A method and system for predicting indicators of a dual riser catalytic reaction process

Abstract

The application provides a double-lift tube catalytic reaction process index prediction method and system, and the method comprises the following steps: obtaining production process data of a target double-lift tube catalytic cracking device at different intervals in at least one complete operation cycle; dividing components of reactant materials according to characteristics of the target double-lift tube catalytic reaction process; fully considering diesel component conversion reaction characteristics in the reaction process to construct a double-lift tube catalytic reaction process dynamics model, and solving dynamics parameters of the model; performing cluster analysis on the production process data, dividing sample subsets according to the cluster results, training a sub-model based on a space-time attention long short-term memory network based on the sample subsets, and constructing a data-driven model; and adaptively adjusting weights of the dynamics model and the data-driven model to construct a hybrid model, and predicting catalytic reaction process indexes. Based on the method, a double-lift tube catalytic reaction process index prediction system is also provided. The application can effectively improve the reliability of index prediction results.

Description

Technical Field

[0001] This invention belongs to the field of petrochemical production process modeling technology, and specifically relates to a method and system for predicting indicators of a dual-riseer catalytic reaction process. Background Technology

[0002] Catalytic cracking units are the most common secondary processing units in refining and chemical enterprises. They can convert heavy crude oil into high-value-added light oil products and chemical feedstocks, such as gasoline, diesel, and low-carbon olefins, in a highly efficient, green, and economical manner. With the further standardization of clean fuel standards for vehicles, increasingly stringent environmental indicators, and the rapid development of the new energy industry, the peak consumption of gasoline and diesel products has arrived earlier than expected, and the demand for chemical products such as low-carbon olefins has increased rapidly. Furthermore, due to the high aromatic content and low cetane number of catalytic diesel, it differs significantly from the standards for automotive diesel. Therefore, reducing the diesel-to-gasoline ratio or achieving high-value utilization of catalytic diesel has significant research and application value. The dual-riseer combined process catalytic cracking unit achieves highly selective cracking of two differentiated feedstocks: heavy feedstock and hydrotreated catalytic diesel, through the use of two riser reactors. The heavy oil reactor employs a process that maximizes the production of isoalkane components. By using a series of variable-diameter reactors, the reaction path of the cracking process is flexibly controlled. In the first reaction zone, heavy crude oil cracking is promoted, while in the second reaction zone, isomerization and hydrogen transfer reactions are facilitated. This reduces the olefin content in the gasoline fraction, converting it as much as possible into isoalkane and aromatic components, thereby increasing the gasoline's octane number. The hydrotreated diesel reactor utilizes a process that produces high-octane gasoline or light aromatics from hydrotreated catalytic diesel. It further catalytically cracks the inferior catalytic diesel fraction after directional hydrotreating, maximizing ring-opening cracking to produce high-octane gasoline rich in aromatics. This process plays a crucial role in the transformation and upgrading of refining enterprises, aiming to reduce the diesel-to-gasoline ratio and increase chemical production.

[0003] The dual-riseer combined process for catalytic cracking is greatly influenced by feed properties, catalyst properties, and operating conditions, resulting in a complex reaction system and a highly nonlinear, strongly coupled, and continuously dynamic production process. Mechanistic models built from process mechanism analysis often rely on some ideal assumptions, leading to a reduction in model accuracy. Data-driven models, on the other hand, extract relationships between input and output variables from massive amounts of plant operation data. Based on feature selection, pattern classification, and the powerful nonlinear fitting capabilities of the underlying network, they predict key process indicators and typically have good prediction accuracy. However, factors such as overfitting may result in low model sensitivity and unreliable predictions under fluctuating input variables. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a method and system for predicting process indicators in dual-elevator catalytic reaction processes. By integrating the characteristics of mechanistic and data-driven models, the weights of the mechanistic and data-driven models are adaptively adjusted based on the similarity between actual input samples and training samples. This constructs an adaptive hybrid model driven by both process mechanism and production data to predict catalytic process indicators, thereby improving the reliability of the model's prediction results.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A method for predicting indicators of a dual-elevator catalytic reaction process includes the following steps:

[0007] Acquire production process data at different time intervals within at least one complete operating cycle of the target dual-rise catalytic cracking unit;

[0008] Based on the characteristics of the target dual-rise catalytic reaction process, the reactants are divided into components. A kinetic model of the dual-rise catalytic reaction process is constructed, taking into full account the characteristics of diesel component conversion during the reaction. The kinetic parameters of the kinetic model are then solved.

[0009] Cluster analysis of production process data is performed based on hierarchical agglomerative clustering. Based on the clustering results, sample subsets are divided. Sub-models are trained based on spatiotemporal attention long short-term memory networks based on the sample subsets to construct a data-driven model.

[0010] Based on the similarity between input samples and training samples, a hybrid model is constructed by adaptively adjusting the weights of the kinetic model and the data-driven model using the Latin hypercube sampling method to predict indicators of the catalytic reaction process; and the prediction results of the hybrid model are verified.

[0011] Furthermore, the production process data includes: feed and product properties of the heavy oil reactor, feed and product properties of the hydrotreated diesel reactor, catalyst properties, and operating conditions.

[0012] Furthermore, the process of classifying the reactants according to the characteristics of the target dual-riseer catalytic reaction process includes: classifying the reactants according to the characteristics of the dual-riseer catalytic reaction process that combines the production of more isoalkanes and more high-octane gasoline or light aromatics to reduce the diesel-gasoline ratio.

[0013] Furthermore, the reaction kinetic model specifically includes:

[0014] The reaction rate equation is expressed as:

[0015] Where C is the mass concentration vector of the material components during the reaction; e is the dimensionless distance; K is the reaction rate constant matrix; ρ is the gas density; S WHis the true weight hourly space velocity; Z represents the correction term for the reaction rate.

[0016] Furthermore, the process of performing cluster analysis on production process data based on hierarchical agglomerative clustering includes:

[0017] The acquired production process data is extracted using time windows and divided into multiple subsequences;

[0018] Hierarchical agglomerative clustering is performed on the subsequence data. One subsequence is selected as the cluster set point to establish the first-level cluster, represented as: D={d1,...d2,...,d n};

[0019] Calculate the distance between each pair of subsequences in the first-level cluster sequentially (d) i ,d j Distance is used to obtain the similarity between each subsequence:

[0020]

[0021] Where, d i For the i-th subsequence; d j Let be the j-th subsequence; p is the dimension of the sequence data; d ip d represents the p-dimensional data in the i-th subsequence; jp This represents the p-dimensional data in the j-th subsequence;

[0022] Clustering the two shortest subsequence pairs as the next layer clusters, integrating features, and building a higher-level cluster;

[0023] The clustering process is repeated, and the classification threshold is selected based on the minimum loss function to complete the cluster analysis.

[0024] Furthermore, the production process data driving model is constructed by training a sub-model based on a spatiotemporal attention long short-term memory network using a subset of samples.

[0025] The encoder and decoder architecture is used to implement sequence-to-sequence data-driven model construction;

[0026] The encoder incorporates a spatial attention mechanism, assigning different spatial attention values ​​to the input variables. Spatial attention-weighted sample data serves as the input to the encoder LSTM; the input and output of the encoder LSTM unit are then determined.

[0027] Introduce a temporal attention mechanism into the decoder structure: assign a temporal attention value to each encoder hidden state;

[0028] Update the encoder intermediate state and decoder hidden state and predict the final output;

[0029] Dynamic optimization of hyperparameters in spatiotemporal attention long short-term memory network model: The optimal parameter combination for each sub-model after pattern classification is found by grid search method, and a data-driven model is constructed under the optimal hyperparameter combination.

[0030] Furthermore, the similarity calculation process between the input sample and the training sample is as follows:

[0031] Distance between samples m The calculation method is as follows:

[0032]

[0033] x i =[x1,x2,...,x n [X] is the input sample. train Let x represent the training sample set. train_min ∈X train It is the input sample in the training sample set that is closest to the current input sample, where m is a constant that controls the calculation of the distance between the given input and the sample, 1≤m≤N. train N train This represents the number of samples in the training sample set.

[0034] Furthermore, the process of adaptively adjusting the weights of the kinetic model and the data-driven model using the Latin hypercube sampling method to construct a hybrid model for predicting catalytic reaction process indicators is as follows:

[0035] The weights w of the dynamic model are measured based on the similarity between the input sample and the training sample, and the similarity between the virtual input obtained by Latin hypercube sampling and the training sample. M :

[0036]

[0037] in, This refers to virtual input collected within a fixed sampling range. The maximum possible distance between the virtual input and the training dataset. m distance;

[0038] μ(X)+3σ(X) is used as the upper limit of sampling, and μ(X)-3σ(X) is used as the lower limit of sampling, where μ(X) and σ(X) represent the mean and standard deviation of the sample set, respectively.

[0039] The weights of the data-driven model are: w D =1-w M ;

[0040] Predicting catalytic reaction process parameters

[0041] ypred =w M ×y pred_mechanism +w D ×y pred_data-driven

[0042] Among them, y pred For the predicted output of the mixture model, y pred_mechanism ,y pred_data-driven These represent the predicted outputs of the dynamic model and the data-driven model, respectively.

[0043] Furthermore, when verifying the prediction results of the hybrid model, the performance of the dynamic model, the data-driven model, and the hybrid model are comprehensively considered, and the mean squared error (MSE) and correlation coefficient (R²) of the prediction results on the test dataset are used as indicators. 2 The correlation coefficient ρ with consistency c Three metrics are used to evaluate model performance;

[0044] The smaller the MSE value, the smaller the sum of squared prediction errors of the mixture model on the prediction sample set, and the higher the accuracy of the mixture model; R 2 The closer the value of ρ is to 1, the higher the accuracy of the mixture model; c The larger the value, the greater the correlation, and the higher the reliability of the prediction results of the hybrid model.

[0045] This invention also proposes a prediction system for indicators of a dual-elevator catalytic reaction process, comprising an acquisition module, a first construction module, a second construction module, and a prediction module;

[0046] The acquisition module is used to acquire production process data at different time intervals within at least one complete operating cycle of the target dual riser catalytic cracking unit;

[0047] The first construction module is used to classify the reactants into components according to the characteristics of the target dual-rise catalytic reaction process, fully consider the diesel component conversion reaction characteristics during the reaction process to construct a kinetic model of the dual-rise catalytic reaction process, and solve the kinetic parameters of the kinetic model;

[0048] The second building module is used to perform cluster analysis on production process data based on hierarchical agglomerative clustering, divide the sample subsets according to the clustering results, and train the sub-model based on the spatiotemporal attention long short-term memory network according to the sample subsets to build a data-driven model;

[0049] The prediction module is used to construct a hybrid model based on the similarity between the input sample and the training sample, combined with the Latin hypercube sampling method to adaptively adjust the weights of the kinetic model and the data-driven model, to predict the indicators of the catalytic reaction process; and to verify the prediction results of the hybrid model.

[0050] The effects described in the invention are merely those of the embodiments, and not all the effects of the invention. One of the above technical solutions has the following advantages or beneficial effects:

[0051] This invention proposes a method and system for predicting indicators of a dual-rise catalytic reaction process. The method includes the following steps: acquiring production process data at different time intervals within at least one complete operating cycle of the target dual-rise catalytic cracking unit; classifying the reactants into components based on the characteristics of the target dual-rise catalytic reaction process, constructing a kinetic model of the dual-rise catalytic reaction process considering the diesel component conversion reaction characteristics during the reaction, and solving the kinetic parameters of the kinetic model; performing cluster analysis on the production process data based on hierarchical agglomerative clustering, dividing the sample subsets according to the clustering results, training a sub-model based on a spatiotemporal attention long short-term memory network according to the sample subsets, and constructing a data-driven model; adaptively adjusting the weights of the kinetic model and the data-driven model based on the similarity between the input samples and the training samples, combined with the Latin hypercube sampling method, constructing a hybrid model to predict the indicators of the catalytic reaction process; and verifying the prediction results of the hybrid model. Based on this method for predicting indicators of a dual-rise catalytic reaction process, a system for predicting indicators of a dual-rise catalytic reaction process is also proposed. This invention integrates the characteristics of kinetic models and data-driven models, and adaptively adjusts the weights of the kinetic model and data-driven model based on the similarity between actual input samples and training samples. It constructs an adaptive hybrid model driven by both process mechanism and production data to predict catalytic process indicators, which helps to improve the reliability of model prediction results. Attached Figure Description

[0052] Figure 1 This is a flowchart of a method for predicting indicators of a dual-lifter catalytic reaction process proposed in Embodiment 1 of the present invention;

[0053] Figure 2 This is a schematic diagram of the catalytic cracking process of the dual riser combined process proposed in Embodiment 1 of the present invention;

[0054] Figure 3 This is a reaction network diagram of the catalytic process of the combined process of heavy oil producing more isoalkanes and hydrogenation catalytic light cycle oil producing more high-octane gasoline or light aromatics proposed in Example 1 of the present invention.

[0055] Figure 4 This is a schematic diagram illustrating the clustering effect of production process data based on hierarchical agglomerative clustering proposed in Embodiment 1 of the present invention;

[0056] Figure 5 This is a schematic diagram of the spatiotemporal attention long short-term memory network model proposed in Embodiment 1 of the present invention;

[0057] Figure 6This is a schematic diagram illustrating the adaptive hybrid prediction process based on the similarity between input samples and training samples, as proposed in Embodiment 1 of the present invention.

[0058] Figure 7 This is a graph showing the predicted results of the kinetic model of the dual-elevator catalytic reaction process proposed in Example 1 of this invention;

[0059] Figure 8 This is a graph showing the prediction results of the data-driven model for the dual-elevator catalytic reaction process proposed in Example 1 of this invention;

[0060] Figure 9 This is a graph showing the adaptive weighted mixing prediction results for the dual-elevator catalytic reaction process proposed in Example 1 of this invention;

[0061] Figure 10 This is a schematic diagram of a dual-elevator catalytic reaction process index prediction system proposed in Embodiment 2 of the present invention;

[0062] Legend: 101 - Heavy oil reactor, 102 - Hydrogenated diesel reactor, 103 - Regenerator, and 105 - Product separation unit. Detailed Implementation

[0063] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific embodiments and in conjunction with the accompanying drawings. The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure of the invention, components and arrangements of specific examples are described below. Furthermore, reference numerals and / or letters may be repeated in different examples. This repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed. It should be noted that the components illustrated in the drawings are not necessarily drawn to scale. Descriptions of well-known components, processing techniques, and processes are omitted in this invention to avoid unnecessarily limiting the invention.

[0064] Example 1

[0065] Example 1 of this invention proposes a method for predicting indicators of dual-riseer catalytic reaction processes. This method addresses the difficulties in constructing models and the low reliability of indicator predictions for dual-riseer catalytic reaction processes in combined processes where heavy oil produces more isoalkanes and hydrotreated light cycle oil produces more high-octane gasoline or light aromatics. Specifically, this invention provides a method for predicting indicators of dual-riseer catalytic reaction processes aimed at reducing the diesel-to-gasoline ratio in combined processes that produce more isoalkanes and more high-octane gasoline or light aromatics.

[0066] Figure 1 This is a flowchart of a method for predicting indicators of a dual-lifter catalytic reaction process proposed in Embodiment 1 of the present invention;

[0067] In step S100, production process data at different time intervals within at least one complete operating cycle of the target dual-rise catalytic cracking unit are acquired;

[0068] Production process data includes: feed and product properties of heavy oil reactor, feed and product properties of hydrotreated diesel reactor, catalyst properties, and operating conditions.

[0069] Figure 2 This is a schematic diagram of the catalytic cracking process using a dual-riseer combined process proposed in Example 1 of the present invention. The heavy oil reactor 101 processes a mixture of heavy feedstocks, including atmospheric residue, vacuum residue, and wax oil, using a process that produces a high proportion of isoalkane. The hydrotreated diesel reactor 102 uses a process that produces a high proportion of high-octane gasoline and aromatics. The two different feedstocks are fed into their respective riser reactors for reaction. During the reaction, the coking and deactivated catalyst enters the regenerator 103 for coke burn-off regeneration. The reacted oil and gas enter the product separation unit 105 to obtain the main products of the catalytic cracking process, including slurry oil, diesel oil, gasoline, liquefied petroleum gas (LPG), and dry gas.

[0070] The feedstock for the heavy oil reactor is a mixture of heavy oil feedstocks such as atmospheric residue, vacuum residue, and wax oil; the feedstock for the hydrotreated diesel reactor is hydrotreated diesel and recycled light waste oil.

[0071] The properties of feedstock and products include: the four-component composition, density, and residual carbon content of heavy oil feedstock SARA; the composition and density of hydrotreated diesel PONA group; and the composition of catalytic gasoline and diesel PONA group. Among them, the four-component composition of heavy oil feedstock SARA includes: saturate, aromatic, resin, and asphaltene; and the composition of hydrotreated diesel PONA group includes: paraffins, olefins, naphthenes, and aromatics.

[0072] Catalyst properties include: carbon content of the precursor and carbon content of the regenerator.

[0073] Operating conditions include: reaction temperature, reaction pressure, reactant-to-oil ratio, catalyst circulation rate, etc.

[0074] A total of 67 model input variables were obtained. Considering the characteristics of the dynamic model and the data model in the calculation process, the average data of the above production process were obtained at intervals of 1 day and 1 hour, respectively, with a time length of 13 months, which can at least cover one complete operating cycle of the production unit. Among them, the average data at intervals of 1 day was used to solve the dynamic model based on the differential equation system, and the average data at intervals of 1 hour was used to train the data-driven model. After removing abnormal samples, 380 samples at intervals of 1 day and 8980 samples at intervals of 1 hour were obtained respectively.

[0075] In step S110, the reactants are divided into components according to the characteristics of the target dual-rise catalytic reaction process, a kinetic model of the dual-rise catalytic reaction process is constructed based on the characteristics of diesel component conversion reaction during the reaction process, and the kinetic model is solved.

[0076] Specifically, this includes: combining the characteristics of the dual-riseer catalytic reaction process for reducing the diesel-to-gasoline ratio using a combination process that produces more isoalkanes and more high-octane gasoline or light aromatics; classifying the reactants into components; proposing reasonable reaction kinetic assumptions; fully considering the characteristics of diesel component conversion reactions during the dual-riseer reaction, constructing a reaction network, and considering the influencing factors such as catalyst deactivation, basic nitrogen oxides, and adsorption deactivation of heavy aromatics, establishing a kinetic model for the dual-riseer catalytic reaction process for reducing the diesel-to-gasoline ratio; and solving the model's kinetic parameters using a genetic algorithm.

[0077] For complex hydrocarbon compounds in the reaction system, the material components are classified according to similar kinetic characteristics by combining distillation range and hydrocarbon group composition analysis.

[0078] For heavy feedstock oil, it is divided into three material components based on the principle of similar kinetic characteristics: saturated component, aromatic component, and resin and asphaltenes.

[0079] For hydrocatalytic diesel feedstock, it is further divided according to the PONA hydrocarbon group composition into five material components: diesel alkanes, diesel olefins, diesel cycloalkanes, diesel monocyclic aromatics, and diesel polycyclic aromatics.

[0080] In the reaction system, the liquid-phase products, catalytic diesel (boiling range 200℃~380℃) and catalytic gasoline (boiling range 40℃~200℃), are classified according to their hydrocarbon group composition. Among the gaseous products, butene, propylene, and ethylene can be used as important low-carbon olefin chemical raw materials and are classified as separate components. The remaining products are classified into liquefied petroleum gas (C3~C4) and dry gas (C1~C2) according to their carbon content. The coke generated during the reaction is treated as a separate material component.

[0081] The dual-riseer catalytic reaction process for reducing the diesel-to-gasoline ratio was divided into 18 material components with similar kinetic characteristics. Figure 3This is a reaction network diagram of the combined process of heavy oil producing more isoalkanes and hydrogenated catalytic light cycle oil producing more high-octane gasoline or light aromatics proposed in Example 1 of this invention. Based on the division of material components and kinetic assumptions, and taking full account of the diesel component conversion reaction characteristics in the dual riser reaction process, a reaction network of 18 material components containing a total of 130 reaction paths was constructed.

[0082] This application proposes reasonable reaction kinetic assumptions, including: ① the reaction in the reaction system is considered as a homogeneous reaction; ② the gas flow state in the reaction system is isothermal, gas-phase, ideal plug flow, and diffusion within material particles is ignored; ③ the effect of catalyst deactivation on the reaction process in the reaction system is characterized by time-varying catalyst deactivation; ④ no coke is generated from the gases in the reaction system.

[0083] The established reaction kinetic model specifically includes:

[0084] The reaction rate equation is expressed as:

[0085] Where C is the mass concentration vector of the material components during the reaction; e is the dimensionless distance; K is the reaction rate constant matrix; ρ is the gas density; S WH is the true weight hourly space velocity; Z represents the correction term for the reaction rate.

[0086] Z = f(A)·f(N)·φ(C) C f(A) is the operator for the effect of catalyst coking and deactivation; C A The residual carbon content of the feed; k A Deactivation factor of feed residual carbon;

[0087] f(N) is the heavy aromatic hydrocarbon adsorption deactivation operator; C N The feed alkali-nitrogen content; k N The feed alkali-nitrogen deactivation factor; φ C / O The reactant-to-oil ratio; t c This refers to the residence time of the catalyst.

[0088] This is a base nitrogen adsorption deactivation operator; C C denoted as , where β is the catalyst coke content; β is the catalyst coke deactivation factor; and M is the exponential constant of the catalyst coking deactivation operator.

[0089] The reaction rate constant conforms to the Arrhenius equation:

[0090] k i,j k is the reaction rate constant corresponding to the reaction from material component i to material component j in the 130 reaction paths of the reaction network; 0i,jEa i,j These are the pre-exponential factor and activation energy of the conversion reaction from material component i to material component j in the 130 reaction paths of the reaction network, respectively; R is the gas molar constant; and T is the reaction temperature.

[0091] Genetic algorithms are used to solve for the dynamic parameters of a model, including: parallel or serial methods can be used to optimize and solve for the dynamic parameters in the dynamic model based on genetic algorithms.

[0092] Genetic algorithms are a method for searching for optimal solutions by simulating the natural evolutionary process. They utilize three basic genetic operators—selection, crossover, and mutation—to perform optimization calculations, and generally do not require external information during the search and evolution process. The process of searching for the optimal solution includes: (1) random initialization of the population; (2) evaluation of individual fitness; (3) selection operation; (4) crossover operation; (5) mutation operation; and (6) termination condition judgment. In this implementation, a serial approach is used to predict the yield of the main products in the catalytic cracking process based on the combination of kinetic parameters in the genetic algorithm optimization mechanism model.

[0093] In step S120, cluster analysis is performed on the production process data based on hierarchical agglomerative clustering. The sample subsets are divided according to the clustering results. The sub-model is trained based on the spatiotemporal attention long short-term memory network according to the sample subsets to construct a data-driven model.

[0094] The process of performing cluster analysis on production process data based on hierarchical agglomerative clustering includes: extracting the acquired production process data using an appropriate time window and dividing it into multiple subsequences; performing hierarchical agglomerative clustering on the subsequence data, selecting one subsequence as the cluster set point, and establishing the first-level cluster, represented as: D={d1,...d2,...,d n}; Calculate sequentially the relationship between every two subsequence pairs in the first-level cluster (d i ,d j Distance is used to obtain the similarity between each subsequence:

[0095]

[0096] Where, d i For the i-th subsequence; d j Let be the j-th subsequence; p is the dimension of the sequence data; d ip d represents the p-dimensional data in the i-th subsequence; jp This represents the p-dimensional data in the j-th subsequence;

[0097] Clustering the two shortest subsequence pairs as the next layer clusters, integrating features, and building a higher-level cluster;

[0098] The clustering process is repeated, and the classification threshold is selected based on the minimum loss function to complete the cluster analysis.

[0099] The reaction system of the dual-riseer catalytic cracking process, which combines high-octane gasoline and aromatic feedstock production, is complex and is affected by fluctuations in operating conditions and changes in feedstock properties. By classifying the production process data and dividing it into sub-models using an unsupervised hierarchical aggregation clustering algorithm, and fitting the model parameters to each sub-model, the accuracy of the data-driven model can be improved. Figure 4 This is a schematic diagram of the clustering effect of production process data based on hierarchical agglomerative clustering proposed in Embodiment 1 of the present invention; after hierarchical agglomerative clustering, the four types of sub-pattern sample data are obtained and the two-dimensional projection distribution map is extracted by principal component analysis.

[0100] Sub-models were constructed by combining spatiotemporal attention long short-term memory networks, and the hyperparameters of each sub-model were optimized using a grid search algorithm. Figure 5 This is a schematic diagram of the spatiotemporal attention long short-term memory network model proposed in Embodiment 1 of the present invention;

[0101] A sequence-to-sequence data-driven model is constructed using an encoder and decoder architecture; the encoder structure is given an input sequence {x1, x2, ..., x} within a given window. T The hidden sequence is {h1, h2, ..., h}. T}, where T is the length of the input sequence; the hidden sequence is transformed into a fixed-length intermediate vector c = f1(h1, h2, ..., h) through a mapping function. T At this point, the decoder outputs the sequence {y1, y2, ..., y...}. l}Depend on The output sequence is obtained; l is the length of the output sequence, and t represents the sequence number of the current variable.

[0102] The encoder incorporates a spatial attention mechanism, assigning different spatial attention values ​​to the input variables, and using spatial attention-weighted sample data as input to the encoder's LSTM.

[0103] A spatial attention mechanism is incorporated into the encoder, assigning different spatial attention values ​​to the input variables. The spatial attention-weighted sample data is used as the input to the encoder's LSTM, where:

[0104]

[0105]

[0106] This represents the level of attention given to the v-th input variable at time t, where n is the dimension of the variable, and s t-1This represents the hidden state before the current original input to the decoder; Q1 is the first weight matrix, W1 is the second weight matrix, W2 is the third weight matrix, b1 is the bias matrix, and the normalized weight matrix is... To correspond to the spatial attention value, the spatial attention-weighted sample representation is as follows:

[0107] In the encoder, a spatial attention mechanism is introduced to selectively distinguish input variables that are related to the predictor at each time step and assign different spatial attention values ​​to the related input variables. The encoder can adaptively identify input variables that are more closely related to the predictor.

[0108] Determine the inputs and outputs of the encoder LSTM unit;

[0109] Input unit:

[0110] Output unit:

[0111] Forgetting Unit:

[0112]

[0113] Long memory unit:

[0114] Short Memory Units:

[0115] Among them, W xi W is the first weight matrix in the input cell of the encoder LSTM unit; hi The second weight matrix, b, is the input matrix of the encoder LSTM unit. i W is the bias matrix in the input cell of the encoder LSTM unit; xo W is the first weight matrix in the output unit of the encoder LSTM unit; ho b is the second weight matrix in the output unit of the encoder LSTM unit; o W is the bias matrix in the output unit of the encoder LSTM unit; xf W is the first weight matrix in the forgetting unit of the encoder LSTM unit; hf b is the second weight matrix in the forgetting unit of the encoder LSTM unit; f W is the bias matrix in the forgetting cell of the encoder LSTM unit; xg W is the first weight matrix corresponding to the intermediate state; hg b is the second weight matrix corresponding to the intermediate state; g This is the deviation matrix corresponding to the intermediate state; This represents point-by-point multiplication of two vectors.

[0116] Introduce a temporal attention mechanism into the decoder structure: assign a temporal attention value to each encoder hidden state;

[0117] The calculation method is as follows:

[0118]

[0119]

[0120] in, This represents the attention level of the u-th encoder hidden state within the t-th time window, after normalization. This corresponds to the temporal attention value. A temporal attention mechanism is introduced into the decoder LSTM to adaptively determine the relevant hidden states generated by the encoder LSTM at all time steps, and the output is calculated by referring to the previous decoder hidden states.

[0121] Update the encoder intermediate states and decoder hidden states and predict the final output: that is, calculate the time attention-weighted sum of all encoder hidden states to obtain the intermediate vector c. t Combined with the given target sequence {y1,y2,...,y... t-1 Update the decoder's hidden state.

[0122]

[0123]

[0124]

[0125] Predicted output y T The calculation method is as follows:

[0126]

[0127] Where F represents the calculation of the fully connected layer.

[0128] The hyperparameters of the spatiotemporal attention long short-term memory network model are dynamically optimized: the optimal parameter combination for each sub-model after pattern classification is found by grid search, including: the number of layers and hidden nodes of the spatial attention network, the size of the time window, and the number of layers and hidden nodes of the temporal attention network; a data-driven model is constructed under the optimal hyperparameter combination.

[0129] Table 1 below shows the grid search parameters set:

[0130] Spatial attention network layers enc_layer_num [1,2,3,4,5,6] 2 Spatial attention hidden node count enc_hidden_size [1,2,3,4,5,6] 32 Number of layers in a time-attention network dec_layer_num [8,16,32,64,128,256] 1 Number of hidden nodes in time attention dec_hidden_size [8,16,32,64,128,256] 128 Prediction window size window_size [5,6,7,8,9,10,11,12] 12

[0131] In step S130, based on the similarity between the input sample and the training sample, the weights of the kinetic model and the data-driven model are adaptively adjusted by combining the Latin hypercube sampling method to construct a hybrid model for predicting the indicators of the catalytic reaction process.

[0132] The process of calculating the similarity between the input sample and the training sample is as follows:

[0133] Distance between samples m The calculation method is as follows:

[0134]

[0135] x i =[x1,x2,...,x n [X] is the input sample. train Let x represent the training sample set. train_min ∈X train It is the input sample in the training sample set that is closest to the current input sample, where m is a constant that controls the calculation of the distance between the given input and the sample, 1≤m≤N. train N train This represents the number of samples in the training sample set.

[0136] The weights w of the dynamic model are measured based on the similarity between the input sample and the training sample, and the similarity between the virtual input obtained by Latin hypercube sampling and the training sample. M :

[0137]

[0138] in, This refers to virtual input collected within a fixed upper and lower sampling limit; Represents the maximum possible distance between the virtual input and the training dataset. m distance;

[0139] μ(X)+3σ(X) is used as the upper limit of sampling, and μ(X)-3σ(X) is used as the lower limit of sampling, where μ(X) and σ(X) represent the mean and standard deviation of the sample set, respectively.

[0140] The weights of the data-driven model are: w D =1-w M ;

[0141] Predicting catalytic reaction process parameters

[0142] y pred =w M ×y pred_mechanism +w D ×y pred_data-driven

[0143] Among them, ypred For the predicted output of the mixture model, y pred_mechanism ,y pred_data-driven These represent the predicted outputs of the dynamic model and the data-driven model, respectively.

[0144] In step S140, the prediction results of the hybrid model are verified. Figure 6 This is a schematic diagram illustrating the adaptive hybrid prediction process based on the similarity between input samples and training samples, as proposed in Embodiment 1 of the present invention. Figure 7 This is a graph showing the predicted results of the kinetic model of the dual-elevator catalytic reaction process proposed in Example 1 of this invention; Figure 8 This is a graph showing the prediction results of the data-driven model for the dual-elevator catalytic reaction process proposed in Example 1 of this invention; Figure 9 This is a graph showing the adaptive weighted hybrid prediction results for the dual-elevator catalytic reaction process proposed in Example 1 of this invention. In Example 1 of this application, the performance of the kinetic model, data model, and adaptive weighted hybrid model are comprehensively considered, and the mean squared error (MSE) and correlation coefficient (R) of the prediction results on the test dataset are used to determine the prediction results. 2 Consistency correlation coefficient ρ c Three metrics are used to evaluate the model's performance.

[0145] The mean square error (MSE) is calculated as follows:

[0146]

[0147] The smaller the MSE value, the smaller the sum of squared prediction errors of the model on the test sample set, and the higher the model accuracy; the correlation coefficient R 2 The calculation method is as follows:

[0148]

[0149] R 2 R represents the correlation between the model's predicted values ​​and the actual values. 2 The closer the value is to 1, the higher the model's prediction accuracy; the consistency correlation coefficient ρ c The calculation method is as follows:

[0150]

[0151] ρ c ρ represents an index that comprehensively considers the error and correlation of model prediction results. c The larger the value, the greater the correlation, indicating higher reliability of the model's prediction results; the above σ represents the actual value, predicted value, and average value of the forecast indicator, respectively; x σ y Let μ represent the standard deviations of the actual and average values ​​of the predicted indicators, respectively. x μy These represent the mean of the actual value and the average value of the predicted indicator, respectively.

[0152] Table 2 below shows the three metrics used by the three models to predict the yield of key products:

[0153]

[0154]

[0155] In Embodiment 1 of this invention, the test sample set contains a total of 130 samples. The data model is based on MSE and R... 2 Better performance on the metrics indicates higher predictive accuracy of the data model; in calculating the model consistency correlation coefficient ρ... c When considering both the error and correlation of the model prediction results, the adaptive weighted hybrid model performs better, indicating that the hybrid model has higher reliability. Moreover, the hybrid model performs well in terms of prediction accuracy, demonstrating that the dual riser catalytic reaction process index prediction method proposed in this invention can effectively improve the reliability of index prediction results.

[0156] Embodiment 1 of this invention proposes a method for predicting indicators of a dual-lifter catalytic reaction process. This method integrates the characteristics of a kinetic model and a data-driven model. It adaptively adjusts the weights of the kinetic model and the data-driven model based on the similarity between the actual input samples and the training samples. This constructs an adaptive hybrid model driven by both process mechanism and production data to predict catalytic process indicators, which helps to improve the reliability of the model prediction results.

[0157] Example 2

[0158] Based on the method for predicting indicators of a dual-riseer catalytic reaction process proposed in Embodiment 1 of this invention, Embodiment 2 of this invention also proposes a system for predicting indicators of a dual-riseer catalytic reaction process. Figure 10 This is a schematic diagram of a dual-elevator catalytic reaction process index prediction system proposed in Embodiment 2 of the present invention; it includes an acquisition module, a first construction module, a second construction module, and a prediction module;

[0159] The acquisition module is used to acquire production process data at different time intervals within at least one complete operating cycle of the target dual-rise catalytic cracking unit;

[0160] The first construction module is used to classify the reactants into components based on the characteristics of the target dual-rise catalytic reaction process, fully consider the characteristics of diesel component conversion reaction during the reaction process to construct a kinetic model of the dual-rise catalytic reaction process, and solve the kinetic parameters of the kinetic model.

[0161] The second building module is used to perform cluster analysis on production process data based on hierarchical agglomerative clustering, divide the sample subsets according to the clustering results, and train the sub-model based on the spatiotemporal attention long short-term memory network according to the sample subsets to build a data-driven model.

[0162] The prediction module is used to construct a hybrid model based on the similarity between the input sample and the training sample, combined with the Latin hypercube sampling method to adaptively adjust the weights of the kinetic model and the data-driven model, to predict the indicators of the catalytic reaction process; and to verify the prediction results of the hybrid model.

[0163] The acquisition module includes production process data such as: feed and product properties of heavy oil reactor, feed and product properties of hydrotreated diesel reactor, catalyst properties, and operating conditions.

[0164] In the first building module: based on the characteristics of the dual-riseer catalytic reaction process for reducing the diesel-gasoline ratio using a combination process that produces more isoalkanes and more high-octane gasoline or light aromatics, the reactants are divided into components.

[0165] The reaction kinetics model specifically includes:

[0166] The reaction rate equation is expressed as:

[0167] Where C is the mass concentration vector of the material components during the reaction; e is the dimensionless distance; K is the reaction rate constant matrix; ρ is the gas density; S WH is the true weight hourly space velocity; Z represents the correction term for the reaction rate.

[0168] A genetic algorithm is used to optimize and solve the kinetic parameters in the reaction kinetic model.

[0169] During the execution of the second construction module, the process of performing cluster analysis on production process data based on hierarchical agglomerative clustering includes:

[0170] The acquired production process data is extracted using time windows and divided into multiple subsequences;

[0171] Hierarchical agglomerative clustering is performed on the subsequence data. One subsequence is selected as the cluster set point to establish the first-level cluster, represented as: D={d1,...d2,...,d n};

[0172] Calculate the distance between each pair of subsequences in the first-level cluster sequentially (d) i ,d j Distance is used to obtain the similarity between each subsequence:

[0173]

[0174] Where, d i For the i-th subsequence; d j Let be the j-th subsequence; p is the dimension of the sequence data; d ip d represents the p-dimensional data in the i-th subsequence; jp This represents the p-dimensional data in the j-th subsequence;

[0175] Clustering the two shortest subsequence pairs as the next layer clusters, integrating features, and building a higher-level cluster;

[0176] The clustering process is repeated, and the classification threshold is selected based on the minimum loss function to complete the cluster analysis.

[0177] Based on a subset of samples, a sub-model is trained using a spatiotemporal attention long short-term memory network to construct a data-driven model for the production process. An encoder and decoder structure is used to construct a sequence-to-sequence data-driven model. A spatial attention mechanism is incorporated into the encoder, assigning different spatial attention values ​​to the input variables. Spatial attention-weighted sample data is used as the input to the encoder's LSTM. The input and output of the encoder's LSTM unit are determined. A temporal attention mechanism is introduced into the decoder structure: a temporal attention value is assigned to each encoder hidden state. The encoder's intermediate states and the decoder's hidden states are updated, and the final output is predicted. The hyperparameters of the spatiotemporal attention long short-term memory network model are dynamically optimized: the optimal parameter combination is found for each sub-model after pattern classification using a grid search method, and a data-driven model is constructed under the optimal hyperparameter combination.

[0178] In the prediction module, the similarity calculation process between the input sample and the training sample is as follows:

[0179] Distance between samples m The calculation method is as follows:

[0180]

[0181] x i =[x1,x2,...,x n [X] is the input sample. train Let x represent the training sample set. train_min ∈X train It is the input sample in the training sample set that is closest to the current input sample, where m is a constant that controls the calculation of the distance between the given input and the sample, 1≤m≤N. train N train This represents the number of samples in the training sample set.

[0182] The weights w of the dynamic model are measured based on the similarity between the input sample and the training sample, and the similarity between the virtual input obtained by Latin hypercube sampling and the training sample. M :

[0183]

[0184] in, This refers to virtual input collected within a fixed upper and lower sampling limit; Represents the maximum possible distance between the virtual input and the training dataset. m distance;

[0185] μ(X)+3σ(X) is used as the upper limit of sampling, and μ(X)-3σ(X) is used as the lower limit of sampling, where μ(X) and σ(X) represent the mean and standard deviation of the sample set, respectively.

[0186] The weights of the data-driven model are: w D =1-w M ;

[0187] Predicting catalytic reaction process parameters y pred =w M ×y pred_mechanism +w D ×y pred_data-driven ;

[0188] Among them, y pred For the predicted output of the mixture model, y pred_mechanism ,y pred_data-driven These represent the predicted outputs of the dynamic model and the data-driven model, respectively.

[0189] When validating the prediction results of the hybrid model, the performance of the dynamic model, the data-driven model, and the hybrid model are considered comprehensively. This is achieved by evaluating the mean squared error (MSE) and correlation coefficient (R²) of the prediction results on the test dataset. 2 The correlation coefficient ρ with consistency c Three metrics are used to evaluate model performance;

[0190] The smaller the MSE value, the smaller the sum of squared prediction errors of the mixture model on the prediction sample set, and the higher the accuracy of the mixture model; R 2 The closer the value of ρ is to 1, the higher the accuracy of the mixture model; c The larger the value, the greater the correlation, and the higher the reliability of the prediction results of the hybrid model.

[0191] Embodiment 2 of this invention proposes a prediction system for catalytic reaction process indicators using a dual-elevator model. This system integrates the characteristics of a kinetic model and a data-driven model, and adaptively adjusts the weights of the kinetic model and the data-driven model based on the similarity between the actual input samples and the training samples. This constructs an adaptive hybrid model driven by both process mechanism and production data to predict catalytic process indicators, which helps to improve the reliability of the model prediction results.

[0192] The description of the relevant part of the dual riser catalytic reaction process index prediction system provided in Embodiment 2 of this application can be found in the detailed description of the corresponding part of the dual riser catalytic reaction process index prediction method provided in Embodiment 1 of this application, and will not be repeated here.

[0193] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that the elements inherent in a process, method, article, or apparatus that includes a list of elements are included. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. Additionally, portions of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of corresponding technical solutions in the prior art have not been described in detail to avoid excessive elaboration.

[0194] While specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art can make other modifications or variations based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for predicting indicators of a dual-elevator catalytic reaction process, characterized in that, Includes the following steps: Acquire production process data at different time intervals within at least one complete operating cycle of the target dual-rise catalytic cracking unit; Based on the characteristics of the target dual-rise catalytic reaction process, the reactants are divided into components. A kinetic model of the dual-rise catalytic reaction process is constructed, taking into full account the characteristics of diesel component conversion during the reaction. The kinetic parameters of the kinetic model are then solved. Cluster analysis of production process data is performed based on hierarchical agglomerative clustering. Based on the clustering results, sample subsets are divided. Sub-models are trained based on spatiotemporal attention long short-term memory networks based on the sample subsets to construct a data-driven model. Based on the similarity between input samples and training samples, a hybrid model is constructed by adaptively adjusting the weights of the kinetic model and the data-driven model using the Latin hypercube sampling method to predict the indicators of the catalytic reaction process. And to validate the prediction results of the hybrid model; The process of calculating the similarity between the input sample and the training sample is as follows: Distance between samples The calculation method is as follows: ； For the input sample, Represents the training sample set, It is the input sample in the training sample set that is closest to the current input sample, and m is a constant that controls the calculation of the distance between the given input and the sample. ; This represents the number of samples in the training sample set; The process of adaptively adjusting the weights of the kinetic model and the data-driven model using the Latin hypercube sampling method to construct a hybrid model for predicting catalytic reaction process indicators is as follows: The weights of the dynamic model are measured based on the similarity between the input samples and the training samples, and the similarity between the virtual input obtained from Latin hypercube sampling and the training samples. : in, This refers to virtual input collected within a fixed upper and lower sampling limit; The maximum possible difference between the virtual input and the training dataset distance; use This is the upper limit of sampling. The lower limit of sampling is , where These represent the mean and standard deviation of the sample set, respectively. The weights of the data-driven model are: ; Predicting catalytic reaction process parameters in, This is the predicted output of the hybrid model. These represent the predicted outputs of the dynamic model and the data-driven model, respectively.

2. The method for predicting indicators of a dual-riseer catalytic reaction process according to claim 1, characterized in that, The production process data includes: feed and product properties of the heavy oil reactor, feed and product properties of the hydrotreated diesel reactor, catalyst properties, and operating conditions.

3. The method for predicting indicators of a dual-riseer catalytic reaction process according to claim 1, characterized in that, The process of classifying the reactants into components based on the characteristics of the target dual-riseer catalytic reaction process includes: classifying the reactants into components based on the characteristics of the dual-riseer catalytic reaction process that combines the production of more isoalkanes and more high-octane gasoline or light aromatics to reduce the diesel-gasoline ratio.

4. The method for predicting indicators of a dual-riseer catalytic reaction process according to claim 1, characterized in that, The reaction kinetics model specifically includes: The reaction rate equation is expressed as: ; in, This represents the mass concentration vector of the material components during the reaction process; The distance is dimensionless. This is the reaction rate constant matrix; The density of the gas; This represents the actual time-space velocity. A correction term representing the reaction rate.

5. The method for predicting indicators of a dual-riseer catalytic reaction process according to claim 1, characterized in that, The process of performing cluster analysis on production process data based on hierarchical agglomerative clustering includes: The acquired production process data is extracted using time windows and divided into multiple subsequences; Hierarchical agglomerative clustering is performed on the subsequence data. One subsequence is selected as the cluster focal point to establish the first-level cluster, which is represented as follows: ; Calculate the relationship between every two subsequence pairs in the first-level cluster sequentially. Distance, to obtain the similarity between each subsequence: ； in, For the first Subsequences; For the first Subsequences; The dimension of the sequence data; Representing the In subsequences Dimensional data; Representing the In subsequences Dimensional data; Clustering the two shortest subsequence pairs as the next layer clusters, integrating features, and building a higher-level cluster; The clustering process is repeated, and the classification threshold is selected based on the minimum loss function to complete the cluster analysis.

6. The method for predicting indicators of a dual-riseer catalytic reaction process according to claim 1, characterized in that, The process data driving model is constructed by training a sub-model based on a spatiotemporal attention long short-term memory network using a subset of samples. The encoder and decoder architecture is used to implement sequence-to-sequence data-driven model construction; The encoder incorporates a spatial attention mechanism, assigning different spatial attention values ​​to the input variables, and using spatial attention-weighted sample data as input to the encoder's LSTM. Determine the inputs and outputs of the encoder LSTM unit; Introduce a temporal attention mechanism into the decoder structure: assign a temporal attention value to each encoder hidden state; Update the encoder intermediate state and decoder hidden state and predict the final output; The hyperparameters of the spatiotemporal attention long short-term memory network model are dynamically optimized: the optimal parameter combination is found for each sub-model after pattern classification using a grid search method, and a data-driven model is constructed under the optimal hyperparameter combination.

7. The method for predicting indicators of a dual-riseer catalytic reaction process according to claim 1, characterized in that, When validating the prediction results of the hybrid model, the performance of the dynamic model, the data-driven model, and the hybrid model are comprehensively considered. This is achieved by evaluating the mean squared error (MSE) and correlation coefficient of the prediction results on the test dataset. Correlation coefficient with consistency Three metrics are used to evaluate model performance; The smaller the MSE value, the smaller the sum of squared prediction errors of the mixture model on the prediction sample set, and the higher the accuracy of the mixture model. The closer the value is to 1, the higher the accuracy of the mixture model; The larger the value, the greater the correlation, and the higher the reliability of the prediction results of the hybrid model.

8. A dual-rise tube catalytic reaction process index prediction system, used to execute the dual-rise tube catalytic reaction process index prediction method according to any one of claims 1 to 7, characterized in that, It includes an acquisition module, a first construction module, a second construction module, and a prediction module; The acquisition module is used to acquire production process data at different time intervals within at least one complete operating cycle of the target dual riser catalytic cracking unit; The first construction module is used to classify the reactants into components according to the characteristics of the target dual-rise catalytic reaction process, fully consider the diesel component conversion reaction characteristics during the reaction process to construct a kinetic model of the dual-rise catalytic reaction process, and solve the kinetic parameters of the kinetic model; The second building module is used to perform cluster analysis on production process data based on hierarchical agglomerative clustering, divide the sample subsets according to the clustering results, and train the sub-model based on the spatiotemporal attention long short-term memory network according to the sample subsets to build a data-driven model; The prediction module is used to construct a hybrid model based on the similarity between the input sample and the training sample, combined with the Latin hypercube sampling method to adaptively adjust the weights of the kinetic model and the data-driven model, and to predict the indicators of the catalytic reaction process. And to verify the prediction results of the hybrid model.

Images