Fluidized bed reactor production process prediction method, apparatus, medium and product
Patent Information
- Application Number
- CN202511626293.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-07
- Publication Date
- 2026-01-30
Smart Images
Figure CN121434741A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of industrial control technology, and in particular to a method, equipment, medium and product for predicting the production process of a fluidized bed reactor. Background Technology
[0002] Fluidized bed reactors are core equipment in polysilicon production, enabling gas-solid reactions of silanes through gas fluidization technology. This equipment utilizes fluidized materials to increase the gas-solid contact area and heat transfer efficiency, thereby improving reaction rates and yields. It is widely used in gas-solid phase reaction fields such as catalysis and combustion.
[0003] However, the complex fluid dynamics behaviors such as particle agglomeration and bubble entrainment in the bed are coupled with multi-scale reaction dynamics, resulting in strong nonlinear time-varying characteristics of key parameters such as fluidization velocity, temperature, and pressure, which are difficult to control.
[0004] Traditional prediction methods, typically based on empirical models or simple statistical analysis, struggle to capture the nonlinear relationships during multi-condition transitions, leading to a surge in prediction errors under different modes such as fluidized, bubbling, and turbulent dynamics. Furthermore, traditional methods suffer from feature extraction bottlenecks when processing multidimensional data, failing to extract key features reflecting the gas-solid two-phase flow state within the bed from high-dimensional data, resulting in insufficient model generalization ability.
[0005] Therefore, there is an urgent need for a method to predict the production process of fluidized bed reactors, so as to improve the accuracy and reliability of predicting various variables in the production process of cold hydrogenated fluidized bed reactors, thereby improving the production efficiency of polysilicon. Summary of the Invention
[0006] This invention provides a method, equipment, storage medium, and program product for predicting the production process of a fluidized bed reactor, in order to improve the accuracy and reliability of predicting various variables in the production process of a cold hydrogenated fluidized bed reactor, thereby improving the production efficiency of polysilicon.
[0007] In a first aspect, this application provides a method for predicting the production process of a fluidized bed reactor, the method comprising:
[0008] Obtain raw industrial data for the fluidized bed reactor, including multiple production data for the fluidized bed equipment and quenching equipment;
[0009] Based on a preset variational autoencoder model, latent features are extracted from the raw industrial data to determine the corresponding latent feature representations; the variational autoencoder model includes an encoder and a decoder, and the latent feature representations characterize the core variational features and potential spatial structure of the fluidized bed reactor;
[0010] Based on the trained convolutional neural network model, production prediction is performed on the latent feature representation to obtain the predicted value of hydrogen feed rate of the fluidized bed reactor; the convolutional neural network model includes a first convolutional layer, a second convolutional layer and a fully connected layer.
[0011] Optionally, the step of extracting latent features from the original industrial data based on a preset variational autoencoder model and determining the corresponding latent feature representation includes:
[0012] Based on the encoder, a nonlinear mapping operation is performed on the original industrial data to obtain potential spatial parameters;
[0013] Based on the latent spatial parameters, reparameterization processing is performed to generate latent feature representations;
[0014] Based on the decoder, the latent feature representation is reconstructed to verify the information integrity of the latent feature representation.
[0015] Optionally, if the convolutional neural network model includes a first convolutional layer, a second convolutional layer, and a fully connected layer, then the step of generating predictions based on the trained convolutional neural network model for the latent feature representation includes:
[0016] The latent feature representation is transformed into a vector dimension to obtain the corresponding two-dimensional tensor;
[0017] Based on the first convolutional layer, local feature extraction is performed on the two-dimensional tensor to obtain a primary feature representation;
[0018] Based on the second convolutional layer, cross-regional feature combination is performed on the primary feature representation to generate a higher-order abstract representation;
[0019] The predicted value of the hydrogen feed rate is obtained by performing regression prediction on the higher-order abstract representation based on the fully connected layer.
[0020] Optionally, after obtaining the raw industrial data of the fluidized bed reactor, the method further includes:
[0021] The raw industrial data is preprocessed;
[0022] Correlation analysis was performed on the preprocessed raw industrial data, and the hydrogen feed rate was determined as the target variable for prediction.
[0023] Based on the predicted target variable, a variational autoencoder model is constructed.
[0024] Optionally, the correlation analysis of the preprocessed raw industrial data includes:
[0025] Based on the preprocessed raw industrial data, calculate the Pearson correlation coefficient matrix;
[0026] Based on the Pearson correlation coefficient matrix, the comprehensive correlation between the variables in the original industrial data is determined;
[0027] Based on the comprehensive correlation, prediction target variables are selected from the original industrial data. The prediction target variables are variables whose correlation with the efficiency of fluidized bed reactors meets preset conditions.
[0028] Optionally, the preprocessing of the raw industrial data includes:
[0029] The raw industrial data is cleaned and outlier processed.
[0030] The raw industrial data is standardized to obtain standardized data.
[0031] Based on a preset ratio, the standardized data is divided into datasets to determine the corresponding training and test sets.
[0032] Optionally, after performing production prediction on the latent feature representation based on the trained convolutional neural network model to obtain the predicted hydrogen feed rate of the fluidized bed reactor, the method further includes:
[0033] Based on the predicted hydrogen feed rate, the parameters of the fluidized bed reactor are adjusted and controlled to maintain the efficiency of the fluidized bed reactor within a preset range; the parameter adjustment and control includes at least one of the following parameters: catalyst dosage, reaction temperature, and pressure of the fluidized bed reactor.
[0034] In a second aspect, this application provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the production process prediction method for any of the fluidized bed reactors described in the first aspect above.
[0035] Thirdly, this application provides a computer storage medium storing computer program instructions, which are executed by a processor using any of the fluidized bed reactor production process prediction methods described in the first aspect above.
[0036] Fourthly, an embodiment of this application provides a computer program product including computer program instructions, which, when executed by a processor, implement the production process prediction method for any of the fluidized bed reactors described in the first aspect above.
[0037] The beneficial effects of this invention are as follows:
[0038] This application provides a method, equipment, medium, and product for predicting the production process of a fluidized bed reactor. The method acquires real-time multidimensional production data, including the circulating hydrogen flow rate, silicon tetrachloride flow rate, and gas velocity of the fluidized bed reactor in a polycrystalline silicon cold hydrogenation fluidized bed. This real-time multidimensional production data is input into a trained temperature prediction model to obtain the predicted reaction temperature output by the model. Based on the predicted reaction temperature, the polycrystalline silicon cold hydrogenation fluidized bed is optimized and controlled to improve the accuracy and reliability of temperature prediction in the cold hydrogenation fluidized bed reactor, thereby increasing the production efficiency of polycrystalline silicon. Attached Figure Description
[0039] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0040] Figure 1 A schematic flow diagram of a production process prediction method for a fluidized bed reactor provided in this application embodiment;
[0041] Figure 2 This is a schematic diagram of the structure of a variational autoencoder model provided in an embodiment of this application;
[0042] Figure 3 This is a schematic diagram of the structure of a convolutional neural network model provided in an embodiment of this application;
[0043] Figure 4 This is a schematic diagram illustrating the processing procedure of a convolutional layer provided in an embodiment of this application;
[0044] Figure 5 This is a schematic diagram of a model prediction result provided in an embodiment of this application;
[0045] Figure 6 This application provides a schematic diagram of model performance comparison.
[0046] Figure 7 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation
[0047] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. Unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.
[0048] The terms "first" and "second" in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the term "comprising" and any variations thereof are intended to cover non-exclusive protection. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices. The term "multiple" in this application can mean at least two, for example, two, three, or more, and this application does not impose limitations.
[0049] The term "and / or" in the embodiments of this application is merely a description of the association relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.
[0050] It is understood that the following specific embodiments of this application involve data related to the polysilicon production process. When the various embodiments of this application are applied to specific products or technologies, relevant licenses or consents are required, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. For example, relevant volunteers can be recruited and agreements can be signed to authorize their data, thereby enabling the implementation using the data of these volunteers; alternatively, implementation can be carried out within an authorized organization, using data from members of the organization to implement the following implementation methods for data management; or, in specific implementations, the relevant data used are all simulated data, such as simulated data generated in a virtual scene.
[0051] The embodiments of this application relate to artificial intelligence and machine learning (ML) technologies, and are primarily designed based on machine learning in artificial intelligence.
[0052] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.
[0053] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, and machine learning / deep learning.
[0054] Machine learning is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory, among others. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning.
[0055] Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence. Its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, and inductive learning. Artificial Neural Networks (ANNs) abstract the neural network of the human brain from an information processing perspective, establishing a simple model and forming different networks with different connection methods. A neural network is a computational model composed of a large number of interconnected nodes (or neurons). Each node represents a specific output function called the activation function. The connection between any two nodes represents a weighted value for the signal passing through that connection, called a weight. This is equivalent to the memory of the artificial neural network. The network's output varies depending on the network's connection methods, weight values, and activation functions. The network itself is usually an approximation of a certain algorithm or function in nature, or it may be an expression of a logical strategy.
[0056] The design concept of the embodiments of this application will be briefly introduced below.
[0057] Fluidized bed reactors are core equipment in polysilicon production, enabling gas-solid reactions of silanes through gas fluidization technology. This equipment utilizes fluidized materials to increase the gas-solid contact area and heat transfer efficiency, thereby improving reaction rates and yields. It is widely used in gas-solid phase reaction fields such as catalysis and combustion.
[0058] However, the complex fluid dynamics behaviors such as particle agglomeration and bubble entrainment in the bed are coupled with multi-scale reaction dynamics, resulting in strong nonlinear time-varying characteristics of key parameters such as fluidization velocity, temperature, and pressure, which are difficult to control.
[0059] Traditional prediction methods, typically based on empirical models or simple statistical analysis, struggle to capture nonlinear relationships during multi-condition transitions, leading to a surge in prediction errors under different modes such as fluidized, bubbling, and turbulent dynamics. Furthermore, traditional methods suffer from feature extraction bottlenecks when processing multidimensional data, failing to extract key features reflecting the gas-solid two-phase flow state within the bed from high-dimensional data, resulting in insufficient model generalization ability. For instance, while traditional mechanistic models can accurately describe physical processes using equations like the Navier-Stokes equations, the numerical solution of the differential equations is too time-consuming, failing to meet the real-time requirements of industrial control. This trade-off between computational efficiency and prediction accuracy restricts their engineering applications.
[0060] In view of the above-mentioned technical problems, this application provides a method, equipment, medium and product for predicting the production process of a fluidized bed reactor. The method acquires the raw industrial data of the fluidized bed reactor, extracts latent features from the raw industrial data through a preset variational autoencoder model, determines the corresponding latent feature representation, and then uses a trained convolutional neural network model to predict the production based on the latent feature representation, thereby obtaining the predicted value of the hydrogen feed rate of the fluidized bed reactor. This improves the prediction accuracy and reliability of various variables in the production process of the cold hydrogenated fluidized bed reactor, thereby improving the production efficiency of polysilicon.
[0061] Furthermore, this application provides an innovative solution to address the challenges of complex dynamic characteristics by combining a Variational Autoencoder (VAE) with a Convolutional Neural Network (CNN). The latent variable spatial learning mechanism of the VAE can extract low-dimensional latent feature representations from high-dimensional data, enabling decoupling analysis of nonlinear coupling relationships between operational variables and providing more accurate and valuable feature representations for the prediction model. The CNN, with its powerful nonlinear modeling capabilities, utilizes local connectivity and weight sharing characteristics to accurately capture the spatial distribution patterns and temporal evolution laws in sensor data, better addressing complex dynamic changes in reactors. Thus, this application establishes a VAE-CNN combined model, achieving robust handling of industrial data sparsity and noise interference through feature extraction and deep processing of multiple operational variables. This model enhances data-driven capabilities while maintaining physical interpretability, exhibiting higher modeling efficiency and lower prediction errors compared to traditional modeling methods, providing strong support for production optimization and decision-making.
[0062] Furthermore, this application embodiment uses a VAE-CNN model to predict the production process of a fluidized bed reactor. By combining a variational autoencoder and a convolutional neural network, it achieves high-precision modeling and accurate prediction of the reactor's dynamic behavior, significantly improving production efficiency and process stability. Moreover, unlike models built based on physical laws, this application employs data-driven deep learning technology, eliminating the need for a specific physical model of the fluidized bed reactor. By mining reactor behavior characteristics from actual production data, it achieves efficient modeling of multivariable systems. Furthermore, this application combines the advantages of VAE and CNN to capture the complex nonlinear dynamic processes in the fluidized bed reactor, enabling the model to more accurately predict and evaluate the dynamic changes of the fluidized bed reactor.
[0063] Please refer to Figure 1 The following is a schematic flowchart of a production process prediction method for a fluidized bed reactor provided in an embodiment of this application. The specific implementation process is as follows:
[0064] Step 101: Obtain raw industrial data for the fluidized bed reactor.
[0065] In this embodiment of the application, the raw industrial data of the fluidized bed reactor includes multiple production data of the fluidized bed equipment and the quenching equipment, such as fluidized bed unit data and quenching unit data.
[0066] Specifically, embodiments of this application can utilize sensors and data acquisition systems to collect raw data of the fluidized bed reactor's production process in real time. Fluidized bed unit data may include: temperature and pressure difference data, flow rate data, gas velocity and proportion data, fluidized bed pressure drop data, pressure indication data, and temperature indication data. Quenching unit data may include: temperature indication data, pressure indication data, pressure difference indication data, and flow rate indication data. For example, embodiments of this application can collect continuous production monitoring data of the fluidized bed reactor from January to December 2023 to cover multi-dimensional industrial data under different operating conditions and time periods, including but not limited to key process parameters such as hydrogen feed rate, tube-side temperature difference, bed temperature, bed pressure difference, pressure, and flow rate, making the data diverse and representative.
[0067] In one possible implementation, the embodiments of this application can preprocess the raw industrial data and perform correlation analysis on the preprocessed raw industrial data to determine the hydrogen feed rate as the prediction target variable, thereby constructing a variational autoencoder model based on the prediction target variable.
[0068] In one possible implementation, embodiments of this application can calculate a Pearson correlation coefficient matrix from preprocessed raw industrial data to determine the comprehensive correlation between variables in the raw industrial data. Based on the comprehensive correlation, predictive target variables that meet preset conditions for correlation with fluidized bed reactor efficiency can be selected from the raw industrial data.
[0069] Specifically, in this application embodiment, the Pearson coefficient can be used to perform correlation analysis on the preprocessed data to screen out key feature variables that have a significant impact on the remaining variables as prediction targets.
[0070] Specifically, this application uses the Pearson correlation coefficient to measure the linear correlation between each pair of feature variables, and the calculation formula is as follows:
[0071]
[0072] Where n is the total number of samples. and For the observed values of the variable, and This is the mean.
[0073] Furthermore, in this embodiment, the Pearson coefficients among all variables are organized into a matrix R to construct a correlation coefficient matrix. Then, the absolute value of the correlation coefficient between each variable and other variables is calculated, and the variable with the highest overall correlation is selected as the prediction target, thus achieving key feature selection. Ultimately, this application determines that the hydrogen feed flow rate exhibits the strongest overall correlation with the remaining variables and has a significant impact on reactor efficiency, and therefore selects it as the prediction target variable for the model.
[0074] Specifically, in this application embodiment, the hydrogen feed rate is the core control parameter of the fluidized bed reactor, which directly affects the reaction rate, product distribution and energy consumption. Therefore, this application uses the Pearson correlation coefficient matrix to calculate that the hydrogen feed rate is strongly correlated with key indicators such as bed temperature difference and product flow rate, and its average absolute correlation coefficient is the highest. Therefore, it is determined as the target variable for prediction.
[0075] In one possible implementation, the embodiments of this application can perform data cleaning and outlier processing on the raw industrial data, and standardize the processed raw industrial data to obtain standardized data. Then, the standardized data can be divided into datasets according to a preset ratio to determine the corresponding training set and test set.
[0076] Specifically, the embodiments of this application can collect and clean the raw data of the production process in real time, use the 3σ algorithm to process outliers, perform correlation analysis on the production data based on the Pearson correlation coefficient to screen out the hydrogen feed rate that has a significant impact on reactor efficiency as the prediction target variable, and standardize the feature data and the prediction target variable respectively. For example, the Z-Score standardization method is used to divide the standardized data into training set and test set in an 8:2 ratio.
[0077] Specifically, the data cleaning in this embodiment mainly targets the modeling needs of fluidized bed reactors. Data integrity is ensured by removing data with incomplete target variables and missing feature columns exceeding 90%. Furthermore, this embodiment employs the 3σ criterion for outlier handling, calculating the mean μ and standard deviation σ of each feature, and deleting abnormal records with feature values exceeding the range [μ-3σ, μ+3σ] to eliminate noise interference caused by data acquisition errors or sudden changes in operating conditions.
[0078] Step 102: Based on the preset variational autoencoder model, extract latent features from the original industrial data and determine the corresponding latent feature representations.
[0079] In this embodiment, the variational autoencoder (VAE) model may include an encoder and a decoder, and the latent feature representation reflects the core variational characteristics and potential spatial structure of the fluidized bed reactor. Thus, by processing the data using the VAE model, this application maps high-dimensional data into a latent spatial representation, capturing the core features of the data and providing a good feature representation for subsequent predictions.
[0080] Specifically, VAE is a generative model primarily used to learn latent variable representations of data and generate new data with a distribution similar to the original data. In this embodiment, standardized data can be used as input to the network, and the VAE model can be used as an unsupervised feature extraction tool to automatically extract features from the original training data, thereby reducing potential information loss.
[0081] In one possible implementation, refer to Figure 2 The diagram shown is a structural schematic of a variational autoencoder (VAE) model provided in an embodiment of this application. Its network structure typically includes an encoder and a decoder, which are connected through latent variables and implemented using a neural network structure to capture the latent spatial structure of the input data in the form of a probability distribution.
[0082] In one possible implementation, embodiments of this application can use the encoder of a variational autoencoder model to perform a nonlinear mapping operation on the raw industrial data to obtain latent spatial parameters, and then perform reparameterization processing on the latent spatial parameters to generate latent feature representations. Next, the decoder of the variational autoencoder model is used to reconstruct the latent feature representations to verify the information integrity of the latent feature representations.
[0083] Specifically, in this embodiment of the application, the VAE model consists of an encoder and a decoder. The encoder compresses high-dimensional features into low-dimensional latent representations through multi-layer nonlinear mapping. While preserving the core variation features of the data, it optimizes the latent space structure through regularization constraints. The decoder reconstructs the input data to verify the information integrity of the latent representation.
[0084] Specifically, in the embodiments of this application, the encoder can use a neural network to process the input high-dimensional data. Two key parameters for transformation into the latent space: mean vector and logarithmic variance , where L is the dimension of the latent variable. The specific calculation is as follows:
[0085]
[0086] in, Let be a nonlinear function fitted to a neural network, with parameters . .
[0087] Furthermore, the output is split into two parts:
[0088] =split( )
[0089] The input dimension D represents the number of features. Validation is performed in the {8, 16, 32, 64} dimensions using a grid search method. The best balanced performance is achieved on the test set when the dimension L of the latent variables is 16.
[0090] The encoder's first layer is a linear fully connected layer with an input dimension of D and an output dimension equal to the hidden layer size of 16, matching the input feature dimension. Experiments demonstrate that a second linear layer with a 32-dimensional output can effectively capture the non-linear features of the input data, including... .
[0091] Furthermore, in this embodiment, the reparameterization technique is used to calculate the latent variable Z as follows:
[0092] σ
[0093] in, , This represents the random error when sampling from a standard normal distribution. This formula implements sampling from the latent distribution.
[0094] Thus, in this embodiment, the decoder can estimate the conditional distribution P(X|Z) using the original sample X and the latent variable Z, and then sample and reconstruct the input data from the conditional distribution. Then, the reconstruction error guides the network update, thereby gradually improving the reconstruction accuracy.
[0095] In one possible implementation, the VAE model in this application embodiment can be jointly optimized using reconstruction loss and KL divergence loss. The reconstruction loss is used to measure the reconstruction error of the input data, and the KL divergence loss is used to constrain the difference between the latent distribution and the standard normal distribution.
[0096] Specifically, in this application, the reconstruction loss can use mean squared error (MSE) to measure the similarity between the reconstructed data and the original input data, and KL divergence loss is used to constrain the latent representation to be close to the standard normal distribution. The total loss of the final model is the weighted sum of the two, where the weight parameter can be 1.
[0097] Specifically, in the embodiments of this application, the training objectives of the VAE model include two parts: reconstruction loss and KL divergence.
[0098] The reconstruction error formula is as follows:
[0099]
[0100] The KL divergence is shown below:
[0101]
[0102] In summary, the comprehensive loss function of the VAE model in this application is:
[0103]
[0104] Where β is the weighting coefficient, the β ablation experiment verified that when β=1, the model performance fluctuates little and the value is reasonable.
[0105] In one possible implementation, the VAE model in this application embodiment can be trained using the Adam optimizer, trained with mini-batch data, and its reconstruction and generalization capabilities can be evaluated using cross-validation.
[0106] Specifically, this application embodiment can use the Adam optimizer for training. Its adaptive learning rate characteristic can effectively handle the noise features in industrial data. At the same time, weight decay is introduced to prevent overfitting. The training process is carried out through iterative optimization on small batches of data. The initial learning rate of 0.001 is determined through learning rate range testing, and the decay parameter of 0.0001 is determined through cross-validation. While ensuring the model convergence speed, this significantly reduces the risk of overfitting.
[0107] Step 103: Based on the trained convolutional neural network model, perform production prediction on the latent feature representation to obtain the predicted value of hydrogen feed rate of fluidized bed reactor.
[0108] In this embodiment, the Convolutional Neural Network (CNN) model includes a first convolutional layer, a second convolutional layer, and a fully connected layer. This application inputs the computation results of the VAE model into the CNN model to further extract data features and perform regression prediction. CNN excels at extracting features from local regions and is suitable for processing industrial data with temporal dependencies. Even after dimensionality reduction in the feature space, CNN can still effectively identify patterns in latent features. Thus, this embodiment can extract the latent spatial representation of data through VAE, and CNN only needs to perform regression tasks on the dimensionality-reduced features, thereby reducing the complexity of high-dimensional data. Furthermore, this application uses the VAE model to compress features of production data, filtering redundant information and retaining key features. Combined with the local perception and weight sharing characteristics of CNN, it extracts the spatiotemporal correlation of latent features and achieves multi-scale feature fusion through a cascaded architecture, breaking through the limitations of traditional models in modeling nonlinear systems and realizing high-precision prediction of complex nonlinear systems.
[0109] For details, please refer to Figure 3 The diagram shows a structural schematic of a convolutional neural network model provided in an embodiment of this application. This CNN model consists of an input layer, an output layer, and multiple intermediate hidden layers (including convolutional layers and fully connected layers). Each node is connected to the node in the previous layer via weights. The nodes perform a weighted sum of the inputs from the previous layer, and then perform a non-linear mapping through an activation function to generate the output of the current layer.
[0110] In one possible implementation, embodiments of this application can perform vector dimension transformation on the latent feature representation to obtain a corresponding two-dimensional tensor. Local features are extracted from the two-dimensional tensor through a first convolutional layer in the convolutional neural network model to obtain a primary feature representation. Then, a second convolutional layer in the convolutional neural network model performs cross-regional feature combination on the primary feature representation to generate a higher-order abstract representation. Finally, a fully connected layer in the convolutional neural network model performs regression prediction on the higher-order abstract representation to obtain the predicted hydrogen feed rate.
[0111] Specifically, in the embodiments of this application, the process of converting the latent features output by the VAE model into the input of the CNN may include: reshaping the one-dimensional latent vector into a two-dimensional tensor to adapt to the spatial input structure of the convolutional layer, and enhancing the feature expression capability by adding channel dimensions, so that the convolutional kernel can capture cross-dimensional correlations.
[0112] Specifically, in this embodiment, the CNN model may include two convolutional layers and one fully connected layer. The first convolutional layer detects abrupt changes in local features through a sliding window, the second convolutional layer captures cross-regional feature combinations, and deep convolutional operations combine shallow features to form a higher-order abstract representation. The fully connected layer is used to unfold all local features into a one-dimensional vector, and uses a learnable weight matrix to perform a weighted summation of feature importance, thereby outputting the final predicted value. The weight coefficients of the weight matrix represent the degree of influence of each feature on the hydrogen feed rate.
[0113] In one possible set of implementations, reference is made to... Figure 4 The diagram illustrates the processing procedure of a convolutional layer according to this application. In this embodiment, the input to the CNN model is a two-dimensional tensor, including batch size and latent dimension, where the latent dimension is the latent spatial feature dimension obtained by the VAE model. In this step, the data needs to undergo feature transformation, expanding the feature dimension of each sample to the input format required by the convolutional layer. This is typically done by adding an extra channel dimension, transforming it into a three-dimensional tensor. In the CNN model, the role of the convolutional layer is to extract local features by performing convolution operations on the input data using convolution kernels.
[0114] Specifically, the convolution operation formula in this application can be as follows:
[0115] =
[0116] in, This represents the convolution operation, where y is the feature map after convolution. These are the features of the input. These are the weights of the convolutional kernel. In this layer, the input data for the convolutional operation is the latent representation of the VAE model, and the output data is the features extracted through the convolutional layer.
[0117] Furthermore, after the convolution operation, the output feature map is flattened into a one-dimensional vector and fed into the output layer for further feature mapping. The output layer corresponds to a fully connected layer, which predicts the target value based on the extracted features and maps the input features to the target output through a linear transformation.
[0118] Specifically, as above Figure 3 As shown, in the CNN model of this embodiment, the goal of the regression task is to predict the hydrogen feed rate; therefore, the output layer has only one neuron. The input layer corresponds to the latent space obtained by the VAE, which has a dimension of 16. Convolutional layers 1 and 2 are both one-dimensional convolutional layers. Feature visualization shows that the first layer has 64 neurons and the second layer has 32 neurons, which reduces the number of parameters required and increases the computational speed compared to symmetric structures. The features extracted by these convolutional layers are then mapped to fully connected layers for the final regression prediction.
[0119] Specifically, in this embodiment, the CNN model uses mean squared error as the loss function and the Adam optimizer to update the network parameters. Model performance is evaluated using MAE, RMSE, and R² metrics. The model's MAE is 283, RMSE is 386, and R² is 0.98. Comparative experiments were also conducted. The prediction metrics using the gradient boosting regression model are as follows: MAE 342, RMSE 628, and R² 0.95. The prediction metrics using the linear regression model are as follows: MAE 542, RMSE 746, and R² 0.92.
[0120] For details, please refer to Figure 5 The image shown is a schematic diagram of a model prediction result provided in this application. Figure 5 This diagram illustrates the comparison between the predicted and actual values of the VAE-CNN model in this embodiment of the application. The vertical axis represents the feed flow rate of the E-05 tube side of the hydrogen primary heat exchanger, and the horizontal axis represents the sample index, with a scale from 0 to 300. The solid blue line in the diagram represents the actual value, i.e., the true measured data of the hydrogen feed rate in the fluidized bed reactor, while the dashed orange line represents the predicted value, i.e., the output result of the VAE-CNN model in this embodiment of the application. Based on the above... Figure 5 In the sample index range of 0 to 300, the dense distribution and changing trend of the two curves are visible. The predicted value curve can closely follow the fluctuation of the actual value curve. Especially at key points such as sample index 50, 100, 150, 200, 250, and 300, the predicted value and the actual value are highly coincident. This verifies that the technical means of extracting potential features through variational autoencoder and then regressing and predicting through convolutional neural network in the embodiments of this application have achieved high-precision fitting of hydrogen feed rate, which significantly improves the accuracy and stability of the prediction of fluidized bed reactor production process.
[0121] Further reference Figure 6The diagram shows a performance comparison of the models provided in this application embodiment. The horizontal axis is labeled with different test points (0.0, 1.0, 1.5, 2.0, 2.5), and the performance of the three models is compared using a three-dimensional bar chart of three colors. The blue bars represent the VAE-CNN model (WE_OWN) of this application embodiment, the orange bars represent the Random Forest model, and the green bars represent the Linear Regression model. At the 0.0 test point on the horizontal axis, the score of our model is 282.57, compared to 341.63 for the random forest model and 541.62 for the linear regression model. At the 1.0 test point, the score is 365.98, compared to 627.93 for the random forest model and 746.04 for the linear regression model. At the 1.5 test point, the score is 746.04, compared to 800.00 for the random forest model and 850.00 for the linear regression model. At the 2.0 and 2.5 test points, our model scores are both 0.98, significantly better than the 0.95 of the random forest model and the 0.97 of the linear regression model. Therefore, the above comparison results fully verify that our VAE-CNN model, in the task of predicting hydrogen feed rate in fluidized bed reactors, has a lower error score and more stable performance compared to traditional random forest and linear regression methods, demonstrating the technical advantages of our application through latent feature extraction and a cascaded convolutional neural network structure.
[0122] In summary, the comparison shows that the model constructed in this application's embodiments matches the actual values well, with small fluctuations in prediction error, demonstrating high prediction accuracy. Furthermore, the prediction method provided in this application's embodiments is simple to implement, computationally efficient, and has good applicability in actual fluidized bed reactor production process prediction scenarios.
[0123] In one possible implementation, after obtaining the predicted value of the hydrogen feed rate of the fluidized bed reactor, the present application embodiment can use the predicted value of the hydrogen feed rate to adjust and control the parameters of the fluidized bed reactor so that the efficiency of the fluidized bed reactor is maintained within a preset range.
[0124] Specifically, this application can achieve industrial scenario control through the above-mentioned prediction model and prediction process. For example, it can adjust the operating parameters such as the amount of catalyst, reaction temperature and pressure in the reactor in real time to keep the reactor efficiency within a preset range. It can also construct a prediction feedback closed-loop mechanism for the quench unit, determine the optimal combination of process parameters through iterative optimization, and continuously verify the applicability of the optimization strategy through historical data.
[0125] Please see Figure 7As shown, based on the same technical concept, this application also provides a computer device 70. In one embodiment, this computer device can be a device specifically used for predicting the production process of a fluidized bed reactor, or it can be a device for overall control of the fluidized bed reactor. The computer device is as follows... Figure 7 As shown, it includes a memory 701, a communication module 703, and one or more processors 702.
[0126] The memory 701 is used to store computer programs executed by the processor 702. The memory 701 may mainly include a program storage area and a data storage area. The program storage area may store the operating system and programs required to run instant messaging functions, etc.; the data storage area may store various instant messaging information and operation instruction sets, etc.
[0127] Memory 701 may be volatile memory, such as random-access memory (RAM); memory 701 may also be non-volatile memory, such as read-only memory, flash memory, hard disk drive (HDD), or solid-state drive (SSD); or memory 701 may be any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but is not limited thereto. Memory 701 may be a combination of the above-described memories.
[0128] The processor 702 may include one or more central processing units (CPUs) or digital processing units, etc. The processor 702 is used to implement the above-described fluidized bed reactor production process prediction method when it calls the computer program stored in the memory 701.
[0129] The communication module 703 is used to communicate with the industrial control system.
[0130] This application embodiment does not limit the specific connection medium between the memory 701, communication module 703, and processor 702 described above. This application embodiment... Figure 7 The memory 701 and the processor 702 are connected via a bus 704, and the bus 704 is in Figure 7 The diagram uses thick lines to describe the connections between other components; these are for illustrative purposes only and should not be considered limiting. The 704 bus can be divided into address bus, data bus, control bus, etc. For ease of description, Figure 7 It is described using only a thick line, but does not indicate that there is only one bus or one type of bus.
[0131] The memory 701 stores a computer storage medium, which stores computer-executable instructions. The computer-executable instructions are used to implement the production process prediction method of the fluidized bed reactor in the embodiments of this application. The processor 702 is used to execute the production process prediction method of the fluidized bed reactor in the above embodiments.
[0132] Based on the same inventive concept, embodiments of this application also provide a storage medium storing a computer program that, when run on a computer, causes the computer to perform the steps in the production process prediction method for a fluidized bed reactor according to various exemplary embodiments of this application described above.
[0133] In some possible implementations, various aspects of the fluidized bed reactor production process prediction method provided in this application can also be implemented in the form of a computer program product, which includes a computer program that, when run on a computer device, causes the computer device to perform the steps in the fluidized bed reactor production process prediction method according to various exemplary embodiments of this application as described above. For example, the computer device can perform the steps of the various embodiments.
[0134] The program product may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.
[0135] The program product of the embodiments of this application may employ a portable compact disc read-only memory (CD-ROM) and include a computer program, and may run on a computer device. However, the program product of this application is not limited thereto. In this application, the readable storage medium may be any tangible medium that contains or stores a program, and the computer program included therein may be used by or in conjunction with a command execution system, apparatus, or device.
[0136] A readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying a readable computer program. This propagated data signal may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable signal medium may also be any readable medium other than a readable storage medium, capable of sending, propagating, or transmitting a program for use by or in conjunction with a command execution system, apparatus, or device.
[0137] Computer programs contained on readable media may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.
[0138] Computer programs for performing the operations of this application can be written in any combination of one or more programming languages, including object-oriented programming languages such as Java and C++, as well as conventional procedural programming languages such as the "C" language or similar programming languages.
[0139] It should be noted that although several units or sub-units of the device have been mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to embodiments of this application, the features and functions of two or more units described above can be embodied in one unit. Conversely, the features and functions of one unit described above can be further divided and embodied by multiple units.
[0140] Furthermore, although the operations of the method of this application are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.
[0141] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0142] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.
[0143] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.
Claims
1. A method of production process prediction for a fluidized bed reactor, characterized by, The method comprises: obtaining original industrial data of a fluidized bed reactor, the original industrial data comprising a plurality of production data of the fluidized bed device and the quenching device; based on a preset variational autoencoder model, latent feature extraction is performed on the original industrial data to determine a corresponding latent feature representation; the variational autoencoder model comprises an encoder and a decoder, and the latent feature representation represents the core variation characteristics and the latent space structure of the fluidized bed reactor; based on a trained convolutional neural network model, production prediction is performed on the latent feature representation to obtain a hydrogen feed amount prediction value of the fluidized bed reactor.
2. The method of claim 1, wherein, The method comprises: based on the encoder, performing a nonlinear mapping operation on the original industrial data to obtain a latent space parameter; based on the latent space parameter, performing reparameterization processing to generate a latent feature representation; based on the decoder, data reconstruction is performed on the latent feature representation to verify the information integrity of the latent feature representation.
3. The method of claim 1, wherein, The convolutional neural network model comprises a first convolutional layer, a second convolutional layer and a fully connected layer, and the production prediction based on the trained convolutional neural network model comprises: performing vector dimension conversion on the latent feature representation to obtain a corresponding two-dimensional tensor; based on the first convolutional layer, performing local feature extraction on the two-dimensional tensor to obtain a primary feature representation; based on the second convolutional layer, performing cross-region feature combination on the primary feature representation to generate a high-order abstract representation; based on the fully connected layer, performing regression prediction on the high-order abstract representation to obtain the hydrogen feed amount prediction value.
4. The method of claim 1, wherein, After obtaining the original industrial data of the fluidized bed reactor, the method further comprises: preprocessing the original industrial data; performing correlation analysis on the preprocessed original industrial data to determine the hydrogen feed amount as a prediction target variable; based on the prediction target variable, constructing a variational autoencoder model.
5. The method of claim 4, wherein, The correlation analysis on the preprocessed original industrial data comprises: based on the preprocessed original industrial data, calculating a Pearson correlation coefficient matrix; based on the Pearson correlation coefficient matrix, determining the comprehensive correlation between variables in the original industrial data; based on the comprehensive correlation, screening the prediction target variable from the original industrial data, the prediction target variable being a variable associated with the fluidized bed reactor efficiency and meeting a preset condition.
6. The method of claim 4, wherein, The preprocessing of the original industrial data comprises: performing data cleaning and outlier processing on the original industrial data; performing standardization processing on the processed original industrial data to obtain standardized data; based on a preset ratio, dividing the standardized data into a training set and a test set.
7. The method of claim 1, wherein, After the production prediction based on the trained convolutional neural network model, the method further comprises: Based on the hydrogen feed amount prediction value, parameter adjustment control is performed on the fluidized bed reactor to maintain the efficiency of the fluidized bed reactor in a preset interval; the parameter adjustment control at least includes: regulating at least one parameter of the catalyst usage, the reaction temperature and the pressure of the fluidized bed reactor. 8.A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein, the processor executes the computer program to implement the steps of the method in any one of claims 1 to 7. 9.A computer storage medium having computer program instructions stored thereon, wherein, the computer program instructions are executed by a processor to implement the steps of the method in any one of claims 1 to 7. 10.A computer program product, comprising computer program instructions, wherein, the computer program instructions are executed by a processor to implement the steps of the method in any one of claims 1 to 7.