Intelligent hybrid modeling method for ebullated-bed residue hydroprocessing unit
By integrating data-driven and reaction mechanism approaches, a hybrid model for a fluidized bed residue hydrotreating unit was established, which solved the problem of low model prediction accuracy and achieved high-precision prediction when the properties of the feedstock change.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- EAST CHINA UNIV OF SCI & TECH
- Filing Date
- 2023-05-22
- Publication Date
- 2026-05-15
AI Technical Summary
In existing technologies, the model prediction accuracy of fluidized bed residue hydrotreating units is low, especially when the feedstock properties change, and there is a lack of effective parameter optimization methods.
A data-driven and reaction mechanism-integrated approach is adopted. A data model is established using a TCN-LSTM network and the Harris Eagle optimization algorithm. Combined with the mechanism model, an adaptive weighting coefficient is used to fuse the data and form a hybrid model.
It improves the prediction accuracy and stability of the model, has strong extrapolation and interpolation capabilities, and can maintain high-precision prediction when the properties of raw materials change.
Smart Images
Figure CN116595880B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial plant modeling technology, and more specifically, to an intelligent hybrid modeling method for fluidized bed residue hydrotreating units that integrates data-driven approaches and reaction mechanisms. Background Technology
[0002] The trend of crude oil becoming heavier and of lower quality is inevitable, and the demand for heavy fuels is gradually decreasing. Residue oil hydrotreating technology has become a focus of attention and investment for refineries because it can achieve clean and efficient conversion of residue oil.
[0003] Fluidized bed residue hydrotreating units are a relatively new process that has emerged in recent years, and research on their models is still limited. Fluidized bed residue hydrotreating units can handle a wide range of feedstock sources and offer advantages such as long service life, online catalyst addition and removal, and flexible operation, significantly improving feedstock conversion rates.
[0004] Data-driven models have higher accuracy. They are machine learning models based on observational data that discover the correlation between input and output, and have stronger universality. However, when the input conditions (such as the properties of raw materials) change greatly, the accuracy of data-driven models will decrease significantly, or even the predicted trend will be completely opposite.
[0005] Mechanism models are precise mathematical models built upon process mechanisms and procedures. Their parameters have very clear physical meanings and are highly stable, capable of handling scenarios with significant changes in operating conditions. However, their accuracy is not as good as that of data-driven models.
[0006] In the past, the parameters of data-driven models were usually determined by random or exhaustive methods. This method of finding parameters is inefficient and highly dependent on experience.
[0007] Therefore, there is currently no model for fluidized bed residue hydrotreating units that can accurately predict the product yield of such units. Summary of the Invention
[0008] The purpose of this invention is to provide an intelligent mixing modeling method for fluidized bed residue hydrotreating units, thereby solving the problem of low prediction accuracy in existing fluidized bed residue hydrotreating unit models.
[0009] To achieve the above objectives, this invention provides an intelligent mixing modeling method for fluidized bed residue hydrotreating units, comprising the following steps:
[0010] Step S1: Obtain data from the fluidized bed residue hydrotreating unit;
[0011] Step S2: Preprocess the acquired data;
[0012] Step S3: Establish a data model for the fluidized bed residue hydrogenation unit;
[0013] Step S4: Establish a mechanism model for a fluidized bed residue hydrotreating unit;
[0014] Step S5: Based on adaptive weighting coefficients, the data model and mechanism model of the fluidized bed residue hydrotreating unit are fused to obtain a hybrid model.
[0015] Step S6: Predict product yield based on a hybrid model.
[0016] In one embodiment, the data for the fluidized bed residue hydrotreating unit in step S1 includes the operating conditions of the fluidized bed residue hydrotreating unit, feedstock information, and product yield.
[0017] In one embodiment, the data preprocessing in step S2 further includes:
[0018] The acquired data undergoes outlier handling, normalization, and dataset partitioning.
[0019] In one embodiment, the outlier handling further includes:
[0020] An anomaly detection method using a smooth temporal window is employed for anomaly identification and processing.
[0021] The expression corresponding to anomaly identification is as follows:
[0022]
[0023] Among them, data raw For the original data, filter avg The filter is the average value within the sliding window. std Here, is the standard deviation within the sliding window, and threshold is the threshold value.
[0024] The expression corresponding to outlier smoothing is as follows:
[0025] data smoothed (i) = influence * data raw +(1-influence)*data smoothed (i-1);
[0026] Among them, data raw The original data, data smoothed The data is smoothed, and influence is the smoothing coefficient.
[0027] In one embodiment, step S3 further includes:
[0028] Step S31: Establish a TCN-LSTM network;
[0029] Step S32: Optimize and obtain hyperparameters based on the whale algorithm and differential evolution algorithm;
[0030] Step S33: Establish a data model for the fluidized bed residue hydrogenation unit based on the optimized hyperparameters.
[0031] In one embodiment, the TCN-LSTM network in step S31 further includes an input layer, a TCN layer, a fully connected layer, an LSTM layer, a DROUT layer, and an output layer:
[0032] The input layer inputs the time series of data from the pre-processed fluidized bed residue hydrotreating unit.
[0033] The TCN layer performs feature extraction and time-series analysis on the input data;
[0034] The fully connected layer maps the feature space calculated by the TCN layer to the LSTM layer;
[0035] The LSTM layer further calculates and extracts the features extracted by the TCN layer.
[0036] The DROUT layer will cause some neurons to stop working with a certain probability;
[0037] The output layer outputs the yield of each product and the operating conditions of the device.
[0038] In one embodiment, step S32 further includes:
[0039] Step S321: Initialize parameters;
[0040] Step S322: Initialize the population;
[0041] Step S323: Fitness calculation. The individual position corresponds to the hyperparameters of the TCN-LSTM network. The preprocessed data and network hyperparameters are fed into the TCN-LSTM network for training. The test set error is calculated and used as the objective function of the whale algorithm and the differential evolution optimization algorithm.
[0042] Step S324: Position Update. Update each position in the population and determine whether the training termination condition is met. If it is met, output the optimal individual position; otherwise, return to step S323.
[0043] In one embodiment, the position update expression in step S324 is as follows:
[0044]
[0045] Where Leader_pos is the current optimal position;
[0046] Leader_pos' is the position after mutation by the differential evolution algorithm;
[0047] cr represents the crossover probability;
[0048] r is a random number;
[0049] A is a random vector;
[0050] D represents the distance between the individual's location and the location of a random individual in the population;
[0051] D p In the whale algorithm, this represents the distance between the individual and the prey.
[0052] l is a random number between -1 and 1;
[0053] b is a constant used to define the shape of the logarithmic spiral.
[0054] ε is the defined threshold.
[0055] In one embodiment, step S4 further includes:
[0056] Step S41: Divide the reaction system of the fluidized bed residue hydrogenation unit into substances;
[0057] Step S42: Set assumptions based on the reaction mechanism and process flow;
[0058] Step S43: Establish a mechanism model for a fluidized bed residue hydrotreating unit;
[0059] Step S44: Solve to obtain the model parameters.
[0060] In one embodiment, the mechanism model of the fluidized bed residue hydrotreating unit in step S43 is expressed as follows:
[0061]
[0062] Where k is the rate constant of the corresponding conversion reaction process;
[0063] C i The content of each component;
[0064] The time-varying function of catalyst deactivation;
[0065] t c This refers to the residence time of the catalyst.
[0066] In one embodiment, step S44 further includes:
[0067] The model parameters were obtained by using the Harris Eagle optimization algorithm and the fourth-order Runge-Kutta method.
[0068] In one embodiment, step S5 further includes:
[0069] Step S51: Calculate the distance between the input data and the training samples, as shown in the following expression:
[0070]
[0071] Among them, sample i It is a vector containing the input variables;
[0072] Sample is the training dataset, and ||·|| is the Euclidean distance;
[0073] sample i.min It is the training dataset and the sample i The average distance to the nearest neighbor;
[0074] Step S52: Perform Latin hypercube sampling in the input space to obtain virtual input. The upper and lower limits of sampling are mean(Sample) ± 3*standard_deviation(Sample), and the number of samples is N, where N is the number of sample data in the training dataset.
[0075] Step S53: Calculate the adaptive weight coefficients, the corresponding expressions are as follows:
[0076]
[0077] k D =1-k M ;
[0078] Where, k M k D These are the weight coefficients for the mechanistic model and the data model, respectively.
[0079] input is the vector of samples to be predicted;
[0080] Samples were obtained by Latin hypercube sampling.
[0081] In one embodiment, after step S5, the method further includes:
[0082] Develop evaluation metrics to validate the constructed model.
[0083] This invention proposes an intelligent hybrid modeling method for a fluidized bed residue hydrotreating unit. By using adaptive weighting coefficients, the established mechanistic model and data model of the fluidized bed residue hydrotreating unit are fused to obtain a hybrid model. Comparative experiments demonstrate that the established hybrid model has strong extrapolation and interpolation capabilities, strong model stability, and high prediction accuracy. Attached Figure Description
[0084] The above and other features, properties and advantages of the present invention will become more apparent from the following description taken in conjunction with the accompanying drawings and embodiments, in which the same reference numerals always denote the same features, wherein:
[0085] Figure 1 A flowchart of an intelligent mixing modeling method for a fluidized bed residue hydrotreating unit according to an embodiment of the present invention is disclosed;
[0086] Figure 2 A process flow diagram of a fluidized bed residue hydrogenation unit according to an embodiment of the present invention is disclosed;
[0087] Figure 3 A flowchart of a data model construction method for a TCN-LSTM fluidized bed residue hydrotreating unit based on WOA-DE optimization according to an embodiment of the present invention is disclosed;
[0088] Figure 4 A flowchart illustrating the fusion of the mechanism model and data model according to an embodiment of the present invention is disclosed;
[0089] Figure 5 A comparison diagram of the effects of various models according to an embodiment of the present invention is disclosed;
[0090] Figure 6 A graph showing the product yield fitting effect of a hybrid model according to an embodiment of the present invention is disclosed. Detailed Implementation
[0091] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.
[0092] Figure 1 A flowchart of a smart mixing modeling method for a fluidized bed residue hydrotreating unit according to an embodiment of the present invention is disclosed, as follows: Figure 1 As shown, the present invention proposes an intelligent mixing modeling method for a fluidized bed residue hydrotreating unit, which includes the following steps:
[0093] Step S1: Obtain data from the fluidized bed residue hydrotreating unit;
[0094] Step S2: Preprocess the acquired data;
[0095] Step S3: Establish a data model for the fluidized bed residue hydrogenation unit;
[0096] Step S4: Establish a mechanism model for a fluidized bed residue hydrotreating unit;
[0097] Step S5: Based on adaptive weighting coefficients, the data model and mechanism model of the fluidized bed residue hydrotreating unit are fused to obtain a hybrid model.
[0098] Step S6: Predict product yield based on a hybrid model.
[0099] The present invention proposes an intelligent hybrid modeling method for fluidized bed residue hydrotreating units that integrates data-driven approaches and reaction mechanism analysis. First, data from the fluidized bed residue hydrotreating unit is acquired based on expert experience and feature selection, and the data is preprocessed. Next, a TCN-LSTM network is established, and the network parameters are optimized using a DE-WOA parallel optimization algorithm. The optimized model is then trained to obtain a data model of the fluidized bed residue hydrotreating unit. Then, a mechanistic model of the fluidized bed residue hydrotreating unit is established based on its reaction mechanism and process flow. Finally, an adaptive weighting coefficient is obtained between the mechanistic model and the data model based on the input and training data to obtain a hybrid model.
[0100] Figure 2 A process flow diagram of a fluidized bed residue hydrotreating unit according to an embodiment of the present invention is disclosed, such as... Figure 2 As shown, the main feedstocks of the fluidized bed residue hydrotreating unit are vacuum residue and hydrogen. The feedstocks are mixed and heated before entering the first and second reactors for reaction. After the reaction, the mixture is separated into low-grade gas and low-grade oil through hot and cold separation. The low-grade oil is then processed through a desulfurization stripping tower, an atmospheric distillation tower, and a vacuum distillation tower to produce diesel, wax oil, and other products. Some unconverted oil is discharged from the bottom of the vacuum separation tower.
[0101] The following text is in the format of Figure 2 The fluidized bed residue hydrotreating unit shown is an example, and these steps of the intelligent mixing modeling method for the fluidized bed residue hydrotreating unit proposed in this invention are described in detail. It should be understood that, within the scope of this invention, the above-mentioned technical features of this invention and the technical features specifically described below (such as in the examples) can be combined and correlated with each other to constitute preferred technical solutions.
[0102] Step S1: Obtain data from the fluidized bed residue hydrogenation unit.
[0103] Data for fluidized bed residue hydrotreating units include operating conditions, feedstock information, and product yield.
[0104] By combining staff experience (expert experience) and feature selection methods, data such as operating conditions, feedstock information, and product yield of the fluidized bed residue oil hydrogenation unit are obtained;
[0105] A total of 32 data points were obtained through a combination of experience screening and feature selection, including operating conditions, raw material properties, and product yield. Some data obtained for the fluidized bed residue hydrotreating unit are shown in Table 1. The data are arranged in chronological order and collected every four hours.
[0106] The feature selection method can employ the classic Person correlation coefficient method, and the corresponding expression is shown below:
[0107]
[0108] In the formula, r is the correlation coefficient, and the magnitude of r is directly proportional to the correlation; n is the sample size; x i y i These are sample observations. It is the average value of the variable.
[0109] Table 1
[0110]
[0111] S2. Preprocess the acquired data.
[0112] The initial data is processed for outlier handling, normalization, and dataset partitioning.
[0113] In this embodiment, the dataset is divided according to the order of data collection, i.e., the chronological order. 70% of the 3,000 collected data are used as the training set, the remaining 30% as the test set, and 10% of the training set is used as the validation set.
[0114] The data at this point is simply arranged by time, and a sliding window needs to be constructed to serve as input to the model.
[0115] The processed sliding window data is shown in the following formula.
[0116]
[0117] Y = (y1 y2 … … y i ) T ;
[0118] Where n is the window size, is the parameter to be optimized, x is the feature vector, and y corresponding to each sliding window is the yield of each product at the corresponding time point.
[0119] The outlier handling of the initial data further includes:
[0120] An anomaly detection method using a smooth temporal window is employed for anomaly identification and processing.
[0121] This outlier handling method uses a sliding window of a certain length. Within this window, the mean and standard deviation are calculated, the calculated data are compared with a set threshold, and finally, the data is smoothed.
[0122] Outlier identification methods calculate the mean and variance within a sliding window (Lag), and identify outliers if the following conditions are met.
[0123] The expression corresponding to anomaly identification is as follows:
[0124]
[0125] In the formula, data raw For the original data, filter avg The filter is the average value within the sliding window. std Let be the standard deviation within the sliding window, and threshold be the threshold value. Preferred, threshold is set to 2.
[0126] The expression corresponding to outlier smoothing is as follows:
[0127] data smoothed (i) = influence * data raw +(1-influence)*data smoothed (i-1);
[0128] Among them, data raw The original data, data smoothed The smoothed data is represented by a smoothing coefficient of 0 to 1. The larger the coefficient, the greater the influence on the original data. Preferably, the influence coefficient is 0.5.
[0129] When the data is a normal value, its smoothed value is the original value. If it is an outlier, it is smoothed according to the formula above.
[0130] Normalization involves normalizing all variables to the interval 0 to 1. The normalization formula is as follows:
[0131]
[0132] Where x represents the data before normalization, and X represents the data after normalization. min x is the minimum value of the data before normalization. max This represents the maximum value of the data before normalization.
[0133] S3. Establish a data model for a fluidized bed residue oil hydrogenation unit.
[0134] Previous literature reports have shown that scholars have applied other optimization algorithms, such as genetic algorithms, to the parameter optimization of weights and thresholds in BP neural networks.
[0135] This invention uses the Whale Algorithm and the Differential Evolution (WOA-DE) algorithm to optimize the hyperparameters of the TCN-LSTM (Temporal Convolutional Network-Long Short-Term Memory Artificial Neural Network) network, such as the sliding window size, number of filters, number of kernels, batch size, number of holes, and number of units. Based on the optimized hyperparameters, the TCN-LSTM network is trained to establish a data model of a fluidized bed residue hydrotreating unit.
[0136] Figure 3 A flowchart illustrating a method for constructing a data model of a TCN-LSTM fluidized bed residue hydrotreating unit based on WOA-DE optimization according to an embodiment of the present invention is disclosed, as follows: Figure 3 As shown, the method for constructing a data model for a TCN-LSTM fluidized bed residue hydrotreating unit based on the WOA-DE algorithm optimization further includes the following steps:
[0137] Step S31: Establish a TCN-LSTM network;
[0138] Step S32: Optimize and obtain hyperparameters based on the whale algorithm and differential evolution algorithm;
[0139] Step S33: Establish a data model for the fluidized bed residue hydrogenation unit based on the optimized hyperparameters.
[0140] Furthermore, the TCN-LSTM network further includes an input layer, a TCN layer, a fully connected layer, an LSTM layer, a DROUT layer, and an output layer:
[0141] The input layer takes a time series of m*n fluidized bed residue hydrotreating unit data after preprocessing, where m is the number of feature variables and n is the size of the sliding time window. The feature variables include the feed flow rate (conditions) of the fluidized bed residue hydrotreating unit and the operating conditions of the unit, totaling 16 sets of feature variables.
[0142] The TCN layer performs feature extraction and temporal analysis on the input data. The TCN consists of a 1*1 convolutional block and two residual blocks in parallel. The residual block is composed of dilated causal convolution, weight normalization, activation function, and DROUOUT.
[0143] The fully connected layer maps the feature space calculated by TCN onto the LSTM layer.
[0144] The LSTM layer further calculates and extracts the features extracted by TCN.
[0145] The DROUT layer allows some neurons to stop working with a certain probability, preventing the model from overfitting. In this embodiment, the activation function is REUL.
[0146] The output layer outputs the yield of each product and the operating conditions of the unit, including wax oil, diesel oil, naphtha, and low-grade gas.
[0147] After determining the TCN-LSTM network structure, it is necessary to determine the optimal values of its parameters. Here, the WOA-DE parallel algorithm is used to optimize the network hyperparameters.
[0148] Furthermore, step S32, which optimizes the hyperparameters based on the WOA-DE algorithm, further includes the following steps:
[0149] Step S321, Initialize parameters for the WOA-DE algorithm:
[0150] Set initial parameters, including population size, maximum number of iterations, and training termination conditions;
[0151] The training termination condition is that optimization can be stopped when the results of two consecutive optimizations are less than a set threshold (ε) and this condition is maintained for more than two consecutive optimizations.
[0152] In this embodiment, the initialization parameters of the WOA-DE algorithm are shown in the table below:
[0153] Table 2
[0154] parameter numerical values parameter numerical values Population size (NP) 100 Variation rate (F) 0.2 Number of iterations (Gm) 130 Crossover rate (Cr) 0.8 Number of variables (N) 6 Threshold (ε) 0.001
[0155] Step S322: Initialize the population:
[0156] Initialize the WOA-DE algorithm population, wherein the population initialization adopts the opposite search operator;
[0157] P i =Lb+(Ub-Lb)*rand;
[0158] Where Lb and Ub are the upper and lower bounds of the feasible solution, i.e. the range of values of the hyperparameters, and rand is a random number between 0 and 1.
[0159] Step S323, Fitness Calculation:
[0160] Individual locations are used as hyperparameters of the TCN-LSTM network. Preprocessed data and network hyperparameters are fed into the network for training. The test set error is calculated and used as the objective function of the WOA-DE optimization algorithm.
[0161] In this embodiment, the error used is the root mean square error (RMSE), and the corresponding calculation formula is as follows:
[0162]
[0163] in, y is the model's predicted value. i The actual values are used to select the individual with the best fitness as the optimal individual.
[0164] Step S324, WOA-DE optimization algorithm position update:
[0165] Update each position in the population and determine whether the training termination condition is met. When the training termination condition is met, stop the optimization and output the optimal individual position; otherwise, return to step S323.
[0166] The network parameters are optimized using a WOA-DE parallel structure. Combining the local search capability of WOA in the later stages with the global search capability of the DE algorithm in the early stages, the corresponding expression for position update is as follows:
[0167]
[0168] Where Leader_pos is the current optimal position, Leader_pos' is the position after mutation by the DE algorithm, cr is the crossover probability, r is a random number between 0 and 1, A is a random vector between -2 and 2, and D is the distance of the individual's position from the position of a random individual in the population. p is the distance between an individual and its prey in WOA, l is a random number between -1 and 1, and b is a constant used to define the shape of the logarithmic spiral.
[0169] ε is a defined threshold used to represent the proportion of accuracy improvement in each iteration. That is, when the proportion of accuracy improvement is greater than ε, the DE algorithm is used to optimize the population; otherwise, the WOA algorithm is used to optimize the population.
[0170] Furthermore, after optimization using the WOA-DE algorithm, step S33 yields the following optimal parameters: a sliding window of 10, a filter count of 64, a training batch size of 32, and an LSTM unit count of 30.
[0171] For the void parameter, it is generally set to a list that grows exponentially by 2 by default. In this embodiment, it is set to [1,2,4,8,16,32,64].
[0172] The optimized hyperparameters are set in the TCN-LSTM network, the model is trained, and the data model of the fluidized bed residue hydrotreating unit is obtained.
[0173] S4. Establishment of a mechanism model for a fluidized bed residue hydrotreating unit:
[0174] Step S4 further includes the following steps:
[0175] Step S41: Divide the reaction system of the fluidized bed residue hydrogenation unit into substances;
[0176] Step S42: Set assumptions based on the reaction mechanism and process flow;
[0177] Step S43: Establish a mechanism model for a fluidized bed residue hydrotreating unit;
[0178] Step S44: Solve to obtain the model parameters.
[0179] By classifying the reaction system of the fluidized bed residue hydrotreating unit into its constituent materials, and based on the reaction mechanism and process flow, and setting assumptions (considering catalytic deactivation), a mechanistic model of the fluidized bed residue hydrotreating unit was established. The model parameters were then solved using the Harris Eagle optimization algorithm and the fourth-order Runge-Kutta method.
[0180] The reaction system of the fluidized bed residue hydrotreating unit was divided into 11 components: feedstock, products (wax oil, diesel oil, naphtha, and low-grade gas), and a mechanism model of the fluidized bed residue hydrotreating unit was established.
[0181] The raw materials and wax oil are further divided into four components according to SARA, and the resulting mechanistic model expression is as follows:
[0182]
[0183] Where k is the rate constant of the corresponding conversion reaction process, and its expression is:
[0184] k0 and E a These are the pre-exponential factor and activation energy, respectively.
[0185] C i R represents the content of each component, R is the universality factor (8.314), and T is the reaction temperature;
[0186] Let be the time-varying function of catalyst deactivation, and its expression is:
[0187] t c denoted as the catalyst residence time, and β and y as catalyst deactivation constants.
[0188] Based on the reaction relationships between the components, k in the expression can be expanded as follows:
[0189]
[0190] The operating conditions of the equipment and the properties of the raw materials are shown in Table 3.
[0191] Table 3 Component Data and Device Operating Conditions
[0192]
[0193] Using the data in the table, combined with the Harris Eagle Algorithm (HHO) and the fourth-order Runge-Kutta method (ODE45), the mechanistic model parameters are obtained, and the parameter matrix k is as follows:
[0194]
[0195] The Harris Eagle algorithm is a swarm intelligence optimization algorithm proposed in 2019 by Heidari et al. from the University of Tehran. The Harris Eagle algorithm has strong optimization capabilities and does not require cumbersome parameter tuning.
[0196] Runge-Kutta methods are an important class of implicit or explicit iterative methods for solving nonlinear ordinary differential equations. These techniques were invented around 1900 by mathematicians Karl Runge and Martin Wilhelm Kutta.
[0197] Step S5: Based on adaptive weighting coefficients, the data model and mechanism model of the fluidized bed residue hydrotreating unit are fused to obtain a hybrid model.
[0198] Figure 4 A flowchart illustrating the fusion of the mechanism model and data model according to an embodiment of the present invention is disclosed, such as... Figure 4 As shown, the input data and training samples yield adaptive weighting coefficients for the mechanistic model and the data model. These coefficients are then combined in parallel to obtain a hybrid model that integrates the data model and the mechanistic model.
[0199] The fusion model is obtained by multiplying the mechanistic model and the data model by their respective weights and then summing them. The adaptive coefficient calculation steps are as follows:
[0200] Step S51: Calculate the distance between the input and the training samples, as shown in the following expression:
[0201]
[0202] Among them, sample i It is a vector containing the input variables;
[0203] Sample is the training dataset, and ||·|| is the Euclidean distance;
[0204] sample i.min It is the training dataset and the sample i The average distance to the nearest neighbor.
[0205] Step S52, Latin hypercube sampling:
[0206] The virtual input is obtained by performing Latin hypercube sampling in the input space, with upper and lower limits of mean(Sample)±3*standard_deviation(Sample) and the number of samples is N, where N is the number of sample data in the training dataset.
[0207] Step S53: Calculate the adaptive weight coefficients. The corresponding expression is as follows:
[0208]
[0209] k D =1-k M ;
[0210] Where, k M k D These are the weight coefficients for the mechanistic model and the data model, respectively.
[0211] input is the vector of samples to be predicted;
[0212] Samples were obtained by Latin hypercube sampling.
[0213] Step S6: Predict product yield based on a hybrid model.
[0214] Furthermore, after step S5 or step S6, the method further includes:
[0215] Develop evaluation metrics to validate the constructed model.
[0216] The hybrid model is compared with data models and mechanistic models, and the consistency correlation coefficient is used to evaluate the extrapolation and interpolation capabilities of the model.
[0217] Taking into account both the model's accuracy and sensitivity, a new evaluation index is constructed to compare the hybrid model with mechanistic and data models, verifying its extrapolation and interpolation capabilities.
[0218] In this embodiment, the consistency correlation coefficient is used to evaluate the model, which reflects both the correlation and the error value. The calculation formula is as follows:
[0219]
[0220] Where, σ x σ y These are the standard deviations of the model's predicted values and the actual values, respectively.
[0221] μ x μ y These are the average of the model's predicted values and the average of the actual values, respectively.
[0222] ρ represents the correlation coefficient between the model's predicted values and the actual values. The larger the value, the stronger the correlation. C The larger the value, the better the model's performance.
[0223] In this embodiment, 200 samples were used to make predictions using a data model, a mechanistic model, and a hybrid model, respectively. The consistency correlation coefficient was used to evaluate the models, and the calculated model indices are shown in Table 4.
[0224] Table 4. Correlation coefficients for model consistency
[0225]
[0226]
[0227] It is easy to see from the above evaluation indicators that the consistency correlation coefficient of the hybrid model is closer to 1, and the stability and accuracy of the model are better than those of the mechanistic model and the data model.
[0228] Figure 5 A comparison chart of the effects of various models according to an embodiment of the present invention is disclosed. Taking the prediction of low-part gas yield as an example, the fitting effects of the three models are compared, such as... Figure 5 As shown, the mechanistic model is very stable, but its accuracy is not high. The data model has high accuracy, but its local fluctuations are large. The hybrid model, on the other hand, performs well in both accuracy and stability.
[0229] Figure 6 A graph showing the product yield fitting effect of a hybrid model according to an embodiment of the present invention is disclosed, such as... Figure 6 The prediction results of some variables in the hybrid model shown can be seen that the hybrid model has a good effect on the yield prediction of various products, and the model accuracy and stability are relatively good.
[0230] This invention proposes an intelligent hybrid modeling method for a fluidized bed residue hydrotreating unit. By using adaptive weighting coefficients, a hybrid model is obtained by fusing the established mechanistic model and data model of the fluidized bed residue hydrotreating unit. This method combines the accuracy of the data-driven model with the stability of the mechanistic model. Comparative experiments demonstrate that the established hybrid model has strong extrapolation and interpolation capabilities, high model stability, and high prediction accuracy.
[0231] Although the methods described above are illustrated and depicted as a series of actions for the sake of simplicity, it should be understood and appreciated that these methods are not limited by the order of the actions, as some actions may occur in a different order and / or concurrently with other actions from the illustrations and descriptions herein or not illustrated and described herein but which may be understood by those skilled in the art, according to one or more embodiments.
[0232] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.
[0233] The above embodiments are provided for those skilled in the art to implement or use the present invention. Those skilled in 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 that conforms to the innovative features mentioned in the claims.
Claims
1. A method for intelligent mixing modeling of a fluidized bed residue hydrotreating unit, characterized in that, Includes the following steps: Step S1: Obtain data from the fluidized bed residue hydrotreating unit; Step S2: Preprocess the acquired data; Step S3: Establish a data model for the fluidized bed residue hydrogenation unit; Step S4: Establish a mechanism model for a fluidized bed residue hydrotreating unit; Step S5: Based on adaptive weighting coefficients, the data model and mechanism model of the fluidized bed residue hydrotreating unit are fused to obtain a hybrid model. Step S6: Predict product yield based on a hybrid model; Step S3 further includes: Step S31: Establish a TCN-LSTM network; Step S32: Optimize and obtain hyperparameters based on the whale algorithm and differential evolution algorithm; Step S33: Establish a data model for the fluidized bed residue hydrotreating unit based on the optimized hyperparameters; The TCN-LSTM network in step S31 further includes an input layer, a TCN layer, a fully connected layer, an LSTM layer, a DROUT layer, and an output layer: The input layer inputs the time series of data from the pre-processed fluidized bed residue hydrotreating unit. The output layer outputs the yield of each product and the operating conditions of the device. Step S4 further includes: Step S41: Divide the reaction system of the fluidized bed residue hydrogenation unit into substances; Step S42: Set assumptions based on the reaction mechanism and process flow; Step S43: Establish a mechanism model for a fluidized bed residue hydrotreating unit; Step S44: Solve to obtain the model parameters; The mechanism model of the fluidized bed residue hydrotreating unit in step S43 is expressed as follows: ； in, k Let be the rate constant for the corresponding conversion reaction process, expressed as: ; and These are the pre-exponential factor and activation energy, respectively. C i For the content of each component, R As a universal factor, T The reaction temperature; Let be the time-varying function of catalyst deactivation, expressed as: ; For catalyst residence time, , This is the catalyst deactivation constant; Based on the reaction relationships between the components, the expression... Expanded to: 。 2. The intelligent mixing modeling method for a fluidized bed residue hydrotreating unit according to claim 1, characterized in that, The data for the fluidized bed residue hydrogenation unit in step S1 includes the operating conditions of the fluidized bed residue hydrogenation unit, feedstock information, and product yield.
3. The intelligent mixing modeling method for a fluidized bed residue hydrotreating unit according to claim 1, characterized in that, The data preprocessing in step S2 further includes: The acquired data undergoes outlier handling, normalization, and dataset partitioning.
4. The intelligent mixing modeling method for a fluidized bed residue hydrotreating unit according to claim 3, characterized in that, The outlier handling further includes: An anomaly detection method using a smooth temporal window is employed for anomaly identification and processing. The expression corresponding to anomaly identification is as follows: ； in, The original data, This is the average value within the sliding window. The standard deviation within the sliding window. For threshold; The expression corresponding to outlier smoothing is as follows: ； in, The original data, For the smoothed data, influence This is the smoothing coefficient.
5. The intelligent mixing modeling method for a fluidized bed residue hydrotreating unit according to claim 1, characterized in that, The TCN layer performs feature extraction and time-series analysis on the input data; The fully connected layer maps the feature space calculated by the TCN layer to the LSTM layer; The LSTM layer further calculates and extracts the features extracted by the TCN layer. The DROUT layer will cause some neurons to stop working with a certain probability.
6. The intelligent mixing modeling method for a fluidized bed residue hydrotreating unit according to claim 1, characterized in that, Step S32 further includes: Step S321: Initialize parameters; Step S322: Initialize the population; Step S323: Fitness calculation. The individual position corresponds to the hyperparameters of the TCN-LSTM network. The preprocessed data and network hyperparameters are fed into the TCN-LSTM network for training. The test set error is calculated and used as the objective function of the whale algorithm and the differential evolution optimization algorithm. Step S324: Position Update. Update each position in the population and determine whether the training termination condition is met. If it is met, output the optimal individual position; otherwise, return to step S323.
7. The intelligent mixing modeling method for a fluidized bed residue hydrotreating unit according to claim 6, characterized in that, In step S324, the position update expression is as follows: ； in, This is the current optimal position; This represents the position after mutation using the differential evolution algorithm; The crossover probability; It is a random number; A It is a random vector; D The distance between an individual's location and the location of a random individual in the population; In the whale algorithm, this represents the distance between the individual and the prey. l A random number between -1 and 1; b This is a constant used to define the shape of the logarithmic spiral; The threshold is defined.
8. The intelligent mixing modeling method for a fluidized bed residue hydrotreating unit according to claim 1, characterized in that, Step S44 further includes: The model parameters were obtained by using the Harris Eagle optimization algorithm and the fourth-order Runge-Kutta method.
9. The intelligent mixing modeling method for a fluidized bed residue hydrotreating unit according to claim 1, characterized in that, Step S5 further includes: Step S51: Calculate the distance between the input data and the training samples, as shown in the following expression: ； in, It is a vector containing the input variables; For the training dataset, Euclidean distance; It is the training dataset and The average distance to the nearest neighbor; Step S52: Perform Latin hypercube sampling in the input space to obtain the virtual input, with upper and lower sampling limits as follows: The number of samples is N, where N is the number of samples in the training dataset; Step S53: Calculate the adaptive weight coefficients, the corresponding expressions are as follows: ； ； in, , These are the weight coefficients for the mechanistic model and the data model, respectively. input The vector of samples to be predicted; Samples were obtained by Latin hypercube sampling.
10. The intelligent mixing modeling method for a fluidized bed residue hydrotreating unit according to claim 1, characterized in that, Following step S5, the following further steps are included: Develop evaluation metrics to validate the constructed model.