missing data in the variable data after the abnormal data are eliminated comprises: screening relevant variables of the missing data; According to the principal component analysis method, the first E variables in descending order of contribution rate are selected from the relevant variables as input variables of the RBF neural network missing compensation model. An RBF neural network missing compensation model is established, wherein the hidden layer structure of the RBF neural network missing compensation model is determined according to input variables and an experimental construction method E + 2 nodes; and an output layer is a compensation output of missing data. optimizing a RBF neural network missing compensation model by using a particle swarm algorithm to obtain an optimized RBF neural network missing compensation model; performing compensation variable soft measurement by using the optimized RBF neural network missing compensation model to obtain a compensation value of the missing data.

2. The intelligent prediction method of micro-activity of catalysts for catalytic cracking units according to claim 1, characterized in that, the determination of the upper and lower limit interval values of each variable data comprises: analyzing an operating state to which the variable data belongs; obtaining upper and lower limit interval values of each variable data according to the operating state; the operating state comprises a low-load operating state, a medium-load operating state and a full-load operating state.

3. The intelligent prediction method of micro-activity of catalysts for catalytic cracking units according to claim 1, characterized in that, the segmentation of the variable data by using the window segmentation method comprises: Let D = {X1, X2, X3,..., Xn} be the dataset of the variable data, and let X = {x1, x2, x3,..., xn} be the nth vector of X. j}, the first j vector of X j = { x j1 , x j2 ,…, x jn}, n be the total amount of data; dividing the variable data by using a clustering algorithm to determine a segmentation threshold; Determining a window size based on a segmentation threshold, a data volume of the variable data, and an operational state of the variable data s ; The window slides over the first data of each vector to obtain data segments S .

4. The intelligent prediction method of micro-activity of a catalyst of a catalytic cracking unit according to claim 1, characterized in that, identifying abnormal data segments and abnormal data in the abnormal data segments by using a local outlier factor algorithm based on characteristic attributes, and eliminating the abnormal data, which comprises: Computing the data segment S and the data segment O the first k distance, formula: d = sqrt((x2 - x1)2 + (y2 - y1)2) wherein I is the number of attributes, f ( S i ) is the value of the first S dimensional attribute of the data segment i , f ( O i ) is the value of the first i dimensional attribute of the data segment O , the attribute value comprising: an entropy value, a mean value, a maximum value, and a peak interval. wherein the entropy value Entropy The expression is: wherein, m = {1, 2, …, M}, M is the number of subspaces, p u m is the probability distribution function of the first u m subspace, expressed as:​ k u m ) is the subspace of the data u m the number of data contained;​ The local anomaly factor of the data segment S The local anomaly factor of the data segment S is evaluated by the function ； Data segment S The local reachable density of the data segment is: ； wherein, lrd k ( S ) is the local reachable density of data segment S , lrd k ( O ) is the local reachable density of data segment O , L k ( S ) is the set of all data segments in data set D whose distance is not more than the first S distance of data segment k , LOF k ( S ) is the local outlier factor value of data segment S ; is the sum of reachable distances of all neighborhood data S within data segment L k ( S ). Based on local anomaly factor LOF k The data segments are classified by grades according to the values. identifying abnormal data in different grade data segments by using a local outlier factor algorithm; discriminating all identified abnormal data and eliminating the abnormal data.

5. The intelligent prediction method of micro-activity of catalysts for catalytic cracking units according to claim 1, characterized in that, the dimension reduction processing of the normalized data to obtain the input variable data group comprises: performing dimension reduction processing on the normalized data by using a partial least squares method; analyzing to obtain, by using a principal component analysis method, a variable group composed of variables whose correlation coefficients and contribution rates are greater than a threshold value as the input variable data group.

6. The intelligent prediction method of micro-activity of catalysts for catalytic cracking units according to claim 1, characterized in that, the construction of the initial prediction model of micro-reaction activity of the catalytic cracking catalyst based on the recursive RBF neural network comprises: constructing a basic model of micro-reaction activity of the catalytic cracking catalyst based on a recursive RBF neural network, wherein the basic model of micro-reaction activity of the catalytic cracking catalyst comprises: Input layer: This layer contains n one input variable u i ( t ), i = 1, 2, … , n ; The hidden layer comprises J neurons, and an output expression of each hidden layer neuron is ； where c j ( t ) is the center vector of the j σ j ( t ) is the width of the j j input vector h j ( t ) is:​​ ； wherein y ( t -1) is t the output of the recurrent RBF neural network at time v j ( t ) is t the feedback connection weight of the output neuron at time j and the i-th hidden layer neuron. an output layer: Among them, w( t ) is the connection weight vector between the hidden layer and the output layer, θ( t ) represents the output vector of the hidden layer neurons. θ j ( t ) is the first j The output of each hidden layer neuron w j ( t ) is the first j The connection weights between each hidden layer neuron and the output neuron. y ( t )for t The output of the recursive RBF neural network at each time step.

7. The intelligent prediction method of micro-activity of a catalyst of a catalytic cracking unit according to claim 1, characterized in that, the method further comprises: After obtaining the micro-reaction activity target prediction model of the catalytic cracking catalyst, the performance of the target prediction model is evaluated according to the root mean square error and the precision, RMSE The expression is: ； a calculation formula of the prediction accuracy is: ； wherein, z= 1, 2,..., Z , Z is the number of test samples, y d ( t ) is the desired output of the catalyst microreactor activity, y ( t ) is the actual output of the catalyst microreactor activity.

8. A catalytic cracking unit catalyst microreactivity intelligent prediction device, characterized in that, the controller is configured to: preprocess variable data to obtain a processed sample set; According to the sample set, an initial prediction model of micro-reaction activity of a catalytic cracking catalyst based on a recursive RBF neural network is constructed; The initial prediction model is optimized by using a fast descent algorithm to obtain a target prediction model of micro-reaction activity of the catalytic cracking catalyst; The target prediction model is used to predict preprocessed to-be-measured data to obtain a predicted value of the micro-reaction activity of the catalyst; The preprocessed sample set is obtained by preprocessing the variable data, and the preprocessing comprises: all variable data is obtained; upper and lower limit interval values of each variable data are determined; the variable data is segmented by using a window segmentation method; abnormal data segments and abnormal data in the abnormal data segments are identified by using a local outlier factor algorithm based on characteristic attributes, and the abnormal data is removed; missing data in the variable data after the abnormal data is removed is compensated; the compensated variable data is normalized to obtain normalized data; dimensionality reduction processing is performed on the normalized data to obtain an input variable data set; a preset number of data sets in the input variable data set are selected as the sample set; the missing data in the variable data after the abnormal data is removed is compensated, and the compensation comprises: related variables of the missing data are screened; According to the principal component analysis method, the first E variables in descending order of contribution rate are selected from the relevant variables as input variables of the RBF neural network missing compensation model. An RBF neural network missing compensation model is established, wherein the hidden layer structure of the RBF neural network missing compensation model is determined according to input variables and an experimental construction method E + 2 nodes; and an output layer is a compensation output of missing data. an RBF neural network missing compensation model is optimized by using a particle swarm algorithm to obtain an optimized RBF neural network missing compensation model; the optimized RBF neural network missing compensation model is used for variable soft measurement to obtain a compensation value of the missing data.

9. The catalytic cracking unit catalyst microreactor activity intelligent prediction apparatus according to claim 8, characterized in that, The upper and lower limit interval values of each variable data are determined, and the determination comprises: an operating state to which the variable data belongs is analyzed; upper and lower limit interval values of each variable data are obtained according to the operating state; the operating state comprises a low-load operating state, a medium-load operating state, and a full-load operating state.

10. The catalytic cracking unit catalyst microreactor activity intelligent prediction apparatus of claim 8, wherein, The variable data is segmented by using a window segmentation method, and the segmentation comprises: The dataset of the variable data is set as D = {X1, X2, X3, …, Xn}, the first vector is X1, the second vector is X2, the third vector is X3, and the nth vector is Xn. j}, the first vector is X1, the second vector is X2, the third vector is X3, and the nth vector is Xn. j j = x j1 x j2 x jn n is the total amount of data.​​​​​ the variable data is divided by using a clustering algorithm to determine a segmentation threshold; Determining a window size based on a segmentation threshold, a data volume of the variable data, and an operational state of the variable data s ; The window slides over the first data of each vector to obtain data segments S .

11. The catalytic cracking unit catalyst microreactor activity intelligent prediction apparatus of claim 8, wherein, abnormal data segments and abnormal data in the abnormal data segments are identified by using a local outlier factor algorithm based on characteristic attributes, and the abnormal data is removed, and the identification comprises: Computing the data segment S and the data segment O of the first k distance, formula: d = sqrt((x2 - x1)2 + (y2 - y1)2) wherein I is the number of attributes, f ( S i ) is the value of the first S dimensional attribute of the data segment i , f ( O i ) is the value of the first O dimensional attribute of the data segment i , the attribute value comprising: an entropy value, a mean value, a maximum value, and a peak interval; wherein the entropy value Entropy The expression is: wherein, m = {1, 2, …, M}, M is the number of subspaces, p u m is the probability distribution function of the first u m subspace, expressed as:​ k u m ) is the subspace of the data u m the number of data contained​ The local anomaly factor of the data segment S The local anomaly factor of the data segment S is evaluated by the function ； Data segment S The local reachable density of the data segment is: ； in, lrd k ( S ) is the data segment S Locally achievable density, lrd k ( O ) is the data segment O Locally achievable density, L k ( S ) represents a data segment in dataset D. S The distance does not exceed its first k The set of all data segments for distance. LOF k ( S ) is the data segment S Local abnormal factor values; For data segment S All neighborhood data L k ( S The sum of reachable distances; Based on local anomaly factor LOF k The data segments are classified by grades according to the values. abnormal data in different grade data segments is identified by using a local outlier factor algorithm; all identified abnormal data is discriminated to remove the abnormal data.

12. The catalytic cracking unit catalyst microreactor activity intelligent prediction apparatus of claim 8, wherein, The normalized data is subjected to dimensionality reduction processing to obtain an input variable data set, and the dimensionality reduction processing comprises: dimensionality reduction processing is performed on the normalized data by using a partial least squares method; a principal component analysis method is used to analyze to obtain a variable group composed of variables with a correlation coefficient and a contribution rate greater than a threshold value, and the variable group constitutes the input variable data set.

13. The catalytic cracking unit catalyst microreactor activity intelligent prediction apparatus of claim 8, wherein, According to the sample set, an initial prediction model of micro-reaction activity of a catalytic cracking catalyst based on a recursive RBF neural network is constructed; The initial prediction model is optimized by using a fast descent algorithm to obtain a target prediction model of micro-reaction activity of the catalytic cracking catalyst; Input layer: This layer contains n one input variable u i ( t ), i = 1, 2, … , n ; The hidden layer comprises J neurons, and an output expression of each hidden layer neuron is ； where c j ( t ) is the center vector of the first j neuron, The target prediction model is used to predict preprocessed to-be-measured data to obtain a predicted value of the micro-reaction activity of the catalyst; j ( t ) is the width of the first j neuron, and the input vector h j of the first hidden layer neuron is given by: j ( t ) is: ； wherein y ( t -1) is t the output of the recursive RBF neural network at time v j ( t ) is t the feedback connection weight of the output neuron at time j to the i-th hidden layer neuron. The preprocessed sample set is obtained by preprocessing the variable data, and the preprocessing comprises: Among them, w( t ) is the connection weight vector between the hidden layer and the output layer, θ( t ) represents the output vector of the hidden layer neurons. all variable data is obtained; j ( t ) is the first j The output of each hidden layer neuron w j ( t ) is the first j The connection weights between each hidden layer neuron and the output neuron. y ( t )for t The output of the recursive RBF neural network at each time step.

14. The catalytic cracking unit catalyst microreactor activity intelligent prediction apparatus of claim 8, wherein, upper and lower limit interval values of each variable data are determined; After obtaining the micro-reaction activity target prediction model of the catalytic cracking catalyst, the performance of the target prediction model is evaluated according to the root mean square error and the precision, the variable data is segmented by using a window segmentation method; The expression is: ； abnormal data segments and abnormal data in the abnormal data segments are identified by using a local outlier factor algorithm based on characteristic attributes, and the abnormal data is removed; ； wherein, z= 1, 2, …, Z , Z is the number of test samples, y d ( t ) is the expected output of the catalyst microreactor activity, y ( t ) is the actual output of the catalyst microreactor activity.

15. A catalytic cracking unit catalyst microreactivity intelligent prediction system characterized by, missing data in the variable data after the abnormal data is removed is compensated; the compensated variable data is normalized to obtain normalized data; dimensionality reduction processing is performed on the normalized data to obtain an input variable data set; a preset number of data sets in the input variable data set are selected as the sample set; the missing data in the variable data after the abnormal data is removed is compensated, and the compensation comprises: related variables of the missing data are screened; an RBF neural network missing compensation model is optimized by using a particle swarm algorithm to obtain an optimized RBF neural network missing compensation model; the optimized RBF neural network missing compensation model is used for variable soft measurement to obtain a compensation value of the missing data. The upper and lower limit interval values of each variable data are determined, and the determination comprises: an operating state to which the variable data belongs is analyzed; upper and lower limit interval values of each variable data are obtained according to the operating state; the operating state comprises a low-load operating state, a medium-load operating state, and a full-load operating state. The variable data is segmented by using a window segmentation method, and the segmentation comprises: the variable data is divided by using a clustering algorithm to determine a segmentation threshold; abnormal data segments and abnormal data in the abnormal data segments are identified by using a local outlier factor algorithm based on characteristic attributes, and the abnormal data is removed, and the identification comprises: Entropy LOF LOF abnormal data in different grade data segments is identified by using a local outlier factor algorithm; all identified abnormal data is discriminated to remove the abnormal data. The normalized data is subjected to dimensionality reduction processing to obtain an input variable data set, and the dimensionality reduction processing comprises: dimensionality reduction processing is performed on the normalized data by using a partial least squares method; a principal component analysis method is used to analyze to obtain a variable group composed of variables with a correlation coefficient and a contribution rate greater than a threshold value, and the variable group constitutes the input variable data set. According to the sample set, an initial prediction model of micro-reaction activity of a catalytic cracking catalyst based on a recursive RBF neural network is constructed; The initial prediction model is optimized by using a fast descent algorithm to obtain a target prediction model of micro-reaction activity of the catalytic cracking catalyst; The target prediction model is used to predict preprocessed to-be-measured data to obtain a predicted value of the micro-reaction activity of the catalyst; The preprocessed sample set is obtained by preprocessing the variable data, and the preprocessing comprises: all variable data is obtained; upper and lower limit interval values of each variable data are determined; the variable data is segmented by using a window segmentation method; abnormal data segments and abnormal data in the abnormal data segments are identified by using a local outlier factor algorithm based on characteristic attributes, and the abnormal data is removed; missing data in the variable data after the abnormal data is removed is compensated; the compensated variable data is normalized to obtain normalized data; dimensionality reduction processing is performed on the normalized data to obtain an input variable data set; a preset number of data sets in the input variable data set are selected as the sample set; the missing data in the variable data after the abnormal data is removed is compensated, and the compensation comprises: related variables of the missing data are screened; an RBF neural network missing compensation model is optimized by using a particle swarm algorithm to obtain an optimized RBF neural network missing compensation model; the optimized RBF neural network missing compensation model is used for variable soft measurement to obtain a compensation value of the missing data. The system comprises: An initial prediction model construction module is configured to construct an initial prediction model of micro-reaction activity of a catalytic cracking catalyst based on a recursive RBF neural network according to the sample set; An initial prediction model optimization module is configured to optimize the initial prediction model by using a fast descent algorithm to obtain a target prediction model of micro-reaction activity of a catalytic cracking catalyst; A data prediction module is configured to predict preprocessed data to be measured by using the target prediction model to obtain a predicted value of catalyst micro-reaction activity. The preprocessed variable data is obtained by: Obtaining all variable data; Determining the upper and lower limit interval values of each variable data; Segmenting the variable data by using a window segmentation method; Identifying abnormal data segments and abnormal data in the abnormal data segments by using a local anomaly factor algorithm based on characteristic attributes, and eliminating the abnormal data; Compensating for missing data in the variable data after eliminating the abnormal data; Normalizing the compensated variable data to obtain normalized data; Reducing the dimension of the normalized data to obtain an input variable data set; Selecting a preset number of data sets from the input variable data set as a sample set; Compensating for missing data in the variable data after eliminating the abnormal data includes: Screening relevant variables of the missing data; According to the principal component analysis method, the first E variables in descending order of contribution rate are selected from the relevant variables as input variables of the RBF neural network missing compensation model. An RBF neural network missing compensation model is established, wherein the hidden layer structure of the RBF neural network missing compensation model is determined according to input variables and an experimental construction method E + 2 nodes; and an output layer is a compensation output of missing data. Optimizing and calculating an RBF neural network missing compensation model by using a particle swarm algorithm to obtain an optimized RBF neural network missing compensation model; Compensating for variable soft measurement by using the optimized RBF neural network missing compensation model to obtain a compensation value of the missing data. 16.A machine readable storage medium, on which instructions are stored for causing a machine to perform the intelligent prediction method of micro-reaction activity of a catalytic cracking device catalyst according to any one of claims 1-7.

Images