Catalytic cracking device product and energy consumption prediction method

Through variational modal decomposition and two-stage self-attention model processing of the data of the catalytic cracking device, the problem of low energy consumption prediction accuracy is solved, the prediction accuracy is improved, and it is suitable for complex catalytic cracking processes.

CN118969112AInactive Publication Date: 2024-11-15EAST CHINA UNIV OF SCI & TECH
View PDF 7 Cites 0 Cited by

Patent Information

Application Number
CN202410957572.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-17
Publication Date
2024-11-15
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

There is a problem of low energy consumption prediction accuracy in catalytic cracking devices, which is mainly due to the nonlinear and spatiotemporal characteristics of industrial data and the great influence of external environmental factors, resulting in inaccurate prediction of product yield and energy consumption process.

Method used

Variable modal decomposition and hyperparameter grid search modal decomposition are used to modal decomposition of predictor variables, combining the two-stage self-attention model of characteristic attention mechanism and temporal attention mechanism to predict the predictive components and process the spatiotemporal characteristics between data. Influence, improve prediction accuracy through error compensation.

Benefits of technology

It significantly improves the prediction accuracy of product yield and energy consumption processes in catalytic cracking devices, solves the adverse effects of nonlinear and spatiotemporal characteristics on data prediction, and has good application value.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118969112A_ABST
    Figure CN118969112A_ABST
Patent Text Reader

Abstract

The invention discloses a catalytic cracking device product and energy consumption prediction method, which comprises the following steps of: firstly, acquiring production process data of a catalytic cracking device based on a preset time period, and preprocessing to obtain an original data set; performing modal decomposition on prediction variables in the original data set by combining variational modal decomposition and hyper-parameter grid search to obtain a plurality of prediction components, and reconstructing a prediction data set of each prediction component according to the prediction components and influence variables in the original data set; performing prediction and result summation on each prediction data set by adopting a two-stage self-attention model based on a feature attention and time attention mechanism to obtain a preliminary prediction value; and finally, comparing the preliminary predicted value with an actual value, combining an error item with an influence variable in an original data set to obtain an error data set, predicting the error data set in combination with variational mode decomposition and a double-stage self-attention model, and compensating the obtained error predicted value into the preliminary predicted value to obtain a final prediction result.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the field of petrochemical production process modeling, and in particular to a method for predicting products and energy consumption of a catalytic cracking unit. Background Art

[0002] In the oil refining industry, catalytic cracking process is a key technology in oil refining. The corresponding catalytic cracking unit is a key device for converting crude oil into high-value oil products and chemical raw materials efficiently, environmentally friendly and economically. It mainly includes four units: reaction regeneration unit, product separation unit, product refining unit and flue gas treatment unit. The products generated include slurry oil, diesel, gasoline, liquefied petroleum gas and dry gas. Among them, diesel and gasoline are its main products. The energy consumption in the catalytic cracking process involves most units, mainly composed of catalyst coking, water consumption, fuel gas consumption, steam heat loss, hot feed heat loss and power consumption of various devices in reaction regeneration. The catalytic cracking unit has the characteristics of large processing capacity, large heat release and large heat loss, and is one of the main energy-consuming devices in the oil refining industry. Therefore, the catalytic cracking process is not only crucial for the production of high-value products, but also of great significance for improving energy utilization efficiency.

[0003] With the continuous development of technology, in order to pursue higher economic benefits, catalytic cracking units have become more and more complex. This complexity has led to increasingly complex industrial data, and the relationship between industrial data variables is nonlinear. In addition, although the chemical reactions in the catalytic cracking unit occur continuously in the time domain, there is a certain time delay in the measurement process, and different reactions occur at different locations in the device, resulting in mutual dependence between measurements at different locations, which gives industrial data spatiotemporal characteristics.

[0004] In addition, the energy consumption in the catalytic cracking unit is not only affected by the internal environment of the equipment, but also by external factors such as seasonal changes. These influences lead to low prediction accuracy for product yield and energy consumption in industrial processes. Summary of the invention

[0005] A brief summary of one or more aspects is given below to provide a basic understanding of these aspects. This summary is not an exhaustive overview of all conceived aspects, and is neither intended to identify the key or critical elements of all aspects nor to define the scope of any or all aspects. Its only purpose is to give some concepts of one or more aspects in a simplified form as a prelude to a more detailed description that will be given later.

[0006] The purpose of the present invention is to solve the above problems. A method for predicting products and energy consumption of a catalytic cracking unit is provided. The method uses variational mode decomposition and hyperparameter grid search to perform mode decomposition on the prediction variables in the original data set, and combines the prediction components with the influencing variables of the original data set to establish a data set of multiple prediction components to solve the impact of the nonlinearity between the prediction variables and the influencing variables on the prediction. A two-stage self-attention model based on feature attention mechanism and time attention mechanism is used to predict the prediction data set of each prediction component respectively, which solves the adverse effects of the high spatiotemporal characteristics between the production process data on data prediction, and has good application value.

[0007] The technical solution of the present invention is:

[0008] The present invention provides a method for predicting products and energy consumption of a catalytic cracking unit, comprising the following steps:

[0009] Step S1: collecting production process data of the catalytic cracking unit based on a preset time period, and preprocessing the acquired production process data to obtain an original data set based on a time series; wherein the original data set includes prediction variables corresponding to product prediction and energy consumption prediction and influencing variable data corresponding to each prediction variable;

[0010] Step S2: Using variational mode decomposition and hyperparameter grid search to perform mode decomposition on the prediction variables in the original data set to obtain multiple prediction components;

[0011] Step S3: reconstructing the prediction data set of each prediction component according to the prediction components obtained after decomposition and the corresponding influencing variable data in the original data set;

[0012] Step S4: using a two-stage self-attention model based on a feature attention mechanism and a temporal attention mechanism to predict the prediction data set of each prediction component respectively, and obtaining the prediction result of each prediction component;

[0013] Step S5: summing up the prediction results of each prediction component to obtain a preliminary prediction value;

[0014] Step S6: Compare the preliminary predicted value with the actual value to obtain an error term, and combine the error term with the influencing variables in the original data set to obtain an error data set;

[0015] Step S7: Combining variational mode decomposition and the two-stage self-attention model to predict the error data set, and compensating the obtained error prediction value to the preliminary prediction value to obtain the final prediction result.

[0016] According to an embodiment of the catalytic cracking unit product and energy consumption prediction method of the present invention, when performing data collection, the catalytic cracking unit product and energy consumption prediction method selects multiple influencing variables for the prediction variables of product prediction and energy prediction, and then collects corresponding data points as production process data according to the selected influencing variables and a preset time period, and performs preprocessing to obtain an original data set for data prediction; wherein the production process data includes real-time data and analytical test data. Real-time measurement data includes operating conditions, product flow rate and raw material flow rate, and analytical test data includes raw material properties and product properties.

[0017] According to one embodiment of the catalytic cracking unit product and energy consumption prediction method of the present invention, when the catalytic cracking unit product and energy consumption prediction method preprocesses the acquired generation process data, a sliding window is used to remove outliers from each influencing variable data, and the real-time measurement data after the outliers are removed is averaged, and then the averaged real-time measurement data is combined with the corresponding analytical test data to obtain the original data set of the same time scale.

[0018] According to one embodiment of the method for predicting products and energy consumption of a catalytic cracking unit of the present invention, when the method for predicting products and energy consumption of a catalytic cracking unit uses variational modal decomposition and hyperparameter grid search to perform modal decomposition on the prediction variables in the original data set, the hyperparameters of the variational modal decomposition are first determined, and then the optimal parameters of the variational modal decomposition are searched using the hyperparameter grid, and finally the prediction variables in the data set are modally decomposed under the optimal parameters of the variational modal decomposition, thereby obtaining multiple prediction components; wherein the hyperparameters of the variational modal decomposition include a modal number parameter K and a bandwidth constraint parameter alpha.

[0019] According to one embodiment of the catalytic cracking unit product and energy consumption prediction method of the present invention, the prediction data set of each prediction component is time series data containing feature and time information. When the catalytic cracking unit product and energy consumption prediction method uses a two-stage self-attention model based on feature attention mechanism and time attention mechanism to predict the prediction data set of each prediction component, an input attention layer is set before the LSTM encoder to assign corresponding input feature attention weights to the input feature vector sequence at each moment, and a time attention layer is set before the LSTM decoder to assign corresponding time attention weights to the feature vector sequence at each moment after the input feature is weighted, so as to combine the input feature attention weights and the time attention weights to perform prediction and obtain corresponding prediction results.

[0020] According to an embodiment of the method for predicting products and energy consumption of a catalytic cracking unit of the present invention, a two-stage self-attention model based on a feature attention mechanism and a time attention mechanism performs prediction by the following steps:

[0021] Step S41: Input the input feature vector sequence at each moment into the input attention layer in chronological order, extract the input features in the input feature vector sequence in combination with the output state of the encoder at the previous moment, and obtain the input feature sequence;

[0022] Step S42: using a softmax function to weight each input feature in the attention feature sequence to obtain an attention weight for each input feature;

[0023] Step S43: multiply each input feature attention weight by the corresponding input feature vector to obtain a feature vector sequence after input feature attention weighting;

[0024] Step S44: Inputting the feature vector sequence at each moment after the input feature attention weighting into the LSTM encoder in chronological order for encoding, to obtain an encoded feature vector sequence;

[0025] Step S45: input the encoded feature vector sequence into the temporal attention layer, and extract the temporal features of the feature vector sequence at each moment in combination with the output state of the decoder at the previous moment to obtain a temporal feature sequence;

[0026] Step S46: using the softmax function to weight each time feature in the time feature sequence to obtain each time attention weight;

[0027] Step S47: multiply each time attention weight by the feature vector sequence at the corresponding moment to obtain a feature vector sequence that fuses the input attention feature weight and the time attention weight;

[0028] Step S48: Input the feature vector sequence that fuses the input attention feature weights and the temporal attention weights into the decoder for decoding, and output the decoded sequence as a prediction value.

[0029] According to an embodiment of the method for predicting products and energy consumption of a catalytic cracking unit of the present invention, the input attention layer extracts the input feature vector sequence of length T at time t by the following formula: The input feature attention weights are applied to the input of the LSTM encoder to obtain the feature vector sequence after the input feature attention weights:

[0030]

[0031] Among them, V α , W α , U α , b α represents the learning parameters of the input attention layer;

[0032] Represents the input attention layer for the kth input feature vector at time t Extracted input features;

[0033] Represents the input attention layer for any input feature vector at time t Extracted input features;

[0034] h t-1 represents the hidden layer state of the input LSTM encoder at time t-1;

[0035] s t-1 Represents the LSTM cell state of the input attention layer at time t-1;

[0036] Represents the extracted input features Weighted input feature attention weights;

[0037] Represents a feature vector sequence of length T at time t after applying the input feature attention weights.

[0038] According to an embodiment of the method for predicting products and energy consumption of a catalytic cracking unit of the present invention, the time attention layer uses the following formula to weight the input feature attention to the feature vector sequence with a length of T at time t. The feature vector sequence encoded by the LSTM encoder is subjected to temporal attention weight extraction and applied to the input of the LSTM decoder to obtain a feature vector sequence that fuses the input attention feature weights and the temporal attention weights:

[0039]

[0040] in, ω d , U d represents the learning parameters of the temporal attention layer,

[0041] It represents the feature vector h after the temporal attention layer encodes the i-th feature vector at time t. i The extracted time features,

[0042] d t-1 represents the hidden layer state of the LSTM decoder at time t-1,

[0043] Indicates that the time attention is on the LSTM cell state at time t-1.

[0044] Represents the extracted time features The weighted temporal attention weight,

[0045] c tRepresents the sequence of feature vectors at time t that fuses the input attention feature weights and the temporal attention weights.

[0046] According to one embodiment of the catalytic cracking unit product and energy consumption prediction method of the present invention, after obtaining the final prediction result, the catalytic cracking unit product and energy consumption prediction method introduces multiple different prediction indicators to evaluate the prediction accuracy of the two-stage self-attention model based on the feature attention mechanism and the time attention mechanism, so as to adjust and train the two-stage self-attention model based on the feature attention mechanism and the time attention mechanism according to the evaluation result.

[0047] The present invention also provides a computer readable medium storing a computer program code, wherein the computer program code implements the method as described above when executed by a processor.

[0048] The present invention also provides a catalytic cracking unit product and energy consumption prediction device based on variational mode decomposition and a two-stage self-attention mechanism integrated model, comprising:

[0049] a memory for storing instructions executable by a processor; and

[0050] A processor is used to execute the instructions to implement the method as described above.

[0051] Compared with the prior art, the present invention has the following beneficial effects: the present invention is aimed at the catalytic cracking unit product and energy consumption prediction method, and adopts variational mode decomposition and hyperparameter grid search to perform mode decomposition on the prediction variables in the original data set, and combines the prediction components with the influencing variables of the original data set, so as to establish a prediction data set of multiple prediction components, so as to solve the adverse effects of the nonlinearity between the prediction variables and the influencing variables on data prediction. At the same time, the present invention adopts a two-stage self-attention model based on the feature attention mechanism and the time attention mechanism to predict the prediction data set of each prediction component respectively, and sums up the prediction results of each prediction component to obtain the prediction value of the corresponding prediction variable, which solves the adverse effects of the high spatiotemporal characteristics between the data in the production process on data prediction. Compared with the prior art, the present invention solves the influence of the nonlinearity and spatiotemporal characteristics of the data of the catalytic cracking unit generation process on data prediction, improves the prediction accuracy of the product yield and energy consumption process in the catalytic cracking unit generation process, and has good application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The above features and advantages of the present invention can be better understood after reading the detailed description of the embodiments of the present disclosure in conjunction with the following drawings. In the drawings, the components are not necessarily drawn to scale, and components with similar related properties or features may have the same or similar reference numerals.

[0053] Figure 1It is a flow chart showing an embodiment of the method for predicting products and energy consumption of a catalytic cracking unit of the present invention.

[0054] Figure 2 It is a data flow diagram showing an embodiment of a method for predicting products and energy consumption of a catalytic cracking unit according to the present invention.

[0055] Figure 3 1 is a process flow diagram showing an embodiment of a catalytic cracking device of the present invention.

[0056] Figure 4 It is a schematic diagram showing an embodiment of the present invention for performing variational modal decomposition on diesel yield and gasoline yield.

[0057] Figure 5 2 is a network structure diagram showing an embodiment of a two-stage self-attention model based on a feature attention mechanism and a temporal attention mechanism of the present invention.

[0058] Figure 6 It is a flowchart showing the steps of an embodiment of the present invention using a two-stage self-attention model based on feature attention mechanism and time attention mechanism for prediction.

[0059] Figure 7 It is a schematic diagram showing search results of an embodiment of the present invention of performing a grid search on model parameters of a diesel yield prediction model and a gasoline yield prediction model.

[0060] Figure 8 It is a schematic diagram showing an embodiment of the diesel yield, gasoline yield and energy consumption prediction effect of the present invention. DETAILED DESCRIPTION

[0061] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for the description of the embodiments. Obviously, the drawings described below are only some examples or embodiments of the present application. For ordinary technicians in this field, the present application can also be applied to other similar scenarios based on these drawings without creative work. Unless it is obvious from the language environment or otherwise explained, the same reference numerals in the figures represent the same structure or operation.

[0062] 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.

[0063] Unless otherwise specifically stated, the relative arrangement, numerical expressions and numerical values ​​of the parts and steps set forth in these embodiments do not limit the scope of the present application. At the same time, it should be understood that, for ease of description, the sizes of the various parts shown in the accompanying drawings are not drawn according to the actual proportional relationship. The technology, method and equipment known to those of ordinary skill in the relevant field may not be discussed in detail, but in appropriate cases, the technology, method and equipment should be considered as a part of the authorization specification. In all examples shown and discussed here, any specific value should be interpreted as being merely exemplary, rather than as a limitation. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters represent similar items in the following drawings, so that once a certain item is defined in an accompanying drawing, it does not need to be further discussed in subsequent drawings.

[0064] When describing the embodiments of the present invention in detail, for the sake of convenience, the cross-sectional views showing the device structure will not be partially enlarged according to the general scale, and the schematic views are only examples, which should not limit the scope of protection of the present invention. In addition, in actual production, the three-dimensional dimensions of length, width and depth should be included.

[0065] In the description of the present application, it should be understood that the directions or positional relationships indicated by directional words such as "front, back, up, down, left, right", "lateral, vertical, perpendicular, horizontal" and "top, bottom" are usually based on the directions or positional relationships shown in the drawings, and are only for the convenience of describing the present application and simplifying the description. Unless otherwise specified, these directional words do not indicate or imply that the device or element referred to must have a specific direction or be constructed and operated in a specific direction, and therefore cannot be understood as limiting the scope of protection of the present application; the directional words "inside and outside" refer to the inside and outside relative to the contours of each component itself.

[0066] Disclosed herein is an embodiment of a method for predicting products and energy consumption of a catalytic cracking unit. Figure 1 is a flow chart showing an embodiment of a method for predicting products and energy consumption of a catalytic cracking unit according to the present invention, Figure 2 This is a data flow diagram showing an embodiment of the method for predicting the product and energy consumption of a catalytic cracking unit of the present invention. Figure 1 and Figure 2 ,The following is a detailed description of each step of the ,product and energy consumption prediction method for catalytic cracking unit.

[0067] Step S1: Collecting production process data of the catalytic cracking unit based on a preset time period, and preprocessing the acquired production process data to obtain an original data set based on a time series; wherein the original data set includes prediction variables corresponding to product prediction and energy consumption prediction and influencing variable data corresponding to each prediction variable.

[0068] Figure 3 FIG. 1 is a process flow diagram showing an embodiment of a catalytic cracking device of the present invention. Figure 3 As shown in the figure, the process of the catalytic cracking unit mainly involves the reaction regeneration unit, product separation unit, product refining unit and flue gas treatment unit, which will generate a large amount of production process data during the production process. Among them, the reaction regeneration unit is the core of the catalytic cracking process. Different from the traditional catalytic cracking unit, this unit includes two reactors, which use the multi-product isoparaffin process (MIP) and the multi-product high-octane gasoline and aromatics process (LTAG), and a regenerator to process its built-in raw materials alone, so that the production strategy of the catalytic cracking unit can be flexibly changed to achieve highly selective raw material conversion. During production, the raw oil enters the reactor from the bottom of the riser, contacts the regenerated catalyst at high temperature, and then rises rapidly under the acceleration of the pre-stripping steam, undergoes a thorough cracking reaction to obtain the reaction oil gas. At the same time, the deactivated catalyst enters the regenerator, burns the surface material in the regenerator to obtain the regenerated catalyst, and then re-enters the reactor for production. During the regeneration process, the regenerator will emit gases that have an impact on the environment, such as nitrogen oxides, sulfur dioxide, carbon monoxide and carbon dioxide, which are cleaned by the flue gas treatment unit. During the treatment, the flue gas discharged from the top of the regenerator passes through the precipitated catalyst particles in the rotary separation device and enters the carbon monoxide combustion furnace, where carbon monoxide is burned into carbon dioxide. At the same time, concentrated ammonia water and sulfur trioxide adsorbent are introduced to remove nitrogen oxides and sulfur trioxide. The gas then enters the flue gas desulfurization absorption tower to remove sulfur dioxide, and is finally discharged into the atmosphere through a chimney equipped with an electrostatic precipitator.

[0069] Next, the reaction oil and gas produced by the reaction regeneration unit enters the product separation unit for separation, and the main products of the catalytic cracking process are obtained, including slurry oil, diesel, gasoline, liquefied petroleum gas and dry gas. Since the separated products (such as gasoline, dry gas and liquefied petroleum gas) contain other impurities, they must enter the product refinement unit for desulfurization and refining, so as to obtain purified gasoline, liquefied petroleum gas and dry gas.

[0070] According to the above-mentioned catalytic cracking unit process flow, product prediction mainly involves reaction regeneration unit, fractionation, absorption stabilization unit and product refining unit, and energy consumption prediction mainly involves reaction regeneration unit. Since the units involved in product prediction and energy prediction are different, when performing data prediction, it is necessary to select corresponding influencing variables for the prediction variables of product prediction and energy prediction respectively, and then collect the data points corresponding to the influencing variables in the time period as production process data according to the selected influencing variables and a preset time period, including real-time data and analytical test data, and pre-process these data to obtain the original data set based on time series for subsequent data prediction.

[0071] Among them, the real-time measurement data includes operating conditions, product flow and raw material flow, and the analytical test data includes raw material properties and product properties. After obtaining these data, a sliding window is used to remove outliers from the data of each influencing variable. At the same time, considering the complexity of the device process and the time lag effect of the data caused by material flow, it is also necessary to average the real-time measurement data after removing outliers, and then combine the averaged real-time measurement data with the corresponding analytical test data to obtain the original data set of the same time scale.

[0072] In one embodiment, 46 influencing variables are selected for product yield prediction variables based on expert experience, and 34 influencing variables are selected for energy consumption prediction variables. The data collection period is set to 1 hour, and the corresponding production process data within 1 hour is collected according to the selected influencing variables, and outliers are removed and averaged to obtain an original data set containing 8760 data points.

[0073] Step S2: Variational mode decomposition and hyperparameter grid search are used to perform mode decomposition on the prediction variables in the original data set to obtain multiple prediction components.

[0074] In this embodiment, when using variational mode decomposition and hyperparameter grid search to perform modal decomposition on the prediction variables in the original data set, firstly, the hyperparameters of variational mode decomposition are determined, then the hyperparameter grid is used to search for the optimal parameters of variational mode decomposition, and finally, the prediction variables in the data set are modally decomposed under the optimal parameters of variational mode decomposition, thereby obtaining multiple prediction components. Among them, the hyperparameters of variational mode decomposition include the mode number parameter K and the bandwidth constraint parameter alpha.

[0075] Variational mode decomposition (VMD) is an adaptive, fully non-recursive method of variational mode decomposition and signal processing. When using variational mode decomposition to decompose time series-based modes, the decomposition process can be expressed as a constrained variational problem as shown below:

[0076]

[0077] Among them, u k ={u1,u2,...,u K} means decomposing the original sequence into K modal components. k ={ω1,ω2,...,ω K} represents the center frequency of each mode. represents derivation, “*” represents convolution, δ(t) represents unit pulse, By summing all modal reconstructions, we can obtain an f value close to the original signal.

[0078] In order to speed up the solution, the Lagrange multiplier λ can be introduced to k and the modal center frequency ω k Strengthen the constraints. Therefore, for the above constrained variational problem, the augmented Lagrangian with a quadratic penalty term can be used to obtain good convergence. The formula is as follows:

[0079]

[0080] Among them, α represents the variance regularization parameter of white noise. The above formula is used to calculate the modal component U k and the modal center frequency ω k After strengthening the constraints, the alternating direction multiplier algorithm is used to continuously update and iterate to obtain the optimal solution to the problem.

[0081]

[0082] Among them, γ represents the update parameter of the Lagrange multiplier, which is used to determine whether the multiplier should be used to enforce the constraint. The above formula is repeatedly iterated to finally meet the convergence condition. Its convergence condition is:

[0083]

[0084] Where ε is the definition coefficient used to determine convergence. When the VMD mode converges in the frequency domain, the solution in the time domain can be obtained by inverse Fourier transform.

[0085] Step S3: reconstructing the prediction data set of each prediction component according to the prediction components obtained after decomposition and the corresponding influencing variable data in the original data set.

[0086] In this embodiment, after the modal decomposition of the prediction variables is completed through the above steps, the prediction data set of each prediction component is reconstructed according to the prediction components obtained after the decomposition and the corresponding influencing variable data in the original data set, and the prediction result of the prediction component can be obtained by predicting the prediction data set of each prediction component and summing the prediction results. Therefore, the influence of the nonlinearity between the prediction variables and the influencing variables on the prediction can be effectively alleviated through variational modal decomposition.

[0087] Taking product yield as an example, the decomposition effect of the predicted variables after decomposition using variational mode decomposition method is as follows: Figure 4 As shown in Figure 2, (a) represents the decomposition process of diesel yield, and (b) represents the decomposition process of gasoline yield. Figure 4 It can be seen that by performing variational mode decomposition on the prediction variables, the input original data set can be disassembled and reconstructed into prediction data sets of multiple prediction components.

[0088] Step S4: Use a two-stage self-attention model based on feature attention mechanism and time attention mechanism to predict the prediction data set of each prediction component respectively, and obtain the prediction result of each prediction component.

[0089] The dual-stage self-attention model based on feature attention mechanism and temporal attention mechanism (DA-LSTM model) is a specialized deep learning model designed for time series prediction. It combines a variant of the recurrent neural network (RNN) long short-term memory network (LSTM) model with an attention mechanism to organically learn from past observations to predict future values. It includes two stages: feature attention mechanism and temporal attention mechanism. Among them, the feature attention mechanism is used to evaluate the inherent characteristics of the data, determine which input variables in the data set have greater significance, obtain more information about these features through the feature attention mechanism, and assign higher weights to them. The temporal attention mechanism helps the prediction model focus on specific time intervals when processing sequence and time series data, and integrates these time intervals with the output of the model. The temporal attention mechanism aims to focus on data related to specific time intervals. For data sets with specific time intervals, the temporal attention mechanism is used to assign corresponding weights to each moment, which can effectively extract the impact of the features of each moment on the final result, thereby enhancing the prediction model's understanding of sequence data and making it more suitable for time series prediction. This two-stage self-attention model combines LSTM-based feature extraction and attention-based relevance weighting, effectively addressing the challenges of time series forecasting by enhancing the model's ability to capture and exploit temporal dependencies and important information in the data. Figure 5 is a network structure diagram showing an embodiment of a dual-stage self-attention model based on a feature attention mechanism and a time attention mechanism of the present invention. Figure 5 , further describing this embodiment in detail.

[0090] like Figure 5 As shown, in this embodiment, when a two-stage self-attention model based on a feature attention mechanism and a time attention mechanism is used to predict the prediction data set of each prediction component, an input attention layer is set before the LSTM encoder to assign corresponding input feature attention weights to the input feature vector sequence at each moment, and a time attention layer is set before the LSTM decoder to assign corresponding time attention weights to the feature vector sequence at each moment after the input feature is weighted, so as to combine the input feature attention weights and the time attention weights to perform prediction and obtain corresponding prediction results. Figure 6 is a flowchart showing an embodiment of the present invention using a two-stage self-attention model based on feature attention mechanism and time attention mechanism for prediction. Figure 6 , detailing the various steps of prediction using a two-stage self-attention model based on feature attention mechanism and temporal attention mechanism.

[0091] Step S41: Input the input feature vector sequence at each moment into the input attention layer in chronological order, extract the input features in the input feature vector sequence in combination with the output state of the encoder at the previous moment, and obtain the input feature sequence.

[0092] In this embodiment, it is assumed that the input feature vector sequence at time t is Its length is T. It is input into the input attention layer and combined with the output state h of the LSTM encoder at time t-1 t-1 , the following formula is used to extract the input features in the input feature vector sequence with a length of T at time t:

[0093]

[0094] Among them, V α , W α , U α , b α represents the learning parameters of the input attention layer, Represents the input attention layer for the kth input feature vector at time t The extracted input features, h t-1 represents the output state of the input LSTM encoder at time t-1, s t-1 Represents the LSTM cell state of the input attention layer at time t-1.

[0095] Step S42: Use the softmax function to weight each input feature in the attention feature sequence to obtain the attention weight of each input feature.

[0096] In this embodiment, for the kth input feature vector at time t, Extracted input features Use the following formula to assign the corresponding input feature attention weight

[0097]

[0098] in, Represents the input attention layer for the kth input feature vector at time t The extracted input features, Represents the input attention layer for any input feature vector at time t The extracted input features. In this way, through this formula, the weight assignment of each feature in the input feature sequence is completed.

[0099] Step S43: multiply each input feature attention weight by the corresponding input feature vector to obtain a feature vector sequence after the input feature attention weighting.

[0100] In this embodiment, after obtaining the input feature attention weights of each input feature vector through the above steps, they are applied to the input of the LSTM encoder to obtain a feature vector sequence after the input feature attention weights are obtained:

[0101]

[0102] in, Represents a feature vector sequence of length T at time t after applying the input feature attention weights.

[0103] Step S44: Input the feature vector sequence at each moment after the input feature attention weighting into the LSTM encoder in chronological order for encoding to obtain the encoded feature vector sequence.

[0104] Step S45: Input the encoded feature vector sequence into the temporal attention layer, and extract the temporal features of the feature vector sequence at each moment in combination with the output state of the decoder at the previous moment to obtain a temporal feature sequence.

[0105] In this embodiment, the weighted feature vector sequence obtained in the above steps is After that, it is input into the LSTM encoder for encoding to obtain the encoded feature vector sequence h i , and then the encoded feature vector sequence h i Input into the time attention layer and extract its time features through the following formula:

[0106]

[0107] in, ω d , U d represents the learning parameters of the temporal attention layer, It represents the feature vector h after the temporal attention layer encodes the i-th feature vector at time t. i The extracted time features, d t-1 represents the output state of the LSTM decoder at time t-1, Indicates the state of the LSTM cell at time t-1. Therefore, the time feature extraction of each feature vector in the encoded feature vector sequence is completed through this formula to obtain the time feature sequence.

[0108] Step S46: Use the softmax function to weight each time feature in the time feature sequence to obtain each time attention weight.

[0109] In this embodiment, the time features extracted in the above steps are The corresponding time attention weight is assigned to it through the following formula

[0110]

[0111] Therefore, through this formula, the weight assignment of each time feature in the time feature sequence is completed.

[0112] Step S47: Multiply each temporal attention weight by the feature vector sequence at the corresponding moment to obtain a feature vector sequence that fuses the input attention feature weight and the temporal attention weight.

[0113] In this embodiment, after obtaining the input feature attention weights of each input feature vector through the above steps, they are applied to the input of the LSTM decoder to obtain a feature vector sequence that fuses the input attention feature weights and the temporal attention weights:

[0114]

[0115] Among them, c t represents the feature vector sequence at time t that fuses the input attention feature weights and the temporal attention weights, Represents the extracted time features The weighted temporal attention weight, h i represents the i-th encoded feature vector.

[0116] Step S48: Input the feature vector sequence that fuses the input attention feature weights and the temporal attention weights into the decoder for decoding, and output the decoded sequence as a prediction value.

[0117] In this embodiment, after the feature extraction and encoding of all input data sequences are completed through the above steps, the acquired feature vector sequence [c1, c1, ..., c t ,…,c T ] is input into the LSTM decoder for decoding to obtain the prediction result of the corresponding prediction component.

[0118] Step S5: Sum up the prediction results of each prediction component to obtain a preliminary prediction value.

[0119] In this embodiment, after the prediction results of each prediction component are obtained through the above steps, they are summed up to obtain the preliminary prediction value of the corresponding prediction variable.

[0120] In addition, in the present embodiment, for different prediction variables, a corresponding DA-LSTM model can be constructed for prediction. Assuming that the product obtained by the catalytic cracking unit is gasoline and liquefied petroleum gas, the corresponding product yield prediction model (diesel yield prediction model and gasoline yield prediction model parameters) and energy consumption prediction model can be established based on the DA-LSTM model architecture. Among them, when using these prediction models for prediction, first, the prediction variables corresponding to these prediction models must be subjected to variational mode decomposition, and the variational mode decomposition parameters are optimized by using a grid search method. At the same time, due to the different data processing difficulties of different prediction models, the model parameters, including window size, encoder hidden layer number, and decoder hidden layer number are also different, so it can also be used to select parameters for these DA-LSTM models.

[0121] Specifically, it is assumed that the penalty factor of the variational mode decomposition ranges from [100, 2000], and the decomposition mode number K ranges from [3, 7], and only contains integers. For the dual-stage self-attention model (DA-LSTM model) that includes the feature attention mechanism and the time attention mechanism, the window size T is set to the range of T∈{3, 5, 7, 10}, the encoder hidden layer m and the decoder hidden layer n range values ​​are m=n∈{16, 32, 64, 128, 256}, and then the grid search method is used for parameter selection.

[0122] In the grid search process, it is necessary to examine the difference between the decomposed result and the actual value. For grid search results, please refer to Figure 7 . Among them, the diesel yield prediction model parameters represented by (a) and (b) have a penalty factor of 100, K of 7, DA-LSTM network structure parameter T of 7, and encoder hidden layer m and decoder hidden layer n of 256 in the process of variational mode decomposition. The gasoline yield prediction model parameters represented by (c) and (d) have a penalty factor of 100, K of 7, DA-LSTM network structure parameter T of 5, and encoder hidden layer m and decoder hidden layer n of 256 in the process of variational mode decomposition. The penalty factor of the energy consumption prediction model in the process of variational mode decomposition is 100, K of 7, DA-LSTM network structure parameter T of 7, and encoder hidden layer m and decoder hidden layer n of 256.

[0123] Step S6: Compare the preliminary predicted value with the actual value to obtain an error term, and combine the error term with the influencing variables in the original data set to obtain an error data set.

[0124] Step S7: Combining variational mode decomposition and the two-stage self-attention model to predict the error data set, and compensating the obtained error prediction value to the preliminary prediction value to obtain the final prediction result.

[0125] In this embodiment, since there is a residual term in the VMD decomposition process, the residual term is usually directly predicted as a component and then added to the final prediction result. However, since the residual term is irregular, the direct prediction effect is poor, which will have an adverse effect on the final prediction. At the same time, the DA-LSTM model will also have errors in the prediction process, so the residual generated by VMD and the DA-LSTM model error are finally considered together, the error term is combined with the influencing variables in the original data set, and an error data set is established. Then, the error data set is subjected to variational modal decomposition, and the modal components after decomposition by the DA-LSTM model are used for prediction, and the prediction results are summed to obtain the error prediction result. The error prediction result is compensated to the preliminary prediction value to obtain the final prediction result, thereby eliminating the influence of the residual generated by the VMD decomposition and the DA-LSTM model. Figure 8 Schematic diagram showing an embodiment of the diesel yield, gasoline yield and energy consumption prediction effect of the present invention. Figure 8 It can be seen that the predicted value obtained by the present invention is basically consistent with the actual value, the prediction effect is excellent, and the prediction accuracy requirement in industrial applications is met.

[0126] In the existing general prediction models, only one or two of the three technologies, VMD, DA-LSTM, and EC, are applied. It has not been confirmed whether the prediction effect of integrating the three technologies is superior. Therefore, the present invention compares the prediction results on the test data set using mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE) and correlation coefficient (R 2 ) Four indicators are used to evaluate the model performance to verify the prediction effect of the present invention, and the calculation formula is as follows:

[0127]

[0128] The MSE, RMSE, and MAE indicators are sensitive to the scale of the original data. Therefore, the lower the values ​​of these three indicators, the better the prediction performance. 2 Not affected by the data size, R 2 The closer the value is to 1, the higher the prediction accuracy. At the same time, different prediction methods are used to predict the same data set, and the results are shown in the following table:

[0129]

[0130] After comparison, it can be found that the prediction method proposed in this invention is better in MSE, RMSE, MAE and R 2 All four indicators showed the best prediction effect for energy consumption prediction.

[0131] This specification also provides a computer-readable medium storing a computer program code, which, when executed by a processor, implements the above-mentioned method for predicting products and energy consumption of a catalytic cracking unit.

[0132] This specification also provides a catalytic cracking unit product and energy consumption prediction device based on a variational mode decomposition and a two-stage self-attention mechanism integrated model, including an instruction memory for storing instructions executable by a processor, and a processor for executing instructions in the instruction memory to implement the catalytic cracking unit product and energy consumption prediction method as described above.

[0133] The previous description of the disclosure is provided to enable any person skilled in the art to make or use the disclosure. Various modifications to the disclosure will be apparent to those skilled in the art, and the general principles defined herein may be applied to other variations without departing from the spirit or scope of the disclosure. Thus, the disclosure is not intended to be limited to the examples and designs described herein, but should be granted the widest scope consistent with the principles and novel features disclosed herein.

[0134] Those skilled in the art will further appreciate that the various illustrative logic blocks, modules, circuits, and algorithm steps described in conjunction with the embodiments disclosed herein may be implemented as electronic hardware, computer software, or a combination of the two. To clearly illustrate this interchangeability of hardware and software, various illustrative components, blocks, modules, circuits, and steps are generally described above in terms of their functionality. Whether such functionality is implemented as hardware or software depends on the specific application and the design constraints imposed on the overall system. The technician may implement the described functionality in different ways for each specific application, but such implementation decisions should not be interpreted as resulting in a departure from the scope of the present invention.

[0135] The various illustrative logic blocks, modules, and circuits described in conjunction with the embodiments disclosed herein may be implemented or performed with a general purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. A general purpose processor may be a microprocessor, but in the alternative, the processor may be any conventional processor, controller, microcontroller, or state machine. The processor may also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors in cooperation with a DSP core, or any other such configuration.

[0136] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. The software module may reside in a RAM memory, a flash memory, a ROM memory, an EPROM memory, an EEPROM memory, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to a processor so that the processor can read and write information from / to the storage medium. In an alternative, a storage medium may be integrated into a processor. The processor and the storage medium may reside in an ASIC. The ASIC may reside in a user terminal. In an alternative, the processor and the storage medium may reside in a user terminal as discrete components.

[0137] In one or more exemplary embodiments, the functions described may be implemented in hardware, software, firmware, or any combination thereof. If implemented as a computer program product in software, each function may be stored on or transmitted by a computer-readable medium as one or more instructions or codes. Computer-readable media include both computer storage media and communication media, including any medium that facilitates the transfer of a computer program from one place to another. Storage media may be any available medium that can be accessed by a computer. As an example and not limitation, such a computer-readable medium may include RAM, ROM, EEPROM, CD-ROM or other optical disk storage, disk storage or other magnetic storage device, or any other medium that can be used to carry or store the desired program code in the form of an instruction or data structure and can be accessed by a computer. Any connection is also properly referred to as a computer-readable medium. For example, if the software is transmitted from a website, a server, or other remote source using a coaxial cable, a fiber optic cable, a twisted pair, a digital subscriber line (DSL), or wireless technologies such as infrared, radio, and microwaves, the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwaves are included in the definition of the medium. Disk and disc as used herein include compact disc (CD), laser disc, optical disc, digital versatile disc (DVD), floppy disk and Blu-ray disc, wherein disk often reproduces data magnetically, while disc reproduces data optically with lasers. Combinations of the above should also be included within the scope of computer-readable media.

Claims

1. A method for predicting products and energy consumption of a catalytic cracking unit, characterized in that: The following steps are involved: Step S1: collecting production process data of the catalytic cracking unit based on a preset time period, and preprocessing the acquired production process data to obtain an original data set based on a time series; wherein the original data set includes prediction variables corresponding to product prediction and energy consumption prediction and influencing variable data corresponding to each prediction variable; Step S2: Using variational mode decomposition and hyperparameter grid search to perform mode decomposition on the prediction variables in the original data set to obtain multiple prediction components; Step S3: reconstructing the prediction data set of each prediction component according to the prediction components obtained after decomposition and the corresponding influencing variable data in the original data set; Step S4: using a two-stage self-attention model based on a feature attention mechanism and a temporal attention mechanism to predict the prediction data set of each prediction component respectively, and obtaining the prediction result of each prediction component; Step S5: summing up the prediction results of each prediction component to obtain a preliminary prediction value; Step S6: Compare the preliminary predicted value with the actual value to obtain an error term, and combine the error term with the influencing variables in the original data set to obtain an error data set; Step S7: Combining variational mode decomposition and the two-stage self-attention model to predict the error data set, and compensating the obtained error prediction value to the preliminary prediction value to obtain the final prediction result.

2. The method for predicting products and energy consumption of a catalytic cracking unit according to claim 1, characterized in that: The method for predicting the product and energy consumption of the catalytic cracking unit selects a plurality of influencing variables for the prediction variables of the product prediction and the energy prediction respectively when collecting data, and then collects corresponding data points as production process data according to the selected influencing variables and a preset time period, and performs preprocessing to obtain an original data set for data prediction; wherein the production process data includes real-time data and analytical test data. The real-time measurement data includes operating conditions, product flow rate and raw material flow rate, and the analytical test data includes raw material properties and product properties.

3. The method for predicting products and energy consumption of a catalytic cracking unit according to claim 2, characterized in that: When the catalytic cracking unit product and energy consumption prediction method preprocesses the acquired generation process data, a sliding window is used to remove outliers from the data of each influencing variable, and the real-time measurement data after the outliers are removed is averaged, and then the averaged real-time measurement data is combined with the corresponding analytical test data to obtain the original data set of the same time scale.

4. The method for predicting products and energy consumption of a catalytic cracking unit according to claim 1, characterized in that: When the catalytic cracking unit product and energy consumption prediction method uses variational modal decomposition and hyperparameter grid search to perform modal decomposition on the prediction variables in the original data set, the hyperparameters of the variational modal decomposition are first determined, then the optimal parameters of the variational modal decomposition are searched using the hyperparameter grid, and finally the prediction variables in the data set are modally decomposed under the optimal parameters of the variational modal decomposition, thereby obtaining multiple prediction components; wherein the hyperparameters of the variational modal decomposition include a modal number parameter K and a bandwidth constraint parameter alpha.

5. The method for predicting products and energy consumption of a catalytic cracking unit according to claim 1, characterized in that: The prediction data set of each prediction component is time series data containing feature and time information. When the catalytic cracking unit product and energy consumption prediction method uses a two-stage self-attention model based on feature attention mechanism and time attention mechanism to predict the prediction data set of each prediction component, an input attention layer is set before the LSTM encoder to assign corresponding input feature attention weights to the input feature vector sequence at each moment, and a time attention layer is set before the LSTM decoder to assign corresponding time attention weights to the feature vector sequence at each moment after the input feature is weighted, so as to combine the input feature attention weights and the time attention weights to perform prediction and obtain corresponding prediction results.

6. The method for predicting products and energy consumption of a catalytic cracking unit according to claim 5, characterized in that: The two-stage self-attention model based on feature attention mechanism and temporal attention mechanism performs prediction through the following steps: Step S41: Input the input feature vector sequence at each moment into the input attention layer in chronological order, extract the input features in the input feature vector sequence in combination with the output state of the encoder at the previous moment, and obtain the input feature sequence; Step S42: using a softmax function to weight each input feature in the attention feature sequence to obtain an attention weight for each input feature; Step S43: multiply each input feature attention weight by the corresponding input feature vector to obtain a feature vector sequence after input feature attention weighting; Step S44: Inputting the feature vector sequence at each moment after the input feature attention weighting into the LSTM encoder in chronological order for encoding, to obtain an encoded feature vector sequence; Step S45: Input the encoded feature vector sequence into the temporal attention layer, and extract the temporal features of the feature vector sequence at each moment in combination with the output state of the decoder at the previous moment to obtain a temporal feature sequence; Step S46: using the softmax function to weight each time feature in the time feature sequence to obtain each time attention weight; Step S47: multiply each time attention weight by the feature vector sequence at the corresponding moment to obtain a feature vector sequence that fuses the input attention feature weight and the time attention weight; Step S48: Input the feature vector sequence that fuses the input attention feature weights and the temporal attention weights into the decoder for decoding, and output the decoded sequence as a prediction value.

7. The method for predicting products and energy consumption of a catalytic cracking unit according to claim 6, characterized in that: The input attention layer extracts the input feature vector sequence of length T at time t using the following formula: The input feature attention weights are applied to the input of the LSTM encoder to obtain the feature vector sequence after the input feature attention weights: Among them, V α , W α , U α 、b a represents the learning parameters of the input attention layer; Represents the input attention layer for the kth input feature vector at time t Extracted input features; Represents the input attention layer for any input feature vector at time t Extracted input features; h t-1 represents the hidden layer state of the input LSTM encoder at time t-1; s t-1 Represents the LSTM cell state of the input attention layer at time t-1; Represents the extracted input features Weighted input feature attention weights; Represents a feature vector sequence of length T at time t after applying the input feature attention weights.

8. The method for predicting products and energy consumption of a catalytic cracking unit according to claim 7, characterized in that: The temporal attention layer uses the following formula to pay attention to the weighted feature vector sequence of length T at time t of the input feature: The feature vector sequence encoded by the LSTM encoder is subjected to temporal attention weight extraction and applied to the input of the LSTM decoder to obtain a feature vector sequence that fuses the input attention feature weights and the temporal attention weights: in, ω d , U d represents the learning parameters of the temporal attention layer, It represents the feature vector h after the temporal attention layer encodes the i-th feature vector at time t. i The extracted time features, d t-1 represents the hidden layer state of the LSTM decoder at time t-1, Indicates that the time attention is on the LSTM cell state at time t-1. Represents the extracted time features The weighted temporal attention weight, c t Represents the sequence of feature vectors at time t that fuses the input attention feature weights and the temporal attention weights.

9. The method for predicting products and energy consumption of a catalytic cracking unit according to claim 1, characterized in that: After obtaining the final prediction results, the catalytic cracking unit product and energy consumption prediction method introduces multiple different prediction indicators to evaluate the prediction accuracy of the two-stage self-attention model based on the feature attention mechanism and the time attention mechanism, so as to adjust and train the two-stage self-attention model based on the feature attention mechanism and the time attention mechanism according to the evaluation results.

10. A computer readable medium storing computer program code, characterized in that: The computer program code implements the method according to any one of claims 1 to 9 when executed by a processor.

11. A catalytic cracking unit product and energy consumption prediction device based on variational mode decomposition and dual-stage self-attention mechanism integrated model, characterized in that: include: a memory for storing instructions executable by a processor; as well as A processor, configured to execute the instructions to implement the method according to any one of claims 1 to 9.

Citation Information

Patent Citations

  • Motor bearing fault diagnosis method based on feature optimization and GWOA-XGBoost

    CN114169110A

  • User electricity consumption prediction method based on VMD-Lasso-CNN-LSTM

    CN114692947A

  • Blood pressure change trend estimation method based on wavelet scattering transformation

    CN116172531A

  • Prediction method based on periodic feature extraction and time sequence attention

    CN116341712A

  • Oil consumption prediction method based on improved double-attention mechanism

    CN117787467A