Ensemble learning online prediction method for key parameters of continuous catalytic reforming process

Abstract

The present invention relates to the field of continuous catalytic reforming process, and more specifically, to an integrated learning online prediction method for key parameters of the continuous catalytic reforming process. The method comprises the following steps: S1, selecting auxiliary variables and main variables of the continuous catalytic reforming process; S2, preprocessing the process data sample set to generate a data set; S3, dividing the data set; S4, selecting spn data samples from the training data set as the overall data set, and the remaining data samples as the correction data set, and respectively using the overall data set as the sub-training set to construct three adaptive regression sub-models; S5, extracting data samples from the correction data set in sequence, and updating the corresponding adaptive regression sub-models; S6, applying the evaluation data set to each adaptive regression sub-model to establish an integrated model; S7, predicting the main variables based on the input auxiliary variables. The present invention effectively improves the prediction accuracy of the model in the continuous catalytic reforming process.

Description

Technical Field

[0001] The present invention relates to the field of continuous catalytic reforming technology, and more specifically, to an integrated learning online prediction method for key parameters of a continuous catalytic reforming process. Background Art

[0002] In recent years, with the increasing requirements for automobile oil and the increasing crisis of energy, catalytic reforming technology has also attracted great attention. The emergence of catalyst continuous regeneration reforming technology has epoch-making significance in the history of catalyst development, providing unlimited possibilities for the efficient production of high-quality gasoline and other chemical raw materials.

[0003] The continuous catalytic reforming process is mainly used to increase the octane number of gasoline refineries and aromatic-rich petrochemical products. It is currently the most popular naphtha reforming process in the world. It realizes the continuous extraction of catalyst during the steam operation of the reactor.

[0004] Figure 1 The continuous catalytic reforming process flow chart is revealed, such as Figure 1 As shown, during the catalytic cracking process, heavy naphtha liquid is sent by pump 300 to heat exchanger 400 and mixed with hydrogen-rich recovery gas from recycle compressor 600.

[0005] Before entering each reactor (reactor No. 1 201 to reactor No. 4 204), the temperature of the liquid gas feed mixture is raised to the reaction temperature in the heater (heater No. 1 101 to heater No. 4 104) to ensure that the mixture can be completely vaporized.

[0006] In the four reactors, several chemical reactions, mainly dehydrogenation, occur stably.

[0007] When the reactants flow through the catalyst bed quickly, the temperature of the reactants drops sharply. The heating and reaction process continues until they flow out of the No. 4 reactor 204 and enter the contact and separation part.

[0008] Among them, the reforming catalyst is continuously consumed during operation, and chemical reactions occur when it comes into contact with various substances, causing the catalyst activity to gradually decrease. Therefore, the catalyst needs to be continuously regenerated to maintain a high and stable catalyst activity.

[0009] The function of the regenerator 500 is to realize the continuous circulation of the catalyst and complete the regeneration of the catalyst at the same time.

[0010] In the separator 700 , the hydrogen-rich gas is separated and returned to the circulation compressor 600 .

[0011] In addition to this, the net hydrogen-rich gas is used for additional hydrogen-consuming petrochemical processes, and the bottom product of the stabilizer is high-octane liquid reformate.

[0012] The catalytic reforming process uses naphtha as raw material and improves its octane number through reforming reactions.

[0013] As a key parameter of the process, in order to improve economic benefits and improve energy pollution and safety issues, online detection of the octane number of reformed gasoline is essential.

[0014] However, the continuous catalytic reforming process is a very complex industrial production technology with the characteristics of large time lag, high coupling and nonlinearity.

[0015] In industrial engineering in recent decades, when variables cannot be directly obtained, soft measurements have been widely used to estimate key performance indicators instead of hard measurements, and have excellent performance in parameter prediction and process monitoring.

[0016] Many innovative soft sensor modeling methods such as artificial neural networks, Gaussian process regression and deep learning have been applied and developed.

[0017] However, traditional multivariate statistical methods such as principal component regression, partial least squares and slow feature analysis are still popular in practical applications.

[0018] Partial least squares is a commonly used modeling technique due to its inherent structure and feature extraction capabilities. It can project process data into low-dimensional latent variables, contain comprehensive information in input and output data, and can handle collinearity between process variables well. However, partial least squares is essentially restricted by its linear characteristics. When the data relationship is nonlinear, partial least squares cannot properly extract data features.

[0019] Therefore, a series of nonlinear improvement methods have been proposed to integrate the nonlinear features in the linear partial least squares framework. These nonlinear techniques combine neural network methods, kernel-based algorithms and local modeling strategies.

[0020] Among them, LWPLS (Locally weighted partial least squares) is based on the combination of sample weighting technology and partial least squares method, and is one of the most widely used local modeling methods.

[0021] However, due to the nonlinear and time-varying characteristics of actual industrial processes, traditional soft sensor models may suffer from severe performance gradients when they do not match the originally designed models.

[0022] Although many new methods have been used to alleviate this problem, each method only focuses on certain aspects of model characteristics, so a comprehensive framework is urgently needed to combine these features. Summary of the invention

[0023] The purpose of the present invention is to provide an integrated learning online prediction method for key parameters of a continuous catalytic reforming process, so as to solve the problems of poor prediction accuracy and poor adaptability of key parameters of the existing continuous catalytic reforming process.

[0024] In order to achieve the above object, the present invention provides an integrated learning online prediction method for key parameters of a continuous catalytic reforming process, comprising the following steps:

[0025] S1. Select auxiliary variables and main variables of the continuous catalytic reforming process, collect data to form a process data sample set, wherein the auxiliary variables reflect the operation status of the continuous catalytic reforming process, and the main variables reflect the quality of the finished product;

[0026] S2. Preprocessing the collected process data sample set of the continuous catalytic reforming process to generate a data set;

[0027] S3, dividing the data set into a training data set, an evaluation data set and a test data set;

[0028] S4, select spn data samples from the training data set as the overall data set, and the remaining data samples as the correction data set, respectively use the overall data set as the sub-training set to construct three adaptive regression sub-models, the three adaptive regression sub-models include a moving window model based on the local weighted partial least squares method, a time difference model based on the local weighted partial least squares method, and an instant learning model based on the local weighted partial least squares method;

[0029] S5, extracting data samples from the correction data set in sequence, verifying each adaptive regression sub-model, adding the current data sample to the sub-training set of the adaptive regression sub-model with the smallest estimated error value, and updating the corresponding adaptive regression sub-model;

[0030] S6, applying the evaluation data set to each adaptive regression sub-model, calculating the root mean square error of the predicted value of each adaptive regression sub-model, using Bayesian estimation to determine the weight coefficient of each adaptive regression sub-model, and establishing an integrated model;

[0031] S7. Use the constructed integrated model to predict the final output value of the main variable based on the input auxiliary variables.

[0032] In one embodiment, in step S1:

[0033] Auxiliary variables, including recontact temperature, knockout drum temperature, hydrocracked naphtha, naphtha, reactor temperature, knockout drum pressure, circulating hydrogen flow rate, primary alkane content, compressor pressure, reflux flow rate, tray temperature, reflux flow rate, tray temperature, bottoms temperature, and reactor No. 1 outlet temperature;

[0034] The main variable is the octane number, and the corresponding expression is as follows:

[0035]

[0036] Among them, Y RON , and is the output measurement parameter of the continuous catalytic reforming process, X Feed It is the feed variable of the continuous catalytic reforming process.

[0037] In one embodiment, the preprocessing method in step S2 further comprises:

[0038] Remove outliers and missing values, and remove measurement noise.

[0039] In one embodiment, in step S4, based on the moving window model of the local weighted partial least squares method, the modeling process is as follows:

[0040] S411, setting {X1, Y1}, i.e., window size H, as the training data matrix in the initial window, wherein X1 is the initial H input variables in the selected overall data set, and Y1 is the initial H output variables in the selected overall data set;

[0041] S412, establishing a local weighted partial least squares regression model for the output of multiple sampling data points behind the window;

[0042] S413, let the window move forward by step size D, so that {X w , Y w}Retrain the local weighted partial least squares regression model in the window, where X w is the H input variable after w D steps in the overall data set, Y w are the H output variables after w D steps in the overall data set.

[0043] In one embodiment, in step S4, the time difference model based on the local weighted partial least squares method has the following modeling process:

[0044] S421, respectively calculating the first-order difference Δx(t) between the input variables and the first-order difference Δy(t) between the output variables of adjacent sampling data points;

[0045] S422, constructing a relationship model between Δx(t) and Δy(t) using a local weighted partial least squares regression model;

[0046] S423. Calculate query sample x q The first-order difference Δx of (t) q(t);

[0047] S424, Δx q (t) Input the relationship model to predict the difference value of the response variable Δy p (t);

[0048] S425, based on adjacent sampling data points y p (t-1) and the difference value Δy p (t), calculate the predicted value y of the response variable p (t), the expression is as follows,

[0049] y p (t) = Δy p (t)+y p (t-1).

[0050] In one embodiment, in step S4, the instant learning model based on the local weighted partial least squares method has a modeling process as follows:

[0051] S431, when traversing and inputting a query sample x q ∈R m When it arrives, calculate x q and x i The distance between

[0052]

[0053] Among them, there are n data sample points (x i ,y i ) i＝1～n The dataset consists of x∈R m and y∈R l Represent the input and output training sets respectively, x i represents the i-th input training value;

[0054] S432, selecting N data samples with the smallest distance to form a related sample set;

[0055] S433. Based on the relevant sample set, a local weighted partial least squares regression model is constructed;

[0056] S434, through the LWPLS regression model, obtain the query sample x q The corresponding response variable y p The predicted value of .

[0057] In one embodiment, in step S5:

[0058] The estimation error is the error between the predicted value and the actual value;

[0059] Add the data sample with the smallest estimated error value to the corresponding sub-training set.

[0060] In one embodiment, the step S6 calculates the root mean square error of the prediction value of each adaptive regression sub-model, and the expression is as follows:

[0061] RMSE of the predicted value of the i-th adaptive regression sub-model i ,

[0062]

[0063] Among them, y d Represents the actual value, y dp，i represents the predicted value of the i-th adaptive regression sub-model, and D represents the number of evaluation samples in the evaluation data set.

[0064] In one embodiment, the output result y of the integrated model in step S6 is p , the corresponding expression is:

[0065]

[0066] Among them, y p，i represents the predicted output value of the i-th adaptive regression sub-model, C is the current process state, P(M i |C) is the i-th adaptive regression sub-model M in the current process state C i The probability of M1, M2, ..., M I There are I different adaptive regression sub-models.

[0067] In one embodiment, in step S6, the current process state C is the i-th adaptive regression sub-model M i The probability P(M i |C), the corresponding expression is:

[0068]

[0069] Among them, P(M i ) is the prior probability of the i-th adaptive regression sub-model, P(C|M i ) is the i-th adaptive regression sub-model M i is the probability of the current process state C, p is the adjustment parameter, RMSE i is the root mean square error of the predicted value of the i-th adaptive regression sub-model;

[0070] P(C|M i ) is,

[0071]

[0072] P(M i) is,

[0073]

[0074] r i The corresponding expression of is,

[0075]

[0076] The present invention provides an integrated learning online prediction method for key parameters of a continuous catalytic reforming process. The MW, TD and JITL adaptive regression sub-models of the LWPLS regression algorithm are combined with a Bayesian method to fuse data to obtain an integrated model. The integrated model solves different process characteristic problems such as nonlinearity, time variation and mutation existing in the soft measurement of key parameters of the continuous catalytic reforming process, greatly improves the performance and efficiency of each adaptive regression sub-model, can adapt to complex environments, and effectively improves the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] The above and other features, properties and advantages of the present invention will become more apparent through the following description in conjunction with the accompanying drawings and embodiments, in which the same reference numerals always represent the same features, wherein:

[0078] Figure 1 A continuous catalytic reforming process flow diagram is revealed;

[0079] Figure 2 A flow chart of an integrated learning online prediction method for key parameters of a continuous catalytic reforming process according to an embodiment of the present invention is disclosed;

[0080] Figure 3 A modeling flow chart of an integrated model according to an embodiment of the present invention is disclosed;

[0081] Figure 4a A flow chart of modeling a moving window model based on a local weighted partial least squares method according to an embodiment of the present invention is disclosed;

[0082] Figure 4b A flow chart of modeling a time difference model based on a local weighted partial least squares method according to an embodiment of the present invention is disclosed;

[0083] Figure 5 A modeling flow chart of a real-time learning model based on a local weighted partial least squares method according to an embodiment of the present invention is disclosed;

[0084] Figure 6 A flow chart of a method for dividing a training data set according to an embodiment of the present invention is disclosed;

[0085] Figure 7 A comparison chart of octane number predictions of five models is revealed.

[0086] The meanings of the reference numerals in the figures are as follows:

[0087] 101 No. 1 heater;

[0088] 102 No. 2 heater;

[0089] 103 No. 3 heater;

[0090] 104 No. 4 heater;

[0091] 201 Reactor No. 1;

[0092] 202 Reactor No. 2;

[0093] 203 Reactor No. 3;

[0094] 204 Reactor No. 4;

[0095] 300 pumps;

[0096] 400 heat exchanger;

[0097] 500 CCR regenerator;

[0098] 600 cycle compressor;

[0099] 700 Divider. DETAILED DESCRIPTION

[0100] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the invention and are not used to limit the invention.

[0101] In view of the different process characteristics existing in the continuous catalytic reforming process, such as nonlinearity, time-varying and sudden changes, the present invention proposes an integrated learning online prediction method for key parameters of the continuous catalytic reforming process based on an adaptive local weighted partial least squares method. By analyzing the mechanism of the continuous catalytic reforming process and selecting appropriate auxiliary variables, the training data set is divided into an overall data set and a correction data set, and a moving window model, a time difference model and a real-time learning model based on the local weighted partial least squares method are trained. The moving window model, the time difference model and the real-time learning model based on the local weighted partial least squares method are combined, and all feature information is comprehensively considered in an integrated manner using the Bayesian theorem. The prediction output is fused and calculated to establish a cost-effective soft measurement integrated model to adapt to complex environments, correctly predict the changes in the octane number of the main variable, and improve the model efficiency and accuracy.

[0102] Figure 2A flow chart of an integrated learning online prediction method for key parameters of a continuous catalytic reforming process according to an embodiment of the present invention is disclosed. Figure 3 A modeling flow chart of an integrated model according to an embodiment of the present invention is disclosed, such as Figure 2 and Figure 3 The present invention provides an integrated learning online prediction method for key parameters of a continuous catalytic reforming process, comprising the following steps:

[0103] S1. Select auxiliary variables and main variables of the continuous catalytic reforming process, collect data to form a process data sample set, wherein the auxiliary variables reflect the operation status of the continuous catalytic reforming process, and the main variables reflect the quality of the finished product;

[0104] S2. Preprocessing the collected process data sample set of the continuous catalytic reforming process to generate a data set;

[0105] S3, dividing the data set into a training data set, an evaluation data set and a test data set;

[0106] S4, select spn data samples from the training data set as the overall data set, and the remaining data samples as the correction data set, respectively use the overall data set as the sub-training set to construct three adaptive regression sub-models, the three adaptive regression sub-models include a moving window model based on the local weighted partial least squares method, a time difference model based on the local weighted partial least squares method, and an instant learning model based on the local weighted partial least squares method;

[0107] S5, extracting data samples from the correction data set in sequence, verifying each adaptive regression sub-model, adding the current data sample to the sub-training set of the adaptive regression sub-model with the smallest estimated error value, and updating the corresponding adaptive regression sub-model;

[0108] S6, applying the evaluation data set to each adaptive regression sub-model, calculating the root mean square error of the predicted value of each adaptive regression sub-model, using Bayesian estimation to determine the weight coefficient of each adaptive regression sub-model, and establishing an integrated model;

[0109] S7. Use the constructed integrated model to predict the final output value of the main variable based on the input auxiliary variables.

[0110] Each step is described in detail below.

[0111] S1. Select auxiliary variables and main variables of the continuous catalytic reforming process, collect data to form a process data sample set, wherein the auxiliary variables reflect the operating status of the continuous catalytic reforming process, and the main variables reflect the quality of the finished product.

[0112] According to the mechanism of continuous catalytic reforming process, appropriate auxiliary variables and main variables are selected.

[0113] The selection of auxiliary variables and main variables can be directly summarized from the actual continuous catalytic reforming process.

[0114] Auxiliary variables should be able to reflect the operating conditions of the continuous catalytic reforming process.

[0115] The selected auxiliary variables include but are not limited to:

[0116] Recontact temperature, separator tank temperature, hydrocracking naphtha, naphtha, reactor temperature, separator tank pressure, circulating hydrogen flow rate, primary alkane content, compressor pressure, reflux flow rate, tray temperature, reflux flow rate, tray temperature, bottom temperature and outlet temperature of reactor No. 1, etc.

[0117] The main variable needs to be able to reflect the quality of the finished product.

[0118] The octane number can reflect the quality of the finished product and is an important performance indicator.

[0119] Therefore, octane number was selected as the main variable.

[0120] The octane number can be derived from the following formula:

[0121]

[0122] Among them, Y RON , and are the three output measurements of the continuous catalytic reforming process, Y RON For research octane number, X Feed It is the feed variable of the continuous catalytic reforming process.

[0123] S2. Preprocessing the collected process data sample set of the continuous catalytic reforming process to generate a data set.

[0124] The measurement of process variables in the continuous catalytic reforming industry often contains measurement noise. Therefore, it is necessary to preprocess the collected continuous catalytic reforming process data sample set.

[0125] Preprocessing methods include: removing outliers and missing values, and removing measurement noise.

[0126] In this embodiment, the 3-sigma criterion is used to remove outliers. The 3-sigma criterion is also called the Laida criterion. It assumes that a set of test data contains only random errors, calculates and processes them to obtain the standard deviation, and determines an interval according to a certain probability. It is believed that any error exceeding this interval is not a random error but a gross error, and the data containing this error should be eliminated.

[0127] In this embodiment, the measurement noise is removed and a Gaussian smoothing filter is used to remove white noise.

[0128] The expression of Gaussian smoothing filter is as follows:

[0129]

[0130] Among them, μ represents the mean of the data sample x, and σ represents the standard deviation of the data sample x.

[0131] S3. Divide the data set into training data set, evaluation data set and test data set.

[0132] By dividing the preprocessed process data sample set, an accurate model can be constructed and the accuracy of the model can be evaluated.

[0133] Generally, 45% of the dataset is divided into a training dataset, 10% is divided into an evaluation dataset, and the remaining 45% is divided into a test dataset.

[0134] Among them, the training data set is used to train the model, give the model input and corresponding output data, and train the model to learn the relationship between input and output data;

[0135] Evaluation data sets are used to verify the training level of the evaluation model, such as the classification accuracy of the classifier, the prediction error, etc. The best model can be selected based on the performance of the evaluation data sets.

[0136] The test data set is used to test the final model after training. The final model simulates "new" input data to obtain prediction results and output them.

[0137] S4. Select spn data samples from the training data set as the overall data set, and the remaining data samples as the correction data set. Use the overall data set as the sub-training set to construct three adaptive regression sub-models. The three adaptive regression sub-models include a moving window model based on local weighted partial least squares (MW-LWPLS), a time difference model based on local weighted partial least squares (TD-LWPLS), and a just-in-time learning model based on local weighted partial least squares (JITL-LWPLS).

[0138] Select spn data samples from the training data set as the overall data set, and the remaining data samples as the correction data set.

[0139] Based on the overall data set, three adaptive regression sub-models are constructed.

[0140] Three adaptive regression sub-models, including:

[0141] Moving window model based on locally weighted partial least squares (MW-LWPLS);

[0142] Time difference model based on local weighted partial least squares (TD-LWPLS);

[0143] Just-in-time learning model based on locally weighted partial least squares (JITL-LWPLS).

[0144] The size of the initial partitioned overall dataset largely determines the performance of the final regression ensemble model, and thus the optimal value of the spn parameter needs to be determined through trial and error.

[0145] Figure 4a A modeling flow chart of a moving window model based on a local weighted partial least squares method according to an embodiment of the present invention is disclosed, such as Figure 4a The moving window model based on local weighted partial least squares shown in the figure has the following modeling process:

[0146] S411, setting {X1, Y1}, i.e., window size H, as the training data matrix in the initial window, wherein X1 is the initial H input variables in the selected overall data set, and Y1 is the initial H output variables in the selected overall data set;

[0147] S412, establishing an LWPLS regression model for the output of multiple sampling data points behind the window;

[0148] S413, let the window move forward by step size D, so that {X w , Y w}Retrain the LWPLS regression model in the window, where X w is the H input variable after w D steps in the overall data set, Y w represents the H output variables after w D steps in the overall data set.

[0149] The key step of the moving window method is to determine the appropriate window size H according to the actual process, which is usually determined by repeated trials.

[0150] In addition, for convenience, the step size D of the moving window is set to 1.

[0151] Figure 4b A modeling flow chart of a time difference model based on a local weighted partial least squares method according to an embodiment of the present invention is disclosed, such as Figure 4bThe time difference model based on local weighted partial least squares is composed of time differences between variables. The modeling process is as follows:

[0152] S421, respectively calculate the first-order difference Δx(t) between the input variables and the first-order difference Δy(t) between the output variables of adjacent sampling data points, and the corresponding expressions are:

[0153] Δx(t)=x(t)-x(t-1);

[0154] Δy(t)=y(t)-y(t-1);

[0155] S422. Use the LWPLS regression model to construct a relationship model between Δx(t) and Δy(t). The corresponding expression is:

[0156] Δy(t)=f(Δx(t));

[0157] Where f represents the LWPLS regression model function called.

[0158] S423. Calculate query sample x q The first-order difference Δx of (t) q (t), the corresponding expression is,

[0159] Δx q (t) = x q (t)-x q (t-1);

[0160] S424, Δx q (t) Input the relationship model to predict the difference value of the response variable Δy p (t), the corresponding expression is as follows,

[0161] Δy p (t) = f(Δx q (t));

[0162] S425, based on adjacent sampling data points y p (t-1) and the difference value Δy p (t), calculate the predicted value y of the response variable p (t), the corresponding expression is as follows,

[0163] y p (t) = Δy p (t)+y p (t-1).

[0164] The most critical step in the just-in-time learning framework is selecting the set of neighborhood samples for model building.

[0165] Generally, a distance-based metric, namely Euclidean distance, is used to evaluate similarity. The larger the distance, the lower the similarity between the current historical sampling point and the query sample, and vice versa.

[0166] Figure 5 A modeling flow chart of a real-time learning model based on a local weighted partial least squares method according to an embodiment of the present invention is disclosed, such as Figure 5 The instant learning model based on local weighted partial least squares shown in the figure has the following modeling process:

[0167] Assume that a data set consists of n data sample points (x i ,y i ) i＝1～n A dataset consisting of x∈R m and y∈R l Represent the input and output training sets respectively.

[0168] S431. Similarity measurement.

[0169] When traversing input a query sample x q ∈R m When it arrives, calculate x q and x i The distance between them is shown as follows:

[0170]

[0171] Among them, x i Represents the i-th input training value.

[0172] The calculated distance d i Sort in ascending order.

[0173] S432, related sample set.

[0174] Select the N data samples with the smallest distance, that is, the highest similarity, to form a relevant sample set for establishing the LWPLS regression model, because these data samples and the query sample x q Most similar.

[0175] Related sample set {X local , Y local}={(x1, y1), ..., (x i ,y i ),...,(x N ,y N )}, where N represents the number of samples in the LWPLS regression model.

[0176] S433. Based on the relevant sample set, the LWPLS regression model is constructed. The expression is as follows:

[0177] y i =f(x i ),

[0178] Among them, (x i ,y i ) is the data sample point of the relevant sample set;

[0179] S434, through the LWPLS regression model, obtain the query sample x q The predicted value of the corresponding response variable is output.

[0180] Query sample x q The LWPLS regression model is expressed as follows:

[0181] y p =f(x q ),

[0182] Among them, x q is the query sample, y p is the predicted value of the response variable.

[0183] The LWPLS regression model needs to be rebuilt for each prediction. After the prediction output is completed, the constructed model is abandoned and a new LWPLS regression model is established when a new query sample arrives.

[0184] It is worth noting that the window size in the MW-LWPLS submodel and the local region size in the JITL-LWPLS submodel should not exceed spn to prevent model failure.

[0185] The optimal values ​​of the window size in the MW-LWPLS submodel and the local region size in the JITL-LWPLS submodel also need to be determined through trial and error.

[0186] When the overall dataset is used as a sub-training set for the three adaptive regression sub-models, it is necessary not only to consider covering all characteristics such as nonlinearity, time variation, and mutation, but also to adapt to the unique characteristics of the three adaptive regression sub-models MW-LWPLS, TD-LWPLS, and JITL-LWPLS.

[0187] For the sub-training set of the MW-LWPLS adaptive regression sub-model, the time-varying features dominate;

[0188] For the sub-training set of TD-LWPLS adaptive regression sub-model, the mutation features are dominant;

[0189] For the JITL-LWPLS adaptive regression sub-model, the nonlinear characteristics of the sub-training set are dominant.

[0190] Therefore, in the embodiment of the present invention, firstly, the overall data set is used as the sub-training set of the three adaptive regression sub-models of MW-LWPLS, TD-LWPLS and JITL-LWPLS, so that the sub-training sets of the three adaptive regression sub-models can cover all features;

[0191] Then, by updating the sub-training sets, the sub-training sets of the three adaptive regression sub-models can respectively obtain data of corresponding features.

[0192] S5. Extract data samples from the correction data set in sequence, verify each adaptive regression sub-model, calculate the corresponding estimation error, add the current data sample to the sub-training set of the adaptive regression sub-model with the smallest estimation error value, and update the corresponding adaptive regression sub-model.

[0193] Figure 6 A flow chart of a method for dividing a training data set according to an embodiment of the present invention is disclosed, such as Figure 6 As shown, the data samples of the correction data set are added to the three adaptive regression sub-models of MW-LWPLS, TD-LWPLS and JITL-LWPLS for training and prediction, and the corresponding estimation errors are calculated and sorted in ascending order.

[0194] The estimated error is the error between the predicted value and the actual value.

[0195] The data sample with the smallest estimated error value among the three adaptive regression sub-models is added to the end of the sub-training set corresponding to the adaptive regression sub-model.

[0196] The estimated error value is used as a performance indicator to describe the model's ability to handle process characteristics. The smaller the error value, the better the corresponding adaptive regression sub-model can improve the interpretability of the corresponding specific features.

[0197] Therefore, the current data sample points with specific characteristics explain the adaptive regression sub-model with the prediction dependency with the minimum estimation error value, and these data sample points are added to the sub-training set of the corresponding adaptive regression sub-model.

[0198] S6. Apply the evaluation data set to each adaptive regression sub-model, calculate the root mean square error of the predicted value of each adaptive regression sub-model, use Bayesian estimation to determine the weight coefficient of each adaptive regression sub-model, and establish an integrated model.

[0199] Apply the evaluation samples of the evaluation dataset to each adaptive regression sub-model.

[0200] The root mean square error (RMSE) of each predicted value is calculated, and the weight coefficient is determined using the Bayesian estimation method to establish an integrated model to predict the output.

[0201] Assume that there are I different adaptive regression sub-models, denoted as M1, M2, ..., M I . In this embodiment, I is 3.

[0202] The current process state is represented by C, then in the current process state C is the i-th adaptive regression submodel M i The probability of P(M i |C), which can be used as the weight value of the fusion prediction output estimated by each adaptive regression sub-model.

[0203] In order to better illustrate the integrated model, M i and C are meaningful characters.

[0204] Then, the output of the integrated model y p , the corresponding expression is:

[0205]

[0206] Among them, y p，i represents the predicted output value of the i-th adaptive regression sub-model, P(M i |C) is the current process state C is the i-th adaptive regression sub-model M i probability.

[0207] P(M i |C) can be calculated by the following formula:

[0208]

[0209] Among them, P(C|M i ) is the i-th adaptive regression sub-model M i is the probability of the current process state C, P(M i ) is the prior probability of the i-th adaptive regression sub-model.

[0210] Assume that the prior probability of each adaptive regression sub-model is the same, that is,

[0211]

[0212] For a regression task, if a higher prediction accuracy is to be obtained, the root mean square error (RMSE) must be small, that is, the prediction performance is inversely proportional to the value of RMSE.

[0213] P(C|M i ) can be obtained by the following formula:

[0214]

[0215] Among them, r iThe corresponding expression is,

[0216]

[0217] The parameter p is an adjustment coefficient determined according to the actual application.

[0218] RMSE of the predicted value of the i-th adaptive regression sub-model i It is calculated by the following formula:

[0219]

[0220] Among them, y d Represents the actual value, y dp，i represents the predicted value of the i-th adaptive regression sub-model, and D represents the number of evaluation samples in the evaluation data set.

[0221] S7. Use the constructed integrated model to predict the final output value of the main variable based on the input auxiliary variables

[0222] In this embodiment, the constructed integrated model is used to predict the main variable octane number based on the test data set to obtain the final output.

[0223] The present invention proposes an integrated learning online prediction method for key parameters of a continuous catalytic reforming process. When performing soft measurement modeling of an integrated model, appropriate auxiliary variables and main variables are selected according to the mechanism of the continuous catalytic reforming process, and the collected process data sample set is preprocessed and divided into a training data set, an evaluation data set and a test data set. The training data set is divided and three adaptive regression sub-models are constructed, and the corresponding sub-training sets are updated so that they not only cover all the features of the three sub-models but also contain their own features respectively. The evaluation data set is applied to each adaptive sub-model to calculate the RMSE of each predicted value, and the Bayesian estimation is used to determine the weight coefficient, and an integrated model is established. Based on the test data set, the integrated model is used for prediction to obtain the final output.

[0224] The following is an example of predicting key parameters in a continuous catalytic reforming process to illustrate the integrated learning online prediction method based on the adaptive local weighted partial least squares method of the present invention. The specific steps include:

[0225] Step S1: Collect data samples of the continuous catalytic reforming process and select auxiliary variables and main variables.

[0226] The collected data samples include 84 input measurements and 3 output measurements, namely RON, C6 (%) and C 7+ (%), all measurements were taken every 30 minutes.

[0227] 84 input measurements were selected as auxiliary variables, including but not limited to recontact temperature, separator temperature, hydrocracking naphtha, naphtha, reactor temperature, separator pressure, circulating hydrogen flow rate, primary alkane content, compressor pressure, reflux flow rate, tray temperature, reflux flow rate, tray temperature, bottom temperature, and outlet temperature of reactor No. 1.

[0228] The octane number is selected as the main variable as the performance indicator. The octane number can be derived by the following formula:

[0229]

[0230] Among them, X Feed Represents the feed variable of the continuous catalytic reforming process.

[0231] Step S2: Data preprocessing.

[0232] Considering that there is measurement noise in the continuous catalytic reforming process, noise elimination is required.

[0233] A Gaussian smoothing filter is used to remove white noise, and the calculation is as follows:

[0234]

[0235] Among them, μ represents the mean of data x, and σ represents the standard deviation of data x.

[0236] The 3-sigma criterion is used to remove outliers and missing values.

[0237] Finally, 737 sample data (about half a month) were obtained.

[0238] Step S3: Divide the sample data into three parts: 350 samples are used as training data set, 50 samples are used as evaluation data set, and 337 samples are used as test data set.

[0239] Steps S4-S7, establish MW-LWPLS, TD-LWPLS, JITL-LWPLS sub-models, and establish the integrated model EA-LWPLS.

[0240] At the same time, the LWPLS model was established for comparison, and the number of latent variables in the LWPLS model was set to 5 and the weight coefficient was set to 1;

[0241] The window size in the MW-LWPLS submodel is set to 28;

[0242] For the integrated model EA-LWPLS, the spn and adjustment coefficient in the weight function are set to 280 and 0.5, respectively.

[0243] Among the three adaptive regression sub-models, the JITL-LWPLS sub-model with the JITL model has the most obvious improvement, with RMSE and R 2 are 0.417 and 0.819 respectively, and the goodness of fit R 2 It can be calculated by the following formula:

[0244]

[0245] Figure 7 The octane number prediction comparison chart of five models is revealed, such as Figure 7 As shown in the figure, the JITL model can better reflect the changing trend of the octane number, but it cannot accurately reflect the changes in the octane number.

[0246] Although the MW-LWPLS and TD-LWPLS sub-models cannot catch up with the JITL-LWPLS sub-model in reflecting the trend of changes, they can reflect the changes in the octane number during the process. The results of the TD-LWPLS sub-model are closer to the actual process than those of the MW-LWPLS sub-model, especially after the 300th sampling point. Although the TD-LWPLS sub-model cannot completely track the amplitude of the process, it reflects the fluctuation changes of the process.

[0247] However, between the 50th and 100th sampling points, when the actual octane number dropped to the minimum value, the predicted value of the TD-LWPLS sub-model did not change significantly but remained relatively stable.

[0248] Compared with the TD-LWPLS sub-model, the MW-LWPLS sub-model can better reflect the slowly changing process of the system.

[0249] Figure 7 Figure (e) shows that the integrated model proposed in this invention not only achieves the best prediction performance, but also has the best RMSE and R 2 They are 0.385 and 0.846 respectively, which also accurately reflects the true trend of octane number changes over time.

[0250] The present invention provides an integrated learning online prediction method for key parameters of a continuous catalytic reforming process, which effectively improves the prediction accuracy of the model in a continuous catalytic reforming process containing nonlinear, sudden change and time-varying characteristics.

[0251] 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.

[0252] As shown in this application and claims, unless the context clearly indicates an exception, the words "a", "an", "an" and / or "the" do not refer to the singular and may also include the plural. Generally speaking, the terms "include" and "comprise" only indicate the inclusion of the steps and elements that have been clearly identified, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements.

[0253] The above embodiments are provided for persons familiar with the art to implement or use the present invention. Personnel familiar with the art can make various modifications or changes to the above embodiments without departing from the inventive concept of the present invention. Therefore, the protection scope of the present invention is not limited to the above embodiments, but should be the maximum scope of the innovative features mentioned in the claims.

Claims

1. An integrated learning online prediction method for key parameters of a continuous catalytic reforming process, characterized in that: The following steps are involved: S1. Select auxiliary variables and main variables of the continuous catalytic reforming process, collect data to form a process data sample set, wherein the auxiliary variables reflect the operation status of the continuous catalytic reforming process, and the main variables reflect the quality of the finished product; S2. Preprocessing the collected process data sample set of the continuous catalytic reforming process to generate a data set; S3, dividing the data set into a training data set, an evaluation data set and a test data set; S4, select spn data samples from the training data set as the overall data set, and the remaining data samples as the correction data set, respectively use the overall data set as the sub-training set to construct three adaptive regression sub-models, the three adaptive regression sub-models include a moving window model based on the local weighted partial least squares method that reflects time-varying characteristics, a time difference model based on the local weighted partial least squares method that reflects mutation characteristics, and an instant learning model based on the local weighted partial least squares method that reflects nonlinear characteristics; S5, extracting data samples from the correction data set in sequence, verifying each adaptive regression sub-model, adding the current data sample to the sub-training set of the adaptive regression sub-model with the smallest estimated error value, and updating the corresponding adaptive regression sub-model; S6, applying the evaluation data set to each adaptive regression sub-model, calculating the root mean square error of the predicted value of each adaptive regression sub-model, using Bayesian estimation to determine the weight coefficient of each adaptive regression sub-model, and establishing an integrated model; S7. Use the constructed integrated model to predict the final output value of the main variable based on the input auxiliary variables; In step S1: Auxiliary variables, including recontact temperature, knockout drum temperature, hydrocracked naphtha, naphtha, reactor temperature, knockout drum pressure, recycle hydrogen flow rate, primary alkane content, compressor pressure, reflux flow rate, tray temperature, bottoms temperature, and reactor No. 1 outlet temperature; The main variable is the octane number, and the corresponding expression is as follows: Among them, Y RON , Y C6 and Y C7+ is the output measurement parameter of the continuous catalytic reforming process, X Feed is the feed variable of the continuous catalytic reforming process; In step S6, the root mean square error of the predicted value of each adaptive regression sub-model is calculated, and the expression is as follows: The root mean square error (RMSE) of the predicted value of the i-th adaptive regression sub-model i , Among them, y d Represents the actual value, y dp,i represents the predicted value of the i-th adaptive regression sub-model, and D represents the number of evaluation samples in the evaluation data set.

2. The integrated learning online prediction method for key parameters of the continuous catalytic reforming process according to claim 1 is characterized in that: The preprocessing method in step S2 further comprises: Remove outliers and missing values, and remove measurement noise.

3. The integrated learning online prediction method for key parameters of the continuous catalytic reforming process according to claim 1 is characterized in that: In step S4, based on the moving window model of the local weighted partial least squares method, the modeling process is as follows: S411, setting {X1, Y1}, i.e., window size H, as the training data matrix in the initial window, wherein X1 is the initial H input variables in the selected overall data set, and Y1 is the initial H output variables in the selected overall data set; S412, establishing a local weighted partial least squares regression model for the output of multiple sampling data points in the window; S413, let the window move forward by step size D, so that {X w ,Y w }Retrain the local weighted partial least squares regression model in the window, where X w is the H input variable after w D steps in the overall data set, Y w are the H output variables after w D steps in the overall data set.

4. The integrated learning online prediction method for key parameters of the continuous catalytic reforming process according to claim 1 is characterized in that: In step S4, the time difference model based on the local weighted partial least squares method has the following modeling process: S421, respectively calculating the first-order difference Δx(t) between the input variables and the first-order difference Δy(t) between the output variables of adjacent sampling data points; S422, constructing a relationship model between Δx(t) and Δy(t) using a local weighted partial least squares regression model; S423. Calculate query sample x q The first-order difference Δx of (t) q (t); S424, Δx q (t) Input the relationship model to predict the difference value of the response variable Δy p (t); S425, based on adjacent sampling data points y p (t-1) and the difference value Δy p (t), calculate the predicted value y of the response variable p (t), the expression is as follows, y p (t)=Δy p (t)+y p (t-1) 5. The integrated learning online prediction method for key parameters of the continuous catalytic reforming process according to claim 1 is characterized in that: In step S4, the instant learning model based on the local weighted partial least squares method has the following modeling process: S431, when traversing and inputting a query sample x q ∈R m When it arrives, calculate x q and x i The distance between Among them, there are n data sample points (x i ,y i ) i＝1～n The dataset consists of x∈R m and y∈R l Represent the input and output training sets respectively, x i represents the i-th input training value; S432, selecting N data samples with the smallest distance to form a related sample set; S433. Based on the relevant sample set, a local weighted partial least squares regression model is constructed; S434, through the LWPLS regression model, obtain the query sample x q The corresponding predicted value of the response variable.

6. The integrated learning online prediction method for key parameters of the continuous catalytic reforming process according to claim 1 is characterized in that: In step S5: The estimation error is the error between the predicted value and the actual value; Add the data sample with the smallest estimated error value to the corresponding sub-training set.

7. The integrated learning online prediction method for key parameters of the continuous catalytic reforming process according to claim 1 is characterized in that: The output result y of the integrated model in step S6 p , the corresponding expression is: Among them, y p,i represents the predicted output value of the i-th adaptive regression sub-model, C is the current process state, P(M i |C) is the i-th adaptive regression sub-model M in the current process state C i The probability of M1,M2,…,M I There are I different adaptive regression sub-models.

8. The integrated learning online prediction method for key parameters of the continuous catalytic reforming process according to claim 7 is characterized in that: In step S6, the current process state C is the i-th adaptive regression sub-model M i The probability P(M i |C), the corresponding expression is: Among them, P(M i ) is the prior probability of the i-th adaptive regression sub-model, P(C|M i ) is the i-th adaptive regression sub-model M i is the probability of the current process state C, p is the adjustment parameter, RMSE i is the root mean square error of the predicted value of the i-th adaptive regression sub-model; P(C|M i ) is, P(M i ) is, r i The corresponding expression of is,

Images