Chemical process parameter prediction method, equipment, medium and product

Abstract

The invention relates to the field of petrochemical industry control, in particular to a chemical process parameter prediction method and device, a medium and a product. According to the method, a process parameter set of target chemical equipment is obtained, the process parameter set is input into a trained dynamic fusion model to obtain a prediction parameter value output by the dynamic fusion model, and the dynamic fusion model can fuse historical process parameter information. In the learning process, the time history correlation between sample data is more comprehensively captured, and the correlation between the hidden state and the history is enhanced, so that the prediction accuracy and efficiency of the chemical process parameters are improved.

Description

Technical Field

[0001] The present application relates to the field of petrochemical industry control, and in particular to a chemical process parameter prediction method, equipment, medium and product. Background Art

[0002] As a national pillar industry, the oil refining industry's core processes (such as catalytic cracking, fractional distillation and absorption) are characterized by strong coupling, nonlinearity, and large hysteresis. In order to improve refining efficiency and quality, it is often necessary to build accurate process models to predict process parameters in chemical processes such as refining.

[0003] Traditional methods primarily rely on physical and chemical equations to construct mechanistic models. However, the refining process involves the interaction of hundreds of variables (e.g., temperature, pressure, flow rate, and other multi-parameter coupling), and the reaction process can take hours to days. This makes mechanistic modeling difficult and results in low predictive accuracy. For example, nonlinear relationships between variables are difficult to accurately express through differential equations, and model errors accumulate over the duration of the process. Furthermore, disturbances in actual production, such as raw material fluctuations and catalyst deactivation, cannot be captured by static equations, resulting in extremely high model maintenance costs. Therefore, traditional modeling methods are significantly limited in terms of predictive accuracy, generalization capabilities, and adaptability to new situations. Summary of the Invention

[0004] The present invention provides a chemical process parameter prediction method, equipment, medium and product, which are used to improve the prediction accuracy and efficiency of chemical process parameters.

[0005] In a first aspect, the present application provides a method for predicting chemical process parameters, the method comprising:

[0006] Obtaining a process parameter set of a target device and inputting the process parameter set into a trained dynamic fusion model; the process parameter set includes a plurality of process parameters generated by the target device in a target chemical process;

[0007] Obtaining a predicted parameter value output by the dynamic fusion model, wherein the predicted parameter value represents a process parameter that may be generated by the target device at a future moment;

[0008] Among them, the dynamic fusion model includes a historical gated recurrent unit (HGRU) encoder and a dynamic decoder, and the HGRU encoder is trained based on an improved teacher forcing strategy and a historical supervision sequence; the dynamic decoder determines the fusion feature based on the difference between the output value of the HGRU encoder and the historical supervision sequence, and performs model training based on the fusion feature.

[0009] Optionally, the dynamics fusion model is trained in the following manner:

[0010] Obtaining a training time series corresponding to the target chemical process and determining a corresponding historical supervision sequence; the training time series represents process parameters generated by the target equipment within a specific time range, and the historical supervision sequence includes multiple process parameters of the target equipment at a priori time;

[0011] Based on a preset improved teacher forcing strategy, the HGRU encoder is iteratively trained in combination with the historical supervision sequence and the first loss function; the improved teacher forcing strategy characterizes the process parameters of the target device in the previous time step as input of the current time step;

[0012] Determining a fusion feature based on the difference between the predicted output of the HGRU encoder and the historical supervision sequence; the fusion feature represents the dynamic deviation between the current encoding state and the historical true value;

[0013] Based on the fusion features, the parameters of the dynamics decoder are optimized to obtain a trained dynamics fusion model.

[0014] Optionally, the iterative training of the HGRU encoder based on the preset improved teacher forcing strategy, combined with the historical supervision sequence and the first loss function, includes:

[0015] Initializing the historical supervision sequence based on a supervision variable estimation strategy; the supervision variable estimation strategy represents initializing each historical supervision variable to the output value of the previous variable;

[0016] Performing forward transfer processing on the initialized historical supervision sequence to obtain a hidden state corresponding to the HGRU encoder, wherein the hidden state represents the historical dependency relationship of the process parameters in the target chemical process;

[0017] Performing a linear change on the hidden state to obtain a predicted output of the HGRU encoder;

[0018] Based on the difference between the predicted output and the historical supervision sequence, combined with the first loss function, the parameters of the HGRU encoder are optimized.

[0019] Optionally, determining the fusion feature based on the difference between the predicted output of the HGRU encoder and the historical supervision sequence includes:

[0020] Determining a predicted output of the HGRU encoder based on a preset forward calculation strategy;

[0021] Based on the first fuser, the fusion feature is determined in combination with the vector difference between the predicted output and the target supervision variable; the target supervision variable is the historical supervision variable corresponding to the last moment in the historical supervision sequence.

[0022] Optionally, the step of optimizing parameters of the dynamics decoder based on the fusion features to obtain a trained dynamics fusion model includes:

[0023] Based on multiple fully connected layers, nonlinear mapping is performed on the fusion features respectively, and weighted merging of the mapping results is performed based on preset weights to obtain a merging result;

[0024] Based on a preset solver, combined with the merged results, a dynamic prediction result is obtained;

[0025] The parameters of the dynamic decoder are optimized based on the dynamic prediction result and the second loss function; the second loss function represents the loss value calculated at the final time point.

[0026] Optionally, obtaining a dynamics prediction result based on a preset solver and combining the merged result includes:

[0027] Determining a difference between a state change predicted by the preset solver and a corresponding supervisory variable based on the state change predicted by the preset solver for the fusion feature;

[0028] The differences are fused based on a second fuser to obtain the dynamics prediction result.

[0029] In a second aspect, the present application provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, any one of the chemical process parameter prediction methods described in the first aspect is implemented.

[0030] In a third aspect, the present application provides a computer storage medium, wherein the computer-readable storage medium stores computer program instructions, and the computer program instructions are executed by a processor to implement any one of the chemical process parameter prediction methods described in the first aspect.

[0031] In a fourth aspect, an embodiment of the present application provides a computer program product, comprising computer program instructions, which, when executed by a processor, implement any one of the chemical process parameter prediction methods described in the first aspect.

[0032] The beneficial effects of the present invention are as follows:

[0033] The present invention provides a chemical process parameter prediction method, device, medium, and product. This method obtains a process parameter set for a target chemical plant and inputs it into a trained kinetic fusion model to obtain predicted parameter values ​​output by the kinetic fusion model. The kinetic fusion model can integrate historical process parameter information, more comprehensively capturing the temporal and historical correlations between sample data during the learning process, and strengthening the correlation between hidden states and history, thereby improving the accuracy and efficiency of chemical process parameter prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are only embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0035] Figure 1 A schematic diagram of a training process of a dynamics fusion model provided in an embodiment of the present application;

[0036] Figure 2 A schematic diagram of a process flow of a debutanizer device provided in an embodiment of the present application;

[0037] Figure 3 A schematic diagram of a process flow of a Tennessee process provided in an embodiment of the present application;

[0038] Figure 4 A schematic diagram of the training and testing process of a dynamic fusion model;

[0039] Figure 5 A schematic diagram of the structure of a dynamic fusion model provided in an embodiment of the present application;

[0040] Figure 6 A schematic diagram of the framework of an HRGU decoder provided in an embodiment of the present application;

[0041] Figure 7 A schematic diagram of a fusion device provided in an embodiment of the present application;

[0042] Figure 8 A flow chart of a method for predicting chemical process parameters provided in an embodiment of the present application;

[0043] Figure 9 A schematic diagram showing a comparison of normalized butane concentrations in different models provided in an embodiment of the present application;

[0044] Figure 10 A scatter plot of prediction differences of different models provided in an embodiment of the present application;

[0045] Figure 11 A schematic diagram of the absolute error of predictions of different models provided in an embodiment of the present application;

[0046] Figure 12 A schematic diagram of the structure of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solutions and advantages of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application. Unless there is a conflict, the embodiments in the present application and the features in the embodiments can be combined with each other in any way. In addition, although a logical order is shown in the flowchart, in some cases, the steps shown or described can be performed in an order different from that here.

[0048] The terms "first" and "second" in the specification and claims of this application and the above-mentioned drawings are used to distinguish different objects, rather than to describe a specific order. In addition, the term "comprising" and any of its variations 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 optionally also includes steps or units that are not listed, or optionally also includes other steps or units inherent to these processes, methods, products or devices. "Multiple" in this application can mean at least two, for example, two, three or more, and the embodiments of this application are not limited thereto.

[0049] The term "and / or" in the embodiments of this application is simply a description of the association relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the following three situations: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this document generally indicates that the related objects are in an "or" relationship.

[0050] It is understood that in the following specific implementation methods of this application, data related to chemical processes, etc. are involved. When the various embodiments of this application are applied to specific products or technologies, relevant licenses or consents need to be obtained, and the collection, use and processing of relevant data need to comply with relevant laws, regulations and standards of relevant countries and regions. For example, it is possible to recruit relevant volunteers and sign relevant agreements for volunteer authorization data, and then use the data of these volunteers for implementation; or, by implementing within the scope of an authorized organization, data management is carried out by implementing the following implementation methods using the data of internal members of the organization; or, the relevant data used in the specific implementation are all simulated data, such as simulated data generated in a virtual scene.

[0051] The following is a brief introduction to the design concept of the embodiments of this application.

[0052] The petrochemical industry is a pillar industry in China and plays a vital role in the domestic economy. As one of the core processes in domestic oil refining, catalytic cracking continues to play a vital role in the industry. After decades of development, my country has become a major petrochemical manufacturing country, and with rapid economic growth, catalytic cracking processing capacity has continued to rise.

[0053] The petrochemical industry currently faces a range of challenges, including overcapacity, rising costs, and declining profitability. In recent years, a range of artificial intelligence technologies have been widely applied to catalytic cracking processes to improve refining efficiency and quality. To further enhance the production efficiency of refining companies, more accurate process models are required.

[0054] However, the entire catalytic cracking process is long, and the reaction process exhibits significant coupling and nonlinear characteristics. Establishing process mechanistic knowledge and constructing a mechanistic model for the catalytic cracking unit is extremely difficult. By capturing the nonlinear relationships between input and output through data-driven learning, a process-descriptive model can be constructed. Therefore, building a highly accurate data-driven model is crucial. Previous data models typically employ recurrent neural networks (RNNs), long short-term memory networks (LSTMs), and gated recurrent units (GRUs). However, RNNs suffer from the vanishing and exploding gradient problems. RNNs rely on the hidden state of the previous moment to predict future data, which limits their ability to capture long-term dependencies. Given the process complexity of refining and the high-dimensional, nonlinear nature of data, traditional machine learning and neural network methods struggle to capture continuous-time nonlinear dynamics and structural-dynamic dependencies, resulting in limitations in prediction accuracy, generalization, and adaptability to new situations.

[0055] In view of the above problems, embodiments of the present application provide a chemical process parameter prediction method, apparatus, medium, and product. This method obtains a set of process parameters for a target chemical plant and inputs the process parameter set into a trained kinetic fusion model to obtain predicted parameter values ​​output by the kinetic fusion model. The kinetic fusion model can integrate historical process parameter information, more comprehensively capturing the temporal and historical correlations between sample data during the learning process, and strengthening the correlation between hidden states and history, thereby improving the accuracy and efficiency of chemical process parameter prediction.

[0056] Below, the method provided by the exemplary embodiment of the present application will be described with reference to the accompanying drawings. It should be noted that the above application scenarios are only shown to facilitate understanding of the spirit and principles of the present application, and the implementation of the present application is not limited in this respect.

[0057] In the embodiment of the present application, a kinetic fusion model can be used to implement the above-mentioned chemical process parameter prediction process. Before the kinetic fusion model is put into use, model training is required in advance to make the kinetic fusion model converge. Next, in order to facilitate the description of the model application process, the training process of the kinetic fusion model is first introduced.

[0058] Please refer to Figure 1 , is a schematic diagram of a training process of a dynamic fusion model provided in an embodiment of the present application. The specific implementation process of the training process is as follows:

[0059] Step 101: Obtain a training time series corresponding to the oil refining industry process and determine a corresponding historical supervision sequence.

[0060] In the embodiment of the present application, the training time series represents the process parameters generated by the target device within a specific time range, and the historical supervision sequence includes multiple process parameters of the target device at a priori time.

[0061] In a possible implementation, the present application may define a supervision sequence through the real historical value history at a priori moment. The historical supervision sequence definition formula is as follows:

[0062] s(t)=y(t)

[0063] S(T)={s(1), s(2),..., s(T)}

[0064] ={y(1), y(2), ..., y(T)}

[0065] Among them, s(t) is the historical supervision vector at time t, y(t) is the true historical value at time t, and S(T) is the historical supervision sequence at time T.

[0066] In one possible implementation, the embodiment of the present application may collect training time series from the oil refining industry process and use the Min-Max Normalization method on the collected time series data to optimize the performance of the network and improve learning efficiency.

[0067] Specifically, the normalization calculation in this application is as follows:

[0068]

[0069] Where x is the original data, min(x) and max(x) are the minimum and maximum values ​​of the feature data, respectively, and x′ is the normalized data.

[0070] In one possible implementation, reference Figure 2 The figure shows a process flow diagram of a debutanizer column device provided in an embodiment of the present application. A debutanizer column (DC) is a device used to separate butane from heavier components during the petroleum refining process. Because butane concentration significantly affects the debutanizer column, predicting butane concentration is crucial for its control and optimization. Therefore, this application uses the DC column as the target device, trains a corresponding kinetic fusion model, and predicts butane concentration in the DC column.

[0071] In one possible implementation, reference Figure 3 Shown is a process flow diagram of a Tennessee process provided in an embodiment of the present application, in which a standard Tennessee process (Tennessee Eastman, TE) device can simulate a variety of complex chemical reaction processes and process mixed raw materials of various chemical components. The mixed raw materials are preliminarily heated in a preheater by heat exchange with the reaction products, and then mixed with a specific reaction gas in a reactor. In the environment created by the high temperature and high pressure of the reactor and the catalyst, a series of complex chemical reactions occur, and some impurities contained in the raw materials are removed at the same time. The product mixture after the reaction is then passed through a series of separation systems and fractionation systems such as a condenser, a compressor, a separator and a stripping tower to produce various chemical products. Therefore, the present application can use the TE device as the target device, train the corresponding kinetic fusion model, and perform component prediction of the TE device.

[0072] Step 102: Based on the preset improved teacher forcing strategy, combined with the historical supervision sequence and the first loss function, the HGRU encoder is iteratively trained.

[0073] In an embodiment of the present application, the improved teacher forcing strategy represents taking the process parameters of the target device in the previous time step as the input of the current time step. Through the improved teacher forcing strategy, the present application can establish a historical gated recurrent unit (HGRU) deep network model, and learn hidden information based on input variables and historical output variables through a predefined loss function and historical supervision sequence. In this way, the encoder HGRU network can be responsible for processing the input information and capturing the historical dependencies therein, generating a feature representation that is sensitive to the current time step.

[0074] In one possible implementation, the calculation formula of the loss function in this application is as follows:

[0075]

[0076] Among them, L is the loss function, T is the time, y(t) is the actual historical value at time t, and y(t) is the predicted value at time t.

[0077] Furthermore, the goal of training the model is to find a set of optimal parameters θ * , which can minimize the total loss of training:

[0078]

[0079] In this way, this application can use the backpropagation algorithm to tune the network parameters and obtain the network weights and bias parameters. The Adam algorithm is used to train the HGRU model, combining the ideas of momentum gradient descent and the root mean square backpropagation algorithm to effectively update the weights in the neural network and calculate the adaptability of different parameters.

[0080] In one possible implementation, the present application can initialize the historical supervision sequence through a supervision variable estimation strategy, and perform forward transfer processing on the initialized historical supervision sequence to obtain the hidden state corresponding to the HGRU encoder, which represents the historical dependency of the process parameters in the target chemical process. The hidden state is then linearly changed to obtain the predicted output of the HGRU encoder, and the difference between the predicted output and the historical supervision sequence can be used to optimize and adjust the parameters of the HGRU encoder to complete the training of the HGRU encoder.

[0081] Specifically, the supervisory variable estimation strategy in the embodiment of the present application initializes the historical supervisory variable to the output value of the previous sample for each x(T+j) from the first to the last sample in the test sequence, and the historical supervisory sequence S(T) is as follows:

[0082] S(T+j)=y(T+j-1)

[0083] S(T+j)={s(T+1), s(T+2),..., s(T+j)}

[0084] ={y(T),y(T+1),...,y(T+j-1)}

[0085] Among them, s(T) is the historical supervision vector at time T, y(T) is the true historical value at time T, and S(T+j) is the historical supervision sequence at time T+j.

[0086] In this way, the application can substitute the initialized data sample into the HGRU network and obtain the hidden state through forward transfer, where the formula for forward transfer is:

[0087] netz=W z x(t)+U z h(t-1)+V z s(t)+b z

[0088] z(t)=σ(netz)

[0089] netr=W r x(t)+U r h(t-1)+V r s(t)+b r

[0090] r(t)=σ(netr)

[0091]

[0092] where z(t) is the update gate of the history-gated recurrent unit, r(t) is the reset gate, h(t) is the candidate hidden state, h(t) is the final hidden state, tanh is the Tanh activation function, σ is the Sigmoid activation function, and · represents the Hadamard product.

[0093] In one possible implementation, the present application can perform a linear transformation on the obtained hidden state through a fully connected layer (FC) layer, thereby extracting a feature representation that matches the dimension of the target variable and obtaining the final prediction output. The calculation of the prediction output is as follows:

[0094] y(t)=Vh(t)+b

[0095] Among them, V and b are the weight and bias terms of the hidden state of the fully connected layer, respectively.

[0096] Step 103: Determine the fusion feature based on the difference between the predicted output of the HGRU encoder and the historical supervision sequence.

[0097] In the embodiment of the present application, the fusion feature represents the dynamic deviation between the current encoding state and the historical true value.

[0098] In one possible implementation, the embodiment of the present application can determine the predicted output of the HGRU encoder through a preset forward calculation strategy, and determine the fusion feature through a first fusion device by combining the vector difference between the predicted output and the target supervision variable, where the target supervision variable is the historical supervision variable corresponding to the last moment in the historical supervision sequence.

[0099] Specifically, the embodiment of the present application can use the fuser 1 to calculate the vector difference between the current encoding output of the HGRU encoder and the historical supervision variable at the last moment, thereby fusing the encoding features of the HGRU to obtain the fused features.

[0100] Specifically, the calculation process of vector difference is as follows:

[0101] Z(t)=X(t)-s(t)

[0102] Among them, is the vector difference between the current encoding output and the historical supervision variable at the last moment, is the current encoding output, and is the historical supervision variable.

[0103] In one possible implementation, the encoding output is obtained by the HGRU forward calculation, as follows:

[0104] X(t)=HGRU(X(t),s(t))

[0105] Among them, s(t) is the historical supervision variable at the prediction moment, and X(t) is the output of the HRGU network coding.

[0106] Fusion 1 transforms the HGRU’s changing features y hgru (t) Fusion, the fusion feature calculation method is:

[0107] Δy(t)=y hgru (t)-s(t)

[0108] Δy(t) is the fusion feature, y hgru (t) is the encoded feature and s(t) is the historical supervision variable.

[0109] In this way, the fused feature Δy(t) not only performs dynamic analysis on the encoding features, but also provides key dynamic indicators for the DODE decoder.

[0110] Step 104: Based on the fusion features, optimize the parameters of the dynamics decoder to obtain a trained dynamics fusion model.

[0111] In the embodiment of the present application, the dynamic decoder based on the neural ordinary differential equation can capture the deep complex dynamic relationships in the fusion features, realize the accurate prediction and simulation of the dynamic changes in the long-term prediction, and significantly improve the model's high-precision prediction of the behavior of complex systems.

[0112] In one possible implementation, the present application performs nonlinear mapping processing on the fusion features through multiple fully connected layers, and performs weighted merging of the mapping results obtained according to the preset weights of each fully connected layer to obtain a merged result, and then optimizes the parameters of the dynamic decoder based on the merged result and the preset second loss function. The second loss function only calculates the loss at the final time point, so that the embodiment of the present application only evaluates the loss at the end of the entire time range, thereby reducing the error accumulated due to multiple numerical integrations.

[0113] Specifically, the present application can perform nonlinear mapping on the fused features through two FC layers, and weight the mapping results output by each FC layer respectively, so as to merge the two mapping results.

[0114] Specifically, the forward formula of the fully connected layer 1 is:

[0115] h1=Tanh(W1x+b1)

[0116] x1=W2h1+b2

[0117] Among them, W1 and W2 are the weights of the fully connected layer 1, and b1 and b2 are biases.

[0118] Specifically, the forward formula of the fully connected layer 2 is:

[0119] h2=Softplus(U1x+b3)

[0120] x2=U2h2+b4

[0121] Among them, U1 and U2 are the weights of the fully connected layer 2, and b3 and b4 are biases.

[0122] Furthermore, the combined calculation formula is:

[0123] y=[x1 x2][αβ] Τ =x1·α+x2·β

[0124] where and are learning parameters.

[0125] In a possible implementation, the present application can solve the ordinary differential equation through an improved ODE solver, integrate the hidden state change rate, and thus obtain the dynamic prediction result.

[0126] Specifically, the calculation formula of the improved ODE solver is:

[0127] Z(t1)=ODESolve(Z(t0),f,t0,t1,θ)

[0128] Where ODESolve represents the numerical solver for the integral, are the parameters of the neural network, and are the system states at time points and respectively.

[0129] In one possible implementation, the present application implements the problem of inaccurate gradient calculation and unstable numerical solution in the training process of processing neural ordinary differential equation networks. The formula for calculating the gradient of the neural network using the adjoint state is:

[0130]

[0131] Because the adjoint method decomposes the gradient calculation process into integrals over successive time steps, it can lead to subtle differences from the gradients obtained directly via the chain rule. Furthermore, the numerical solution algorithms of ODE solvers are inherently unstable, further increasing the computational error at each time point. These small errors can accumulate over time steps, ultimately affecting the accuracy of the overall model gradient and the stability of training.

[0132] Therefore, to reduce the above errors, the embodiment of the present application further adjusts the integration processing strategy of the loss function. The traditional method is to accumulate losses at multiple time points, thereby calculating the loss of state z(t) at each time point t and accumulating these losses. However, solving at each time point may introduce additional errors, especially when the numerical integration method using the ODE solver is not stable enough.

[0133] In order to improve this process, the present embodiment defines a new loss function L final , which calculates the loss at the final time point t1, and its loss function is as follows:

[0134] L final =L(Z(t1))

[0135] In this way, the embodiment of the present application only evaluates the loss at the end of the entire time range, thereby reducing the error accumulated by multiple numerical integrations. At the same time, ODESlove only needs to be applied to the final state, reducing the error that may be introduced in the intermediate states. In this way, the embodiment of the present application achieves more stable and accurate gradient calculation in the neural dynamics network, while maintaining the high accuracy of the model's prediction of the system.

[0136] In one possible implementation, the embodiment of the present application can process the difference between the state change of the fused feature and the supervisory variable through the second fuser and the DODE decoder to enhance the dynamic prediction at the current moment, thereby generating the final prediction output.

[0137] Specifically, the embodiment of the present application can utilize the fuser 2 to fuse the difference between the state change of the fused feature and the supervisory variable according to the DODE decoder to enhance the dynamic prediction at the current moment and generate the final prediction output. The calculation formula of the output is as follows:

[0138] Z(t)=DODE(Z(t),θ)

[0139] Among them, Z(t) is the state change predicted by the decoder DODE based on the state difference and model parameters.

[0140] The DODE decoder further processes the fusion features output by the fuser 1 and predicts the micro-changes in the state, which reflect the natural evolution of the state from the current moment to the next moment. The update expression of the fuser 2 is:

[0141] y(t)=s(t)+Δy dode (t)

[0142] Among them, s(t) is the prediction result of the previous moment, y(t) is the updated prediction value of the current moment, and y(t) is the predicted state change.

[0143] In this way, the model combines these predicted changes with the prediction results of the previous moment and updates the prediction value at the current moment, thereby strengthening the model's pre-judgment of the future state.

[0144] In one possible implementation, reference Figure 4 The figure shows a training and testing process diagram of a dynamic fusion model provided by an embodiment of the present application. Figure 4In the present application, multi-dimensional time series data (x represents process parameters, y represents target variables) can be collected from industrial processes (such as catalytic cracking units), separated into training sets (for model learning) and test sets (for performance verification) by dividing the data modules, and performing maximum-minimum normalization on all data to eliminate dimensional differences. In the training process, the present application can configure the parameters of the model, such as setting hyperparameters such as the number of network layers and the number of neurons, and determining the network structure (such as the hidden layer dimension of the HGRU encoder). And define the historical supervision variable as the true value (s(t) = y(t)), and input the process parameter x(t) and the supervision variable s(t) into the HGRU deep network for training within the time step t = 1, 2...T. At the end of the training stage, the supervision sequence is expanded according to the rule s(T+j) = y(T+j-1) (j is the number of prediction steps), and the test data x(T+1) and s(T+1) are input into the trained model, and the model outputs the predicted value Continuous integration operations on state variables are performed to realize dynamic evolution modeling of process parameters. The final prediction result y is compared with the actual industrial process data to verify the model accuracy and generalization ability.

[0145] In one possible implementation, reference Figure 5The figure shows a structural diagram of a dynamic fusion model provided by an embodiment of the present application, including three core modules: HGRU calculation, dynamic fusion and neural ordinary differential equations. The HGRU calculation process includes: starting from the input sequence (Input) X1, X2, ..., Xn, processing it step by step through the HGRU network to generate the encoding feature) X(t) and the historical supervision sequence S, wherein the gating mechanism of the HGRU (update gate, reset gate) enhances the temporal correlation by fusing historical true values. Dynamic fusion includes fuser 1 and fuser 2, wherein fuser 1 inputs the difference Z(t) between the encoding feature X(t) output by the HGRU and the historical supervision variable S into the fully connected layer 1, maps it to a high-dimensional space through a nonlinear activation function (Tanh and Softplus), and generates dynamic features [x1, x2]. Based on the dynamic features, the ODE solver is driven to integrate the state differential equation and output the state change. Fusher 2 generates the final prediction value y by weighted superposition of the state change predicted by the ODE and the historical supervision variable to achieve residual correction. The fully connected layer 2 and the neural network then perform a secondary nonlinear mapping on the fusion result to optimize the prediction accuracy. In this way, after the input data X, S is encoded by HGRU, it interacts with the historical variables through the fuser 1, drives the DODE decoder to model continuous dynamics, and finally the fuser 2 outputs the predicted value y, forming a closed loop of "time series encoding → dynamic fusion → residual prediction". In summary, the embodiment of the present application processes the input information through the history gated recurrent unit neural network encoder, captures the historical dependencies therein, and generates a feature representation that is sensitive to the current time step. Then, it is passed to the decoder neural dynamics network through the feature fusion module. Finally, the dynamic network decoder based on the neural ordinary differential equation captures the deep complex dynamic relationships in the fusion features, and realizes accurate prediction and simulation of dynamic changes in long-term predictions. Experimental results show that compared with other methods, the dynamic fusion model based on the neural ordinary differential equation has superior long-term prediction capabilities. In addition, the proposed nonlinear history gated recurrent unit (HGRU) deep network model can fuse historical information, so that the time history correlation between sample data can be more comprehensively captured during the learning process, and the correlation between hidden states and history is strengthened.

[0146] In one possible implementation, reference Figure 6 The figure shows a schematic diagram of the framework of a HRGU decoder provided by an embodiment of the present application. Figure 6In this application, the process parameters X(t) (such as temperature and pressure) at the current time step are used as the input of the neural network and feature encoded through multiple layers of neurons (represented by gray dots). The HGRU encoder (neural network part) is composed of multiple layers of neuron nodes, which receives X(t) and integrates historical supervision variables to generate encoded features through an improved gating mechanism (update gate / reset gate). The ODE solver receives the intermediate state features output by the neural network and solves the ordinary differential equation through numerical integration to predict the state evolution at the next time step.

[0147] In one possible implementation, reference Figure 7 The figure shows a framework diagram of a fusion device provided in an embodiment of the present application, wherein the fusion device 1 receives an input 1 containing process parameters and historical supervision variables and an independent process parameter input 2, performs gated cyclic encoding processing through the HGRU unit, generates state difference features and retains the updated supervision sequence, and the fusion device 2 receives a dual-path state difference feature input (input 1 and input 2) from the fusion device 1, performs dynamic integral solution on the ordinary differential equation through the DODE unit, and finally outputs the dual-path fusion result as a predicted value.

[0148] After obtaining the trained kinetic fusion model, the embodiment of the present application can use the kinetic fusion model to predict and process the chemical process parameters to obtain process parameter information that may be generated by the target equipment at a future time.

[0149] refer to Figure 8 FIG. 1 is a flow chart of a method for predicting chemical process parameters provided in an embodiment of the present application. The specific implementation process of the method is as follows:

[0150] Step 801: Obtain a process parameter set of a target chemical equipment, and input the process parameter set into a trained dynamic fusion model.

[0151] In the embodiment of the present application, the process parameter set includes multiple process parameters generated by the target equipment in the target chemical process.

[0152] Step 802: Obtain the predicted parameter values ​​output by the dynamics fusion model.

[0153] In the embodiment of the present application, the predicted parameter value represents the process parameter that may be generated by the target device at a future moment.

[0154] In a possible embodiment, the present application embodiment takes the debutanization separation process as an example of the target chemical process to be predicted. Figure 2The DC process flow diagram shown in the figure has seven secondary variables. U1 is the tower top temperature, U2 is the tower top pressure, U3 is the reflux flow rate, U4 is the split flow rate, U5 is the tower temperature, U6 is the tower bottom temperature A, and U7 is the tower bottom temperature B. Table 1 below shows the secondary variables of the debutanizer soft sensor. The DC dataset contains approximately 2400 samples with a sampling time of 6 minutes and is normalized. 1560 samples were randomly selected from the dataset as training samples, and the remaining 840 samples were used as testing samples.

[0155] Table 1

[0156]

[0157] The configuration settings of the HGRU deep network model proposed in the embodiment of the present application are as follows: the number of iterations is set to 60, the optimizer is Adam, the loss function is MSE, and the batch size is 64. The configuration settings of the HGRU-DODE model are as follows: the number of iterations is set to 48, the number of hidden layer neurons of the HGRU model is 128, the number of hidden layer neurons of the ODE solver is 512, and the batch size is 64. In addition, in the model comparison, this time we compared the standard RNN, LSTM, GRU, SLSTM supervised models, as well as DiPLS, KDiPLS, MLP, OTKR, N-Beats, MDRNN, BiLSTM, CNN-BiL1STM and other models. Refer to Table 2 below for a comparison of the performance indicators of butane concentration prediction of different models.

[0158] Table 2

[0159]

[0160] As shown in Table 2 above, the proposed HGRU-DODE model significantly improves the various prediction performance indicators of HGRU, demonstrating its high prediction accuracy and strong generalization capabilities. Experimental results show that the introduction of the DODE approach to HGRU enhances the model's ability to capture historical data and improves its prediction and generalization capabilities for unknown data.

[0161] For further reference, Figure 9 The figure shows a schematic diagram of the normalized comparison of butane concentrations of different models provided in the embodiment of the present application. It can be seen that the HGRU-DODE model of the present application has the highest prediction accuracy on the training set, with R2 reaching 0.998, and R2 on the test set also reaching 0.99. Figure 10The figure shows a scatter plot of prediction differences of different models provided in an embodiment of the present application. The comparison of the scatter plot shows that the prediction accuracy of HRRU-DODE on the training set is higher than that of the SLSTM and HGRU models, and the test prediction effect is consistent with the training effect.

[0162] In a possible implementation, the present application embodiment takes the TE process as an example of the target chemical process to be predicted, and the TE process is a typical industrial process. Figure 2 The following is a schematic diagram of the Tennessee process, showing the reaction of feeds A and C with D and E, respectively, to produce products G and h. The process primarily consists of five process units: a two-phase reactor where the exothermic reaction occurs, a separator, a stripping column, a compressor, and a condenser. FI represents the flow indicator, PI represents the pressure indicator, LI represents the liquid level indicator, and TI represents the temperature indicator. It can simulate a variety of complex chemical reactions and process mixed feedstocks with diverse chemical compositions. The mixed feedstock is initially heated in the preheater through heat exchange with the reaction products before being mixed with the specified reaction gas in the reactor. Within the high temperature, high pressure, and catalyst-generated environment of the reactor, a series of complex chemical reactions occur, simultaneously removing some impurities from the feedstocks. The product mixture then passes through a series of separation and fractionation systems, including a condenser, compressor, separator, and stripping column, to produce various chemical products.

[0163] Specifically, see Table 3 below for the input variables of the TE process. The TE process dataset contains 7200 samples, collected over 72 hours with a sampling interval of 0.01 hours. 5760 samples were randomly selected from the dataset as training samples, and the remaining 1440 samples were used as test samples.

[0164] Table 3

[0165]

[0166] Table 4 below compares the performance of different models for the Tennessee process C component prediction. HGRU-DODE achieves higher prediction metrics than other models for the Tennessee process, and significantly improves on all performance indicators compared to the HGRU model. These results demonstrate that HGRU-DODE further improves HGRU's dynamics capture capabilities, enhancing the model's predictive power.

[0167] Table 4

[0168]

[0169] For further reference, Figure 11The figure shows the absolute error of predictions of different models provided in the embodiment of the present application, which further proves the superiority of the HGRU-DODE model of the present application. The scatter points of the predictions and actual fittings of the HGRU-DODE training set are all located on the diagonal line, and the R2 reaches 0.999. Figure 10 In the test set predictions shown, the scatter plots that closely match the actual results are also tightly wrapped around the diagonal line, with an R² of 0.997, indicating high prediction accuracy. This indicates that the HGRU-DODE model has good consistency in its predictions for the training and test sets. These results demonstrate the strong predictive accuracy of HGRU-DODE.

[0170] See Figure 12 As shown, based on the same technical concept, the embodiment of the present application further provides a computer device 120. In one embodiment, the computer device can be a target reaction device or a control device dedicated to controlling the target reaction device. Figure 12 As shown, it includes a memory 1201 , a communication module 1203 and one or more processors 1202 .

[0171] Memory 1201 is used to store computer programs executed by processor 1202. Memory 1201 may primarily include a program storage area and a data storage area. The program storage area may store an operating system and programs required for running instant messaging functions, while the data storage area may store various instant messaging messages and operating instruction sets.

[0172] Memory 1201 may be a volatile memory, such as random-access memory (RAM); a non-volatile memory, such as read-only memory, flash memory, a hard disk drive (HDD), or a solid-state drive (SSD); or 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 1201 may be a combination of the aforementioned memories.

[0173] The processor 1202 may include one or more central processing units (CPUs) or digital processing units, etc. The processor 1202 is configured to implement the above-mentioned chemical process parameter prediction method when calling the computer program stored in the memory 1201 .

[0174] The communication module 1203 is used to communicate with terminal devices or other servers.

[0175] The specific connection medium between the memory 1201, the communication module 1203 and the processor 1202 is not limited in the embodiment of the present application. Figure 12 In the embodiment, the memory 1201 and the processor 1202 are connected via a bus 1204. The bus 1204 is connected to the processor 1202 via a bus 1204. Figure 12 The connections between the other components are shown in bold lines for illustration only and are not intended to be limiting. The bus 1204 can be divided into an address bus, a data bus, a control bus, etc. For ease of description, Figure 12 The diagram shows a single thick line, but this does not indicate that there is only one bus or one type of bus.

[0176] The memory 1201 stores computer storage media, which stores computer executable instructions. The computer executable instructions are used to implement the chemical process parameter prediction method of the embodiment of the present application. The processor 1202 is used to execute the chemical process parameter prediction method of the above-mentioned embodiments.

[0177] Based on the same inventive concept, an embodiment of the present application also provides a storage medium, which stores a computer program. When the computer program runs on a computer, the computer executes the steps of the chemical process parameter prediction method according to various exemplary embodiments of the present application described above in this specification.

[0178] In some possible embodiments, various aspects of the chemical process parameter prediction method provided in the present application can also be implemented in the form of a computer program product, which includes a computer program. When the program product is run on a computer device, the computer program is used to enable the computer device to execute the steps of the chemical process parameter prediction method according to various exemplary embodiments of the present application described above in this specification. For example, the computer device can execute the steps of each embodiment.

[0179] The program product may employ any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or component, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0180] The program product of the embodiment of the present application may be a portable compact disc read-only memory (CD-ROM) and include a computer program, and can be run on a computer device. However, the program product of the present application is not limited thereto. In the present application, a readable storage medium may be any tangible medium containing or storing a program, and the computer program included therein may be used by or in conjunction with a command execution system, apparatus, or device.

[0181] A readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, which carries a readable computer program. Such a propagated data signal may take a variety of 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 that can transmit, propagate, or transfer a program for use by or in conjunction with a command execution system, apparatus, or device.

[0182] The computer program embodied on the readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, RF, etc., or any suitable combination of the foregoing.

[0183] Computer programs for performing the operations of the present application may be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java, C++, etc., and conventional procedural programming languages ​​such as "C" or similar programming languages.

[0184] It should be noted that although several units or subunits of the device are mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, depending on the embodiment of the application, the features and functions of two or more units described above can be embodied in a single unit. Conversely, the features and functions of a single unit described above can be further divided and embodied by multiple units.

[0185] Furthermore, although the operations of the method of the present application are described in a particular order in the accompanying drawings, this does not require or imply that the operations must be performed in this particular order, or that all illustrated operations must be performed to achieve the desired results. Additionally or alternatively, some steps may be omitted, multiple steps may be combined into one step, and / or one step may be decomposed into multiple steps.

[0186] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0187] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present application.

[0188] Obviously, those skilled in the art may make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is intended to include these modifications and variations.

Claims

1. A chemical process parameter prediction method, characterized in that: The method comprises: Obtaining a process parameter set of a target device and inputting the process parameter set into a trained dynamic fusion model; the process parameter set includes a plurality of process parameters generated by the target device in a target chemical process; Obtaining a predicted parameter value output by the dynamic fusion model, wherein the predicted parameter value represents a process parameter that may be generated by the target device at a future moment; Among them, the dynamic fusion model includes a historical gated recurrent unit (HGRU) encoder and a dynamic decoder, and the HGRU encoder is trained based on an improved teacher forcing strategy and a historical supervision sequence; the dynamic decoder determines the fusion feature based on the difference between the output value of the HGRU encoder and the historical supervision sequence, and performs model training based on the fusion feature.

2. The method according to claim 1, wherein The dynamics fusion model is trained in the following way: Obtaining a training time series corresponding to the target chemical process and determining a corresponding historical supervision sequence; the training time series represents process parameters generated by the target equipment within a specific time range, and the historical supervision sequence includes multiple process parameters of the target equipment at a priori time; Iteratively training the HGRU encoder based on a preset improved teacher forcing strategy, in combination with the historical supervision sequence and a first loss function; The improved teacher forcing strategy characterization uses the process parameters of the target equipment in the previous time step as input of the current time step; Determining a fusion feature based on a difference between a predicted output of the HGRU encoder and the historical supervision sequence; The fusion feature represents the dynamic deviation between the current encoding state and the historical true value; Based on the fusion features, the parameters of the dynamics decoder are optimized to obtain a trained dynamics fusion model.

3. The method according to claim 2, wherein The method of iteratively training the HGRU encoder based on the preset improved teacher forcing strategy, in combination with the historical supervision sequence and the first loss function, comprises: Initializing the historical supervision sequence based on a supervision variable estimation strategy; the supervision variable estimation strategy represents initializing each historical supervision variable to the output value of the previous variable; Performing forward transfer processing on the initialized historical supervision sequence to obtain a hidden state corresponding to the HGRU encoder, wherein the hidden state represents the historical dependency relationship of the process parameters in the target chemical process; Performing a linear change on the hidden state to obtain a predicted output of the HGRU encoder; Based on the difference between the predicted output and the historical supervision sequence, combined with the first loss function, the parameters of the HGRU encoder are optimized.

4. The method according to claim 2, wherein Determining the fusion feature based on the difference between the predicted output of the HGRU encoder and the historical supervision sequence includes: Determining a predicted output of the HGRU encoder based on a preset forward calculation strategy; Based on the first fuser, the fusion feature is determined in combination with the vector difference between the predicted output and the target supervision variable; the target supervision variable is the historical supervision variable corresponding to the last moment in the historical supervision sequence.

5. The method according to claim 2, wherein Optimizing parameters of the dynamics decoder based on the fusion features to obtain a trained dynamics fusion model includes: Based on multiple fully connected layers, nonlinear mapping is performed on the fusion features respectively, and weighted merging of the mapping results is performed based on preset weights to obtain a merging result; Based on a preset solver, combined with the merged results, a dynamic prediction result is obtained; The parameters of the dynamic decoder are optimized based on the dynamic prediction result and the second loss function; the second loss function represents the loss value calculated at the final time point.

6. The method according to claim 5, wherein The method of obtaining a dynamic prediction result based on a preset solver and combining the combined result includes: Determining a difference between a state change predicted by the preset solver and a corresponding supervisory variable based on the state change predicted by the preset solver for the fusion feature; The differences are fused based on a second fuser to obtain the dynamics prediction result.

7. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

8. A computer storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer program product comprising computer program instructions, characterized in that When the computer program instructions are executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images