Quality index prediction method and system in catalytic cracking process

Abstract

The invention provides a catalytic cracking process quality index prediction method and system, and the method comprises the steps: obtaining the actual production data of a catalytic cracking process in a preset time, carrying out the data preprocessing and feature selection, constructing a convolutional neural network input data set for prediction, and confirming a prediction index; inputting the data set into a pre-established convolutional neural network, firstly extracting input feature information by using one-dimensional convolution, retaining feature information most related to a predicted value in a convolutional layer by using a maximum pooling technology, and processing output information of the pooling layer according to two directions before and after time by taking a bidirectional gating cycle unit as a basic unit, so as to obtain a predicted value; and capturing time dependence between features, selectively paying attention to key information extracted by a convolution layer and a pooling layer by adopting an attention layer, delivering the key information to a full connection layer, and fitting to obtain a predicted value. Based on the method, the invention further provides a catalytic cracking process quality index prediction system, and the catalytic cracking process quality index prediction accuracy is greatly improved.

Description

Technical Field

[0001] The invention belongs to the technical field of petrochemical industry, and in particular relates to a method and system for predicting quality indicators of a catalytic cracking process. Background Art

[0002] Catalytic cracking is one of the most commonly used technical methods in the petrochemical and refining industries. It mainly controls the reaction process by adjusting factors such as reaction temperature, reaction pressure, catalyst type and loading amount to convert heavy petroleum fractions into light fractions, such as gasoline and liquefied gas. In order to achieve green production, energy saving and consumption reduction, and improve corporate benefits, it is of great significance to realize timely monitoring and control of catalytic cracking process performance. Soft measurement is a common method in related fields. It estimates key quality variables by establishing mathematical models and using easily measurable auxiliary variables to provide assistance for process monitoring. It has the advantages of fast response speed, high prediction accuracy, and low cost.

[0003] Soft sensing methods are mainly divided into mechanism models and data-driven models. The mechanism model is based on the understanding and description of the physical and chemical principles of the modeled process, and predicts the behavior and performance of the process by modeling and simulating the process. It can deeply analyze and understand the internal mechanism of the process, and can provide quantitative physical and chemical parameters, with high interpretability and reliability. However, the mechanism model requires an in-depth understanding of the physical and chemical principles of the modeled process. The modeling process is complicated, and for complex systems, the mechanism model may be difficult to solve, and the model accuracy is low; while the data-driven model is based on statistical learning theory and machine learning algorithms, and predicts the behavior and performance of the process by learning and fitting existing data. It can model and predict the process without an in-depth understanding of the physical and chemical principles of the modeled process, and has high flexibility. Such as principal component regression (PCR), partial least squares regression (PLSR), support vector machine (SVR), etc. In the past few years, many successful applications of soft sensing have been proposed in chemical engineering, biochemical engineering, metallurgy and pharmaceutical industries. However, with the continuous development of industrial technology, the data generated in the system control process has the following characteristics: huge data volume, multiple data sources and high data dimension. This makes it impossible for traditional data-driven methods to make more accurate predictions due to limited representation learning capabilities. Summary of the invention

[0004] In order to solve the above technical problems, the present invention proposes a method and system for predicting quality indicators of a catalytic cracking process, and the present invention improves the prediction accuracy of quality indicators of a catalytic cracking process.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for predicting quality indicators of a catalytic cracking process comprises the following steps:

[0007] Acquire actual production data of the catalytic cracking process at a preset time, construct a convolutional neural network input data set for prediction after preprocessing and feature selection of the actual production data, and confirm the prediction index;

[0008] The data set is input into a pre-established convolutional neural network. First, one-dimensional convolution is used to extract input feature information. The maximum pooling technique is used to retain the feature information in the convolution layer that is most relevant to the predicted value. The bidirectional gated recurrent unit is used as the basic unit to process the output information of the pooling layer in both the forward and backward directions in time to capture the temporal dependency between features. The attention layer is used to selectively focus on the key information extracted by the convolution layer and the pooling layer, which is then handed over to the fully connected layer to fit the predicted value.

[0009] Furthermore, the actual production data of the catalytic cracking process includes: the mass flow rate of each raw oil input in the catalytic cracking process, the pressure and temperature at different positions of each device during the reaction process, and the content of components in the mixed raw material.

[0010] Furthermore, the process of constructing a convolutional neural network input data set for prediction after preprocessing and feature selection of the actual production data includes:

[0011] Detect missing values ​​and outliers in the data, remove data that does not meet the error threshold, and use linear interpolation to supplement the outlier point data that has been removed;

[0012] A preset proportion of feature variables is randomly selected to build a subtree, which is then summed and averaged according to feature importance and sorted in descending order. Finally, the process variables are retained as the convolutional neural network input data set for prediction.

[0013] Furthermore, the prediction indicators include gasoline, diesel, liquefied gas, dry gas and oil slurry yields.

[0014] Furthermore, the method further includes: using a random forest algorithm to calculate the importance of each feature prediction indicator, and the calculation formula is:

[0015]

[0016] Where N is the total number of subtrees generated; err 1i is the characteristic error of the current subtree without adding random noise; err 2i The characteristic error of adding random noise to the current subtree; i is the sequence number of the current subtree;

[0017] According to the importance of each feature prediction indicator, the corresponding proportion of features are eliminated, and then the feature set with the lowest error rate is selected as the convolutional neural network input data set for prediction.

[0018] Furthermore, the method also includes performing data denoising on the data set using wavelet transform, and inputting the denoised data set into a pre-established convolutional neural network.

[0019] Furthermore, the bidirectional gated recurrent unit is composed of an update gate and a reset gate, which is specifically expressed as follows:

[0020] z t =σ(W z [x t ,h t-1 ]);

[0021] r t =σ(W r [x t ,h t-1 ]);

[0022]

[0023]

[0024] Among them, x t is the current input; z t is the update gate, r t is the reset gate; h t-1 is the hidden layer state before the current time direction; h t is the output of the hidden layer before the current time direction; is the current candidate state; W Z is the weight coefficient of the update gate; W r is the weight coefficient of the reset gate; W is the weight coefficient of the current candidate state; σ is the sigmoid function.

[0025] Furthermore, the bidirectional gated recurrent unit is divided into two layers, and the temporal dependency of features is extracted in two different directions in the time sequence, and finally two sets of hidden variables are obtained, which are calculated and expressed as:

[0026]

[0027]

[0028]

[0029] in, Output hidden variables for the backward layer; Output hidden variables for the forward layer.

[0030] Furthermore, the attention layer is expressed as:

[0031] e t =V Ttanh(W e h t +b e );

[0032]

[0033]

[0034] Among them, V is the first weight parameter, W e is the second weight parameter, b e is the third weight parameter, e t is the probability distribution value of the attention mechanism at time t; a t is the weight of the attention mechanism corresponding to the hidden variable output by the bidirectional gated recurrent unit at time t; s t is the output value of the attention mechanism at time t.

[0035] The present invention also proposes a catalytic cracking process quality index prediction system, including a pretreatment module and a prediction module;

[0036] The preprocessing module is used to obtain actual production data of the catalytic cracking process at a preset time, construct a convolutional neural network input data set for prediction after preprocessing and feature selection of the actual production data, and confirm the prediction index;

[0037] The prediction module is used to input the data set into a pre-established convolutional neural network. First, one-dimensional convolution is used to extract input feature information, and the maximum pooling technology is used to retain the feature information in the convolution layer that is most relevant to the predicted value. The bidirectional gated recurrent unit is used as the basic unit to process the output information of the pooling layer in the two directions of time, forward and backward, to capture the temporal dependency between features, and then the attention layer is used to selectively focus on the key information extracted by the convolution layer and the pooling layer, which is then handed over to the fully connected layer to fit the predicted value.

[0038] The effects provided in the content of the invention are only the effects of the embodiments, not all the effects of the invention. One of the above technical solutions has the following advantages or beneficial effects:

[0039] The present invention proposes a method and system for predicting quality indicators of a catalytic cracking process. The method comprises the following steps: obtaining actual production data of a catalytic cracking process at a preset time, constructing a convolutional neural network input data set for prediction after preprocessing and feature selection of the actual production data, and confirming the prediction indicators; inputting the data set into a pre-established convolutional neural network, firstly extracting input feature information by using one-dimensional convolution, retaining the feature information most relevant to the prediction value in the convolution layer by using the maximum pooling technology, and using the bidirectional gated recurrent unit as the basic unit, processing the output information of the pooling layer in the two directions of time before and after, capturing the time dependency between the features, and selectively focusing on the key information extracted by the convolution layer and the pooling layer by using the attention layer to hand it over to the fully connected layer, and fitting to obtain the prediction value. Based on a method for predicting quality indicators of a catalytic cracking process, a system for predicting quality indicators of a catalytic cracking process is also proposed. The prediction method proposed by the present invention is simple, fast and practical. Compared with the traditional measurement method, it greatly shortens the detection time and reduces the waste of manpower and material resources. Since the process variables have the characteristics of nonlinearity and high coupling, the data feature information loss is large, which makes the model fitting effect poor. The present invention greatly improves the prediction accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 A flow chart of a method for predicting quality indicators of a catalytic cracking process proposed in Example 1 of the present invention;

[0041] Figure 2 This is the catalytic cracking process proposed in Example 1 of the present invention;

[0042] Figure 3 The basic structure of the CNN-BIGRU-ATTENTION model proposed in Example 1 of the present invention;

[0043] Figure 4 The prediction results of gasoline yield by three different methods proposed in Example 1 of the present invention are as follows;

[0044] Figure 5 The prediction errors of gasoline yield by three different methods proposed in Example 1 of the present invention;

[0045] Figure 6 This is the diesel yield prediction result proposed in Example 1 of the present invention;

[0046] Figure 7 The diesel yield prediction performance proposed in Example 1 of the present invention;

[0047] Figure 8 This is the prediction result of the liquefied gas yield proposed in Example 1 of the present invention;

[0048] Fig. 9 The liquefied gas yield prediction performance proposed in Example 1 of the present invention;

[0049] Fig.10 The dry gas yield prediction result proposed in Example 1 of the present invention;

[0050] Fig.11 The dry gas yield prediction performance proposed in Example 1 of the present invention;

[0051] Fig.12 This is the prediction result of the oil slurry yield proposed in Example 1 of the present invention;

[0052] Fig.13 The oil slurry yield prediction performance proposed in Example 1 of the present invention;

[0053] Fig.14 This is a schematic diagram of a catalytic cracking process quality index prediction system proposed in Example 2 of the present invention. DETAILED DESCRIPTION

[0054] In order to clearly illustrate the technical features of the present solution, the present invention is described in detail below through specific implementation methods and in conjunction with the accompanying drawings. The disclosure below provides many different embodiments or examples for realizing different structures of the present invention. In order to simplify the disclosure of the present invention, the components and settings of specific examples are described below. In addition, the present invention may repeat reference numbers and / or letters in different examples. This repetition is for the purpose of simplification and clarity, and does not itself indicate the relationship between the various embodiments and / or settings discussed. It should be noted that the components illustrated in the accompanying drawings are not necessarily drawn to scale. The present invention omits the description of known components and processing techniques and processes to avoid unnecessary limitations on the present invention.

[0055] Example 1

[0056] The method for predicting quality indicators of a catalytic cracking process proposed in Example 1 of the present invention can more effectively extract industrial process data features, reduce model losses, and improve prediction accuracy.

[0057] Figure 1 A flow chart of a method for predicting quality indicators of a catalytic cracking process proposed in Example 1 of the present invention;

[0058] In step S100, actual production data of the catalytic cracking process at a preset time is obtained, a convolutional neural network input data set for prediction is constructed after preprocessing and feature selection of the actual production data, and prediction indicators are confirmed;

[0059] Figure 2The catalytic cracking workflow proposed in Example 1 of the present invention; the catalytic cracking process is one of the important means of lightening heavy oil in the petroleum refining industry, and its products are mainly catalytic gasoline with low olefin content, propylene-rich liquefied gas, catalytic diesel and a small part of slurry oil. It mainly includes a reaction-regeneration system, a fractionation system, and an absorption and stabilization system. Reaction-regeneration system: In the reactor, the raw material contacts the catalyst and undergoes a cracking reaction to produce hydrocarbons with a smaller molecular weight. The products after the reaction include light oil, heavy oil and catalyst. In the regenerator, the used catalyst is first baked to remove the gas adsorbed on the catalyst, and then the activity of the catalyst is regenerated by a hydrogenation reaction. Fractionation system: In the fractionation tower, the product is heated and separated into different components; the absorption-stabilization system absorbs unstable components and removes these undesirable components.

[0060] The data collected in this application are taken from the data in the real-time database of the refinery over a period of time. 50 input variables are selected as the key variables of the real continuous catalytic cracking process, such as the mass flow rate of the feed (crude oil), the pressure and temperature at the inlet, top or bottom of each key device during the reaction process, and other key parameters. The following Table 1 is the catalytic cracking process variables.

[0061] Table 1: Catalytic Cracking Process Variables

[0062]

[0063]

[0064]

[0065] The present invention finally obtains 3662 samples, which are divided into a training set and a test set. Usually, 80% of the data is used for training the model, and the remaining 20% ​​is used for testing the performance of the model. The present invention uses the first 3062 samples as training samples, the last 600 samples as prediction samples, and finally determines the gasoline, diesel, liquefied gas, dry gas, and oil slurry yields as prediction indicators.

[0066] The process of preprocessing actual production data includes: first detecting missing values ​​and outliers in historical data, eliminating process variable data with a large number of missing data and data that does not meet the Raida criterion, and using linear interpolation to supplement the abnormal point data that has been eliminated in the data. The average value of the two original data before and after the abnormal point in time sequence is used to replace the abnormality. The abnormal data that cannot be averaged is replaced by the average value of the original data.

[0067] A preset proportion of feature variables is randomly selected to build a subtree, which is then summed and averaged according to feature importance and sorted in descending order. Finally, the process variables are retained as the convolutional neural network input data set for prediction.

[0068] In this application, 1 / 3 of the feature variable data is randomly selected to construct a subtree. In this embodiment, the number of subtrees is set to 100, and models are constructed for five prediction indicators respectively. The features are added and averaged according to their importance and sorted in descending order. Finally, 33 process variables are retained as the input data set of the model.

[0069] The importance of each feature prediction indicator is calculated using the random forest algorithm. The calculation formula is:

[0070]

[0071] Where N is the total number of subtrees generated; err 1i is the characteristic error of the current subtree without adding random noise; err 2i The characteristic error of adding random noise to the current subtree; i is the sequence number of the current subtree;

[0072] According to the importance of each feature prediction indicator, the corresponding proportion of features are eliminated, and then the feature set with the lowest error rate is selected as the convolutional neural network input data set for prediction.

[0073] This application also includes the use of wavelet transform to reduce the noise of the data set, and inputting the denoised data set into a pre-established convolutional neural network. First, the noise that may be contained in the data set is eliminated, and the denoising effects of the mean filter method and the wavelet transform are compared. The Wavelet function of the pywt package in Python is selected for wavelet transform to reduce the noise of the data. It decomposes the data into a set of wavelet coefficients, which represent the different frequency components of the signal at different scales. The high-frequency coefficients are selected and set to 0, the noise is eliminated, and the data is regenerated. Then the denoised data set is normalized to the maximum and minimum, and input into the pre-established convolutional neural network.

[0074] In step S200, the data set is input into a pre-established convolutional neural network. First, one-dimensional convolution is used to extract input feature information. The maximum pooling technique is used to retain the feature information in the convolution layer that is most relevant to the predicted value. The bidirectional gated recurrent unit is used as the basic unit to process the output information of the pooling layer in both the forward and backward directions in time to capture the temporal dependency between features. The attention layer is used to selectively focus on the key information extracted by the convolution layer and the pooling layer, which is then handed over to the fully connected layer to fit the predicted value.

[0075] Figure 3The basic structure of the CNN-BIGRU-ATTENTION model proposed in Example 1 of the present invention; the basic structure of the convolutional neural network is set to convolution layer-pooling layer-convolution layer-pooling layer, and the input data shape is [none,33,1], where none is the batch data length. The convolution layers are all one-dimensional convolutions, and the convolution kernel length is 16. The convolution kernel reads the input data and performs a dot product operation to learn the local information of the feature. The filter sizes are set to 32 and 16 respectively, the pooling layer uses maximum pooling, the pooling window size is set to 2, the nonlinear activation function is "relu", and the convolutional neural network output matrix size is [none,1,16], where none is the batch data length.

[0076] The bidirectional gated recurrent unit uses the gated recurrent unit as the basic unit, and is divided into two layers, the forward layer and the backward layer. Each layer has 20 gated recurrent units. It learns the time dependency of the upper layer output matrix in different directions of the time sequence, extracts the time series features, and outputs the hidden variable shape size of [none, 1, 40], where none is the batch data length. The bidirectional gated recurrent unit consists of an update gate and a reset gate, which is specifically expressed as:

[0077] z t =σ(W z [x t ,h t-1 ]);

[0078] r t =σ(W r [x t ,h t-1 ]);

[0079]

[0080]

[0081] Among them, x t is the current input; z t is the update gate, r t is the reset gate; h t-1 is the hidden layer state before the current time direction; h t is the output of the hidden layer before the current time direction; is the current candidate state; W Z is the weight coefficient of the update gate; W r is the weight coefficient of the reset gate; W is the weight coefficient of the current candidate state; σ is the sigmoid function.

[0082] The bidirectional gated recurrent unit is divided into two layers, and the temporal dependency of features is extracted in two different directions in the chronological order, and finally two sets of hidden variables are obtained, which are calculated and expressed as:

[0083]

[0084]

[0085]

[0086] in, Output hidden variables for the backward layer; Output hidden variables for the forward layer.

[0087] The attention layer is represented as:

[0088] e t =V T tanh(W e h t +b e );

[0089]

[0090]

[0091] Among them, V is the first weight parameter, W e is the second weight parameter, b e is the third weight parameter, e t is the probability distribution value of the attention mechanism at time t; a t is the weight of the attention mechanism corresponding to the hidden variable output by the bidirectional gated recurrent unit at time t; s t is the output value of the attention mechanism at time t.

[0092] The model runs for 200 iterations, and the length of each batch of data is 49. Gasoline is widely used and has high economic benefits. In order to verify the effectiveness of the method we proposed, three different methods are used to compare the results of gasoline yield. The following table 2 is a comparison of the gasoline yield prediction effect; Figure 4 The prediction results of gasoline yield by three different methods proposed in Example 1 of the present invention are as follows;

[0093] Table 2: Comparison of gasoline yield prediction results

[0094] FNN LSTM-ATTENTION CNN-BIGRU-ATTENTION RMSE 0.46328 0.19734 0.09127 R2 0.77872 0.90574 0.95640

[0095] Table 3 below shows the prediction results of other indicators.

[0096] Table 3: Prediction results of other indicators

[0097] Training set (RMSE) Training set (R2) Test set (RMSE) Test set (R2) diesel fuel 0.98 0.223 0.96 0.246 Liquefied gas 0.98 0.184 0.979 0.214 Dry gas 0.96 0.045 0.92 0.081 Oil slurry 0.91 0.119 0.90 0.131

[0098] Among them, "RMSE" and "R2" are used to quantify the effectiveness of the method of the present invention, and the specific calculation formula is as follows:

[0099]

[0100]

[0101] Among them, m is the number of samples in the test data, y i and are the actual value and predicted value of the prediction indicator respectively; is the average value of the actual value of the predicted indicator in the test data set

[0102] In terms of model prediction results, the fully connected neural network (FNN) structure is relatively simple. When the input-output relationship is complex and the input data has multi-scale feature information, it is difficult to establish an effective mapping relationship. Therefore, the prediction error in the test set is relatively large.

[0103] Figure 5 The prediction errors of gasoline yield by three different methods proposed in Example 1 of the present invention; the LSTM-ATTENTION model extracts the temporal dependency of feature context by learning and memorizing long sequences, thereby further improving the model performance. The CNN-BIGRU-ATTENTION model combines the advantages of CNN and BIGRU, and can effectively capture the local features and long-distance dependencies of the data. At the same time, by introducing the attention mechanism, the model can autonomously learn the important parts of the data, thereby showing superior performance when processing industrial production data, and maintaining a high prediction accuracy for different prediction indicators. Specifically: Figure 6 This is the diesel yield prediction result proposed in Example 1 of the present invention; Figure 7 The diesel yield prediction performance proposed in Example 1 of the present invention; Figure 8 This is the prediction result of the liquefied gas yield proposed in Example 1 of the present invention; Fig. 9 The liquefied gas yield prediction performance proposed in Example 1 of the present invention; Fig.10 The dry gas yield prediction result proposed in Example 1 of the present invention; Fig.11 The dry gas yield prediction performance proposed in Example 1 of the present invention; Fig.12 This is the prediction result of the oil slurry yield proposed in Example 1 of the present invention; Fig.13 This is the oil slurry yield prediction performance proposed in Example 1 of the present invention.

[0104] The above data prove that the method proposed in this application has been verified in the industrial catalytic cracking process, and the average mean square error of the proposed method is much lower than that of other methods.

[0105] The method for predicting the quality index of the catalytic cracking process proposed in Example 1 of the present invention is simple, fast and practical. Compared with the traditional measurement method, it greatly shortens the detection time and reduces the waste of manpower and material resources. Since the process variables have the characteristics of nonlinearity and high coupling, the data feature information loss is large, resulting in poor model fitting effect. The present invention greatly improves the prediction accuracy of the model.

[0106] Example 2

[0107] Based on the method for predicting quality indexes of a catalytic cracking process proposed in Example 1 of the present invention, Example 2 of the present invention further proposes a system for predicting quality indexes of a catalytic cracking process. Fig.14 This is a schematic diagram of a catalytic cracking process quality index prediction system proposed in Example 2 of the present invention, including a pretreatment module and a prediction module;

[0108] The preprocessing module is used to obtain the actual production data of the catalytic cracking process at a preset time, construct a convolutional neural network input data set for prediction after preprocessing and feature selection of the actual production data, and confirm the prediction indicators;

[0109] The prediction module is used to input the data set into a pre-established convolutional neural network. First, one-dimensional convolution is used to extract input feature information, and the maximum pooling technology is used to retain the feature information in the convolution layer that is most relevant to the predicted value. The bidirectional gated recurrent unit is used as the basic unit to process the output information of the pooling layer in the two directions of time, capturing the temporal dependency between features. The attention layer is used to selectively focus on the key information extracted by the convolution layer and the pooling layer, which is then handed over to the fully connected layer to fit the predicted value.

[0110] The processes performed by the pretreatment module include: the actual production data of the catalytic cracking process includes: the mass flow rate of each raw oil input in the catalytic cracking process, the pressure and temperature at different positions of each device during the reaction process, and the component content in the mixed raw material.

[0111] The process of constructing a convolutional neural network input data set for prediction after preprocessing and feature selection of the actual production data includes:

[0112] Detect missing values ​​and outliers in the data, remove data that does not meet the error threshold, and use linear interpolation to supplement the outlier point data that has been removed;

[0113] A preset proportion of feature variables is randomly selected to build a subtree, which is then summed and averaged according to feature importance and sorted in descending order. Finally, the process variables are retained as the convolutional neural network input data set for prediction.

[0114] Predicted indicators include gasoline, diesel, liquefied gas, dry gas and slurry oil yields.

[0115] The importance of each feature prediction indicator is calculated using the random forest algorithm. The calculation formula is:

[0116]

[0117] Where N is the total number of subtrees generated; err 1i is the characteristic error of the current subtree without adding random noise; err 2i The characteristic error of adding random noise to the current subtree; i is the sequence number of the current subtree;

[0118] According to the importance of each feature prediction indicator, the corresponding proportion of features are eliminated, and then the feature set with the lowest error rate is selected as the convolutional neural network input data set for prediction.

[0119] Wavelet transform is used to perform data denoising on the data set, and the denoised data set is input into the pre-established convolutional neural network.

[0120] The process of executing the prediction module includes: the bidirectional gated recurrent unit is composed of an update gate and a reset gate, which is specifically expressed as:

[0121] z t =σ(W z [x t ,h t-1 ]);

[0122] r t =σ(W r [x t ,h t-1 ]);

[0123]

[0124]

[0125] Among them, x t is the current input; z t is the update gate, r t is the reset gate; h t-1 is the hidden layer state before the current time direction; h t is the output of the hidden layer before the current time direction; is the current candidate state; W Z is the weight coefficient of the update gate; W r is the weight coefficient of the reset gate; W is the weight coefficient of the current candidate state; σ is the sigmoid function.

[0126] The bidirectional gated recurrent unit is divided into two layers, and the temporal dependency of features is extracted in two different directions in the chronological order, and finally two sets of hidden variables are obtained, which are calculated and expressed as:

[0127]

[0128]

[0129]

[0130] in, Output hidden variables for the backward layer; Output hidden variables for the forward layer.

[0131] The attention layer is represented as:

[0132] e t =V T tanh(W e h t +b e );

[0133]

[0134]

[0135] Among them, V is the first weight parameter, W e is the second weight parameter, b e is the third weight parameter, e t is the probability distribution value of the attention mechanism at time t; a t is the weight of the attention mechanism corresponding to the hidden variable output by the bidirectional gated recurrent unit at time t; s t is the output value of the attention mechanism at time t.

[0136] The catalytic cracking process quality index prediction system proposed in Example 2 of the present invention is simple, fast and practical. Compared with the traditional measurement method, it greatly shortens the detection time and reduces the waste of manpower and material resources. Since the process variables have the characteristics of nonlinearity and high coupling, the data feature information loss is large, resulting in poor model fitting effect. The present invention greatly improves the prediction accuracy of the model.

[0137] The description of the relevant parts of a catalytic cracking process quality index prediction system provided in the embodiment of the present application can be found in the detailed description of the corresponding parts of a catalytic cracking process quality index prediction method provided in Example 1 of the present application, and will not be repeated here.

[0138] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between these entities or operations. Moreover, the term "include", "comprise" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or equipment that includes a series of elements are inherent to the elements. In the absence of more restrictions, the elements limited by the sentence "comprise one..." do not exclude the presence of other identical elements in the process, method, article or equipment that includes the elements. In addition, the above-mentioned technical solution provided in the embodiment of the present application is consistent with the corresponding technical solution in the prior art in principle, and the part is not described in detail, so as not to repeat too much.

[0139] Although the above describes the specific implementation of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. For those skilled in the art, other different forms of modifications or deformations can be made on the basis of the above description. It is not necessary and impossible to list all the implementation methods here. On the basis of the technical solution of the present invention, various modifications or deformations that can be made by those skilled in the art without creative work are still within the scope of protection of the present invention.

Claims

1. A method for predicting quality indicators of a catalytic cracking process, characterized in that: The following steps are involved: Acquire actual production data of the catalytic cracking process at a preset time, construct a convolutional neural network input data set for prediction after preprocessing and feature selection of the actual production data, and confirm the prediction index; The data set is input into a pre-established convolutional neural network. First, one-dimensional convolution is used to extract input feature information. The maximum pooling technique is used to retain the feature information in the convolution layer that is most relevant to the predicted value. The bidirectional gated recurrent unit is used as the basic unit to process the output information of the pooling layer in both the forward and backward directions in time to capture the temporal dependency between features. The attention layer is used to selectively focus on the key information extracted by the convolution layer and the pooling layer, which is then handed over to the fully connected layer to fit the predicted value.

2. A method for predicting quality indicators of a catalytic cracking process according to claim 1, characterized in that: The actual production data of the catalytic cracking process includes: the mass flow rate of each raw material oil input in the catalytic cracking process, the pressure and temperature at different positions of each device during the reaction process, and the content of components in the mixed raw material.

3. A method for predicting quality indicators of a catalytic cracking process according to claim 1, characterized in that: The process of constructing a convolutional neural network input data set for prediction after preprocessing and feature selection of the actual production data includes: Detect missing values ​​and outliers in the data, remove data that does not meet the error threshold, and use linear interpolation to supplement the outlier point data that has been removed; A preset proportion of feature variables is randomly selected to build a subtree, which is then summed and averaged according to feature importance and sorted in descending order. Finally, the process variables are retained as the convolutional neural network input data set for prediction.

4. A method for predicting quality indicators of a catalytic cracking process according to claim 3, characterized in that: The prediction indicators include gasoline, diesel, liquefied gas, dry gas and slurry oil yields.

5. A method for predicting quality indicators of a catalytic cracking process according to claim 4, characterized in that: The method further includes: using a random forest algorithm to calculate the importance of each feature prediction indicator, and the calculation formula is: Where N is the total number of subtrees generated; err 1i is the characteristic error of the current subtree without adding random noise; err 2i The characteristic error of adding random noise to the current subtree; i is the sequence number of the current subtree; According to the importance of each feature prediction indicator, the corresponding proportion of features is eliminated, and then the feature set with the lowest error rate is selected as the convolutional neural network input data set for prediction.

6. A method for predicting quality indicators of a catalytic cracking process according to claim 1, characterized in that: The method also includes performing data denoising on the data set using wavelet transform, and inputting the denoised data set into a pre-established convolutional neural network.

7. A method for predicting quality indicators of a catalytic cracking process according to claim 1, characterized in that: The bidirectional gated recurrent unit is composed of an update gate and a reset gate, which is specifically expressed as: z t =σ(W z [x t ,h t-1 ]); r t =σ(W r [x t ,h t-1 ]); Among them, x t is the current input; z t is the update gate, r t is the reset gate; h t-1 is the hidden layer state before the current time direction; h t is the output of the hidden layer before the current time direction; is the current candidate state; W Z is the weight coefficient of the update gate; W r is the weight coefficient of the reset gate; W is the weight coefficient of the current candidate state; σ is the sigmoid function.

8. A method for predicting quality indicators of a catalytic cracking process according to claim 7, characterized in that: The bidirectional gated recurrent unit is divided into two layers, and the temporal dependency of features is extracted in two different directions in the time sequence, and finally two sets of hidden variables are obtained, which are calculated and expressed as: in, Output hidden variables for the backward layer; Output hidden variables for the forward layer.

9. A method for predicting quality indicators of a catalytic cracking process according to claim 1, characterized in that: The attention layer is expressed as: e t =V T tanh(W e h t +b e ); Among them, V is the first weight parameter, W e is the second weight parameter, b e is the third weight parameter, e t is the probability distribution value of the attention mechanism at time t; a t is the weight of the attention mechanism corresponding to the hidden variable output by the bidirectional gated recurrent unit at time t; s t is the output value of the attention mechanism at time t.

10. A catalytic cracking process quality index prediction system, characterized in that: Includes preprocessing module and prediction module; The preprocessing module is used to obtain actual production data of the catalytic cracking process at a preset time, construct a convolutional neural network input data set for prediction after preprocessing and feature selection of the actual production data, and confirm the prediction index; The prediction module is used to input the data set into a pre-established convolutional neural network. First, one-dimensional convolution is used to extract input feature information, and the maximum pooling technology is used to retain the feature information in the convolution layer that is most relevant to the predicted value. The bidirectional gated recurrent unit is used as the basic unit to process the output information of the pooling layer in the two directions of time, forward and backward, to capture the temporal dependency between features, and then the attention layer is used to selectively focus on the key information extracted by the convolution layer and the pooling layer, which is then handed over to the fully connected layer to fit the predicted value.

Images