Prediction method and system for starting transient temperature field of photo-thermal steam turbine rotor

Abstract

The invention discloses a method and a system for predicting a starting transient temperature field of a photo-thermal steam turbine rotor. The method comprises the following steps of: obtaining material setting, contact construction and boundary conditions of the photo-thermal steam turbine rotor to be predicted, and obtaining steady-state temperature field data of the rotor through finite elements; preprocessing the acquired steady-state temperature field data; exporting a steady-state temperature field of the result, converting the steady-state temperature field into steady-state temperature values of all nodes sorted according to node numbers and adding and generating the steady-state temperature values to obtain a steady-state temperature field sequence; and inputting the preprocessed steady-state temperature field sequence data into a preset transient temperature field prediction model to obtain a to-be-predicted starting transient temperature field of the photo-thermal steam turbine rotor. The system comprises a data acquisition module, a data preprocessing module and a prediction module. According to the invention, rapid and accurate prediction of the temperature field is realized, and a rapid decision basis is provided for operation optimization, life calculation and reliability maintenance of the unit.

Description

Technical Field

[0001] The present invention belongs to the technical field of steam turbines, and in particular relates to a method and system for predicting the transient temperature field of a solar thermal steam turbine rotor when it is started. Background Art

[0002] The turbine rotor is the core component of the turbine. It can convert the thermal energy of steam into the required mechanical energy and plays an important role in maintaining the stable operation of the solar thermal power generation system. During the high-speed rotation of the rotor, it is often subject to the combined effects of the centrifugal force of the impeller, the steam flow force and the thermal stress caused by the internal temperature difference. During the startup of the unit, the thermal shock caused by the rapid change of steam parameters makes the surface temperature field distribution extremely uneven, resulting in inconsistent expansion inside and outside the rotor, which in turn causes serious thermal stress. At the same time, the centrifugal force changes caused by the speed change, the internal non-uniform heat conduction and the temperature dependence of the material cause the temperature and stress field of the rotor to undergo significant changes during the startup of the turbine. Ultimately, it aggravates material fatigue and loss, affects structural integrity, and causes damage or failure of the equipment. Therefore, accurate prediction of the transient temperature field of the turbine rotor during different startup conditions is helpful to improve operational reliability and extend the service life of the turbine, and provide a basis for the optimization of rotor structure and materials.

[0003] At present, the analysis of the transient temperature field of thermal-solid coupling during the startup of steam turbine rotors is mainly divided into two methods: finite element numerical simulation and data-driven. Finite element simulation can simulate complex geometric structures and set multiple boundaries. However, as the unit model structure becomes more complex, the boundary conditions become more and more numerous, and the coupling nonlinear effect becomes more and more obvious, finite element simulation faces problems such as long numerical solution time, poor model convergence, and high resource consumption, which is not conducive to the rapid optimization of the structure and the accurate decision-making of the unit operation. The data-driven method mainly reduces the computational cost and time by constructing a simplified proxy model. However, traditional models such as polynomial response surface models, kriging models, and Gaussian processes often suffer from "dimensionality disaster" in complex situations. Although deep learning models represented by neural networks have powerful nonlinear fitting capabilities and can learn complex mappings of high-dimensional input and output data, they require a large amount of data as a prerequisite for model training, which is difficult to achieve in real industrial production.

[0004] In this case, constructing a transient temperature field result without a complex rotor model to achieve accurate transient temperature field prediction model training is of great practical significance for the reliability maintenance of solar thermal turbine rotors.

[0005] The basic meaning of meta-learning is to allow the network model to learn a priori knowledge. Through the general knowledge learned in different tasks or environments, the model has the ability to analyze and solve problems, so that it can quickly adapt to unknown new tasks. Meta-learning consists of an upper-level meta-learner and an inner-level base learner, which is essentially a hierarchical optimization problem. From the perspective of mapping construction, the mapping construction of meta-learning can be considered as an outer-layer mapping (often an inter-domain mapping) and an inner-layer mapping (often an inter-domain mapping). Patent CN202410593048.1 proposes a small-sample rotating machinery fault diagnosis method based on meta-learning and cyclic reconstruction, which realizes signal fault classification and sample screening and reconstruction through a meta-learning framework. Patent CN202410724493.7 proposes a small-sample machine tool processing quality prediction method based on meta-learning, which directly obtains the mapping between the workpiece roughness and the data to be measured by constructing a model. Meta-learning has achieved certain results in the field of industrial reliability analysis. However, due to the higher complexity of turbine rotors in operating conditions and structural connections, there are currently few studies based on the meta-learning framework in the field of turbine rotors. However, obtaining accurate results using a small number of samples is of great significance for the reliable operation of turbine rotors.

[0006] Transient temperature field prediction is to use existing parameters or easily obtained results to directly predict transient temperature fields that are difficult to monitor and complex to simulate. Patent CN202011008953.4 constructs a method for predicting the temperature field of the low-pressure final stage of a steam turbine. The historical operating state parameters and temperature field data are trained by the K nearest neighbor algorithm to directly obtain the temperature field. Patent CN202311753891.3 uses the temperature field data at the first moment to construct a ConvLSTM network to further obtain temperature field data at subsequent moments. In numerical calculations, the steady-state temperature field is an important input for transient temperature field calculations. Compared with the state parameter input in between, the steady-state temperature field is more closely related to the transient temperature field. However, accurately obtaining temperature field results is of great significance for the life prediction and optimization of steam turbine rotors. Summary of the invention

[0007] In view of the shortcomings of the existing methods, the purpose of the present invention is to provide a method and system for predicting the transient temperature field of a solar thermal steam turbine rotor during startup. The purpose is to construct a prediction model for the transient temperature field of a solar thermal steam turbine rotor during startup under a small sample condition, to achieve rapid and accurate prediction of the temperature field, and to provide a rapid decision-making basis for unit operation optimization, life calculation, and reliability maintenance.

[0008] In order to achieve the above object, the present invention adopts the following technical solution: A method for predicting transient temperature field of a solar thermal steam turbine rotor during startup, comprising: 1) Obtain the material setting, contact construction and boundary conditions of the solar thermal turbine rotor to be predicted, and obtain the steady-state temperature field data of the rotor through finite element analysis; 2) Preprocess the acquired steady-state temperature field data; export the resulting steady-state temperature field, convert it into adding and generating steady-state temperature values ​​of all nodes sorted by node numbers, and obtain a steady-state temperature field sequence; 3) The preprocessed steady-state temperature field sequence data is input into the preset transient temperature field prediction model to obtain the startup transient temperature field of the solar thermal steam turbine rotor to be predicted; the preset transient temperature field prediction model includes a base learner constructed by a residual network, a domain classifier and a meta-learner constructed by Transforme, which is obtained by training in combination with an adversarial feature alignment strategy.

[0009] A further improvement of the present invention is to obtain the material setting, contact construction and boundary conditions of the solar thermal steam turbine rotor to be predicted, and obtain the steady-state temperature field data of the rotor through finite element analysis, including: By setting materials, building contacts, and adding boundary conditions to the actual rotor model, the corresponding steady-state temperature field is obtained through ANSYS AP DL.

[0010] A further improvement of the present invention is that the acquired steady-state temperature field data is preprocessed, including: The steady-state temperature values ​​of all nodes sorted by node number are converted into a complete sequence , N represents the number of nodes, and the steady-state temperature field sequence is obtained.

[0011] A further improvement of the present invention is that the preset transient temperature field prediction model includes: The blackone of the base learner of the meta-learning network model is constructed by the residual network. In order to complete the training of the network model under the premise of the transient temperature field of the target domain that is not needed, a domain classifier is further constructed to make the features of the source domain and the target domain difficult to distinguish in the model. The blackone of the meta-learner is composed of a Transformer network, which specifically includes the following steps: S211: Input the steady-state temperature field vector of the source domain , through a base learner built from ResNet-34 , whose parameters are , the goal of the base learner is to predict the corresponding transient temperature field ; This process is expressed as:

[0012] Among them, the base learner It is composed of multiple residual blocks, and the output of each residual block is expressed as:

[0013] in, It is l The input features of the layer, It is l The residual function of the layer is expressed as:

[0014] in, is the activation function, stands for batch normalization, represents the convolution kernel, The supervised training process of the bias-based learner is achieved by minimizing the following loss function:

[0015] Where N is the number of source domain samples, It is a loss function used to measure the difference between the predicted value and the true value; the loss function used is the mean absolute error MAE, as follows:

[0016] in, Represents the number of training batches. Represents the true value of a batch of samples, Represents the predicted value of a batch of samples; S212: Domain Classifier It is a simple neural network consisting of two fully connected layers and a ReLU activation function; it accepts the features extracted from the base learner as input and outputs a scalar value representing the probability that the sample belongs to the source domain; specifically

[0017] in, p Represents input features h The probability of belonging to the source domain, are weight matrices, are bias matrices, Represents the Sigmoid activation function, which is expressed as:

[0018] The training process of the domain classifier aims to enable it to distinguish whether the input features are from the source domain or the target domain, so the optimization objective is defined as maximizing the probability that the domain classifier correctly classifies the source domain and target domain samples; defined as the negative value of the cross entropy loss of a binary classification:

[0019] in, It is i The characteristics of source domain samples, It is j The characteristics of the target domain samples, and are the number of samples in the source domain and the target domain respectively; S213: Adversarial feature alignment; the base learner trains the model in an "adversarial" way so that the features it generates are more consistent between the source domain and the target domain; and the successful training of the domain classification enables it to have basic judgment capabilities. Therefore, in order to enable the entire model to achieve the established goals, adversarial loss is further added to the loss of the base learner, so that the features of the source domain and the target domain become more consistent under the judgment of the domain classifier, thereby achieving feature alignment and domain adaptation. The adversarial loss function is as follows:

[0020] The loss function of the final base learner consists of the supervised loss of the source domain and the adversarial loss of feature alignment, namely:

[0021] in, is a hyperparameter used to balance the impact of supervised learning and adversarial learning; Meta-learner is a model constructed by a Transformer network, whose parameters are ; The input of the meta-learner is the steady-state temperature field of the source domain and the pseudo-steady-state temperature field of the target domain , the output is the transient temperature field prediction of the target domain; this process is expressed as:

[0022] Each layer in the Transformer network consists of a multi-head self-attention mechanism and a feed-forward neural network; l Layer, output features It is expressed as:

[0023] in, is a feed-forward neural network, It is a multi-head attention mechanism, which is calculated as:

[0024] Among them, each is the result of self-attention calculation:

[0025] in, By inputting features , the query, key and value matrices obtained by linear transformation, is the dimension of the key vector.

[0026] A further improvement of the present invention is that the goal of meta-learning is to find the parameter , so that it can be quickly adapted to the target domain through fine-tuning; this process is represented by the following optimization objective:

[0027] in, is a set of tasks, each task includes a support set and queryset .

[0028] A prediction system for transient temperature field of a solar thermal steam turbine rotor during startup, comprising: The data acquisition module acquires the material setting, contact construction and boundary conditions of the solar thermal steam turbine rotor to be predicted, and obtains the steady-state temperature field data of the rotor through finite element analysis; The data preprocessing module preprocesses the acquired steady-state temperature field data; the steady-state temperature field of the exported result is converted into steady-state temperature values ​​of all nodes sorted by node numbers, and a steady-state temperature field sequence is obtained; The prediction module inputs the preprocessed steady-state temperature field sequence data into the preset transient temperature field prediction model to obtain the startup transient temperature field of the solar thermal steam turbine rotor to be predicted; wherein the preset transient temperature field prediction model includes a base learner constructed by a residual network, a domain classifier and a meta-learner constructed by Transforme, which is obtained by training in combination with an adversarial feature alignment strategy.

[0029] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a method for predicting a transient temperature field of a solar thermal steam turbine rotor when starting up.

[0030] Compared with the prior art, the present invention has at least the following beneficial technical effects: The present invention provides a method and system for predicting the transient temperature field of a solar thermal steam turbine rotor during startup, which can realize the prediction of the transient temperature field of a steam turbine rotor during startup; a two-layer meta-learning and adversarial feature alignment strategy are adopted to solve the problem of small sample data and difficult construction of the rotor transient temperature field. The two-layer meta-learning effectively eliminates the difference between the source domain and the target domain, and the adversarial feature alignment makes up for the problem of missing transient temperature field in the target domain. It enables the prediction of the rotor transient temperature field to be realized under the premise of consuming less computing resources and insufficient sample data. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 is an overall flow chart of the method of the present invention; Figure 2 Schematic diagram of each network model in the method of the present invention; Figure 3 This is a simplified schematic diagram of the rotor in the present invention; Figure 4 It is a schematic diagram of a real rotor in the present invention; Figure 5 It is a schematic diagram of the steady-state temperature field of a real rotor under a certain working condition in the present invention; Figure 6 It is a schematic diagram of the actual rotor transient temperature field under a certain working condition in the present invention.

[0032] Figure 7 The present invention is a structural block diagram of a prediction system for transient temperature field of a solar thermal steam turbine rotor during startup. DETAILED DESCRIPTION

[0033] In the following, only some exemplary embodiments are briefly described. As those skilled in the art will appreciate, the described embodiments may be modified in various ways without departing from the spirit or scope of the present invention. Therefore, the drawings and descriptions are considered to be exemplary and non-restrictive in nature.

[0034] It should be understood that when used in this specification and the appended claims, the terms "include" and "comprises" indicate the presence of described features, integers, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.

[0035] It should also be understood that the terms used in the present specification are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the present specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0036] It should be further understood that the term "and / or" used in the present description and the appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0037] Various structural schematic diagrams of the embodiments disclosed in the present invention are shown in the accompanying drawings. These figures are not drawn to scale, and some details are magnified and some details may be omitted for the purpose of clear expression. The shapes of various regions and layers shown in the figures and the relative sizes and positional relationships therebetween are only exemplary, and may deviate in practice due to manufacturing tolerances or technical limitations, and those skilled in the art may additionally design regions / layers with different shapes, sizes, and relative positions according to actual needs.

[0038] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0039] Example 1 The present invention provides a method for predicting the transient temperature field of a solar thermal steam turbine rotor during startup, comprising: 1) Obtain the material setting, contact construction and boundary conditions of the solar thermal turbine rotor to be predicted, and obtain the steady-state temperature field data of the rotor through finite element analysis; 2) Preprocess the acquired steady-state temperature field data; export the resulting steady-state temperature field, convert it into adding and generating steady-state temperature values ​​of all nodes sorted by node numbers, and obtain a steady-state temperature field sequence; 3) The preprocessed steady-state temperature field sequence data is input into the preset transient temperature field prediction model to obtain the startup transient temperature field of the solar thermal steam turbine rotor to be predicted; the preset transient temperature field prediction model includes a base learner constructed by a residual network, a domain classifier and a meta-learner constructed by Transforme, which is obtained by training in combination with an adversarial feature alignment strategy.

[0040] In this embodiment, the material setting, contact construction and boundary conditions of the solar thermal steam turbine rotor to be predicted are obtained, and the steady-state temperature field data of the rotor is obtained by finite element analysis, including: By setting materials, building contacts, and adding boundary conditions to the actual rotor model, the corresponding steady-state temperature field is obtained through ANSYS AP DL.

[0041] In this embodiment, the acquired steady-state temperature field data is preprocessed, including: The steady-state temperature values ​​of all nodes sorted by node number are converted into a complete sequence , N represents the number of nodes, and the steady-state temperature field sequence is obtained.

[0042] In this embodiment, the preset transient temperature field prediction model includes: The network model includes base learning, adversarial feature alignment and meta-learner. Its main structure is as follows Figure 2The blackone of the base learner of the meta-learning network model is constructed by the residual network. In order to complete the training of the network model under the premise of the transient temperature field of the target domain that is not needed, a domain classifier is further constructed to make the features of the source domain and the target domain difficult to distinguish in the model. The blackone of the meta-learner is composed of a Transformer network, which specifically includes the following steps: S211: Input the steady-state temperature field vector of the source domain , through a base learner built from ResNet-34 , whose parameters are , the goal of the base learner is to predict the corresponding transient temperature field ; This process is expressed as:

[0043] Among them, the base learner It is composed of multiple residual blocks, and the output of each residual block is expressed as:

[0044] in, It is l The input features of the layer, It is l The residual function of the layer is expressed as:

[0045] in, is the activation function, stands for batch normalization, represents the convolution kernel, The supervised training process of the bias-based learner is achieved by minimizing the following loss function:

[0046] Where N is the number of source domain samples, is a loss function used to measure the difference between the predicted value and the true value; the loss function used is the mean absolute error (MAE), as follows:

[0047] in, Represents the number of training batches. Represents the true value of a batch of samples, Represents the predicted value of a batch of samples; S212: Domain Classifier It is a simple neural network consisting of two fully connected layers and a ReLU activation function; it accepts the features extracted from the base learner as input and outputs a scalar value representing the probability that the sample belongs to the source domain; specifically

[0048] in, p Represents input features h The probability of belonging to the source domain, are weight matrices, are bias matrices, Represents the Sigmoid activation function, which is expressed as:

[0049] The training process of the domain classifier aims to enable it to distinguish whether the input features are from the source domain or the target domain, so the optimization objective is defined as maximizing the probability that the domain classifier correctly classifies the source domain and target domain samples; defined as the negative value of the cross entropy loss of a binary classification:

[0050] in, It is i The characteristics of source domain samples, It is j The characteristics of the target domain samples, and are the number of samples in the source domain and the target domain respectively; S213: Adversarial feature alignment; the base learner trains the model in an "adversarial" way so that the features it generates are more consistent between the source domain and the target domain; and the successful training of the domain classification enables it to have basic judgment capabilities. Therefore, in order to enable the entire model to achieve the established goals, adversarial loss is further added to the loss of the base learner, so that the features of the source domain and the target domain become more consistent under the judgment of the domain classifier, thereby achieving feature alignment and domain adaptation. The adversarial loss function is as follows:

[0051] The loss function of the final base learner consists of the supervised loss of the source domain and the adversarial loss of feature alignment, namely:

[0052] in, is a hyperparameter used to balance the impact of supervised learning and adversarial learning; Meta-learner is a model constructed by a Transformer network, whose parameters are ; The input of the meta-learner is the steady-state temperature field of the source domain and the pseudo-steady-state temperature field of the target domain , the output is the transient temperature field prediction of the target domain; this process is expressed as:

[0053] Each layer in the Transformer network consists of a multi-head self-attention mechanism and a feed-forward neural network; l Layer, output features It is expressed as:

[0054] in, is a feed-forward neural network, It is a multi-head attention mechanism, which is calculated as:

[0055] Among them, each is the result of self-attention calculation:

[0056] in, By inputting features , the query, key and value matrices obtained by linear transformation, is the dimension of the key vector; The goal of meta-learning is to find the parameters , so that it can be quickly adapted to the target domain with a small amount of fine-tuning; this process is represented by the following optimization objective:

[0057] in, is a set of tasks, each task includes a support set and queryset .

[0058] Example 2 like Figure 1 As shown, the present invention provides a method for predicting the transient temperature field of a solar thermal steam turbine rotor during startup, comprising: 1) Dataset construction process: S11: Source domain dataset construction. Use ANSYS APDL to batch build a simplified rotor unit with multi-level slots as follows Figure 3As shown. The simplified rotor removes part of the structural shape, but retains the slot-shaped unit and basic configuration that are the focus of the rotor temperature. By giving centrifugal force, blade load and thermal boundary conditions that are consistent with the actual situation, the data is ensured to be physically realistic. Then the steady-state temperature field and transient temperature field of the simplified rotor unit are calculated. By setting the range of variation of the simplified rotor unit length and the depth of the slot, the length of the simplified rotor unit and the depth of the slot are randomly changed through parametric programming, and the output type is set to the steady-state temperature value of all nodes sorted by node number and the transient temperature value at the end time. The data type is in txt format. The total number of data samples is set to 2000.

[0059] S12: Target domain data set construction. By setting materials, building contacts, and adding boundary conditions to the actual rotor model, the steady-state temperature values ​​of all nodes sorted by node numbers and the transient temperature values ​​at the end time are generated. The actual rotor model is as follows: Figure 4 The obtained steady-state temperature field and transient temperature field are respectively Figure 5 , Figure 6 The saved transient temperature field is used to verify the results.

[0060] S13: Target domain steady-state temperature field data expansion: By using a generative adversarial network, the data samples are expanded with the given single steady-state temperature field results to be the same as the samples of the source domain dataset.

[0061] 2) Model training and testing: S21: The steady-state temperature field and transient temperature field of the source domain are used as the input and label quantity of the inner base learner, and the support set and query set are divided according to the ratio of 8:2. At the same time, the steady-state temperature field of the source domain and the pseudo-steady-state temperature field of the target domain are used as the input and label quantity of the meta-learner.

[0062] S31: Generate a transient temperature field by inputting the real complex turbine rotor steady-state temperature field into the meta-learning model.

[0063] S32: Compare the predicted transient temperature field with the actual transient temperature field.

[0064] Furthermore, the specific method for expanding the target domain steady-state temperature field data in step S13 is as follows: first, a generator is constructed and the discriminator Assume that the source domain generator is used to accept a random noise vector z As input, and output a synthetic sample , the discriminator accepts a sample (real sample or generated sample) as input and outputs a probability , represents the probability that the input sample is a real sample. Assume that the steady-state temperature field dataset of the target domain is ,Generator Accepts a random noise vector , and generate a new sample The generator and discriminator are optimized through adversarial training, and the objective function is:

[0065] in, is the expectation of the discriminator to correctly identify the real samples.

[0066] is the expectation of the discriminator correctly identifying the generated samples as fake.

[0067] Furthermore, the blackone of the base learner of the meta-learning network model constructed in step S21 is mainly constructed by a residual network (ResNet). In order to complete the training of the network model under the premise of the transient temperature field of the target domain that is not required, a domain classifier is further constructed to make the features of the source domain and the target domain difficult to distinguish in the model. The blackone of the meta-learner is mainly composed of a Transformer network. Specifically, the following steps are included: S211: Input the steady-state temperature field vector of the source domain , through a base learner built from ResNet-34 , whose parameters are , the goal of the base learner is to predict the corresponding transient temperature field . This process can be expressed as:

[0068] Among them, the base learner It is composed of multiple residual blocks, and the output of each residual block can be expressed as:

[0069] in, It is l The input features of the layer, It is l The residual function of the layer is expressed as:

[0070] in, is the activation function, stands for batch normalization, represents the convolution kernel, Represents the bias term The supervised training process of the base learner can be achieved by minimizing the following loss function:

[0071] Where N is the number of source domain samples, It is a loss function used to measure the difference between the predicted value and the true value. The loss function used in the present invention is the mean absolute error (MAE), as follows:

[0072] in, Represents the number of training batches. Represents the true value of a batch of samples, Represents the predicted value of a batch of samples.

[0073] S212: Domain Classifier It is a simple neural network consisting of two fully connected layers and a ReLU activation function. It accepts features extracted from the base learner as input and outputs a scalar value representing the probability that the sample belongs to the source domain.

[0074] in, p Represents input features h The probability of belonging to the source domain, are weight matrices, are bias matrices, Represents the Sigmoid activation function, which is expressed as:

[0075] The training process of the domain classifier aims to enable it to distinguish whether the input features are from the source domain or the target domain, so the optimization objective can be defined as maximizing the probability that the domain classifier correctly classifies the source domain and target domain samples. It is defined as the negative value of the cross entropy loss of a binary classification:

[0076] in, It is i The characteristics of source domain samples, It is j The characteristics of the target domain samples, and are the number of samples in the source domain and the target domain, respectively.

[0077] S213: Adversarial feature alignment. The base learner trains the model in an adversarial way to make the features it generates more consistent between the source domain and the target domain. The successful training of the domain classification enables it to have basic judgment capabilities, so that the entire model can achieve the established goals. The goal of the adversarial loss is to make the features extracted by the base learner difficult to distinguish from the domain classifier, that is, to make the domain classifier It is impossible to determine whether these features are from the source domain or the target domain, so as to achieve feature alignment. Therefore, it is necessary to further add adversarial loss to the loss of the base learner, so that the features of the source domain and the target domain become more consistent under the judgment of the domain classifier, so as to achieve feature alignment and domain adaptation. The adversarial loss function is as follows:

[0078] The loss function of the final base learner consists of the supervised loss of the source domain and the adversarial loss of feature alignment, namely:

[0079] in, is a hyperparameter used to balance the impact of supervised learning and adversarial learning.

[0080] Meta-learner is a model constructed by a Transformer network, whose parameters are The input of the meta-learner is the steady-state temperature field of the source domain and the pseudo-steady-state temperature field of the target domain , the output is the transient temperature field prediction of the target domain. This process can be expressed as:

[0081] Each layer in the Transformer network consists of a multi-head self-attention mechanism and a feed-forward neural network. l Layer, output features It can be expressed as:

[0082] in, is a feed-forward neural network, It is a multi-head attention mechanism, which is calculated as:

[0083] Among them, each is the result of self-attention calculation:

[0084] in, By inputting features , the query, key and value matrices obtained by linear transformation, is the dimension of the key vector.

[0085] The goal of meta-learning is to find the parameters , so that it can be quickly adapted to the target domain with a small amount of fine-tuning. This process can be expressed by the following optimization objective:

[0086] in, is a set of tasks, each task includes a support set and queryset .

[0087] Finally, aggregate the losses of all tasks and optimize the initial parameters , so that the loss after inner layer update is minimized on all tasks:

[0088] in, is the outer learning rate.

[0089] In the present invention, the gradient descent method adopts the Adam algorithm, and its adaptive process is:

[0090]

[0091] in, W To optimize the parameters, is the learning rate, , is a hyperparameter. .

[0092] 3) Model Validation: S31: By inputting the steady-state temperature field of the target domain, based on the transient temperature field output by the network, the MAE, mean absolute percentage error (MAPE), root mean square error (RMSE) and R 2 The results were verified by the coefficient of determination.

[0093] Furthermore, the MAE, MAPE, RMSE, and R 2 The formula for the coefficient of determination is:

[0094] Example 3 like Figure 7 As shown, the present invention provides a prediction system for transient temperature field of a solar thermal steam turbine rotor during startup, comprising: The data acquisition module acquires the material setting, contact construction and boundary conditions of the solar thermal steam turbine rotor to be predicted, and obtains the steady-state temperature field data of the rotor through finite element analysis; The data preprocessing module preprocesses the acquired steady-state temperature field data; the steady-state temperature field of the exported result is converted into steady-state temperature values ​​of all nodes sorted by node numbers, and a steady-state temperature field sequence is obtained; The prediction module inputs the preprocessed steady-state temperature field sequence data into the preset transient temperature field prediction model to obtain the startup transient temperature field of the solar thermal steam turbine rotor to be predicted; wherein the preset transient temperature field prediction model includes a base learner constructed by a residual network, a domain classifier and a meta-learner constructed by Transforme, which is obtained by training in combination with an adversarial feature alignment strategy.

[0095] Example 4 The present invention provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the steps of the method for predicting the transient temperature field of a solar thermal steam turbine rotor during startup are implemented.

[0096] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0097] The present application is described with reference to the flowcharts and / or block diagrams of the methods, systems and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of the processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A system that specifies the functions of a box or multiple boxes.

[0098] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0099] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0100] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic features of the present invention. Therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is defined by the attached claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention. Any figure mark in the claims should not be regarded as limiting the claims involved.

[0101] In addition, it should be understood that although this specification is described in accordance with the implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation modes that can be understood by those skilled in the art. The above content is only to illustrate the technical idea of ​​the present invention, and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution according to the technical idea proposed by the present invention shall fall within the protection scope of the claims of the present invention.

Claims

1. A method for predicting transient temperature field of a solar thermal steam turbine rotor during startup, characterized in that: include: 1) Obtain the material setting, contact construction and boundary conditions of the solar thermal turbine rotor to be predicted, and obtain the steady-state temperature field data of the rotor through finite element analysis; 2) Preprocess the acquired steady-state temperature field data; The steady-state temperature field of the exported result is converted into the steady-state temperature values ​​of all nodes sorted by node numbers and added to obtain the steady-state temperature field sequence; 3) The preprocessed steady-state temperature field sequence data is input into the preset transient temperature field prediction model to obtain the startup transient temperature field of the solar thermal steam turbine rotor to be predicted; the preset transient temperature field prediction model includes a base learner constructed by a residual network, a domain classifier and a meta-learner constructed by Transforme, which is obtained by training in combination with an adversarial feature alignment strategy.

2. The method for predicting transient temperature field of a solar thermal steam turbine rotor during startup according to claim 1, characterized in that: Obtain the material setting, contact construction and boundary conditions of the solar thermal turbine rotor to be predicted, and obtain the steady-state temperature field data of the rotor through finite element analysis, including: By setting materials, building contacts, and adding boundary conditions to the actual rotor model, the corresponding steady-state temperature field is obtained through ANSYS APDL.

3. The method for predicting transient temperature field of a solar thermal steam turbine rotor during startup according to claim 1, characterized in that: Preprocess the acquired steady-state temperature field data, including: The steady-state temperature values ​​of all nodes sorted by node number are converted into a complete sequence , N represents the number of nodes, and the steady-state temperature field sequence is obtained.

4. The method for predicting transient temperature field of a solar thermal steam turbine rotor during startup according to claim 1, characterized in that: Preset transient temperature field prediction models, including: The blackone of the base learner of the meta-learning network model is constructed by the residual network. In order to complete the training of the network model under the premise of the transient temperature field of the target domain that is not needed, a domain classifier is further constructed to make the features of the source domain and the target domain difficult to distinguish in the model. The blackone of the meta-learner is composed of a Transformer network, which specifically includes the following steps: S211: Input the steady-state temperature field vector of the source domain , through a base learner built from ResNet-34 , whose parameters are , the goal of the base learner is to predict the corresponding transient temperature field ; This process is expressed as: Among them, the base learner It is composed of multiple residual blocks, and the output of each residual block is expressed as: in, It is l The input features of the layer, It is l The residual function of the layer is expressed as: in, is the activation function, stands for batch normalization, represents the convolution kernel, The supervised training process of the bias-based learner is achieved by minimizing the following loss function: Where N is the number of source domain samples, It is a loss function used to measure the difference between the predicted value and the true value; the loss function used is the mean absolute error MAE, as follows: in, Represents the number of training batches. Represents the true value of a batch of samples, Represents the predicted value of a batch of samples; S212: Domain Classifier It is a simple neural network consisting of two fully connected layers and a ReLU activation function; it accepts the features extracted from the base learner as input and outputs a scalar value representing the probability that the sample belongs to the source domain; specifically in, p Represents input features h The probability of belonging to the source domain, are weight matrices, are bias matrices, Represents the Sigmoid activation function, which is expressed as: The training process of the domain classifier aims to enable it to distinguish whether the input features are from the source domain or the target domain, so the optimization objective is defined as maximizing the probability that the domain classifier correctly classifies the source domain and target domain samples; defined as the negative value of the cross entropy loss of a binary classification: in, It is i The characteristics of source domain samples, It is j The characteristics of the target domain samples, and are the number of samples in the source domain and the target domain respectively; S213: Adversarial feature alignment; the base learner trains the model in an "adversarial" way so that the features it generates are more consistent between the source domain and the target domain; and the successful training of the domain classification enables it to have basic judgment capabilities. Therefore, in order to enable the entire model to achieve the established goals, adversarial loss is further added to the loss of the base learner, so that the features of the source domain and the target domain become more consistent under the judgment of the domain classifier, thereby achieving feature alignment and domain adaptation. The adversarial loss function is as follows: The loss function of the final base learner consists of the supervised loss of the source domain and the adversarial loss of feature alignment, namely: in, is a hyperparameter used to balance the impact of supervised learning and adversarial learning; Meta-learner is a model constructed by a Transformer network, whose parameters are ; The input of the meta-learner is the steady-state temperature field of the source domain and the pseudo-steady-state temperature field of the target domain , the output is the transient temperature field prediction of the target domain; this process is expressed as: Each layer in the Transformer network consists of a multi-head self-attention mechanism and a feed-forward neural network; l Layer, output features It is expressed as: in, is a feed-forward neural network, It is a multi-head attention mechanism, which is calculated as: Among them, each is the result of self-attention calculation: in, By inputting features , the query, key and value matrices obtained by linear transformation, is the dimension of the key vector.

5. The method for predicting transient temperature field of a solar thermal steam turbine rotor during startup according to claim 4, characterized in that: The goal of meta-learning is to find the parameters , which enables fast adaptation through fine-tuning on the target domain; This process is represented by the following optimization objective: in, is a set of tasks, each task includes a support set and queryset .

6. A prediction system for transient temperature field of a solar thermal steam turbine rotor during startup, characterized in that: include: The data acquisition module acquires the material setting, contact construction and boundary conditions of the solar thermal steam turbine rotor to be predicted, and obtains the steady-state temperature field data of the rotor through finite element analysis; A data preprocessing module is used to preprocess the acquired steady-state temperature field data; The steady-state temperature field of the exported result is converted into the steady-state temperature values ​​of all nodes sorted by node numbers and added to obtain the steady-state temperature field sequence; The prediction module inputs the preprocessed steady-state temperature field sequence data into the preset transient temperature field prediction model to obtain the startup transient temperature field of the solar thermal steam turbine rotor to be predicted; wherein the preset transient temperature field prediction model includes a base learner constructed by a residual network, a domain classifier and a meta-learner constructed by Transforme, which is obtained by training in combination with an adversarial feature alignment strategy.

7. A prediction system for transient temperature field of a solar thermal steam turbine rotor startup according to claim 6, characterized in that: In the data acquisition module, the material setting, contact construction and boundary conditions of the solar thermal steam turbine rotor to be predicted are obtained, and the steady-state temperature field data of the rotor is obtained through finite element analysis, including: By setting materials, building contacts, and adding boundary conditions to the actual rotor model, the corresponding steady-state temperature field is obtained through ANSYS APDL.

8. The prediction system for transient temperature field of a solar thermal steam turbine rotor startup according to claim 6 is characterized in that: In the data preprocessing module, the acquired steady-state temperature field data is preprocessed, including: The steady-state temperature values ​​of all nodes sorted by node number are converted into a complete sequence , N represents the number of nodes, and the steady-state temperature field sequence is obtained.

9. The prediction system for transient temperature field of a solar thermal steam turbine rotor during startup according to claim 6, characterized in that: In the prediction module, the preset transient temperature field prediction model includes: The blackone of the base learner of the meta-learning network model is constructed by the residual network. In order to complete the training of the network model under the premise of the transient temperature field of the target domain that is not needed, a domain classifier is further constructed to make the features of the source domain and the target domain difficult to distinguish in the model. The blackone of the meta-learner is composed of a Transformer network, which specifically includes the following steps: S211: Input the steady-state temperature field vector of the source domain , through a base learner built from ResNet-34 , whose parameters are , the goal of the base learner is to predict the corresponding transient temperature field ; This process is expressed as: Among them, the base learner It is composed of multiple residual blocks, and the output of each residual block is expressed as: in, It is l The input features of the layer, It is l The residual function of the layer is expressed as: in, is the activation function, stands for batch normalization, represents the convolution kernel, The supervised training process of the bias-based learner is achieved by minimizing the following loss function: Where N is the number of source domain samples, It is a loss function used to measure the difference between the predicted value and the true value; the loss function used is the mean absolute error MAE, as follows: in, Represents the number of training batches. Represents the true value of a batch of samples, Represents the predicted value of a batch of samples; S212: Domain Classifier It is a simple neural network consisting of two fully connected layers and a ReLU activation function; it accepts the features extracted from the base learner as input and outputs a scalar value representing the probability that the sample belongs to the source domain; specifically in, p Represents input features h The probability of belonging to the source domain, are weight matrices, are bias matrices, Represents the Sigmoid activation function, which is expressed as: The training process of the domain classifier aims to enable it to distinguish whether the input features are from the source domain or the target domain, so the optimization objective is defined as maximizing the probability that the domain classifier correctly classifies the source domain and target domain samples; defined as the negative value of the cross entropy loss of a binary classification: in, It is i The characteristics of source domain samples, It is j The characteristics of the target domain samples, and are the number of samples in the source domain and the target domain respectively; S213: Adversarial feature alignment; the base learner trains the model in an "adversarial" way so that the features it generates are more consistent between the source domain and the target domain; and the successful training of the domain classification enables it to have basic judgment capabilities. Therefore, in order to enable the entire model to achieve the established goals, adversarial loss is further added to the loss of the base learner, so that the features of the source domain and the target domain become more consistent under the judgment of the domain classifier, thereby achieving feature alignment and domain adaptation. The adversarial loss function is as follows: The loss function of the final base learner consists of the supervised loss of the source domain and the adversarial loss of feature alignment, namely: in, is a hyperparameter used to balance the impact of supervised learning and adversarial learning; Meta-learner is a model constructed by a Transformer network, whose parameters are ; The input of the meta-learner is the steady-state temperature field of the source domain and the pseudo-steady-state temperature field of the target domain , the output is the transient temperature field prediction of the target domain; this process is expressed as: Each layer in the Transformer network consists of a multi-head self-attention mechanism and a feed-forward neural network; l Layer, output features It is expressed as: in, is a feed-forward neural network, It is a multi-head attention mechanism, which is calculated as: Among them, each is the result of self-attention calculation: in, By inputting features , the query, key and value matrices obtained by linear transformation, is the dimension of the key vector.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a method for predicting a transient temperature field of a solar thermal steam turbine rotor upon startup according to any one of claims 1 to 5.

Images