Dynamic evaluation model attention optimization method oriented to specific instability mode
By introducing physical information of key nodes and attention optimization into the Transformer model, the evaluation logic problem of deep learning models that is difficult to explain and correct in TVSA is solved, achieving more accurate instability mode evaluation and stronger controllability.
Patent Information
- Application Number
- CN202510949696.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-10
- Publication Date
- 2025-10-17
AI Technical Summary
Existing deep learning models struggle to accurately capture critical nodes in transient voltage stability assessment (TVSA), making the assessment logic difficult to interpret and correct, thus affecting its practical application in power systems.
By introducing the physical information of key nodes and constructing an attention-guided loss function, the attention weight distribution of the Transformer model is optimized to make it focus on key features. The evaluation rules of other unstable modes are protected by the attention-preserving loss function, thereby improving the interpretability and controllability of the model.
The model's evaluation accuracy and interpretability for specific instability modes are enhanced, while the evaluation logic for other instability modes is protected, improving the model's controllability and generalization capabilities.
Smart Images
Figure CN120804710A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the field of power systems, and particularly relates to a dynamic evaluation model attention optimization method for specific unstable modes. BACKGROUND
[0002] In the power system, machine learning based on data mining gradually becomes a research hotspot in various power system fields, and has high precision and fast computing speed, and can quickly solve complex problems in actual production applications. However, the "black box" characteristics of the machine learning make the internal decision logic of the model difficult to be analyzed and verified by the operating personnel, and the power grid operating personnel cannot judge whether the evaluation logic of the deep learning model is reasonable, and it is also difficult to modify the model for specific fault conditions, which greatly limits the practical application of the deep learning model in the field of transient voltage stability assessment (TVSA).
[0003] In the transient voltage stability assessment (TVSA) problem, the accuracy of the evaluation is closely related to the correct capture of the key nodes in the unstable mode. However, the reflection of the attention weight on the importance of the features is the result of the spontaneous training of the model. If the key nodes are not paid enough attention to, it will be difficult to fully mine the relationship between the input measurements and the transient voltage stability. Therefore, if the model is trained with physical information as a guide, the attention distribution can be reasonably adjusted and optimized. Specifically, higher weights can be assigned to the key nodes in the unstable mode. Through this adjustment, the unstable mode can be accurately captured by the model. In guided training, the model can more deeply extract information related to stability in the sample, so that the originally incorrect TVSA result can be effectively corrected, thereby obtaining a more reliable and accurate TVSA.
[0004] Based on the above analysis, the application proposes a dynamic evaluation model attention optimization method for specific unstable modes to improve the explainability and controllability of the model. The physical information of the key unstable nodes is used to correct the evaluation rules of the model, so that the model can learn the information related to stability in the sample and make correct evaluation. SUMMARY
[0005] Based on the above, the purpose of the application is to propose a dynamic evaluation model attention optimization method for specific unstable modes, to guide the model to pay attention to the key features that dominate the instability by using the information of the key nodes, to correct the originally incorrect evaluation rules of the specific unstable mode, to make the evaluation rules of the model more consistent with the physical principles, to correct the specific unstable mode while ensuring the evaluation rules of other unstable modes, and to improve the explainability and controllability of the model.
[0006] In order to achieve the above technical purpose, the technical scheme adopted by the application is:
[0007] A dynamic evaluation model attention optimization method for specific instability modes, comprising:
[0008] 1) Generate training samples of multiple different fault types through simulation, extract the feature information of each bus node after fault clearing and the stability label of the sample as original data;
[0009] 2) Input the generated original data training samples into the TVSA model based on the Transformer for training, and collect instability mode samples that are difficult for the model to identify during training;
[0010] 3) Measure the collected instability mode samples, identify the key nodes in the samples according to the severity of the influence of the fault on each node in the sample, and mark the key nodes in the sample;
[0011] 4) According to the training set samples with marked key nodes, optimize the TVSA model based on the Transformer, construct an attention guide loss function, and guide the model to accurately capture the key features of the instability mode;
[0012] 5) Construct an attention preservation loss function, and realize model evaluation rule protection by constraining the attention weight fluctuation range of the key nodes in the non-selected instability mode samples.
[0013] Further, step 1) is as follows:
[0014] According to different fault information of multiple types, a large number of training samples are obtained through simulation;
[0015] The feature information of each bus node is extracted, and the feature information of 0-4 sampling periods after fault clearing and the stability label of the sample are used as original data.
[0016] Further, the feature information includes voltage amplitude U, voltage phase angle φ, etc.
[0017] Further, step 3) is as follows:
[0018] Measure the collected instability mode samples, and evaluate the key nodes of each sample;
[0019] Introduce the fault delay voltage recovery index η m,Busn to evaluate the severity of the influence of the fault on each node in the sample;
[0020] For the nth node in the mth sample, the formulas are as follows formulas (1)-(3):
[0021] η m,Busn = VDA m,Busn +FDT m,Busn (1)
[0022]
[0023] wherein, m is the total number of samples in the dataset, m = 1, 2, …, M; n is the total number of nodes in the system, n = 1, 2, …, N; VDA m,Busn represents the severity of voltage drop of the nth node of the mth sample after failure; FDT m,Busn represents the penalty term of whether the voltage of the nth node of the mth sample can recover to stability within 10 seconds; t cr is the fault clearing time; Δt is the duration after fault clearing; U th is the threshold value for judging whether the voltage is recovered; U m,Busn represents the voltage amplitude of the nth node of the mth sample after failure;
[0024] By evaluating the fault delay voltage recovery index of each node in the sample, the key nodes are identified, and the nodes with the largest index value in each sample are determined as key nodes, and all training set samples are labeled;
[0025] Based on the labeling result of the key nodes, the labeling matrix b is defined as follows (4):
[0026]
[0027] wherein, b M is the key node annotation state of the Mth sample; b M，BusN is the key node identifier corresponding to the Nth node in the Mth sample; if the nth node of the mth sample is the key node of the sample that occurs instability failure, then b m，Busn is 1, otherwise 0; thus the training set sample with labeled key nodes is obtained.
[0028] Further, step 4) is specifically as follows:
[0029] The sample set with labeled key nodes to be optimized is input into the Transformer-based TVSA model training, and the attention of the model to the features is output in the form of feature probability distribution through its attention weight; the attention weight matrix Aaw of the model is defined as follows (5):
[0030]
[0031] wherein, A aw is the attention weight obtained by averaging all attention heads; the matrix A aw is the attention weight of all samples; A M is the attention weight of the Mth sample; a M，BusNThe attention weight of the Nth node in the Mth sample, the sum of the attention weights of each sample is 1; the matrix A aw The sum of each row in the matrix is 1.
[0032] An attention guide loss term is introduced in the loss function, and the attention weight of the key node is integrated into the loss function through the term.
[0033] The attention guide loss term of the mth sample is shown in the following formulas (6)-(8):
[0034]
[0035] Where λ is an attention correction factor, indicating the optimization strength of the attention weight of the key node in the training sample; χ(m) is a logic factor. is the attention weight value of the key node in the training sample; a m,Busn is the attention weight value of the remaining nodes in the training sample; ζ m is a mode factor, which corrects the selected specific instability mode through the mode factor, adjusts the evaluation rule of the model for the instability mode, and X represents the feature sample and B represents the selected instability mode sample.
[0036] The logic factor filters the sample through the sgn symbol function, and filters out the sample whose attention weight value of the key node in the training sample is not the largest.
[0037] Further, step 5) is specifically as follows:
[0038] An attention keeping term is added to the loss function, and the fluctuation range of the attention weight of the key node in the non-selected instability mode sample is constrained to realize the protection of the evaluation rule of the model; the specific expression is as follows formula (9):
[0039]
[0040] In the formula, is the attention weight of the nth node in the mth non-selected instability mode sample before model optimization; is the attention weight of the nth node in the mth non-selected instability mode sample after model optimization; θ represents the threshold value for measuring the reasonable fluctuation of the key node attention of the non-selected instability mode sample before and after optimization; only when b m,Busn =1 (i.e. the node is the key node of the corresponding instability mode), the attention of the key node in the non-selected instability mode sample is limited within the reasonable threshold value of the weight fluctuation.
[0041] The overall optimization loss function is defined as shown in the following formula (10):
[0042]
[0043] In the formula, L CE is a cross-entropy function used to measure the classification error of the model in the training process, which is defined as formula (11) as follows:
[0044]
[0045] In the formula, y m represents the true label of the mth sample; represents the predicted label of the mth sample.
[0046] Further, the dynamic evaluation model attention optimization method for the specific instability mode further comprises: 6) inputting the measured feature data into the TVSA model with completed guided training to accurately capture the key features in the sample, evaluate the rules in accordance with the physical actuality, and accurately evaluate the sample.
[0047] Compared with the prior art, the technical scheme has the beneficial effects that:
[0048] 1. The physical information of the key node is used to correct the evaluation rules of the model, so that the model can learn the information related to stability in the sample and make correct evaluation, thereby enhancing the controllability and interpretability of the model.
[0049] 2. The loss function incorporating the key node information is constructed to realize the optimization guidance of the node attention distribution of the specific instability mode, thereby enhancing the controllability and accuracy of the model.
[0050] 3. The attention preservation term is introduced to protect the attention degree of the key node in the non-target instability mode sample, avoid the interference of the specific instability mode guidance on the evaluation logic of the remaining instability modes, and ensure the physical consistency and generalization ability of the overall evaluation logic of the model. BRIEF DESCRIPTION OF DRAWINGS
[0051] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.
[0052] Figure 1 is a brief implementation flowchart of the dynamic evaluation model attention optimization method for the specific instability mode of the present application;
[0053] Figure 2 is a wiring diagram of IEEE39 node system;
[0054] Figure 3is the attention weight distribution diagram of the optimized Transformer in the node 28-29 dominant instability mode;
[0055] Figure 4 is the attention weight distribution diagram of the optimized Transformer in the node 28-29 dominant instability mode (i.e. the attention weight distribution of the Transformer model optimized by the method proposed in the application). DETAILED DESCRIPTION
[0056] The application will be described in further detail below with reference to the drawings and embodiments. It is particularly pointed out that the following embodiments are only used to illustrate the application, but do not limit the scope of the application. Similarly, the following embodiments are only part of the embodiments of the application, not all embodiments, and all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of the application.
[0057] Referring to the accompanying Figure 1 The application provides a dynamic evaluation model attention optimization method for a specific instability mode, which comprises:
[0058] 1) A plurality of training samples of different fault types are generated by simulation, and the voltage, phase angle and other characteristic quantities of each bus node after fault clearing, as well as the stability label of the sample, are extracted as original data.
[0059] 2) The generated original data training samples are input into the Transformer-based TVSA model training, and instability mode samples that are difficult for the model to identify during training are collected.
[0060] 3) The severity of the impact of each node on the instability mode samples collected is measured to identify the key nodes in the samples, and the key nodes in the samples are labeled.
[0061] 4) The Transformer-based TVSA model is optimized according to the training set samples with labeled key nodes; an attention guide loss function is constructed to guide the model to accurately capture the key features of the instability mode, so that the attention distribution of the model is more in line with the physical reality, thereby optimizing the model evaluation rule.
[0062] 5) In order to avoid the influence of attention guide on other instability mode samples, an attention preservation loss function is constructed to ensure that the optimization model corrects the specific instability mode samples while not affecting the evaluation rules of other instability mode samples, thereby improving the generalization of the model.
[0063] 6) The measured characteristic data is input into the optimized TVSA model, which can accurately capture the key features in the sample, and the evaluation rule is in line with the physical reality while accurately evaluating the sample.
[0064] Embodiment 1
[0065] The embodiment provides a dynamic evaluation model attention optimization method for a specific instability mode, and specifically comprises the following steps:
[0066] 1) A large number of training samples are obtained through simulation according to different fault information of multiple types; characteristic information such as voltage amplitude U and voltage phase angle Φ of each bus node is extracted, and the characteristic information of 0-4 sampling periods after fault clearing and the stability label of the sample are taken as original data.
[0067] 2) The generated original data training sample is input into a TVSA model based on the Transformer for training, and samples of instability modes that are difficult for the model to identify in the training process are collected, and these samples that need to be optimized are optimized.
[0068] 3) The collected instability mode samples are measured and evaluated to obtain key nodes of each sample, and a fault delay voltage recovery index (η m,Busn ) is introduced to evaluate the severity of the influence of each node in the sample caused by the fault, which quantifies the severity of the influence of each node after the fault occurs.
[0069] Specifically, for the nth node in the mth sample, the formulas are as follows formulas (1)-(3):
[0070] η m,Busn =VDA m,Busn +FDT m,Busn (1)
[0071]
[0072] Wherein, m is the total number of samples in the data set, m = 1, 2,..., M; n is the total number of nodes in the system, n = 1, 2,..., N; VDA m,Busn represents the severity of voltage drop of the nth node of the mth sample after the fault; FDT m,Busn represents a penalty term for whether the voltage of the nth node of the mth sample can recover to stability within 10 seconds; t cr is the fault clearing time; Δt is the duration after fault clearing; U th is the threshold value (0.8 p.u.) for judging whether the voltage is recovered; U m,Busn represents the voltage amplitude of the nth node of the mth sample after the fault occurs.
[0073] The key nodes are identified by evaluating the failure delay voltage recovery index of each node in the sample, the node with the largest index value in each sample is determined as the key node, and all training set samples are labeled. Based on the labeling result of the key nodes, the labeling matrix b is defined as follows (4):
[0074]
[0075] Wherein, b M is the key node annotation state of the Mth sample; b M，BusN is the key node identifier corresponding to the Nth node in the Mth sample; if the nth node of the mth sample is the key node of the sample that occurs instability failure, b m，Busn is 1, otherwise 0. Thus, the training set sample with labeled key nodes is obtained.
[0076] 4) The sample set with labeled key nodes to be optimized is input into the Transformer-based TVSA model training, and the attention of the model to the features can be output in the form of feature probability distribution through its attention weight. Since the input feature information is constantly learned by the model, the output attention weight is also dynamically changing. The attention weight matrix A aw of the model is defined as follows (5):
[0077]
[0078] Wherein, A aw is the attention weight obtained by averaging all attention heads; the matrix A aw is the attention weight of all samples; A M is the attention weight of the Mth sample; specifically, a M，BusN is the attention weight of the Nth node in the Mth sample, and the sum of the attention weights of each sample is 1; the sum of each row in the matrix A aw is 1.
[0079] In order to enable the model to accurately capture the key features of the instability mode, the embodiment introduces an attention guiding loss term in the loss function, which integrates the attention weight of the key node into the loss function. Through the round-by-round optimization of the loss function, the attention of the model to the selected key features of the instability mode is strengthened, so that these instability failure samples are more accurately evaluated.
[0080] The attention guiding loss term of the mth sample is shown in the following formulas (6)-(8):
[0081]
[0082] Among them, λ is the attention correction factor, which represents the optimization strength of the attention weight of the key nodes in the training sample; χ(m) is the logical factor; is the attention weight value of the key node in the training sample; a m,Busn Represents the attention weight value of the remaining nodes in the training sample; ζ m Represents the mode factor, which is used to correct the selection of a specific instability mode and adjust the model's evaluation rule for the instability mode. X represents the feature sample, and B represents the selected instability mode sample.
[0083] Specifically, the logic factor filters samples through the sgn symbol function, filtering out samples whose attention weight values of key nodes in the training samples are not the largest, thereby reducing the cost of optimization training and effectively bypassing samples with correct evaluation rules.
[0084] The attention weights of key nodes can be adjusted during the optimization process as L AG As is reduced, a higher attention weight is adaptively assigned.,After the attention to key nodes in a specific instability mode is increased,,fault samples in this instability mode can be accurately,evaluated.
[0085] 5) In order to reduce the interference of the attention guidance process on the evaluation rules of other unstable modes, an attention retention loss term is designed to protect the evaluation rules of samples affected by the guidance. Samples with different unstable modes often correspond to different evaluation rules. Therefore, the attention retention term is added to the loss function to achieve model evaluation rule protection by constraining the fluctuation range of the attention weight of key nodes in non-selected unstable mode samples. Its specific expression is as follows (9):
[0086]
[0087] Where, represents the attention weight of the nth node in the mth non-selected unstable mode sample before model optimization; represents the attention weight of the nth node in the mth non-selected unstable mode sample after model optimization; θ represents the threshold for measuring the reasonable fluctuation of the attention of the key nodes of the non-selected unstable mode samples before and after optimization; only when b m,Busn =1 (i.e., the node is the key node of the corresponding instability mode), the attention to the key nodes in the non-selected instability mode samples is limited to a reasonable threshold of weight fluctuation, ensuring that the attention distribution of the key nodes in the non-selected instability mode samples remains stable during the optimization process, avoiding the interference of the evaluation rules due to the correction of the selected instability mode.
[0088] Therefore, the overall optimization loss function is defined as shown in the following formula (10):
[0089]
[0090] where L CE is the cross-entropy function, which measures the classification error of the model during the training process, and is defined as follows (11):
[0091]
[0092] where y m represents the true label of the mth sample; represents the predicted label of the mth sample.
[0093] With the optimization of the loss function, the model can improve the attention to the selected unstable key nodes and strengthen the ability of the evaluation model to mine unstable pattern samples, so that the model can more accurately evaluate these specific unstable pattern samples, thereby improving the prediction accuracy of the model. At the same time, in order to reduce the influence of attention guidance on the evaluation rules of other samples, the remaining unstable pattern samples are protected by using the attention preservation term. Only the evaluation rules of the selected unstable patterns are changed, and the remaining unstable patterns are not affected as much as possible, thereby ensuring the generalization ability of the optimized model.
[0094] 6) The measured feature data is input into the TVSA model completed by guided training, which can accurately capture the key features in the sample, and the evaluation rules conform to the physical actuality while accurately evaluating the sample.
[0095] The following provides a specific application example:
[0096] The simulation software uses PSD-BPA, and the test system uses the IEEE-39 node system (as shown in the attached Figure 2 The system has five load levels: 90%, 95%, 100%, 105%, and 110%. The fault occurs at 10%, 50%, and 90% of each line. The fault is a three-phase short circuit. The fault duration is 6 to 18.5 cycles, with an interval of 0.5 cycles, for a total of 26. Finally, a total of 32967 fault samples are generated in the IEEE-39 node system, including 12259 stable samples and 20708 unstable samples. The samples are extracted from the fault samples in a 1:1 ratio of stable samples and unstable samples to form the training set and the test set. The accuracy (Pacc), the false alarm rate (Pfa), and the false alarm rate (Pfa) are used as evaluation indicators to measure the performance of the model. Pacc provides an overall indicator of the overall prediction ability of the model, while Pfd and Pfa represent the misclassification rate of unstable samples and the misclassification rate of stable samples, respectively.
[0097] In order to further analyze the improvement of the method in the explainability and controllability of the Transformer-based TVSA model, the output attention distribution is used to analyze and adjust the TVSA rule learned by the model. The ability of the model to capture unstable patterns can be reflected by the attention weight of the key nodes in the sample. Therefore, the unstable mode information of the selected nodes 28-29 is taken as an example, as shown in Table 1.
[0098] Table 1 Unstable mode information
[0099]
[0100] In this unstable mode, the Transformer model and the Transformer model optimized by the method are compared, and the evaluation results are shown in Table 2. In order to more specifically present the test results, the attention weight distribution of the Transformer and the Transformer model optimized by the method (Proposed method group) is shown in Figure 3 and Figure 4 .
[0101] Table 2 Model performance in the case of node 28-29 dominant unstable mode
[0102]
[0103] From Table 2, Figure 3 and Figure 4 , when the attention weight of the key nodes (28, 29) is small, the accuracy of the model in evaluating the unstable mode is only 97.84%. This result shows that the model does not accurately identify, in contrast, after optimizing the training combined with the key node information, the key nodes (28, 29) have received enough attention, and the model accuracy reaches 98.70%, indicating that the model can more effectively capture the key features of the unstable mode.
[0104] From this, it can be inferred that the visualized attention weight distribution can be used to verify the correctness of the model in capturing the key features of the unstable mode. Therefore, the enhancement of the key node attention weight helps to correct the deviation between the model explainability and the physical concept, thereby improving the controllability of the model.
[0105] The above only describes some embodiments of the present application, and does not limit the protection scope of the present application, and any equivalent device or equivalent process transformation using the content of the specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application.
Claims
1. A method for optimizing attention of a dynamic evaluation model for a specific instability pattern, characterized in that: include: 1) Generate training samples of various fault types through simulation, extract the characteristic information of each bus node after the fault is cleared and the stability label of the sample as the original data; 2) Input the generated raw data training samples into the Transformer-based TVSA model training to collect unstable mode samples that are difficult for the model to identify during training; 3) Measure the collected instability mode samples, identify the key nodes in the samples according to the severity of the fault impact on each node in the samples, and mark the key nodes in the samples; 4) Based on the training set samples with labeled key nodes, the Transformer-based TVSA model is optimized and an attention-guided loss function is constructed to guide the model to accurately capture the key features of the instability pattern; 5) Construct an attention preservation loss function to protect the model evaluation rules by constraining the attention weight fluctuation range of key nodes in non-selected unstable pattern samples.
2. The method for optimizing attention of a dynamic evaluation model for a specific instability mode according to claim 1, characterized in that: Step 1) is as follows: Obtain a large number of training samples through simulation based on various types of fault information; The characteristic information of each bus node is extracted, and the characteristic information of 0-4 sampling cycles after the fault is cleared and the stability label of the sample are used as the original data.
3. The method for optimizing attention of a dynamic evaluation model for a specific instability mode according to claim 1, characterized in that: The characteristic information includes voltage amplitude U and voltage phase angle Φ.
4. The method for optimizing attention of a dynamic evaluation model for a specific instability mode according to claim 1, characterized in that: Step 3) is as follows: Measure the collected unstable mode samples and evaluate the key nodes of each sample; Introducing fault delay voltage recovery index η m,Busn To evaluate the severity of the impact of the failure on each node in the sample; For the nth node in the mth sample, the formulas are as follows (1) to (3): η m,Busn =VDA m,Busn +FDT m,Busn (1) Where m is the total number of samples in the dataset, m = 1, 2, ..., M; n is the total number of nodes in the system, n = 1, 2, ..., N; VDA m,Busn Indicates the severity of voltage drop at the nth node of the mth sample after the fault; FDT m,Busn The penalty term indicating whether the voltage of the nth node of the mth sample can recover stability within 10 seconds; t cr is the fault clearing moment; Δt is the duration after the fault is cleared; U th It is the threshold for judging whether the voltage has recovered; U m,Busn represents the voltage amplitude of the nth node after the fault of the mth sample occurs; Key nodes are identified by evaluating the fault delay voltage recovery index of each node in the sample. The nodes with the largest index values in each sample are determined as key nodes, and all training set samples are marked. Based on the labeling results of key nodes, the labeling matrix b is defined as follows (4): Among them, b M is the key node annotation status of the Mth sample; b M，BusN is the key node identifier corresponding to the Nth node in the Mth sample; if the nth node of the mth sample is the key node where the unstable fault occurs in the sample, then b m，Busn is 1, otherwise it is 0; thus, the training set samples with key nodes marked are obtained.
5. The method for optimizing attention of a dynamic evaluation model for a specific instability mode according to claim 1, characterized in that: Step 4) is as follows: The sample set that needs to be optimized with the key nodes marked is input into the Transformer-based TVSA model training. The model's attention to the features is output in the form of feature probability distribution through its attention weights; the model's attention weight matrix Aaw is defined as follows (5): Among them, A aw is the average attention weight of all attention heads; matrix A aw is the attention weight of all samples; A M is the attention weight of the Mth sample; a M，BusN is the attention weight of the Nth node in the Mth sample, and the sum of the attention weights of each sample is 1; the matrix A aw The sum of the rows in is 1; An attention-guided loss term is introduced into the loss function, through which the attention weights of key nodes are integrated into the loss function; The attention-guided loss term for the mth sample is shown in the following formulas (6) to (8): Among them, λ is the attention correction factor, which represents the optimization strength of the attention weight of the key nodes in the training sample; χ(m) is the logical factor; is the attention weight value of the key node in the training sample; a m,Busn Represents the attention weight value of the remaining nodes in the training sample; ζ m Represents the mode factor, which is used to correct the selection of a specific instability mode and adjust the model's evaluation rule for the instability mode. X represents the feature sample, and B represents the selected instability mode sample. The logical factor filters the samples through the sgn symbol function to filter out the samples whose attention weight values of the key nodes in the training samples are not the largest.
6. The method for optimizing attention of a dynamic evaluation model for a specific instability mode according to claim 1, characterized in that: Step 5) is as follows: An attention maintenance term is added to the loss function to implement model evaluation rule protection by constraining the fluctuation range of attention weights of key nodes in non-selected unstable pattern samples. Its specific expression is as follows: Where, represents the attention weight of the nth node in the mth non-selected unstable mode sample before model optimization; represents the attention weight of the nth node in the mth non-selected unstable mode sample after model optimization; θ represents the threshold for measuring the reasonable fluctuation of attention of key nodes of non-selected unstable mode samples before and after optimization; only when b m,Busn = 1, the attention to the key nodes in the non-selected unstable mode samples is limited to a reasonable threshold of weight fluctuation; The overall optimization loss function is defined as follows: Where, L CE is the cross entropy function, which is used to measure the classification error of the model during training. It is defined as follows: Where y m Represents the true label of the mth sample; represents the predicted label of the mth sample.
7. The method for optimizing attention of a dynamic assessment model for a specific instability mode according to any one of claims 1 to 6, characterized in that: Also includes: 6) Input the measured feature data into the trained TVSA model to accurately capture the key features of the sample, ensure that the evaluation rules are consistent with physical reality and accurately evaluate the sample.