Prediction Method and Device for Product Yield in Catalytic Cracking Process

Abstract

The present invention provides a method for predicting the product yield in the catalytic cracking process, a prediction device, and a computer-readable storage medium. The prediction method includes the following steps: obtaining multiple sets of production process data of a target product at multiple sampling times; performing cyclic sampling according to the multiple sets of production process data and pre-trained latent variable conditional probability parameters to determine multiple latent variables; and unbiasedly estimating the yield of the target product according to the multiple latent variables and pre-trained quality index distribution parameters. By performing these steps, the prediction method can use the discriminant probability of the quality index to reduce irrelevant partial hypotheses, thereby reducing the risk of performance degradation, efficiently processing a large amount of, multi-source, high-dimensional chemical production process data, and accurately predicting the yield of the target product.

Description

Technical Field

[0001] The present invention relates to a prediction technology for product yields in the catalytic cracking process, and particularly to a method for predicting product yields in the catalytic cracking process, a device for predicting product yields in the catalytic cracking process, and a computer-readable storage medium. Background Art

[0002] Hydrocracking is one of the most commonly used technical methods in the petrochemical and refining industries. It refers to a class of process technologies in which crude oil undergoes hydrocracking reactions under high temperature, high pressure, and the condition of a catalyst to meet the product requirements. In this field, traditional soft-sensing methods that do not rely on prior knowledge are generally used to predict quality indicators such as the output and yield of hydrocracking products. This soft-sensing method realizes a relatively accurate and rapid prediction of the output variable by establishing a soft-sensing model between easily measurable input variables and difficult-to-measure quality output variables. Compared with traditional non-linear modeling, the static relationship between variables and the characterization of dynamic characteristics can be taken into account during the soft-sensing modeling process, and many successful applications of soft-sensing have been achieved in the fields of chemical engineering, biochemical engineering, metallurgy, and the pharmaceutical industry.

[0003] However, with the continuous development of industrial technologies, the amount of data generated during the system control process is increasing day by day, the data sources are increasing, and the data dimension is also increasing. As a result, traditional data-driven methods are limited by their representation learning ability and cannot make accurate predictions. In addition, in the chemical process, many important quality variables such as oil viscosity and oil components are often difficult to measure in real time, which also causes difficulties in the effective process control and optimization of the hydrocracking process.

[0004] In order to overcome the above-mentioned defects existing in the prior art, there is an urgent need in this field for a prediction technology for product yields in the catalytic cracking process, which is used to efficiently process a large amount of, multi-source, and high-dimensional chemical production process data and accurately predict the yield of target products. Summary of the Invention

[0005] The following presents a brief overview of one or more aspects to provide a basic understanding of these aspects. This overview is not an exhaustive survey of all contemplated aspects, and is neither intended to identify key or decisive elements of all aspects nor to define the scope of any or all aspects. Its sole purpose is to present some concepts of one or more aspects in a simplified form as a prelude to a more detailed description to follow.

[0006] To overcome the above-mentioned defects existing in the prior art, the present invention provides a method for predicting the product yield in the catalytic cracking process, a device for predicting the product yield in the catalytic cracking process, and a computer-readable storage medium, which can use the discrimination probability of quality indicators to reduce irrelevant partial hypotheses, thereby reducing the risk of performance degradation, so as to efficiently process a large amount of, multi-source, and high-dimensional chemical production process data and accurately predict the yield of the target product.

[0007] Specifically, the method for predicting the product yield in the catalytic cracking process provided by the first aspect of the present invention includes the following steps: obtaining multiple sets of production process data x of the target product at multiple sampling times ≤t ; performing loop sampling according to the multiple sets of production process data x ≤t and the pre-trained latent variable conditional probability parameter φ to determine multiple latent variables z i ; unbiasedly estimating the yield μ of the target product according to the multiple latent variables z i and the pre-trained quality indicator distribution parameter θ y .

[0008] Furthermore, in some embodiments of the present invention, the step of performing loop sampling according to the multiple sets of production process data x ≤t and the pre-trained latent variable conditional probability parameter φ to determine multiple latent variables z i includes: performing the first round of loop sampling according to the multiple sets of production process data x ≤t and the pre-trained latent variable conditional probability parameter φ to determine the first latent variable z0; entering the next round of loop and determining whether the next round of loop reaches a preset number of loop rounds; in response to the judgment result that the next round of loop does not reach the preset number of loop rounds, performing loop sampling of the next round of loop according to the previous latent variable z i-1 , the multiple sets of production process data x ≤t and the latent variable conditional probability parameter φ to determine the latent variable z i of the next round of loop, and entering the next round of loop again; and in response to the judgment result that the next round of loop reaches the preset number of loop rounds, ending the loop sampling.

[0009] Furthermore, in some embodiments of the present invention, the step of unbiasedly estimating the yield μ of the target product according to the multiple latent variables z i and the pre-trained quality indicator distribution parameter θ y includes: respectively calculating the yield estimate value μ i corresponding to the loop round i according to each of the latent variables z y,i and the pre-trained quality indicator distribution parameter θ; and according to the yield estimate value μ of each loop round iy,i The mean value is used to unbiasedly estimate the yield μ of the target product y .

[0010] Further, in some embodiments of the present invention, before cyclically sampling according to the multi - group production process data x ≤t and the pre - trained latent variable conditional probability parameter φ to determine a plurality of latent variables z i , the prediction method further includes the following steps: collecting multi - group production process data samples x ≤t ' of the target product at multiple sampling times and their corresponding yield samples y t '; using a regression model based on the latent variable z to construct the corresponding relationship between each production process data sample x ≤t ' and its corresponding yield sample y t ', where the corresponding relationship is characterized by the quality index distribution parameter θ and the latent variable conditional probability parameter φ; and training the quality index distribution parameter θ and the latent variable conditional probability parameter φ according to each production process data sample x ≤t ' and its corresponding yield sample y t '.

[0011] Further, in some embodiments of the present invention, the step of using a regression model based on the latent variable z to construct the corresponding relationship between each production process data sample x ≤t ' and its corresponding yield sample y t ' includes: constructing a regression model based on the latent variable z and characterizing the distribution probability p t of the yield y θ and the conditional probability q φ of the latent variable z, where the parameter of the distribution probability p θ is θ, which is used to estimate the yield y t , and the parameter of the conditional probability q φ is φ, which is used to estimate the latent variable z; parameterizing the regression model according to the distribution probability p θ and the conditional probability q φ .

[0012] Further, in some embodiments of the present invention, the regression model based on the latent variable z satisfies:

[0013] q φ (z i |x ≤t ,z ≤i-1 )=q φ (z i |x i ,z i-1 )

[0014] q φ (z j |z ≤j-1 ) = q φ (z j |z j-1 )

[0015] p θ (y|x ≤t ,z ≤t ) = p θ (y|x t ,z t-1 )

[0016] where x = [x ≤t and y = [y ≤t are the input and output of the regression model respectively; z = [z ≤t is the latent variable introduced by the regression model.

[0017] Furthermore, in some embodiments of the present invention, the regression model is parameterized as:

[0018] z0 = f μ0 (x; φ) + f σ0 (x; φ) ⊙ ∈

[0019]

[0020] where f σ0 (x; φ) = exp(0.5f logvar0 (x; φ)), f μ0 (x; φ) and f logvar0 (x; φ) represent the logarithm of the mean and variance of the posterior probability of the first latent variable respectively, and exp(·) is the exponential function; and represent the logarithm of the mean and variance of the posterior probability of the remaining latent variables respectively, z q is the quality-related feature of the predicted yield y t ⊙ represents element-wise multiplication.

[0021] Furthermore, in some embodiments of the present invention, the step of training the quality index distribution parameter θ and the latent variable conditional probability parameter φ according to each of the production process data samples x ≤t ′ and its corresponding yield sample y t ′ includes: initializing the quality index distribution parameter θ to be trained and the latent variable conditional probability parameter φ to be trained; according to the multiple groups of production process data samples x ≤t′ and perform cyclic sampling on the latent variable conditional probability parameter φ to be trained to determine a plurality of latent variables z i ; Calculate the yield estimate value μ i according to the plurality of latent variables z y ′ and the quality index distribution parameter θ to be trained; according to the yield estimate value μ y ′ and the corresponding yield sample y t ′, calculate the loss function value and adjust the quality index distribution parameter θ to be trained and the latent variable conditional probability parameter φ to be trained according to the loss function value to determine the pre-trained quality index distribution parameter θ and the pre-trained latent variable conditional probability parameter φ.

[0022] Further, in some embodiments of the present invention, the step of calculating the loss function value y according to the yield estimate value μ′ t and the corresponding yield sample yt includes: determining the following objective function according to a parameterized regression model:

[0023]

[0024] where is the lower bound of the logarithmic conditional probability expectation;

[0025] Perform Monte Carlo estimation on the latent variable model of the objective function according to the regression model based on the latent variable z:

[0026]

[0027] where d y is the dimension of the yield y, ε y is the variance of the yield estimate value μ′ y and the corresponding yield sample y′ t , c1 and c2 are constants related to ε y , c1 is non-positive; and

[0028] Maximize the objective function to determine the minimized loss function value

[0029]

[0030] where g y (z; θ) is the prediction function of the yield y.

[0031] Further, in some embodiments of the present invention, the according to the loss function value The steps of adjusting the quality index distribution parameter θ to be trained and the latent variable conditional probability parameter φ to be trained to determine the pre-trained quality index distribution parameter θ and the pre-trained latent variable conditional probability parameter φ include: according to the loss function value Calculate the loss function gradient of the quality index distribution parameter θ respectively And the loss function gradient of the latent variable conditional probability parameter φ According to the loss function gradient And the loss function gradient Update the quality index distribution parameter θ and the latent variable conditional probability parameter φ to be trained; and according to the updated quality index distribution parameter θ and the updated latent variable conditional probability parameter φ, perform the next round of training until the preset training stop criterion is met.

[0032] Further, in some embodiments of the present invention, the target product includes at least one of the combustible gas, gasoline, diesel, and kerosene produced in the catalytic cracking process. Correspondingly, the production process data includes at least one of the mass flow rate, nitrogen content, sulfur content of the feed, and the boiling point at different stages, and / or at least one of the pressure, temperature, and liquid level at the inlet, top, and / or bottom of at least one reaction tower in the catalytic cracking process.

[0033] In addition, the prediction device for the product yield of the above-mentioned catalytic cracking process provided by the second aspect of the present invention includes a memory and a processor. The processor is connected to the memory and is configured to implement the prediction method for the product yield of the above-mentioned catalytic cracking process provided by the first aspect of the present invention.

[0034] In addition, the above-mentioned computer-readable storage medium provided by the third aspect of the present invention stores computer instructions. When the computer instructions are executed by a processor, the prediction method for the product yield of the above-mentioned catalytic cracking process provided by the first aspect of the present invention is implemented. Brief Description of the Drawings

[0035] After reading the detailed description of the embodiments of the present disclosure in conjunction with the following drawings, the above features and advantages of the present invention can be better understood. In the drawings, the components are not necessarily drawn to scale, and components with similar related characteristics or features may have the same or similar reference numerals.

[0036] Figure 1 Shows a schematic flow chart of training a regression model based on latent variables provided by some embodiments of the present invention.

[0037] Figure 2 Shows a schematic diagram of a catalytic cracking process provided by some embodiments of the present invention.

[0038] Figure 3 Shows a schematic flowchart of a method for predicting the product yield of a catalytic cracking process provided according to some embodiments of the present invention.

[0039] Figure 4 Shows a comparison chart of the prediction effects of a hydrocracking process provided according to some embodiments of the present invention.

[0040] Figure 5 Shows a schematic diagram of the prediction performance of a hydrocracking process provided according to some embodiments of the present invention.

[0041] Figure 6 Shows a histogram of errors of a hydrocracking process provided according to some embodiments of the present invention.

[0042] Figure 7 Shows a probability schematic diagram of a hydrocracking process provided according to some embodiments of the present invention.

[0043] Figure 8 Shows a schematic diagram of the error autocorrelation of a hydrocracking process provided according to some embodiments of the present invention.

[0044] Figure 9 Shows a comparison chart of the differential prediction effects of a hydrocracking process provided according to some embodiments of the present invention. Detailed Embodiments

[0045] The following specific embodiments illustrate the implementation manners of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Although the description of the present invention will be introduced in conjunction with preferred embodiments, this does not mean that the features of this invention are limited to this implementation manner. On the contrary, the purpose of introducing the invention in conjunction with the implementation manner is to cover other alternatives or modifications that may be extended based on the claims of the present invention. In order to provide a deep understanding of the present invention, many specific details will be included in the following description. The present invention can also be implemented without using these details. In addition, in order to avoid confusing or obscuring the key points of the present invention, some specific details will be omitted in the description.

[0046] In the description of the present invention, it should be noted that, unless otherwise clearly specified and defined, the terms "installed", "connected", and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected, or indirectly connected through an intermediate medium, and it can be the communication inside two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations.

[0047] In addition, the terms "upper", "lower", "left", "right", "top", "bottom", "horizontal", and "vertical" used in the following description should be understood as the orientations shown in the relevant paragraphs and accompanying drawings. Such relative terms are for convenience of description only and do not represent that the devices described need to be manufactured or operated in a specific orientation, and thus should not be construed as a limitation on the present invention.

[0048] It is understood that although terms such as "first", "second", "third", etc. may be used herein to describe various components, regions, layers, and / or parts, these components, regions, layers, and / or parts should not be limited by these terms, and these terms are only used to distinguish different components, regions, layers, and / or parts. Therefore, the first component, region, layer, and / or part discussed below may be referred to as the second component, region, layer, and / or part without departing from some embodiments of the present invention.

[0049] As described above, with the continuous development of industrial technology, the amount of data generated in the system control process is increasing, the data sources are increasing, and the data dimensions are also increasing, making the traditional data-driven methods limited by the representation learning ability and unable to make accurate predictions. In addition, in the chemical process, many important quality variables such as oil viscosity and oil components are often difficult to measure in real time, which also causes difficulties in the effective process control and optimization of the hydrocracking process.

[0050] In order to overcome the above-mentioned defects existing in the prior art, the present invention provides a method for predicting the product yield of a catalytic cracking process, a device for predicting the product yield of a catalytic cracking process, and a computer-readable storage medium, which can use the discrimination probability of quality indicators to reduce irrelevant partial hypotheses, thereby reducing the risk of performance degradation, so as to efficiently process a large amount of, multi-source, and high-dimensional chemical production process data and accurately predict the yield of the target product.

[0051] In some non-limiting embodiments, the method for predicting the product yield of a catalytic cracking process provided in the first aspect of the present invention can be implemented by the device for predicting the product yield of a catalytic cracking process provided in the second aspect of the present invention. Specifically, the prediction device may be configured with a memory and a processor. The memory includes but is not limited to the computer-readable storage medium provided in the third aspect of the present invention, on which computer instructions are stored. The processor is connected to the memory and is configured to execute the computer instructions stored on the memory to implement the method for predicting the product yield of a catalytic cracking process provided in the first aspect of the present invention.

[0052] Furthermore, in some embodiments, the method for predicting the product yield of the catalytic cracking process provided by the first aspect of the present invention can be implemented in two stages: offline training and online prediction. Correspondingly, the device for predicting the product yield of the catalytic cracking process can also be divided into two parts: an offline training device and an online prediction device, which are used to execute the relevant steps of the offline training stage and the online prediction stage of the prediction method respectively. Optionally, the offline training device and the online prediction device can be integrated into the same device, or can be separately configured on multiple devices, which is not limited herein.

[0053] First, please refer to Figure 1 , Figure 1 which shows a schematic flow chart of training a regression model based on latent variables according to some embodiments of the present invention.

[0054] As Figure 1 shown, in the offline training stage of the regression model based on latent variables, the offline training device can first collect the production process data samples x ≤t ' of the actual factory of the target product over a period of time and its corresponding yield samples y t '. Here, the target product includes but is not limited to at least one of the combustible gas, gasoline, diesel, and kerosene produced by the catalytic cracking process.

[0055] Taking the product aviation kerosene of the hydrocracking process (HCP) as an example, the offline training device can first collect Figure 2 the production process data of a certain refinery from April 2017 to February 2019 (sampling period: 4 months) as shown, and select 55 input variables among them as the key variables of the actual continuous catalytic reforming process.

[0056] Table 1

[0057]

[0058]

[0059]

[0060] As shown in Table 1, the selected input variables include, but are not limited to, the mass flow rate, nitrogen content, sulfur content, and boiling points at different stages of the feedstock (crude oil), and / or the key parameters such as pressure, temperature, and / or liquid level at the inlets, tops, and bottoms of each key reaction tower during the reaction process. In this embodiment, the offline training device actually collected 2052 production process data samples, used the first 1500 samples as training samples, the subsequent 549 samples as prediction samples, and the remaining 3 samples to generate the input data corresponding to the first test label. In addition, the offline training device also collected the product quality samples (e.g., kerosene yield samples) corresponding to each production process data sample as the prediction indicators for training and detecting the model performance.

[0061] In some embodiments, after obtaining multiple production process data samples, the offline training device can preferably preprocess the collected raw data to improve the training efficiency of the model and the prediction accuracy of the trained model.

[0062] For example, the offline training device can first use the Aspen HYSYS process simulation software to solve the problem of insufficient output data labels obtained from the refinery. After that, the offline training device can divide the data set and then standardize the divided data to eliminate the influence of variable dimensions and variation ranges. Specifically, the offline training device can perform z-score standardization on the original value x of each attribute, and the z-score standardization formula is as follows:

[0063]

[0064] where μ is the mean of all sample data, σ is the standard deviation of all sample data, x is the original value of the sample, and x * is the standardized result.

[0065] In this way, the standardized variable value x* will fluctuate around 0. If x* is greater than 0, it means that the original value x of this attribute is higher than the average level. On the contrary, if x* is less than 0, it means that the original value x of this attribute is lower than the average level.

[0066] For another example, the offline training device can also preprocess data through the time difference method. In this time difference method, the input feature at each moment is the difference between the input variables at the current moment and the previous moment, and the prediction target is the difference between the output variables at the current moment and the previous moment. Assume that the normalized process variables are v(1), v(2), v(3),... v(n),..., and the time series output of the normalized kerosene is k(1), k(2), k(3),... k(n),..., then the input of the model is x(j) = v(j) - v(j - 1), and the output of the model is the time difference of the normalized label as y(j) = k(j) - k(j - 1). When predicting the time series output, the offline training device can, according to predict the normalized label value and use the normalized label value to predict the label of the model.

[0067] Please continue to refer to Figure 1 , after collecting and preprocessing the production process data x ≤t , the offline training device can use the collected and preprocessed data to construct the corresponding relationship between each production process data sample x ≤t ′ and its corresponding yield sample y t ′ through a regression model based on the latent variable z. This corresponding relationship can be characterized by the quality index distribution parameter θ and the latent variable conditional probability parameter φ.

[0068] Specifically, the offline training device can first construct a probabilistic discriminative time series model based on a latent variable regression model:

[0069] q φ (z i |x ≤t ,z ≤i-1 ) = q φ (z i |x i ,z i-1 )

[0070] q φ (z j |z ≤j-1 ) = q φ (z j |z j-1 )

[0071] p θ (y|x ≤t ,z ≤t ) = p θ (y|x t ,z t-1 )

[0072] where x = [x ≤t and y = [y≤t are the input and output of the regression model respectively; z = [z ≤t is the latent variable introduced by the regression model; q φ represents the conditional probability of the latent variable, with parameters φ, used to estimate the latent variable z; p θ is the parameter for estimating the quality index distribution, with parameters θ, used to estimate the yield y t .

[0073] After that, the model can be further parameterized as:

[0074] z0 = f μ0 (x; φ) + f σ0 (x; φ) ⊙ ∈

[0075]

[0076] where f σ0 (x; φ) = exp(0.5f logvar0 (x; φ)), f μ0 (x; φ) and f logvar0 (x; φ) respectively represent the logarithm of the mean and variance of the posterior probability of the first latent variable, exp(·) is the exponential function; and respectively represent the logarithm of the mean and variance of the posterior probability of the remaining latent variables, z q is the quality-related feature of the predicted yield y t . ⊙ represents element-wise multiplication.

[0077] Here, the above parameterized models f μ0 (x), f logvar0 (x), g y (z), g x (z) can be any parameterized model with gradients. In this embodiment, the offline training device uses a Python neural network model and uses the neural network model in PyTorch to model the parameterized model. PyTorch is an open-source Python machine learning library based on Torch. When building a neural network model, all established model structures inherit from the base class nn.Module. The offline training device can first define the required network initial function, initialize the parameters such as the weights of the neural network, and then construct the forward propagation function forward, thus establishing a complete network framework. This forward function is used to implement the functions of the model and is the core of realizing the connection relationship between each layer.

[0078] Furthermore, the specific parameter model structure of the above parametric model is as follows:

[0079] f μ0 (x) is modeled as nn.Sequential(nn.Linear(t * n x , n h ), nn.Tanh(), nn.Linear(n h , n h ), nn.Tanh(), nn.Linear(n h , n z )) where nn.Sequential(·, ·) is a container that adds modules to the module in the order passed in the constructor. It can also contain other modules and apply these modules in sequence to produce its output. This method can automatically initialize the parameter functions required by the neural network; nn.Linear(·, ·) is used to set the fully connected layer (fc) and can fix the number of output channels; nn.Tanh(·, ·) provides support for the hyperbolic tangent function in the neural network, where n x represents the tensor size of the input sample, n h represents the tensor size used for passing in the neural network, and n z represents the tensor size of the output sample.

[0080] f logvar0 (x) is modeled as nn.Sequential(nn.Linear(t * n x , n h ), nn.Tanh(), nn.Linear(n h , n h ), nn.Tanh(), nn.Linear(n h , n z )) where the built-in functions of the neural network are the same as those in the above f μ0 (x).

[0081] is modeled as nn.GRU(n x , n z ), nn.GRU(n x , n z ) where GRU(·, ·) is a gated unit. The calculation of a GRU cell h o = GRU(h i , a) can be written in the following form:

[0082] r = Sig(W r [h i , a])

[0083] b = Sig(W b [h i , a])

[0084]

[0085] where ⊙ is element-wise multiplication and Sig is the sigmoid function.

[0086] is modeled as nn.Sequential(nn.Linear(n z + n x , n h ), nn.Tanh(), nn.Linear(n h , n h ), nn.Tanh(), nn.Linear(n h , n z )) where the meanings of the built-in neural network functions are the same as those in f μ0 (x).

[0087] g y (z) is modeled as nn.Sequential(nn.Linear(n z , n y )) where the meanings of the built-in neural network functions are the same as those in f μ0 (x).

[0088] The structure of the dynamic model of the method mentioned above is shown in Appendix Table 2, where Concentrate(x i , z i ) means concatenating the input dimensions of the function and outputting an intermediate variable with a dimension of n x + n z . The hyperparameters of the model are t, n x , n y , n h , n z .

[0089] Table 2

[0090]

[0091] After that, the offline training device can train the quality index distribution parameter θ and the latent variable conditional probability parameter φ according to each production process data sample x ≤t ′ and its corresponding yield sample y t ′.

[0092] Specifically, the offline training device can first initialize the quality index distribution parameter θ to be trained and the latent variable conditional probability parameter φ to be trained, and set parameters such as the time series window length size of the model, the latent variable dimension, the number of hidden layer neurons, and the stop criterion for the training phase. In this embodiment, the offline training device can set the time series window length to 3, the latent variable dimension to 10, the number of hidden layer neurons to 20, and the stop criterion for the training phase to stop when the number of repeated training times for all samples reaches 100 times.

[0093] After that, the offline training device can perform cyclic sampling according to multiple sets of production process data samples x ≤t ′ and the latent variable conditional probability parameter φ to be trained to determine multiple latent variables z i . For example, after completing the preprocessing operation, the input data and output data in this embodiment can be respectively denoted as x(1)′, x(2)′, x(3)′,…x(n)′,… and y(1)′, y(2)′, y(3)′,…y(n)′,…. For a given time window t and a randomly sampled moment T (T≥t), the offline training device can set the value of x ≤t ′ to the production process data samples x T ′ at moments T, T + 1…T + t - 1, and set the value of y T+1 ′ to the yield data sample y T+t-1 ′ at moment T + t - 1. At this time, this x T , y T+t-1 ′ is a set of sampling values of the distributions X, Y. A set of sampling values [x1′, x2′,…, x ≤t , y T ′] can be represented by [x(T - t + 1)′, x(T - t + 2)′,…, x(T)′, y(T)′], that is, x t , y t ′~X, Y can be realized by randomly sampling T and assigning [x(T - t + 1)′, x(T - t + 2)′,…, x(T)′, y(T)′] to [x1′, x2′,…, x ≤t , y t ′, y t ′, y t ′].

[0094] After that, the offline training device can perform the first round of cyclic sampling according to multiple sets of production process data samples x ≤t ′ and the latent variable conditional probability parameter φ to be trained, and sample z0 using the following formula:

[0095] z0 = f μ0 (x; φ) + f σ0 (x; φ)⊙∈

[0096] After that, the offline training device can enter the next loop iteration and determine whether the loop iteration i satisfies the stopping criterion. If the loop iteration i does not satisfy the stopping criterion, the offline training device can perform loop sampling for each j in (1, 2, …, t - 1), and according to the previous latent variable z i-1 , multiple groups of production process data x ≤t ′ and the latent variable conditional probability parameter φ to perform loop sampling for this round of loop, and sample z using the following formula j :

[0097]

[0098] where i indicates the serial number of the sampling round, used to estimate the average value of the sampling, and j indicates the corresponding moment of the latent variable z j .

[0099] After that, the offline training device can calculate the quality-related feature z of the prediction quality index using the following formula q :

[0100]

[0101] Then, in combination with the quality index distribution parameter θ to be trained, calculate the yield estimate value μ y ′ using the following formula

[0102]

[0103] In this way, the offline training device can calculate the loss function value according to the yield sample y y ′ corresponding to the yield estimate value μ t ′ (i.e., the actual value) and adjust the quality index distribution parameter θ to be trained and the latent variable conditional probability parameter φ to be trained according to this loss function value to train the quality index distribution parameter θ and the latent variable conditional probability parameter φ.

[0104] Specifically, the present invention first proposes the following theory that the parameterized log conditional probability is a lower bound of the log conditional probability of the intractable distribution, that is

[0105]

[0106] where is the log conditional probability expectation, and the formula is obtained by Bayes' formula In the formula is a constant, and p θ (y|x) refers to the parameterized conditional probability is the lower bound of the log conditional probability expectation.

[0107] Based on the above theory, the offline training device can use the above model to calculate the following optimization objectives, and its objective function is as follows:

[0108]

[0109] After that, the offline training device can use the above latent variable-based regression model to perform Monte Carlo estimation on this objective function:

[0110]

[0111] where d y is the dimension of y, and ε y is the variance of the yield estimate μ′ y and the corresponding yield sample y′ t ; c1 and c2 are constants related to ε y , and c1 is non-positive. The aim is to obtain the expectation of the quality index, which is independent of ε y_tm . Then maximizing the objective function is equivalent to minimizing the prediction, that is

[0112]

[0113] where g y (z; θ) is the prediction function of the quality index y.

[0114] After that, the offline training device can calculate the loss function gradient of the quality index distribution parameter θ and the loss function gradient of the latent variable conditional probability parameter φ respectively according to the calculated loss function value Then, according to this loss function gradient and this loss function gradient update the quality index distribution parameter θ and the latent variable conditional probability parameter φ.

[0115] Furthermore, after determining the latent variable z i in this round of loop and completing the update of the quality index distribution parameter θ and the latent variable conditional probability parameter φ, the offline training device can enter the next round of loop again, and perform the next round of training according to the updated quality index distribution parameter θ and the updated latent variable conditional probability parameter φ until the preset training stop criterion is met, and return the final quality index distribution parameter θ and the latent variable conditional probability parameter φ to complete the training process of the latent variable-based regression model.

[0116] Please continue to refer to Figure 3 , Figure 3 which shows a schematic flowchart of a method for predicting the product yield of a catalytic cracking process according to some embodiments of the present invention.

[0117] As Figure 3 shown, in the online prediction stage of the latent variable-based regression model, the online prediction device can collect the production process data of the actual factory over a period of time as described above, and screen multiple sets of production process data x related to the target product (e.g., aviation kerosene) from it. ≤t . In some embodiments, the multiple sets of production process data x ≤t may include one or more of the 55 input variables shown in Table 1.

[0118] Furthermore, the online prediction device can preprocess the collected raw data as described above to improve the prediction efficiency and accuracy of the model. The specific process is similar to the above training stage and will not be elaborated here.

[0119] After that, the online prediction device can perform cyclic sampling on the basis of the preprocessed multiple sets of production process data x ≤t , and the pre-trained latent variable conditional probability parameter φ to determine multiple latent variables z i . Then, according to the multiple latent variables z i and the pre-trained quality index distribution parameter θ, unbiasedly estimate the yield μ of the target product. y .

[0120] Specifically, in the cyclic sampling process of the latent variable z i , the online prediction device can first set the index variables i and j. Among them, i indicates the serial number of the sampling round and is used to estimate the average value of the sampling, and j indicates the moment corresponding to the latent variable z i .

[0121] After that, for each i in (1, 2,..., t - 1), the online prediction device can first perform the first-round cyclic sampling according to the multiple sets of production process data x ≤t and the pre-trained latent variable conditional probability parameter φ, and use the following formula to determine the first latent variable z0:

[0122] z0 = f μ0 (x; φ) + f σ0 (x; φ) ⊙ ∈

[0123] Then, the online prediction device can enter the next cyclic round and determine whether the cyclic round i meets the stop criterion. If the cyclic round i does not meet the stop criterion, the online prediction device can perform cyclic sampling for each j in (1, 2,..., t - 1), and perform cyclic sampling for this round according to the previous latent variable z i-1 , the multiple sets of production process data x ≤t and the latent variable conditional probability parameter φ, and sample z j using the following formula:

[0124]

[0125] After that, the online prediction device can calculate the quality-related feature z of the prediction quality index using the following formula q :

[0126]

[0127] Combined with the pre-trained quality index distribution parameter θ, the yield estimate value μ for each round of cycle is calculated using the following formula y,i :

[0128]

[0129] Furthermore, after determining the yield estimate value μ for this round of cycle y,i , the online prediction device can enter the next round of cycle again and repeat the above judgment and sampling operations until the cyclic sampling of the entire sampling period and all rounds of cycles is completed. After that, the online prediction device can calculate the average of the yield estimate values μ for each round of cycle i y,i to unbiasedly estimate the yield μ of the target product y .

[0130] To verify the effectiveness of the present invention, this paper provides three different prediction methods: partial least squares regression (PLSR), feedforward neural network (NN), and the probability discriminant time series model based on recurrent neural network (PDTM-RNN) of the present invention to compare the prediction results

[0131] Specifically, this paper uses the root mean square error (RMSE) and the coefficient of determination (R 2 ) to quantify the effectiveness of the above three different prediction methods, and their specific expressions are as follows

[0132]

[0133] where N T is the number of samples in the test data, y i and are the actual and predicted outputs respectively, is the average value of the actual outputs in the test dataset. The RMSE and R 2 indexes predicted by various different methods on the test set are shown in Table 3

[0134] Table 3

[0135] Method PLSR NN PDTM-RNN RMSE 0.1377 0.1101 0.0909 <![CDATA[R 2 > 0.2276 0.5064 0.6640

[0136] As can be seen from Table 3, in the time-independent regression model and the time series regression model, the feedforward neural network (NN) has good prediction ability, while the probabilistic discriminant time series model based on the recurrent neural network (PDTM-RNN) has worse prediction ability because it is a linear model. The neural network based on likelihood probability maximization (PLSR) shows a decline in prediction performance when discriminating quality indicators, manifested as the deterioration of both the RMSE and R 2 statistics. The probabilistic discriminant neural network (PDTM-RNN) has better prediction effect compared with the deterministic neural network (NN), which is reflected in lower RMSE and higher R 2 . Additionally, it can also be summarized from the appendix Figure 4 that the bias generated by the probabilistic discriminant time series method (PDTM-RNN) provided by the present invention is the smallest. Since the deterministic neural network is a special case where the variance of the latent variable in the probabilistic discriminant model (PDTM-RNN) proposed by the present invention is 0, the performance of the deterministic model and the probabilistic discriminant model is relatively close. Due to the fact that the latent variable of the deterministic model is deterministic and overfitting occurs in the training set, it has worse prediction performance compared with the probabilistic discriminant model, which is also reflected in higher RMSE and lower R compared with the probabilistic discriminant model 2 .

[0137] To verify the effectiveness of the method provided by the present invention in numerical simulation, Figure 5 the scatter plot of the prediction results was further recorded. Meanwhile, some statistical test results are presented in Figures 6 - 8 . Specifically, Figure 5 indicates that the difference between the prediction results and the true labels is small, Figure 6 and Figure 7 indicate that the prediction error is approximately normally distributed, while Figure 8 gives the autocorrelation of the error, which indicates that the errors are independent of each other. Further, the HCP difference prediction effects of each method can be seen from Figure 9 . Compared with the deterministic model and the probabilistic generative model, the probabilistic discriminant model provided by the present invention has lower RMSE and higher R 2 , reflecting better prediction results.

[0138] In summary, the probabilistic discriminant time series model with latent variables based on deep learning provided by the present invention can use the discrimination probability of quality indicators in the optimization process. This probabilistic discriminant model reduces some assumptions unrelated to the calculation of the discrimination probability, thus reducing the risk of performance decline. Compared with traditional prediction methods, this probabilistic method can predict the distribution of quality indicators, efficiently process a large amount of multi-source and high-dimensional chemical production process data, and accurately predict the yield of target products.

[0139] Although the foregoing methods are illustrated and described as a series of acts for simplicity of explanation, it should be understood and appreciated that the methods are not limited by the order of the acts, as some acts may occur in different orders and / or concurrently with other acts not illustrated and described herein or other acts that are understandable to those skilled in the art according to one or more embodiments.

[0140] 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 readily apparent to those skilled in the art, and the generic 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 is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting the product yield of a catalytic cracking process, characterized in that, It includes the following steps: Collect multiple groups of production process data samples \(x\) of the target product at multiple sampling times ≤t ' and their corresponding yield samples \(y\) t '; Build a regression model based on the latent variable z to characterize the yield y t of the distribution probability p θ and the conditional probability q of the latent variable z φ , where the parameter of the distribution probability p θ is θ, which is used to estimate the yield y t , and the parameter of the conditional probability q φ is φ, which is used to estimate the latent variable z; Using a regression model based on the latent variable z, construct each of the production process data samples x ≤t ′ and its corresponding yield sample y t ′, where the corresponding relationship is characterized by the quality index distribution parameter θ and the latent variable conditional probability parameter φ; Based on each of the production process data samples x ≤t ′ and its corresponding yield sample y t ′, train the quality index distribution parameter θ and the latent variable conditional probability parameter φ; Obtain multiple sets of production process data x of the target product at multiple sampling times ≤t , and perform loop sampling according to the multiple sets of production process data x ≤t and the pre-trained latent variable conditional probability parameter φ to determine multiple latent variables z i ; and Based on the plurality of latent variables z i and the pre-trained quality metric distribution parameter θ, unbiasedly estimate the yield μ of the target product y .

2. The prediction method according to claim 1, characterized in that, The step of cyclically sampling according to the multi-group production process data x ≤t and the pre-trained latent variable conditional probability parameter φ to determine a plurality of latent variables z i comprises: Based on the described multiple sets of production process data x ≤t and the pre-trained latent variable conditional probability parameter φ, perform the first-round cyclic sampling to determine the first latent variable z0; Enter the next round of loop and determine whether the next round of loop reaches a preset number of loop rounds; In response to the judgment result that the next round of loop does not reach the preset number of loop rounds, according to the previous latent variable z i-1 , the multiple sets of production process data x ≤t and the latent variable conditional probability parameter φ perform loop sampling for the next round of loop to determine the latent variable z i of the next round of loop, and enter the next round of loop again; and In response to the judgment result that the next round of loop reaches the preset number of loop rounds, end the loop sampling.

3. The prediction method according to claim 2, wherein According to the multiple latent variables z i and the pre-trained quality index distribution parameter θ, unbiasedly estimate the yield μ of the target product y The steps of According to each of the latent variables z i and the pre-trained quality index distribution parameter θ, calculate the yield estimate μ for the corresponding cycle round i respectively y,i ; and Based on the estimated yield μ for each of the said cycles i y,i to estimate the mean value and unbiasedly estimate the yield μ of the target product y .

4. The prediction method according to claim 1, wherein The step of constructing the corresponding relationship between each of the production process data samples x ≤t ′ and its corresponding yield sample y t ′ further includes: According to the distribution probability p θ and the conditional probability q φ , parameterize the regression model.

5. The prediction method according to claim 4, wherein The regression model based on the latent variable z satisfies: q φ (z i |x ≤t ,z ≤i-1 ) = q φ (z i |x i ,z i-1 ) q φ (z j |z ≤j-1 ) = q φ (z j |z j-1 ) p θ (y|x ≤t ,z ≤t ) = p θ (y|x t ,z t-1 ) where x = [x ≤t and y = [y ≤t are the input and output of the regression model respectively; z = [z ≤t is the latent variable introduced by the regression model.

6. The prediction method according to claim 5, characterized in that, The regression model is parameterized as: z0 = f μ0 (x; φ) + f σ0 (x; φ) ⊙ ∈ where, f σ0 (x; φ) = exp(0.5f logvar0 (x; φ)), f μ0 (x; φ) and f logvar0 (x; φ) represent the logarithm of the mean and variance of the posterior probability of the first latent variable respectively, and exp(·) is the exponential function; and represent the logarithm of the mean and variance of the posterior probability of the remaining latent variables respectively, z q is the quality-related feature of the predicted yield y t , and ⊙ represents element-wise multiplication.

7. The prediction method according to claim 6, characterized in that, The step of training the quality index distribution parameter θ and the latent variable conditional probability parameter φ according to each production process data sample x ≤t ′ and its corresponding yield sample y t ′ includes: Initialize the quality index distribution parameter θ to be trained and the latent variable conditional probability parameter φ to be trained; According to the multiple groups of production process data samples x ≤t ′ and the latent variable conditional probability parameter φ to be trained, perform cyclic sampling to determine multiple latent variables z i ; Based on the plurality of latent variables z i and the quality index distribution parameter θ to be trained, calculate the yield estimate μ y ′; Based on the estimated yield value μ y ′ and the corresponding yield sample y t ′, calculate the loss function value and According to the loss function value Adjust the quality metric distribution parameter θ to be trained and the latent variable conditional probability parameter φ to be trained to determine the pre-trained quality metric distribution parameter θ and the pre-trained latent variable conditional probability parameter φ.

8. The prediction method according to claim 7, wherein Based on the estimated yield value μ′ y and the corresponding yield sample y′ t , the step of calculating the loss function value includes: According to the parameterized regression model, determine the following objective function: Among them, is the lower bound of the logarithmic conditional probability expectation; According to the regression model based on the latent variable z, perform Monte Carlo estimation on the latent variable model of the objective function: where d y is the dimension of the production rate y, ε y is the estimated value μ ′ y of the production rate and the corresponding production rate sample y ′ t variance, c1 and c2 are constants related to ε y and c1 is non-positive; and Maximize the objective function to determine the minimized loss function value Among them, g y (z; θ) is the prediction function of the yield y.

9. The prediction method according to claim 8, wherein According to the loss function value The step of adjusting the quality metric distribution parameter θ to be trained and the latent variable conditional probability parameter φ to be trained to determine the pre-trained quality metric distribution parameter θ and the pre-trained latent variable conditional probability parameter φ includes: According to the loss function value Calculate the loss function gradient of the quality index distribution parameter θ respectively And the loss function gradient of the latent variable conditional probability parameter φ According to the gradient of the loss function and the gradient of the loss function update the quality metric distribution parameter θ and the conditional probability parameter φ of the latent variable to be trained; and According to the updated quality index distribution parameter θ and the updated latent variable conditional probability parameter φ, perform the next round of training until the preset training stop criterion is met.

10. The prediction method according to any one of claims 1 to 9, characterized in that, The target product includes at least one of the combustible gas, gasoline, diesel, and kerosene generated in the catalytic cracking process, The production process data includes at least one of the mass flow rate, nitrogen content, sulfur content, and boiling point at different stages of the feed, and / or at least one of the pressure, temperature, and liquid level at the inlet, top, and / or bottom of at least one reaction tower in the catalytic cracking process during the reaction process.

11. A prediction device for the product yield of a catalytic cracking process, characterized in that, It includes: A memory; And A processor, the processor is connected to the memory and is configured to implement the prediction method for the product yield of the catalytic cracking process according to any one of claims 1 to 10.

12. A computer-readable storage medium having computer instructions stored thereon, characterized in that, When the computer instruction is executed by the processor, it implements the prediction method for the product yield of the catalytic cracking process according to any one of claims 1 to 10.

Images