Dynamic main steam pressure prediction method based on structured state space model

Abstract

The invention discloses a dynamic main steam pressure prediction method based on a structured state space model. Historical data is divided into a plurality of modals through modal division; extracting latent variables containing dynamic information in different modals through a structured state space module; the distribution between every two domains is aligned through a domain generalization method, and the distribution difference between different domains is reduced by constructing domain generalization loss; constructing a regression module composed of full connection layers, and inputting aligned latent variables to obtain a predicted value of the dynamic main steam pressure; and constructing a joint loss function training model to realize advanced prediction of the dynamic main steam pressure. According to the method, a structured state space module and a domain generalization method are combined, the mechanism and dynamic characteristics of the system are captured, meanwhile, the difference generated by dynamic changes of data distribution is reduced, common knowledge of different periods is learned by matching distribution of the different periods, and accurate and advanced prediction of the dynamic main steam pressure is achieved.

Description

Technical Field

[0001] The present invention relates to the field of main steam pressure prediction of thermal power units, and particularly to a dynamic main steam pressure prediction method based on a structured state space model. Background Art

[0002] During the power generation process of thermal power units, the main steam pressure is an important parameter in the production process and also a key parameter to ensure the stable and safe operation of the boiler. Therefore, one of the main objectives of unit control is to maintain the main steam pressure within a specified range.

[0003] The main steam pressure control of thermal power units is usually achieved by adjusting the fuel quantity and air supply quantity. However, this process has the characteristics of large lag and large inertia, and it is necessary to make an advance prediction for control decisions. In addition, due to the main steam pressure being affected by internal and external disturbances and having strong dynamic characteristics, it adds difficulty to control decisions. Traditional deep learning models such as recurrent neural networks and long short-term memory networks have problems of poor interpretability and poor generalization, and it is difficult to meet the actual application requirements.

[0004] In addition, to accommodate new energy, thermal power units need to frequently adjust the peak load, which makes the units often in the state of load change, resulting in significant time-varying characteristics of the statistical characteristics of process data, such as mean, variance, etc. The traditional structured state space model lacks the ability to cope with the change of data distribution. During the dynamic process of variable load, due to the change of data distribution, model drift phenomenon will occur, and it is difficult to guarantee the model accuracy. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a dynamic main steam pressure prediction method based on a structured state space model to achieve accurate and advanced prediction of the main steam pressure during multi-condition and variable-condition processes.

[0006] To solve the above technical problems, the technical solution adopted by the present invention is as follows:

[0007] A dynamic main steam pressure prediction method based on a structured state space model, comprising the following steps:

[0008] Step S1: Obtain the historical data of the thermal power plant unit, including the true value of the main steam pressure at the corresponding moment;

[0009] Step S2: Perform modal division on the historical data obtained in step S1 to obtain multi-source domain data, and the multi-source domain data is multi-modal data that retains the time series characteristics but has distribution differences;

[0010] Step S3: Input the multi-source domain data obtained in step S2 into the structured state space module, and respectively extract the latent variables containing dynamic information in different modes;

[0011] Step S4: Using the domain generalization method based on difference measurement, pairwise alignment is performed on the latent variables obtained from step S3 to obtain domain-invariant features with similar distributions, and at the same time, the domain generalization loss function is calculated;

[0012] Step S5: Input the domain-invariant features with similar distributions obtained from step S4 into the regression module to obtain the predicted values of the main steam pressure for multiple source domains;

[0013] Step S6: Construct a prediction loss function using the true value of the main steam pressure obtained from step S1 and the predicted value of the main steam pressure obtained from step S5, and jointly construct a combined loss function with the domain generalization loss obtained from step S4. A trained model is obtained by minimizing the combined loss function;

[0014] Step S7: Collect the unit load, main steam flow rate, regulating stage pressure, main steam temperature on the machine side, total fuel quantity, boiler heat, boiler drum pressure, boiler drum liquid level, main feed water flow rate, AGC command, load setting, load command, main steam pressure setting, boiler combustion command, and throttle opening command of the thermal power plant unit at the current moment as input variables, and input the input variables into the model trained in step S6 to obtain the predicted value of the main steam pressure.

[0015] As a preferred embodiment, in step S1, the historical data includes unit load, main steam flow rate, regulating stage pressure, main steam temperature on the machine side, total fuel quantity, boiler heat, boiler drum pressure, boiler drum liquid level, main feed water flow rate, AGC command, load setting, load command, main steam pressure setting, boiler combustion command, throttle opening command, and the true value of the main steam pressure.

[0016] As a preferred embodiment, in step S2, a multi-source domain data is obtained by using a modal partitioning method based on a sliding window. The multi-source domain data includes historical thermal power unit data of K working conditions, and the multi-source domain data is expressed as:

[0017]

[0018] where S i :X p is the past data X in the i-th mode, with continuous labels and are the n-th sample and its corresponding label in the i-th mode respectively, and m is the number of sample features.

[0019] As a preferred embodiment, in step S3, the structured state space module includes a fully connected layer, a normalization layer, a state space model, a residual link, and a convolutional layer;

[0020] First, the multi-source domain data is input into the fully connected layer, and the extracted features are input into the normalization layer to obtain the normalized features. Then, the normalized features are input into the state space model, and the state space equation is solved based on the data-driven method;

[0021] For continuous time, the state space equation maps the input x to y through the hidden state h and is written as:

[0022] h = Ah + Bx, y = Ch + Dx;

[0023] Based on the discrete characteristics of historical data, it is written as a linear recursive state space equation through discretization, and the formula is expressed as follows:

[0024]

[0025] x t represents the input at time t, h t-1 and h t respectively represent the states at the corresponding times, y t represents the output at time t, and represent the parameter matrices after discretization, and the formulas are expressed as:

[0026]

[0027] The state space equation is derived into the form of a convolutional neural network and written as:

[0028]

[0029] The historical data x is input into the structured state space model to obtain features with dynamic characteristics, where the output latent variable H of the state space model p is expressed as:

[0030]

[0031] In the formula, CNN represents the convolutional layer, LN represents the normalization layer, and LL represents the linear layer of the neural network.

[0032] As a preferred implementation, in step S4, the domain generalization method is used to align the distributions of different modalities after modality partitioning and extract domain-invariant features; the maximum mean discrepancy MMD is used to measure the distance between different distributions, and the formula is expressed as:

[0033]

[0034] p and q are the distributions corresponding to domains i and j respectively, n and m are the numbers of samples in domains i and j respectively. To solve the distance between distributions, the latent variable is mapped into the reproducing kernel Hilbert space through the mapping function φ(·), Hi and H j are the latent variables corresponding to domains i and j respectively, H represents the reproducing kernel Hilbert space, and the obtained MMD is a domain generalization loss function for measuring the distance between distributions.

[0035] As a preferred embodiment, MMD is extended to multiple source domains, and the final domain generalization loss function is the sum of the MMD losses between domains, and the formula is as follows: The sum, the formula is as follows:

[0036] L MMD = k∑ 1≤i,j≤K MMD(H i , H j );

[0037] By minimizing the MMD loss function, the data distributions of the latent variables H i and H j in the common latent variable space are made similar, and the domain-invariant feature H is obtained.

[0038] As a preferred embodiment, the step S5 includes constructing a regression module, and the regression module consists of a single fully-connected layer; the domain-invariant feature H is input into the fully-connected layer to obtain the predicted main steam pressure value:

[0039]

[0040] represents the predicted main steam pressure value of the source domain, W and B represent the weight and bias of the regression module respectively, and the weight and bias are trained and updated by the backpropagation algorithm.

[0041] As a preferred embodiment, the step S6 includes: constructing a joint loss function; the joint loss function consists of a domain generalization loss function and a prediction loss, and the joint loss function L is expressed as:

[0042] L = L MMD + L MSE ;

[0043] L MMD represents the domain generalization loss function; L MSE is the prediction loss, and the loss is constructed from the labels and predicted values corresponding to the source domain data, and the formula is expressed as:

[0044]

[0045] According to minimizing the joint loss function, a trained model is obtained.

[0046] The present invention proposes a dynamic main steam pressure prediction method based on a structured state space model. Historical data is divided into multiple modes through mode partitioning; latent variables containing dynamic information in different modes are extracted through a structured state space module; the distributions between two domains are aligned through a domain generalization method, and by constructing a domain generalization loss, the distribution differences between different domains are reduced; a regression module composed of fully connected layers is constructed, and the aligned latent variables are input to obtain the predicted value of the main steam pressure; a joint loss function is constructed to train the model to achieve the advanced prediction of the dynamic main steam pressure.

[0047] Compared with the prior art, the present invention combines a structured state space module and a domain generalization method, reduces the differences caused by the dynamic changes in the data distribution while capturing the mechanism and dynamic characteristics of the system, and learns the common knowledge in different periods by matching the distributions in different periods, realizing the advanced prediction of the dynamic main steam pressure. Brief Description of the Drawings

[0048] Figure 1 It is the overall framework diagram of the prediction method described in the present invention;

[0049] Figure 2 It is the framework diagram of the structured state space model of the present invention.

[0050] Figure 3 It is the domain generalization framework diagram of the present invention. Detailed Embodiments

[0051] A typical embodiment of the present invention provides a dynamic prediction method based on a structured state space model for the main steam pressure prediction of a thermal power unit.

[0052] As Figure 1 shown, the basic idea of the above prediction method is: obtaining historical data and using it as the training set of the model; performing mode partitioning on the training set data to obtain multi-modal data that retains the time series characteristics but has distribution differences; inputting the multi-modal data as multi-source domain data into the structured state space model to capture the latent variables containing dynamic information in different modes; using a difference measurement method to align the two source domains respectively to obtain domain-invariant features; inputting the domain-invariant features after aligning the distributions into the regression module to obtain the predicted value of the main steam pressure; constructing a joint loss function, training the network by minimizing the loss function to obtain a trained model; collecting the data at the current moment and inputting it into the trained model to calculate the advanced predicted value of the main steam pressure.

[0053] As a grey-box model, the structured state space model can deeply describe the internal relationship between the system input variables and output variables. In addition, the structured state space model also has the advantage of capturing the dynamic characteristics of long time series similar to that of recurrent neural networks. Therefore, the present invention uses the structured state space model as the infrastructure to capture the long-term dependence of data and combine the operating mechanism of the system to construct a main steam pressure prediction model.

[0054] In this embodiment, by using the structured state space model, features containing dynamic information are extracted, and domain generalization is carried out by means of difference measurement to achieve the advanced prediction of the dynamic main steam pressure.

[0055] In order to enable those skilled in the art to better understand the present invention, the following further clearly and completely describes the present invention in conjunction with the accompanying drawings and relatively specific embodiments. It should be noted that, without conflict, the embodiments and features in the embodiments in the present application can be combined with each other.

[0056] S1. Obtain the historical data of the thermal power plant unit.

[0057] In this embodiment, the historical data collected during the actual operation of the thermal power unit is used to train the model. The input variables are 15-dimensional operation parameters obtained by real-time monitoring of the unit, specifically including: unit load, main steam flow rate, regulating stage pressure, main steam temperature (on the turbine side), total fuel quantity, boiler heat, boiler steam drum pressure, boiler steam drum level, main feed water flow rate, AGC command, load setting, load command, main steam pressure setting, boiler combustion command, throttle opening command; the output variable is the true value of the main steam pressure collected synchronously to train the model.

[0058] S2. Perform modal partitioning on the historical data obtained in S1 to obtain multi-source domain data.

[0059] The multi-source domain data is obtained by using a modal partitioning method based on a sliding window. The multi-source domain data contains historical thermal power unit data of multiple working conditions, a total of K, and the value of K is determined according to the modal partitioning result. These data are multi-modal data that retain the time series characteristics but have distribution differences, and are used as multi-source domain data to train the model.

[0060] Set the window size to W and the sliding step size to H. Use the maximum mean difference as the difference measurement method for modal partitioning. Calculate the maximum mean difference value between the data of adjacent two windows. If the calculated maximum mean difference value is less than the threshold A, it is considered that these two windows belong to the same source domain; if the calculated maximum mean difference value is greater than or equal to the threshold A, it is considered that these two windows belong to different source domains.

[0061] In this embodiment, the window size W is 1000, the sliding step H is 500, the threshold A is 3, and the number of multi-source domains K obtained after modal division is 3. The multi-source domain data is expressed as:

[0062]

[0063] where S i :X p is the past data X in the i-th modality, with continuous labels and are the n-th sample and its corresponding label in the i-th modality respectively, and m is the number of sample features.

[0064] S3, as Figure 2 shown, input the source domain data of multiple modalities obtained from S2 into the structured state space module, and extract the latent variables containing dynamic information corresponding to different modalities respectively;

[0065] The structured state space module consists of a fully connected layer, a normalization layer, a state space model, a residual connection, and a convolutional layer.

[0066] First, input the multiple source domain data into the fully connected layer; in order to improve the generalization of the model, input the extracted features into the normalization layer to obtain the normalized features; then input the normalized features into the state space model, and solve the state space equation based on the data-driven method. For continuous time, the state space equation maps the input x to y through the hidden state h, written as:

[0067] h = Ah + Bx, y = Ch + Dx;

[0068] Based on the discrete characteristics of historical data, the structured state space model is written as a linear recursive state space equation through discretization, and the formula is expressed as follows:

[0069]

[0070] x t represents the input at time t, h t-1 and h t represent the states at the corresponding times respectively, y t represents the output at time t, and represent the discretized parameter matrices, and the formulas are expressed as:

[0071]

[0072] Furthermore, the state space equation is derived into the form of a convolutional neural network, written as:

[0073]

[0074] Inputting historical data \(x\) into a structured state - space model can obtain features with dynamic characteristics. Among them, the output latent variable \(H\) of the state - space model p , can be expressed as:

[0075]

[0076] In the formula, CNN represents the convolutional layer, LN represents the normalization layer, and LL represents the linear layer of the neural network.

[0077] S4, as Figure 3 shown, for the latent variables obtained from S3, use the domain generalization method to match the distributions of pairwise domains in the latent variable space, obtain domain - invariant features with similar distributions, and at the same time calculate the domain generalization loss function.

[0078] The main steam pressure of a thermal power unit is measured in real - time. It is necessary to establish a model to predict the main steam pressure under multiple working conditions at future moments in advance. Traditional unsupervised domain adaptation aligns unlabeled target - domain data with labeled source - domain data, and adjusts the model trained on source - domain data by learning the features of the target domain during the training process, so that the model can adapt to new working conditions and solve the problem of model mismatch caused by the source domain and the target domain not being independently and identically distributed. However, traditional domain adaptation is for a constant target domain. As data continuously arrives, domain adaptation needs to be repeated for multiple target domains generated by sequential data. Therefore, applying the traditional unsupervised domain adaptation method to predict the main steam pressure has low efficiency.

[0079] The present invention adopts a domain generalization method based on difference measurement, maps the data into a high - dimensional space, calculates the distribution distances between pairwise modalities respectively, obtains the domain generalization loss function, and learns the common knowledge under different working conditions by approximating the distributions of different modalities.

[0080] The present invention is different from traditional unsupervised domain adaptation. It uses the domain generalization method to align the distributions of different modalities after modality partitioning, extracts domain - invariant features, and improves the generalization of the main steam pressure prediction model under multiple working conditions.

[0081] The basic assumption of domain generalization in the present invention is that there is a feature space for multiple known modalities with different distributions. On this feature space, a model can be trained by learning the features of different modalities with known distributions, and has good generalization for predicting data with any distribution. The data distribution in the feature space should be domain - invariant, thereby improving the generalization ability of the model in downstream tasks.

[0082] Domain generalization can be modeled through the joint distribution \(P(X,Y)\) and can be decomposed into:

[0083] P(X,Y) = P(Y|X)P(X) = P(X|Y)P(Y);

[0084] Assume that the distribution shift occurs in the marginal distribution P(X), i.e., P i (X) ≠ P j (X), and the posterior probability P i (Y|X) = P j (Y|X). Based on the domain generalization method using the difference metric, the generalization error is reduced by reducing the difference between two domains. Therefore, based on this assumption, through causal inference, when X is the cause of Y, domain generalization is achieved by aligning the marginal distributions, so as to obtain a domain-invariant distribution in the feature space.

[0085] The domain generalization method based on the difference metric reduces the generalization error by reducing the difference between two domains. This type of method adapts at the feature level, projects data from different domains into a common latent variable space, calculates the distance between distributions as the loss, trains to obtain a feature extractor, and aligns the marginal distributions of the two domains by minimizing the loss function, transforming the problem into an optimization problem, as shown in the following formula:

[0086]

[0087] D i and D j represent domain i and j respectively, dis is the distance between the distributions of domain i and j, p and q are the corresponding distributions of domain i and j respectively, and the purpose of domain generalization is achieved by minimizing dis.

[0088] The domain generalization methods based on the difference metric include but are not limited to KL divergence, maximum mean discrepancy, correlation alignment distance, central distance difference, etc. The domain generalization method used in the present invention adopts the maximum mean discrepancy MMD, maps the data of two domains into a reproducing kernel Hilbert space using a mapping function, calculates the supremum of the expected difference of the data in this space as the distribution distance, thereby measuring the distance between different distributions. The basic definition formula of MMD is:

[0089]

[0090] where f is the Gaussian kernel function that maps the data x and y of different domains, calculates the expectations respectively after mapping, and then calculates the supremum of the expected interpolation. The regeneration property of the reproducing Hilbert space and the properties of the inner product are used for transformation to obtain the final result.

[0091] Since μ p cannot be obtained directly, the mean value is used for calculation instead, and the MMD formula for domain i and j is expressed as:

[0092]

[0093] To solve the distance between distributions, the latent variables are mapped into the reproducing kernel Hilbert space through the mapping function φ(·). p and q are the distributions corresponding to domains i and j respectively, n and m are the number of samples in domains i and j, and H i and H j are the output latent variables of the state space models corresponding to domains i and j respectively. H p represents the reproducing kernel Hilbert space. The obtained MMD loss between domains i and j is the domain generalization loss function for measuring the distance between distributions.

[0094] Generalize MMD to multiple source domains. The final MMD loss function is the sum of the MMD losses between domains , and the formula is as follows:

[0095] L MMD = k∑ 1≤i,j≤K MMD(H i , H j );

[0096] By minimizing the MMD loss function, the purpose of domain generalization is achieved, making the data distributions of the latent variables H i and H j in the common latent variable space similar, and obtaining domain-invariant features H.

[0097] S5. Input the aligned features obtained in S4 into the regression module to obtain the predicted main steam pressure values of multiple source domains, and calculate the prediction loss.

[0098] This step includes constructing a regression module, which consists of a single fully connected layer. Input the domain-invariant features H after domain alignment into the fully connected layer to obtain the predicted main steam pressure value:

[0099]

[0100] represents the predicted main steam pressure value, W and B represent the weights and biases of the regression module respectively, and the weights and biases are trained and updated by the backpropagation algorithm.

[0101] S6. Construct a prediction loss function using the true main steam pressure values obtained in S1 and the predicted main steam pressure values obtained in S5, and construct a joint loss function together with the domain generalization loss obtained in S4. Obtain the trained model by minimizing the joint loss function.

[0102] This step includes constructing a joint loss function, which consists of a domain generalization loss function and a prediction loss. The joint function L is expressed as:

[0103] L = L MMD + L MSE ;

[0104] L MMD represents the domain generalization loss function; L MSE is the prediction loss, which constructs the loss from the true label and the predicted value of the main steam pressure corresponding to the source domain data. The formula is expressed as:

[0105]

[0106] By minimizing the joint loss function, a trained model is obtained.

[0107] S7 collects the unit load, main steam flow rate, regulating stage pressure, main steam temperature (on the turbine side), total fuel quantity, boiler heat, boiler drum pressure, boiler drum level, main feed water flow rate, AGC command, load setpoint, load command, main steam pressure setpoint, boiler combustion command, and throttle valve opening command at the current moment, and uses them as input variables. The input variables can be represented as T:X f , T:X f is the data to be predicted, which is the future data X under an unknown distribution. The input variables are input into the model trained in S6 to obtain the predicted value of the main steam pressure

[0108] The scope of protection required by the present invention is not limited to the above specific embodiments. For those skilled in the art, the present invention can have various deformations and modifications. Any modification, improvement, and equivalent replacement made within the concept and principle of the present invention should be included within the scope of protection of the present invention.

Claims

1. A dynamic main steam pressure prediction method based on a structured state space model, characterized in that, It includes the following steps: Step S1: Obtain the historical data of the thermal power plant unit, including the true value of the main steam pressure; Step S2: Perform modal partitioning on the historical data obtained in Step S1 to obtain multi-source domain data, where the multi-source domain data is multi-modal data that retains time series characteristics but has distribution differences; Step S3: Input the multi-source domain data obtained in Step S2 into the structured state space module, and extract the latent variables containing dynamic information in different modes respectively; Step S4: Adopt a domain generalization method based on difference measurement to align the latent variables obtained in Step S3 pairwise to obtain domain-invariant features with similar distributions, and calculate the domain generalization loss function at the same time; Step S5: Input the domain-invariant features with similar distributions obtained in Step S4 into the regression module to obtain the predicted values of the main steam pressure for multiple source domains; Step S6: Construct a prediction loss function using the true value of the main steam pressure obtained in Step S1 and the predicted value of the main steam pressure obtained in Step S5, and construct a joint loss function together with the domain generalization loss obtained in Step S4. Obtain a trained model by minimizing the joint loss function; Step S7: Collect the unit load, main steam flow, regulating stage pressure, main steam temperature on the machine side, total fuel quantity, boiler heat, boiler steam drum pressure, boiler steam drum liquid level, main feed water flow, AGC command, load setting, load command, main steam pressure setting, boiler combustion command, and throttle opening command of the thermal power plant unit at the current moment as input variables, and input the input variables into the model trained in Step S6 to obtain the predicted value of the main steam pressure.

2. The dynamic main steam pressure prediction method based on a structured state space model according to claim 1, wherein, In Step S1, the historical data includes unit load, main steam flow, regulating stage pressure, main steam temperature on the machine side, total fuel quantity, boiler heat, boiler steam drum pressure, boiler steam drum liquid level, main feed water flow, AGC command, load setting, load command, main steam pressure setting, boiler combustion command, throttle opening command, and true value of the main steam pressure.

3. The dynamic main steam pressure prediction method based on the structured state space model according to claim 1, characterized in that In Step S2, a modal partitioning method based on a sliding window is used to obtain multi-source domain data. The multi-source domain data includes historical thermal power unit data of K working conditions, and the multi-source domain data is expressed as: Among them, S i : X p is the past data X in the i-th modality, with consecutive labels and are the n-th sample and its corresponding label in the i-th modality respectively, and m is the number of sample features.

4. The dynamic main steam pressure prediction method based on the structured state space model according to claim 1, characterized in that In Step S3, the structured state space module includes a fully connected layer, a normalization layer, a state space model, a residual connection, and a convolutional layer; First, input the multi-source domain data into the fully connected layer, and input the extracted features into the normalization layer to obtain the normalized features. Then, input the normalized features into the state space model and solve the state space equation based on the data-driven method; For continuous time, the state space equation maps the input x to y through the hidden state h, written as: h = Ah + Bx, y = Ch + Dx; Based on the discrete characteristics of the historical data, it is written as a linear recursive state space equation through discretization. The formula is expressed as follows: x t represents the input at time t, h t-1 and h t represent the states at the corresponding times respectively, y t represents the output at time t, and represent the parameter matrices after discretization, and are expressed by the formulas as follows: Derive the state space equation into the form of a convolutional neural network, written as: Input historical data x into the structured state space model to obtain features with dynamic characteristics, where the output latent variable H of the state space model p is expressed as: In the formula, CNN represents the convolutional layer, LN represents the normalization layer, and LL represents the linear layer of the neural network.

5. The dynamic main steam pressure prediction method based on the structured state space model according to claim 1, wherein, In Step S4, use the domain generalization method to align the distributions of different modes after modal partitioning and extract domain-invariant features; The maximum mean discrepancy (MMD) is used to measure the distance between different distributions, and the formula is expressed as: Let \(p\) and \(q\) be the distributions corresponding to domains \(i\) and \(j\) respectively, and \(n\) and \(m\) be the number of samples in domains \(i\) and \(j\) respectively. To solve the distance between distributions, the latent variables are mapped into the reproducing kernel Hilbert space through the mapping function \(\varphi(\cdot)\), \(H\) i and \(H\) j are the latent variables corresponding to domains \(i\) and \(j\) respectively, \(H\) represents the reproducing kernel Hilbert space, and the obtained MMD is the domain generalization loss function for measuring the distance between distributions.

6. The dynamic main steam pressure prediction method based on the structured state space model according to claim 5, wherein, Generalize MMD to multiple source domains, and the final domain generalization loss function is the sum of the MMD losses between domains, as shown in the following formula: The formula is as follows: L MMD = k∑ 1≤i，j≤K MMD(H i , H j ); By minimizing the MMD loss function, the latent variables H i and H j have similar data distributions in the common latent variable space, and the domain-invariant feature H is obtained.

7. The dynamic main steam pressure prediction method based on the structured state space model according to claim 7, wherein The step S5 includes constructing a regression module, which consists of a single fully connected layer; inputting the domain-invariant feature H into the fully connected layer to obtain the predicted main steam pressure value: It represents the main steam pressure value predicted by the source domain. W and B represent the weight and bias of the regression module respectively, and the weight and bias are trained and updated by the backpropagation algorithm.

8. The dynamic main steam pressure prediction method based on a structured state space model according to claim 1, wherein, The step S6 includes: constructing a joint loss function; the joint loss function is composed of a domain generalization loss function and a prediction loss, and the joint loss function L is expressed as: L = L MMD + L MSE ; L MMD represents the domain generalization loss function; L MSE is the prediction loss, which constructs the loss from the labels and predicted values corresponding to the source domain data, and is expressed by the formula: According to minimizing the joint loss function, a trained model is obtained.

Images