Product prediction method and system for boiling bed residual oil hydrogenation device, and storage medium
Through the mechanism-data joint model and adaptive weight parallel technology, combined with catalyst deactivation and hydrogen partial pressure factors, the product parameters of the boiling bed residual oil hydrogen refueling device are predicted, which solves the problem of insufficient prediction accuracy in the existing technology and realizes the More accurate product properties and traffic forecasts.
Patent Information
- Application Number
- CN202410957574.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-17
- Publication Date
- 2024-11-08
AI Technical Summary
The prior art is difficult to accurately predict the product parameters of the boiling bed residual oil hydrogenation device, and the model fails to effectively reflect factors such as catalyst wear and hydrogen partial pressure, which affects the accuracy of the prediction.
Using the mechanism-data joint model, combined with the mechanism model and data model of parallel adaptive weights, through cluster analysis and CNN-BiLSTM-Attention model, the product properties and flow rate of the boiling bed residual oil hydrogenation device are predicted, and the catalyst deactivation is considered. Factors such as wear and partial hydrogen pressure.
The accuracy of the product parameter prediction of boiling bed residual oil hydrogenation device is improved, the adaptability and robustness of the model are enhanced, and the actual industrial production conditions can be better reflected.
Smart Images

Figure CN118918966A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of industrial device modeling, and in particular to a product prediction method for an ebullated bed residue oil hydrogenation device, a product prediction system for an ebullated bed residue oil hydrogenation device, and a computer-readable storage medium. Background Art
[0002] Residue hydrotreating is a key process in the oil refining industry, which involves reacting residue oil with hydrogen under high temperature and pressure conditions to produce high-quality fuel oil. The process involves multiphase flow, heat transfer, material transfer, and complex chemical reactions, which interact with each other, making the process modeling and optimization extremely complex.
[0003] In actual operation, the performance of the unit is affected by many factors, including catalyst wear, hydrogen partial pressure, etc. These factors must be reflected in the model to ensure that the model can accurately reflect the actual industrial production situation. With the increasingly stringent environmental regulations and the growing market demand for high-quality fuels, higher requirements are placed on the optimization and control of the ebullated bed residue oil hydrotreating process.
[0004] In order to overcome the above-mentioned defects of the prior art, there is an urgent need in the art for a product prediction technology for an ebullated bed residue oil hydroprocessing unit, so as to achieve the prediction of product parameters of an ebullated bed residue oil hydroprocessing unit by establishing a more accurate industrial model. Summary of the invention
[0005] A brief summary of one or more aspects is given below to provide a basic understanding of these aspects. This summary is not an exhaustive overview of all conceived aspects, and is neither intended to identify the key or decisive elements of all aspects nor to define the scope of any or all aspects. Its only purpose is to give some concepts of one or more aspects in a simplified form as a prelude to a more detailed description that will be given later.
[0006] In order to overcome the above-mentioned defects of the prior art, the present invention provides a product prediction method for an ebullated bed residue oil hydrogenation unit, a product prediction system for an ebullated bed residue oil hydrogenation unit, and a computer-readable storage medium, so as to achieve the prediction of product parameters of an ebullated bed residue oil hydrogenation unit by establishing a more accurate industrial model.
[0007] Specifically, the product prediction method for an ebullated bed residue oil hydroprocessing unit provided according to the first aspect of the present invention comprises the following steps: obtaining raw material data and operating condition data of a sample to be tested of the ebullated bed residue oil hydroprocessing unit; and inputting the raw material data and operating condition data of the sample to be tested into a pre-constructed and trained mechanism-data joint model, so as to predict the output product properties and / or output product flow of the ebullated bed residue oil hydroprocessing unit according to the first product properties and / or the first product flow output by the mechanism model and the data model in parallel with adaptive weights in the mechanism-data joint model, and the second product properties and / or the second product flow.
[0008] Further, in some embodiments of the present invention, the step of predicting the output product property and / or output product flow of the ebullated bed residue oil hydroprocessing unit according to the first product property and / or the first product flow rate, and the second product property and / or the second product flow rate output by the mechanism model and the data model in the mechanism-data joint model with adaptive weights in parallel includes: inputting the raw material data and operating condition data of the sample to be tested into a pre-constructed and trained mechanism model to determine the first product property and / or the first product flow rate via the mechanism model, wherein the mechanism model is a multi-set reaction kinetic model constructed based on multiple operating condition data sets of the ebullated bed residue oil hydroprocessing unit. ; Inputting the raw material data and operating condition data of the sample to be tested into a pre-constructed and trained data model to determine the second product property and / or the second product flow rate via the data model, wherein the data model is a CNN-BiLSTM-Attention prediction model constructed and trained based on a CNN model, a bidirectional LSTM model, and an Attention model; and performing a weighted summation of the first product property and the second product property, and / or a weighted summation of the first product flow rate and the second product flow rate according to pre-fitted adaptive weights, to predict the output product property and / or output product flow rate of the boiling bed residue oil hydrogenation unit.
[0009] Furthermore, in some embodiments of the present invention, the step of constructing and training the mechanism model includes: obtaining multiple groups of historical sample data of the boiling bed residue oil hydrogenation unit, wherein each group of the historical sample data respectively includes raw material data, operating condition data and product data of the corresponding historical sample; clustering each of the historical sample data with respect to the residue oil hydrogenation reaction mechanism and product distribution characteristics to divide it into multiple sub-operating condition data sets; and constructing a corresponding reaction kinetic model according to each of the divided sub-operating condition data sets as the mechanism model.
[0010] Furthermore, in some embodiments of the present invention, before performing cluster analysis on each of the historical sample data regarding the residue hydrogenation reaction mechanism and product distribution characteristics, the product prediction method also includes the following steps: removing outliers in each group of the historical sample data via a k-nearest neighbor algorithm; and / or removing noise in each group of the historical sample data via an SG filtering method.
[0011] Furthermore, in some embodiments of the present invention, the step of clustering each of the historical sample data regarding the residue hydrogenation reaction mechanism and product distribution characteristics to divide it into multiple sub-operating condition data sets includes: performing kernel principal component analysis on each group of the historical sample data, projecting each group of the historical sample data in the input space to a high-dimensional feature space through a nonlinear mapping; and in the high-dimensional feature space, performing principal component analysis on the projected data of each group of the historical sample data, and clustering via fuzzy C-means to divide and obtain the multiple sub-operating condition data sets.
[0012] Further, in some embodiments of the present invention, the step of performing clustering via fuzzy C-means to divide and obtain the plurality of sub-operating condition data sets includes: initializing a membership matrix satisfying preset constraints:
[0013]
[0014] Among them, U ij is the membership degree of the jth sample data to the ith category, k is the number of clusters, and n is the number of groups of the historical sample data; according to the membership matrix, the cluster center matrix is calculated:
[0015]
[0016] Among them, m is the exponential weight factor of membership, C i is the cluster center matrix of the i-th category; the sum of squares of the deviations between the predicted values and the actual values of the output product properties and / or output product flow rates is minimized and determined as the objective function of the clustering model:
[0017]
[0018] Wherein, m is the number of groups of historical sample data, is the predicted value of the four components of wax oil, A j (Y) is the actual value of the four components of wax oil, are the predicted values of product flow rates of diesel, naphtha, and low-fraction gas, respectively. i,j (Y)~ C i,j(Y) are the actual values of the product flow rates of diesel, naphtha and low-fraction gas, respectively, and Y is the kinetic parameter of the clustering model; the cluster center matrix C i As input, calculate the value of the objective function:
[0019]
[0020] Among them, U ij is the membership degree of the jth sample data to the i-th category, x j is the k samples closest to the missing point sample, m is the exponential weight factor of the membership, λ j is a regularization parameter, J is the value of the objective function; and the clustering model is repeatedly iterated until the number of iterations reaches a preset number, or the difference between the objective function values of two consecutive iterations is less than a preset threshold.
[0021] Furthermore, in some embodiments of the present invention, the multiple sub-operating condition data sets include RES lumps, ASP lumps, SAT lumps and ARO lumps about the crude oil components, and VGO-ASP lumps, VGO-RES lumps, VGO-ARO lumps, VGO-SAT lumps, DIE lumps, NAP lumps and GAS lumps about the product components. The mechanism model involves 55 reaction paths and 110 kinetic parameters, among which hydrogenation reactions and cracking reactions proceed simultaneously, all reactions are first-order irreversible reactions, condensation reactions are ignored, and piston flow reactions do not consider intra-particle diffusion.
[0022] Furthermore, in some embodiments of the present invention, the mechanism model further includes: a catalyst time-varying deactivation function:
[0023]
[0024] Among them, θ(t c ) is the catalyst deactivation function, t c is the catalyst residence time, β and γ are the deactivation coefficients; catalyst wear function:
[0025]
[0026] Among them, δ(t c ) is the catalyst wear function, a and n are constants related to the catalyst material; carbon deposition deactivation function:
[0027] φ(t c )=exp(-α·ω)
[0028]
[0029] Among them, φ(t c) is the carbon deposition deactivation function, α is the carbon deposition coefficient, ω is the carbon deposition amount, r 0 is the initial carbon deposition rate, ω 0 is the initial carbon deposit amount; and the hydrogen partial pressure index:
[0030] μ=HP λ
[0031] Wherein, μ is the hydrogen partial pressure index, HP is the hydrogen partial pressure, λ is the hydrogen pressure index, and the mechanism model is expressed as:
[0032] Among them, k satisfies the Arrhenius equation: R is the molar gas constant, k 0 and E a are the pre-exponential factor and activation energy respectively, C is the concentration vector, δ(t c ) is the catalyst wear influence function.
[0033] Furthermore, in some embodiments of the present invention, the data model is composed of an input layer, a multi-layer CNN layer, a BiLSTM layer, an attention layer and an output layer, wherein the CNN layer is used for feature extraction, the BiLSTM layer is trained and predicted through hidden layers in two directions, the attention layer assigns different weights to different features to obtain the correlation between each data in the time series and the predicted value, and uses the Softmax function to perform numerical conversion on the attention score, and the output layer is used to convert the output of the attention layer to the output dimension of the output product properties and / or the output product flow.
[0034] Furthermore, in some embodiments of the present invention, the input layer performs a normalization operation on the historical sample data of length T to obtain multi-dimensional input data X=[x 1 ,x 2 ,…x n ], where n is the number of groups of historical sample data, the CNN layer includes a convolution layer, a pooling layer and a fully connected layer, wherein the convolution layer performs a convolution operation through a sliding convolution window to capture the high-level features of the input data and construct a corresponding feature map, the pooling layer performs dimensionality reduction processing on the feature map output by the convolution layer, and performs a pooling operation through the pooling layer to reduce network parameters, the fully connected layer integrates the extracted dimensionality reduction features and outputs them in the form of a vector, and the BiLSTM layer is composed of a forward long short-term memory network and a backward long short-term memory network. The calculation process is as follows:
[0035] A i =f 1 (ω 1 x i +ω 2A i-1 )
[0036] B j =f 2 (ω 3 x i +ω 5 B i+1 )
[0037] Y i =f 3 (ω 4 A i +ω 6 B i )
[0038] Among them, f 1 、f 2 、f 3 is the activation function between the long short-term memory networks in each layer, A is the forward LSTM hidden state, B is the backward LSTM hidden state, ω 1 ~ω 6 is the weight, x i is the output feature of the CNN layer, Y i is the output feature of the BiLSTM layer. The attention layer is used to dynamically weight the output features at different positions to improve the performance of the data model when processing long sequences and to better focus on key information. The output layer weights the output feature S of the attention layer at time t. t As input, and output the output product properties and / or output product flow rate y of the ebullated bed residue oil hydroprocessing unit t :
[0039] y t =sigmoid(w o s t +b o )
[0040] Among them, w o is the weight of the output product properties and / or output product flow, b o A bias for output product properties and / or output product flow.
[0041] Further, in some embodiments of the present invention, the step of fitting the adaptive weights includes: initializing the deviation e in the PID algorithm, and the coefficient K p , K i , K d , and set the initial weights of the mechanism model and the data model M and weight D; Input multiple groups of historical sample data into the pre-built and trained mechanism model and data model to obtain their output results Pre M and Pre D ; According to the output results of the mechanism model and the data model Pre M and Pre D , and the corresponding initial weight weight M and weight D , calculate the output of the mechanism-data joint model:
[0042] result=weight M *Pre M +weight D *Pre D ;as well as
[0043] According to the output result of the mechanism-data joint model and the corresponding product data target in the historical sample data, the deviation e of the current round is determined, and the PID output u of the current round is calculated accordingly, so as to update the weights of the mechanism model and the data model in the next round through the PID algorithm:
[0044] e=target-Pre M *weight M -Pre D *weight D
[0045] u=Kp*e+Ki*∫edt+Kd*d(e)dt
[0046] weight M =weight M +u
[0047] weight D =1-weight M .
[0048] In addition, the product prediction system for the ebullated bed residue oil hydroprocessing unit provided according to the second aspect of the present invention includes: a memory on which computer instructions are stored; and a processor, connected to the memory and configured to execute the computer instructions stored on the memory to implement the product prediction method for the ebullated bed residue oil hydroprocessing unit as described in any one of the first aspect of the present invention.
[0049] In addition, the computer-readable storage medium provided according to the third aspect of the present invention stores computer instructions, which, when executed by a processor, implement the product prediction method for an ebullated bed residue hydroprocessing unit as described in any one of the first aspect of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0050] The above features and advantages of the present invention can be better understood after reading the detailed description of the embodiments of the present disclosure in conjunction with the following drawings. In the drawings, the components are not necessarily drawn to scale, and components with similar related properties or features may have the same or similar reference numerals.
[0051] Figure 1 A process flow chart of an ebullated bed residue oil hydrogenation unit according to some embodiments of the present invention is shown.
[0052] Figure 2 A method flow chart of an ebullated bed residue oil hydrotreating unit based on an adaptive weighted hybrid model according to some embodiments of the present invention is shown.
[0053] Figure 3A to Figure 3D A comparison of data before and after SG filtering according to some embodiments of the present invention is shown.
[0054] Figure 4 Silhouette coefficient graphs for different numbers of clusters provided according to some embodiments of the present invention are shown.
[0055] Figure 5A to Figure 5C A diagram showing the fitting performance of the mechanism model provided according to some embodiments of the present invention on a validation set is shown.
[0056] Figure 6A to Figure 6C A fitting error diagram of a mechanism model provided according to some embodiments of the present invention on a validation set is shown.
[0057] Figure 7 The CNN-BiLSTM-Attention model structure provided according to some embodiments of the present invention is shown.
[0058] Figure 8A to Figure 8E A product prediction mean absolute error graph of a data model provided according to some embodiments of the present invention is shown.
[0059] Fig. 9 A flow chart of an adaptive weighted hybrid model algorithm provided according to some embodiments of the present invention is shown.
[0060] Figures 10A to 10E A product prediction mean absolute error graph of an adaptive weighted hybrid model provided according to some embodiments of the present invention is shown. DETAILED DESCRIPTION
[0061] The following specific embodiments illustrate the implementation of the present invention, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. Although the description of the present invention will be introduced in conjunction with the preferred embodiment, this does not mean that the features of this invention are limited to this implementation. On the contrary, the purpose of introducing the invention in conjunction with the implementation is to cover other options or modifications that may be extended based on the claims of the present invention. In order to provide a deep understanding of the present invention, the following description will include many specific details. The present invention can also be implemented without using these details. In addition, in order to avoid confusion or blurring the focus of the present invention, some specific details will be omitted in the description.
[0062] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.
[0063] In addition, the terms "upper", "lower", "left", "right", "top", "bottom", "horizontal" and "vertical" used in the following description should be understood as the directions shown in the paragraph and the related drawings. Such relative terms are only used for the convenience of description and do not mean that the device described therein must be manufactured or operated in a specific direction, and therefore should not be understood as limiting the present invention.
[0064] It is understood that although the terms "first", "second", "third", etc. may be used herein to describe various components, regions, layers and / or parts, these components, regions, layers and / or parts should not be limited by these terms, and these terms are only used to distinguish different components, regions, layers and / or parts. Therefore, the first component, region, layer and / or part discussed below may be referred to as a second component, region, layer and / or part without departing from some embodiments of the present invention.
[0065] As mentioned above, in actual operation, the performance of the unit is affected by many factors, including catalyst wear, hydrogen partial pressure, etc. These factors must be reflected in the model to ensure that the model can accurately reflect the actual industrial production situation. With the increasingly stringent environmental regulations and the growing market demand for high-quality fuels, higher requirements are placed on the optimization and control of the ebullated bed residue oil hydrotreating process.
[0066] In order to overcome the above-mentioned defects of the prior art, the present invention provides a product prediction method for an ebullated bed residue oil hydrogenation unit, a product prediction system for an ebullated bed residue oil hydrogenation unit, and a computer-readable storage medium, so as to achieve the prediction of product parameters of an ebullated bed residue oil hydrogenation unit by establishing a more accurate industrial model.
[0067] First, please refer to Figure 1 , Figure 1 A process flow chart of an ebullated bed residue oil hydrogenation unit according to some embodiments of the present invention is shown.
[0068] like Figure 1 As shown, Figure 1 The process flow of ebullating bed residue oil hydrogenation in this case is explained. The process raw materials include residue oils of different sources and grades, vacuum residue oil, catalytic cycle oil and hydrogen, which enter the reactor after heating and mixing. The mixed raw materials drive the catalyst to flow from bottom to top in the reactor. Under high temperature and high pressure conditions, the residue oil and hydrogen undergo cracking, hydrogenation, reforming, isomerization and other reactions, involving the interaction of multiple intermediate products and catalysts. The sulfur, nitrogen and heavy components in the residue oil are converted into lighter hydrocarbon compounds. The reacted substances enter the separator, and the products are separated and recovered through processes such as distillation, condensation and extraction. Multiple reaction conditions need to be adjusted during the production process, such as temperature, pressure, hydrogen flow, catalyst type and flow, etc.
[0069] In some non-limiting embodiments, the product prediction method of the ebullated bed residue oil hydrogenation unit provided in the first aspect of the present invention can be implemented based on the product prediction system of the ebullated bed residue oil hydrogenation unit provided in the second aspect of the present invention. Specifically, the product prediction system of the ebullated bed residue oil hydrogenation unit is configured with a memory and a processor. The memory includes but is not limited to the above-mentioned computer-readable storage medium provided in the third aspect of the present invention, on which computer instructions are stored. The processor is connected to the memory and is configured to execute the computer instructions stored on the memory to implement the product prediction method of the ebullated bed residue oil hydrogenation unit provided in the first aspect of the present invention.
[0070] Specifically, the product prediction method of the ebullated bed residue oil hydroprocessing unit can first obtain the raw material data (e.g., raw material properties, raw material flow rate) and operating condition data of the sample to be tested of the ebullated bed residue oil hydroprocessing unit. Then, the raw material data and operating condition data of the sample to be tested are input into a pre-built and trained mechanism-data joint model, so as to predict the output product properties and / or output product flow rate of the ebullated bed residue oil hydroprocessing unit according to the first product properties and / or first product flow rate output by the mechanism model and the data model in parallel with adaptive weights in the mechanism-data joint model, and the second product properties and / or second product flow rate.
[0071] The following will describe the working principle of the product prediction system of the above-mentioned ebullated bed residue oil hydrogenation unit in combination with some embodiments of the product prediction method of the ebullated bed residue oil hydrogenation unit. Those skilled in the art will understand that the embodiments of the product prediction method of these ebullated bed residue oil hydrogenation units are only some non-restrictive implementation methods provided by the present invention, which are intended to clearly demonstrate the main concept of the present invention and provide some specific solutions that are convenient for the public to implement, rather than to limit all functions or all working modes of the product prediction system of the ebullated bed residue oil hydrogenation unit. Similarly, the product prediction method of the ebullated bed residue oil hydrogenation unit is also only some restrictive implementation methods provided by the present invention, and does not constitute a limitation on the execution subject or execution order of each step in the product prediction method of these ebullated bed residue oil hydrogenation units.
[0072] Please refer to Figure 2 And Table 1, Figure 2 The flowchart of the method of the ebullated bed residue oil hydrogenation unit based on the adaptive weighted hybrid model provided according to some embodiments of the present invention is shown. Table 1 shows the production process data.
[0073] Table 1 Production process data
[0074]
[0075] like Figure 2 As shown, the product prediction system may first execute step S1: data acquisition and preprocessing.
[0076] Specifically, the production process data of a refinery's boiling bed residue oil hydrotreating unit is collected. The production process data includes real-time data analysis and testing data.
[0077] As shown in Table 1, real-time data includes operating conditions, product and raw material flow rates; analytical data includes raw material properties and catalyst properties. The acquired factory data is processed. First, the k-nearest neighbor algorithm is used to calculate the distance between each group of data and all groups of data in the data set. The distances of the k groups of data closest to each group of data are selected and averaged. When the average exceeds a certain threshold, it is judged as an outlier. Here, we take k as 50.
[0078] Please refer to Figure 3A to Figure 3D , Figure 3A to Figure 3D A comparison of data before and after SG filtering according to some embodiments of the present invention is shown.
[0079] like Figure 3A to Figure 3DAs shown in the figure, the data after outlier processing still has various noises due to sensor errors, communication interference and other reasons. Filtering can smooth the data, highlight the overall trend of the data, and provide cleaner and more reliable input for subsequent analysis, modeling or optimization. The Savitzky-Golay (SG) filtering algorithm is a method for smoothing data. It is an improvement of the moving average algorithm. It can effectively remove noise and retain the trend of the signal. The parameters that need to be adjusted in the SG filtering algorithm are the window size and the order of the polynomial. The larger the time window, the smoother the curve, resulting in signal distortion. Lower-order polynomials will also make the filtered curve smoother, while higher-order polynomials can better adapt to data changes. The effect of the filtered signal is observed through cross-validation, and then the optimal parameter value is selected based on the smoothing effect and data characteristics. When the window is 9 and the polynomial is 3rd order, the filtered signal can effectively remove noise, and the overall trend of the signal is clear and smooth.
[0080] Please continue to refer to Figure 2 , the prediction system can proceed to step S2: constructing a boiling bed residue oil hydrogenation mechanism model based on the 11 lumped model.
[0081] First, KPCA-FCM is used to divide the working conditions and obtain the data sets of each sub-working condition. KPCA is used for dimensionality reduction: the data in the original space is projected into a high-dimensional feature space through a nonlinear mapping, and then the data processing of principal component analysis is performed in the high-dimensional space. Fuzzy C-means (FCM) is then used for clustering. The clustering algorithm steps are as follows:
[0082] a. Initialize the membership matrix that satisfies the constraint condition formula (1), and calculate the cluster center matrix according to formula (2).
[0083]
[0084] Among them, U ij is the membership degree of the jth sample data to the ith category, j is the number of samples, k is the number of clusters, n is the number of groups of historical sample data, m is the exponential weight factor of the membership degree, C i is the cluster center matrix of the i-th class.
[0085] b. Taking the minimum sum of squares of the deviation between the predicted yield and the actual yield as the objective function, the expression is:
[0086]
[0087] Where m is the number of groups of sample data, is the predicted value of the four components of wax oil, A j (Y) is the corresponding actual value, are the predicted values of diesel, naphtha and low-gas yields, respectively, and B i,j(Y)~C i,j (Y) corresponds to the actual value, and Y is the kinetic parameter of the model.
[0088] c. Take the obtained cluster center matrix as input, take the minimum sum of squares of the deviation between the predicted yield and the actual yield as the objective function, calculate the membership matrix according to formula (4), and calculate the value of the objective function according to formula (5), and repeat this cycle.
[0089]
[0090] Among them, U ij is the membership degree of the jth sample data to the i-th category, x j is the k samples closest to the missing point sample, m is the exponential weight factor of the membership, λ j is the regularization parameter, and J is the value of the objective function.
[0091] Here, due to the failure of the factory DCS system, human error, and other reasons, there will be abnormal values (missing points and outliers) in the data set. By filling the missing points and removing the outliers in the data set, the integrity and high quality of the data can be ensured, thereby improving the credibility of modeling and analysis. K-Nearest Neighbor (KNN) is used here to fill the missing values. This method is an instance-based learning method that uses the similarity between samples to estimate the missing values. For each missing value in the sample, the algorithm uses the Euclidean distance to calculate the closest K neighbors, and uses the average of these k neighbors as the estimated value of the missing point.
[0092] The calculation formula is shown in formula (6).
[0093]
[0094] in, is the missing point sample, x j Yes The most recent k samples, w j is the weight coefficient.
[0095] d. When the number of iterations reaches the set number or the difference between two consecutive iterations is less than the set threshold, the objective function reaches the minimum value and the iteration terminates; otherwise, repeat step c.
[0096] Here, the clustering results use the silhouette coefficient to evaluate the quality of clustering. Its value range is [-1, 1]. The closer it is to 1, the better the clustering effect. Its calculation formula is shown in formula (7):
[0097]
[0098] Among them, i is the sample vector after dimensionality reduction, a(i) is the distance from i to all other points in the cluster to which it belongs, b(i) is the average distance from i to all points in a cluster that does not contain it, and the average of all sample silhouette coefficients is the total silhouette coefficient of the clustering result.
[0099] Please refer to Figure 4 , Figure 4 Silhouette coefficient graphs for different numbers of clusters provided according to some embodiments of the present invention are shown.
[0100] like Figure 4 As shown in the figure, the enhanced data set is normalized and then divided into working conditions, and the number of clusters is set to 2 to 10. When the number of clusters is 2, the boundaries between clusters are clearer and the classification effect is best.
[0101] Furthermore, 11 lumped reaction kinetic models were constructed in combination with the hydrogenation reaction mechanism and taking into account factors such as catalyst wear, precipitate formation, catalyst deactivation and hydrogen partial pressure. After obtaining multiple sub-operating condition data sets, the crude oil was divided into four lumps according to its components: RES, ASP, SAT and ARO. The ebullated bed residue hydrogenation reaction system mainly involves hydrogenation, saturation and cracking reactions among resins (RES), asphaltene (ASP), aromatic fractions (ARO) and saturated fractions (SAT). The hydrogenation reaction of resins and asphaltene produces saturated fractions and aromatic fractions, and the saturation and cracking of aromatic fractions produces saturated fractions.
[0102] From the process flow of residual oil hydrogenation ebullated bed, we can know that the main products of hydrogenation reaction are wax oil (VGO), diesel (DIE), naphtha (NAP), and low-fraction gas (GAS). In order to meet the actual industrial needs, each product is regarded as a lump. Among them, wax oil (VGO) is a heavy component in the product. After fractionation, it is used as the raw material for wax oil hydrogenation. Its properties and composition have a significant impact on the subsequent production process. Therefore, wax oil is also divided into asphaltene (VGO-ASP), colloid (VGO-RES), aromatic fraction (VGO-ARO), and saturated fraction (VGO-SAT) according to its composition. Therefore, the lump division is realized according to the composition, and the raw materials and products are divided into 11 lumps in total.
[0103] At the same time, the following assumptions are considered: a. Hydrogenation and cracking reactions occur simultaneously during the reaction; b. All reactions are first-order irreversible reactions; c. The main reaction that affects the boiling range is the cracking reaction, and the condensation reaction is ignored; d. The reaction is a piston flow reaction, and the diffusion within the particle is not considered.
[0104] According to the above 11-lumped division method and model assumptions, an 11-lumped reaction network of residue hydrogenation can be obtained, which contains a mechanism model of an ebullated bed residue hydrogenation unit with 55 reactions and 110 parameters. In order to improve the accuracy of the model, the effects of catalyst deactivation, catalyst wear, precipitate formation and hydrogen partial pressure on the model are considered when constructing the model.
[0105] Specifically, a. Catalyst deactivation. Due to the adsorption of the catalyst, its activity must gradually decrease as the reaction proceeds, which in turn has a significant impact on the hydrogenation reaction rate. This model introduces a time-varying deactivation function, which is expressed as:
[0106]
[0107] Among them, θ(t c ) is the catalyst deactivation function, t c is the residence time of catalyst, β and γ are the deactivation coefficients.
[0108] b. Catalyst wear. In the fluidized bed process, the raw materials drive the catalyst from bottom to top in the reactor to produce violent back mixing, which will inevitably lead to the loss of catalysts due to collisions between the catalysts and the reactor wall. Therefore, the catalyst wear function is introduced:
[0109]
[0110] Among them, δ(t c ) is the catalyst wear function, and a and n are constants related to the catalyst material.
[0111] c. Sediment generation. In the ebullating bed residue oil hydrogenation process, coke-like substances will be formed. Carbon deposits will affect the operating performance of the unit and cause the catalyst to deactivate rapidly. Therefore, the coke deactivation function is introduced:
[0112] φ(t c )=exp(-α·ω) (10)
[0113]
[0114] Among them, φ(t c ) is the carbon deposition deactivation function, α is the carbon deposition coefficient, ω is the carbon deposition amount, r 0 is the initial carbon deposition rate, ω 0 is the initial carbon deposit amount.
[0115] d. Hydrogenation factor. Hydrogen partial pressure has a great influence on the saturation reaction of unsaturated components. For the hydrogen partial pressure factor, a hydrogen partial pressure index is added, which is expressed as formula (12). Among them, HP is the hydrogen partial pressure and λ is the hydrogen pressure index.
[0116] μ=HP λ (12)
[0117] Wherein, μ is the hydrogen partial pressure index, HP is the hydrogen partial pressure, and λ is the hydrogen pressure index.
[0118] Furthermore, considering the above factors at the same time, the mechanism model is expressed as:
[0119]
[0120] Among them, k satisfies the Arrhenius equation: R is the molar gas constant, k 0 and E a are the pre-exponential factor and activation energy respectively, C is the concentration vector, δ(t c ) is the catalyst wear influence function.
[0121] Please refer to Figure 5A to Figure 5C , Figure 6A to Figure 6C . Figure 5A to Figure 5C A diagram showing the fitting performance of the mechanism model provided according to some embodiments of the present invention on a validation set is shown. Figure 6A to Figure 6C A fitting error diagram of a mechanism model provided according to some embodiments of the present invention on a validation set is shown.
[0122] like Figure 5A to Figure 5C As shown, after data enhancement, the fitting performance of the model on the validation set is obtained. Figure 5A to Figure 5C As shown in Figure 2, the ratio of fitting data to validation data is 1:1, and the fitting data and validation data are randomly selected from the augmented data set.
[0123] like Figure 6A to Figure 6C As shown in the figure, the error of 75% of the validation data of the main products predicted by the sub-model is below 1.5%. Compared with the single model, the prediction performance of the sub-model for each product has been improved, among which the prediction result of wax oil production has been improved most significantly, with MAE reduced by about 0.5%, and the MAE of each main product (except unconverted oil) is lower than 1.1%. The prediction effect of the sub-model under different working conditions is better than that of the single model.
[0124] Please refer to Figure 2 and Figure 7 , Figure 7 The CNN-BiLSTM-attention model structure provided according to some embodiments of the present invention is shown.
[0125] like Figure 2 As shown, the prediction system can continue to execute step S3: constructing a data model of an ebullient bed residue oil hydrotreating unit based on a CNN-BiLSTM-Attention model.
[0126] like Figure 7As shown in the figure, the long short-term memory unit of the LSTM model and the attention mechanism of the Transformer model are combined with the recurrent neural network (CNN), bidirectional LSTM, and attention to establish a CNN-BiLSTM-attention framework for predicting the yield of residual oil hydroprocessing products in the ebullient bed. The model consists of an input layer, a multi-layer CNN, a BiLSTM, an attention layer, and an output layer. The CNN layer is used for feature extraction; the BiLSTM is trained and predicted through hidden layers in two directions; the attention layer assigns different weights to different features, obtains the correlation between the various data in the time series and the predicted value, and uses the Softmax function to convert the attention score into a numerical value; the output layer is a dense layer, which converts the output of the attention layer into the dimension of the predicted yield to obtain the final predicted value.
[0127] The principles of each layer of the CNN-BiLSTM-Attention model are as follows:
[0128] a. Input layer: Normalize the data with a length of T after SG filtering, expressed as: X = [x 1 ,x 2 ,...x n ].
[0129] b. CNN layer: CNN contains convolutional layers, pooling layers, and fully connected layers. The convolutional layer is its core, which performs inner product operations by sliding convolution windows to capture high-level features of input data. The convolution operation can be expressed as:
[0130]
[0131] The pooling layer reduces the dimension of the feature map output by the convolutional layer, reduces the network parameters through the pooling operation, thereby simplifying the calculation process and enhancing the extrapolation ability of the model. Pooling selects the maximum pooling method:
[0132] P 1 =max(C 1 )+b 2 (15)
[0133] The fully connected layer integrates the extracted features and outputs them in the form of vectors:
[0134] X c =Sigmoid(P 1 W 2 +b 3 ) (16)
[0135] Among them, W 1 , W 2 is the weight, b 1 , b 2, b 3 For deviation.
[0136] c. BiLSTM layer: This layer is a recurrent neural network model composed of forward and backward long short-term memory networks. It not only considers the positive information of historical data in the time dimension, but also the negative information. This means that when processing sequence data, BiLSTM is able to utilize both past and future information. In this way, the model can understand the sequence data more comprehensively and better capture its long-term dependencies. The calculation process of BiLSTM is as follows:
[0137]
[0138] Among them, f 1 、f 2 、f 3 is the activation function between each layer of LSTM network, A is the forward LSTM hidden state, B is the backward LSTM hidden state, ω 1 ~ω 6 is the weight, x i is the output feature of the CNN layer, Y i is the output feature of the BiLSTM layer.
[0139] d. Attention layer: It is used to dynamically weight features at different positions so that the model can focus more on the important parts of the sequence. This helps improve the performance of the model when processing long sequences and enables the model to better focus on key information. The attention transfer process has been introduced in the Transformer model.
[0140] e. Output layer: The output s of the attention layer at time t t As input, the output layer calculates the expression:
[0141] y t =sigmoid(w o s t +b o ) (18)
[0142] Among them, w o is the weight of the output product properties and / or output product flow, b o A bias for output product properties and / or output product flow.
[0143] The experimental model uses a two-layer CNN, the convolution kernel size of both layers is 3*3, the size of the pooling window is 2, and the number of filters in the convolution layer is 64 and 32 respectively; the number of hidden layers of the two BiLSTM layers is consistent with the above LSTM model, which is 100 and 50 respectively; the dimension of the attention layer is 128, and the dimension of the key is 32. The optimizer is Adam, the number of data (samples) passed to the program for training at a time (batch_size) is 32, the total number of training rounds is 100, the time step is 5, and the model is trained using the mean square error loss function and the Adam optimizer.
[0144] Please refer to Table 2, which shows the prediction performance of the model on the same dataset based on the above parameter settings.
[0145] Table 2 CNN-BiLSTM-Attention prediction results
[0146]
[0147] As shown in the validation data set in Table 2, compared with the Transformer model, this model makes up for the defect that the Transformer model has a particularly poor diesel yield prediction effect. The R 2 The average absolute error has been reduced by about half from 0.234 to 0.839, and the prediction effects of other products have also been improved to a certain extent. Compared with the LSTM model, this model has greatly improved the accuracy of the LSTM model for unconverted oil and naphtha, with the average absolute error reduced by 20% and 89% respectively.
[0148] Please refer to Figure 8A to Figure 8E , Figure 8A to Figure 8E A product prediction mean absolute error graph of a data model provided according to some embodiments of the present invention is shown.
[0149] like Figure 8A to Figure 8E As shown in the figure, the model's improvement effect can be clearly seen through the average absolute error of the product prediction of this model and all the above data models, and the error of all products of the CNN-BiLSTM-Attention model is within 0.8%, which can be applied to the actual process. Using Attention to assign different weights to the features generated by the BiLSTM hidden layer allows the network to be trained more effectively, solving the problem of ignoring important feature information of data at some moments, thereby effectively improving the model's prediction accuracy, indicating that it has good applicability in predicting the product yield of the boiling bed residue oil hydrotreating unit.
[0150] Please refer to Figure 2 and Fig. 9 , Fig. 9A flow chart of an adaptive weighted hybrid model algorithm provided according to some embodiments of the present invention is shown.
[0151] like Figure 2 As shown, the prediction system can continue to execute step S4: fusing the mechanism model and the data model based on adaptive weights.
[0152] like Fig. 9 As shown in the figure, the prediction system uses multiple methods such as feedback control, machine learning algorithm, evolutionary algorithm, model predictive control, etc., and uses PID algorithm as the adaptive weight implementation method to realize the modeling of the adaptive weight parallel model. The main process is as follows:
[0153] a. Initialization: Set the deviation e and coefficient K in the PID algorithm p , K i , K d , set the initial weights of the mechanism model and the data model M and weight D .
[0154] b. Model calculation: The preprocessed data is sent to the model for calculation to obtain the mechanism and data model results Pre M and Pre D , and according to the output results of the mechanism model and data model Pre M and Pre D , and the corresponding initial weight weight M and weight D .
[0155] c. Fusion model output: Output results of the computational-data joint model:
[0156] result=weight M *Pre M +weight D *Pre D (19)
[0157] d. Adaptive weight update: According to the output of the mechanism-data joint model and the corresponding product data target in the historical sample data, the deviation e of the current round is determined, and the PID output u of the current round is calculated accordingly, so as to update the weights of the mechanism model and the data model in the next round through the PID algorithm:
[0158] e=target-Pre M *weight M -Pre D *weight D
[0159] u=Kp*e+Ki*∫edt+Kd*d(e)dt
[0160] weight M =weight M +u
[0161] weight D =1-weight M (20)
[0162] Among them, target is the analysis and test data.
[0163] In addition, based on the product prediction system of the ebullating bed residue oil hydrotreating unit, those skilled in the art also verified the method using an adaptive weight parallel model (AWM).
[0164] Specifically, the adaptive weighted parallel model (AWM) is a method that combines the prediction results of multiple models to improve the prediction accuracy by dynamically adjusting the weights of different models. The main idea is to automatically adjust the weights of each model according to the performance of the model in the current situation, so that the prediction results of the combined model are closer to the actual situation. The mechanism and data parallel model can achieve a better balance in the description of system behavior. The adaptive weighted mechanism model can provide an understanding and explanation of the internal mechanism of the system, while the data model can capture the complexity or difficult-to-model parts that the mechanism model cannot fully cover. By connecting these two models in parallel, their respective advantages can be utilized.
[0165] Please refer to Figure 8A to Figure 8E , Figures 10A to 10E And Table 3, Figures 10A to 10E The product prediction mean absolute error diagram of the adaptive weighted hybrid model provided according to some embodiments of the present invention is shown. Table 3 shows the prediction performance of the data model, the mechanism model and the hybrid model on the same data set.
[0166] like Figure 8A to Figure 8E and Figures 10A to 10E , we can clearly see the improvement effect of the fusion model. As shown in the validation data set in Table 3, except for the gas yield, the prediction accuracy of the other four product yields of the fusion model is better than that of the data model and the mechanism model. The gas yield originally accounts for a small proportion, so it can be verified that the product yield prediction of the fusion model has a higher accuracy.
[0167] Table 3 Verification results of data model, mechanism model and fusion model
[0168] Model predictors Data Model Mechanistic model Fusion Model Wax oil yield 0.809 1.028 0.715 Unconverted oil yield 0.645 1.26 0.451 Diesel yield 0.29 0.721 0.258 Naphtha yield 0.372 0.429 0.15 Gas yield 0.057 0.523 0.136
[0169] In summary, the adaptive weighted parallel model is an advanced modeling method that improves the accuracy of prediction by combining the prediction results of multiple models and dynamically adjusting their weights. The core idea of this method is to automatically adjust the weights of each model according to its performance in the current situation, so that the prediction results of the combined model are closer to the actual situation. In the ebullated bed residue oil hydrogenation process, the model based on physical and chemical mechanisms can be used in parallel with the data-based model. The mechanism model can provide an understanding and explanation of the internal mechanism of the system, while the data model can capture the complexity or difficult-to-model parts that the mechanism model cannot fully cover. Through this parallel connection, the advantages of the two models can be fully utilized to achieve a more comprehensive and accurate process description. Based on the industrial measured data of a certain factory's ebullated bed hydrogenation unit, the present invention adopts an adaptive weighted parallel model to successfully overcome the difficulties in traditional modeling, and verifies the accuracy of the model through factory data. This provides a new and effective method for the optimization and control of the ebullated bed residue oil hydrogenation process, and realizes the prediction of product parameters of the ebullated bed residue oil hydrogenation unit.
[0170] Although the above methods are illustrated and described as a series of actions for simplicity of explanation, it should be understood and appreciated that these methods are not limited by the order of the actions, because according to one or more embodiments, some actions may occur in a different order and / or concurrently with other actions from those illustrated and described herein or not illustrated and described herein but understandable to those skilled in the art.
[0171] The previous description of the disclosure is provided to enable any person skilled in the art to make or use the disclosure. Various modifications to the disclosure will be apparent to those skilled in the art, and the general principles defined herein may be applied to other variations without departing from the spirit or scope of the disclosure. Thus, the disclosure is not intended to be limited to the examples and designs described herein, but should be granted the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A product prediction method for an ebullated bed residue oil hydrogenation unit, characterized in that: The following steps are involved: Obtaining raw material data and operating condition data of a sample to be tested of the ebullated bed residue oil hydrogenation device; as well as The raw material data and operating condition data of the sample to be tested are input into a pre-constructed and trained mechanism-data joint model to predict the output product properties and / or output product flow of the boiling bed residue oil hydrotreating unit based on the first product properties and / or first product flow, and the second product properties and / or second product flow output by the mechanism model and the data model in parallel with adaptive weights in the mechanism-data joint model.
2. The product prediction method according to claim 1, characterized in that: The step of predicting the output product properties and / or output product flow rate of the ebullated bed residue oil hydrotreating unit according to the first product properties and / or the first product flow rate, and the second product properties and / or the second product flow rate output by the mechanism model and the data model in parallel with the adaptive weight in the mechanism-data joint model comprises: Inputting the raw material data and operating condition data of the sample to be tested into a pre-built and trained mechanism model to determine the first product property and / or the first product flow rate via the mechanism model, wherein the mechanism model is a multi-set total reaction kinetic model constructed based on multiple operating condition data sets of the ebullated bed residue oil hydrotreating unit; Inputting the raw material data and operating condition data of the sample to be tested into a pre-built and trained data model to determine the second product property and / or the second product flow rate via the data model, wherein the data model is a CNN-BiLSTM-Attention prediction model built and trained based on a CNN model, a bidirectional LSTM model, and an Attention model; and According to the pre-fitted adaptive weights, the first product property and the second product property are weighted summed, and / or the first product flow rate and the second product flow rate are weighted summed to predict the output product property and / or output product flow rate of the boiling bed residue oil hydroprocessing unit.
3. The product prediction method according to claim 2, characterized in that: The steps of constructing and training the mechanism model include: Acquire multiple groups of historical sample data of the ebullated bed residue oil hydrogenation unit, wherein each group of the historical sample data includes raw material data, operating condition data and product data of the corresponding historical sample; Clustering each of the historical sample data with respect to the residue hydrogenation reaction mechanism and product distribution characteristics to divide it into a plurality of sub-operating condition data sets; and According to the divided sub-operating condition data sets, a corresponding reaction kinetic model is constructed to serve as the mechanism model.
4. The product prediction method according to claim 3, characterized in that: Before performing cluster analysis on the residual oil hydrogenation reaction mechanism and product distribution characteristics on each of the historical sample data, the product prediction method further comprises the following steps: Removing outliers from each group of historical sample data using a k-nearest neighbor algorithm; and / or The noise in each group of historical sample data is removed through the SG filtering method.
5. The product prediction method according to claim 3, characterized in that: The step of clustering each of the historical sample data with respect to the residual oil hydrogenation reaction mechanism and product distribution characteristics to divide it into a plurality of sub-operating condition data sets comprises: Performing kernel principal component analysis on each group of the historical sample data, projecting each group of the historical sample data in the input space into a high-dimensional feature space through a nonlinear mapping; and In the high-dimensional feature space, principal component analysis is performed on the projection data of each group of the historical sample data, and clustering is performed via fuzzy C-means to divide and obtain the multiple sub-operating condition data sets.
6. The product prediction method according to claim 5, characterized in that: The step of performing clustering by fuzzy C-means to divide and obtain the plurality of sub-operating condition data sets comprises: Initialize the membership matrix that satisfies the preset constraints: Among them, U ij is the membership degree of the jth sample data to the ith category, k is the number of clusters, and n is the number of groups of the historical sample data; According to the membership matrix, the cluster center matrix is calculated: Among them, m is the exponential weight factor of membership, C i is the cluster center matrix of the i-th category; The objective function of the clustering model is to minimize the sum of squares of the deviations between the predicted values of the output product properties and / or the output product flow and the actual values: Wherein, m is the number of groups of historical sample data, is the predicted value of the four components of wax oil, A j (Y) is the actual value of the four components of wax oil, are the predicted values of product flow rates of diesel, naphtha, and low-fraction gas, respectively. i,j (Y)~ C i,j (Y) are the actual values of the product flow rates of diesel, naphtha, and low-fraction gas, respectively, and Y is the kinetic parameter of the clustering model; Each cluster center matrix C i As input, calculate the value of the objective function: Among them, U ij is the membership degree of the jth sample data to the i-th category, x j is the k samples closest to the missing point sample, m is the exponential weight factor of the membership, λ j is the regularization parameter, J is the value of the objective function; and The clustering model is iterated repeatedly until the number of iterations reaches a preset number, or the difference between the objective function values of two consecutive iterations is less than a preset threshold.
7. The product prediction method according to claim 3, characterized in that: The plurality of sub-operating condition data sets include RES lumps, ASP lumps, SAT lumps and ARO lumps about feedstock components, and VGO-ASP lumps, VGO-RES lumps, VGO-ARO lumps, VGO-SAT lumps, DIE lumps, NAP lumps and GAS lumps about product components, The mechanism model involves 55 reaction paths and 110 kinetic parameters, in which hydrogenation reaction and cracking reaction proceed simultaneously, all reactions are first-order irreversible reactions, condensation reactions are ignored, and the piston flow reaction does not consider the diffusion within the particle.
8. The product prediction method according to claim 7, characterized in that: The mechanism model further includes: Time-varying deactivation function of the catalyst: Among them, θ(t c ) is the catalyst deactivation function, t c is the catalyst residence time, β and γ are the deactivation coefficients; Catalyst wear function: Among them, δ(t c ) is the catalyst wear function, and a and n are constants related to the catalyst material; Carbon deposition deactivation function: φ(t c )=exp(-α·ω) Among them, φ(t c ) is the carbon deposition deactivation function, α is the carbon deposition coefficient, ω is the carbon deposition amount, r0 is the initial carbon deposition rate, ω0 is the initial carbon deposition amount; and Hydrogen partial pressure index: μ=HP λ Where, μ is the hydrogen partial pressure index, HP is the hydrogen partial pressure, and λ is the hydrogen pressure index. The mechanistic model is expressed as: Among them, k satisfies the Arrhenius equation: R is the molar gas constant, k0 and E a are the pre-exponential factor and activation energy respectively, C is the concentration vector, δ(t c ) is the catalyst wear influence function.
9. The product prediction method according to claim 2, characterized in that: The data model consists of an input layer, a multi-layer CNN layer, a BiLSTM layer, an attention layer and an output layer, wherein the CNN layer is used for feature extraction, the BiLSTM layer is trained and predicted through hidden layers in two directions, the attention layer assigns different weights to different features to obtain the correlation between each data in the time series and the predicted value, and uses the Softmax function to perform numerical conversion on the attention score, and the output layer is used to convert the output of the attention layer to the output dimension of the output product properties and / or the output product flow.
10. The product prediction method according to claim 9, characterized in that: The input layer performs normalization operation on the historical sample data of length T to obtain multi-dimensional input data X = [x1, x2, ... x n ], where n is the number of groups of historical sample data, The CNN layer includes a convolution layer, a pooling layer and a fully connected layer, wherein the convolution layer performs a convolution operation through a sliding convolution window to capture the high-level features of the input data and construct a corresponding feature map, the pooling layer performs dimensionality reduction processing on the feature map output by the convolution layer, and performs a pooling operation through the pooling layer to reduce network parameters, and the fully connected layer integrates the extracted dimensionality reduction features and outputs them in the form of a vector. The BiLSTM layer consists of a forward long short-term memory network and a backward long short-term memory network, and its calculation process is as follows: A i =f1(ω1x i +ω2A i-1 ) B j =f2(ω3x i +ω5B i+1 ) Y i =f3(ω4A i +ω6B i ) Among them, f1, f2, f3 are activation functions between the long short-term memory networks in each layer, A is the forward LSTM hidden state, B is the backward LSTM hidden state, ω1~ω6 are weights, x i is the output feature of the CNN layer, Y i is the output feature of the BiLSTM layer, The attention layer is used to dynamically weight the output features at different positions to improve the performance of the data model when processing long sequences and enable it to better focus on key information. The output layer takes the output feature S of the attention layer at time t t As input, and output the output product properties and / or output product flow rate y of the ebullated bed residue oil hydroprocessing unit t : y t =sigmoid(w o s t +b o ) Among them, w o is the weight of the output product properties and / or output product flow, b o A bias for output product properties and / or output product flow.
11. The product prediction method according to claim 2, characterized in that: The step of fitting the adaptive weights comprises: Initialize the deviation e and coefficient K in the PID algorithm p , K i , K d , and set the initial weights of the mechanism model and the data model M and weight D ; Multiple sets of historical sample data are input into the pre-built and trained mechanism model and data model to obtain their output results Pre M and Pre D ; According to the output results of the mechanism model and the data model Pre M and Pre D , and the corresponding initial weight weight M and weight D , calculate the output of the mechanism-data joint model: result = weight M *Pre M + weight D *Pre D ; and According to the output result of the mechanism-data joint model and the corresponding product data target in the historical sample data, the deviation e of the current round is determined, and the PID output u of the current round is calculated accordingly, so as to update the weights of the mechanism model and the data model in the next round through the PID algorithm: e=target-Pre M *weight M -Pre D *weight D u=Kp*e+Ki*∫edt+Kd*d(e)dt weight M =weight M +u weight D =1-weight M 。 12. A product prediction system for an ebullated bed residue oil hydrogenation unit, characterized in that: include: a memory having computer instructions stored thereon; as well as A processor is connected to the memory and is configured to execute computer instructions stored in the memory to implement the product prediction method for the ebullated bed residue hydroprocessing unit according to any one of claims 1 to 11.
13. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the computer instructions are executed by the processor, the product prediction method for the ebullated-bed residue hydroprocessing unit according to any one of claims 1 to 11 is implemented.
