Robust transient stability evaluation model obtaining method based on time sequence transfer learning

Through the time-series transfer learning method, combined with key curve selection, clustering and space-time importance evaluation, the problem of performance degradation of deep learning models in the power system during the transfer process is solved, and a robust transient stability evaluation of different system scales is achieved. , improves the generalization ability and efficiency of the model.

CN118940093AActive Publication Date: 2024-11-12STATE GRID CHONGQING ELECTRIC POWER CO ELECTRIC POWER RES INST +2

Patent Information

Application Number
CN202411142365.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-20
Publication Date
2024-11-12
Estimated Expiration
2044-08-20

AI Technical Summary

Technical Problem

The existing transient stability analysis method based on deep learning has deteriorated performance under the changes in the power system, and the timing characteristics are easily destroyed during the transfer learning process, making it difficult to be applicable to evaluations of different system scales.

Method used

Using a robust transient stability evaluation model acquisition method based on time-sequence transfer learning, we ensure cross-system feature consistency and distribution matching, and enhance feature representation through technologies such as key curve selection, clustering, spatial and temporal importance evaluation algorithm and random upsampling. Reduce intra-class differences and improve generalization ability.

Benefits of technology

Effectively retain the timing characteristics of transient curves, improve the generalization ability of data-driven transient stability analysis, is suitable for evaluation of different system scales, reduces the work burden of simulation analysts, and improves the performance of the model in new scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118940093A_ABST
    Figure CN118940093A_ABST
Patent Text Reader

Abstract

The invention discloses a robust transient stability evaluation model acquisition method based on time sequence transfer learning, which comprises the following steps: in an initial stage, selecting a key curve set from training samples of an original scene and taking the key curve set as a training sample of a data-driven transient stability evaluation model; applying a z-score normalization method to a training sample, inputting the normalized training sample into a time sequence neural network, and pre-training a data-driven transient stability evaluation model by minimizing classification loss; and in the second stage, the training samples are divided into a plurality of categories according to the labels, in each category, the training samples are divided into a plurality of groups through clustering, the comprehensive training loss is obtained by integrating the classification loss and the time sequence distribution matching loss, and the transient stability evaluation model is finally obtained by training by minimizing the comprehensive training loss. According to the method, the generalization performance of the data-driven model on the transient stability evaluation of the power system under different scales can be enhanced without a target domain sample, and the robustness is high.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the intersection of power systems and artificial intelligence, and in particular to a method for acquiring a robust transient stability assessment model based on time series transfer learning. Background Art

[0002] When the system is subject to major disturbances, ensuring the stability of the power system is a basic requirement for grid operation. Therefore, power companies must conduct comprehensive simulations regularly to evaluate the stability of the system under expected emergencies. In simulation analysis, it is necessary to check whether a large number of transient curves in the form of multivariate time series are stable. This leads to the challenge of heavy computational burden for existing manual analysis methods. In addition, with the rapid development of high-penetration renewable energy, the stability characteristics of the power system have become increasingly complex, resulting in a significant increase in the workload and density of transient stability simulation analysis. There is an urgent need to improve the efficiency of transient curve evaluation and reduce the workload of simulation analysts.

[0003] Data-driven methods are ideal tools for intelligent simulation analysis by identifying transient curves. Existing data-driven methods can generally be divided into shallow learning and deep learning. Compared with shallow learning methods, deep learning methods have attracted widespread attention because they can extract more abstract and complex features from data. Various deep learning methods, such as deep fully connected neural networks, convolutional neural networks, time series neural networks, and graph neural networks, have been applied to transient stability analysis. However, the excellent performance of deep learning methods relies on the assumption that the training dataset and the test dataset follow the same distribution. In fact, due to the continuous changes in the power system, such as load demand growth and system expansion, this assumption is often not true. Since the training dataset cannot cover all possible variable operating conditions in the power system, the performance of the well-trained deep model will inevitably deteriorate in new unlearned operating scenarios. In order to be applicable to new operating scenarios (target domains), a natural approach is to retrain the deep learning model for the new scenario. However, this is time-consuming and requires a large number of samples of new scenarios (target samples).

[0004] Transfer learning can leverage the knowledge of the original scene (source domain) to improve retraining efficiency and alleviate the sample requirements of the new operating scene (target domain), which is expected to solve the above problems. Existing transient stability evaluation methods based on feature transfer are mainly based on distribution alignment based on the maximum mean difference of marginal distribution, conditional distribution or joint distribution. However, during the migration process, the temporal features of transient stability may be destroyed. Therefore, existing feature transfer methods can only be applied to small-scale cross-system transient stability problems, or rely on hundreds of new scene samples. How to retain temporal features during the migration process and improve the generalization ability of data-driven transient stability analysis needs further study. Summary of the invention

[0005] The present invention proposes a method for acquiring a robust transient stability assessment model based on timing transfer learning, which can effectively retain the timing characteristics of the transient curve and make the data-driven intelligent transient stability assessment model learned in the source domain suitable for transient stability assessment under different system scales.

[0006] The present invention discloses a method for acquiring a robust transient stability assessment model based on time series transfer learning, which comprises:

[0007] Step 1: In the initial stage, a set of key curves is selected from the training samples of the original scene and used as the training samples of the data-driven transient stability assessment model; the z-score normalization method is applied to the training samples, and the normalized training samples are input into the time series neural network to pre-train the data-driven transient stability assessment model by minimizing the classification loss;

[0008] Step 2: In the second stage, the training samples are divided into multiple categories according to the labels. In each category, the training samples are divided into multiple groups by clustering. The comprehensive classification loss and time series distribution matching loss are combined to obtain the comprehensive training loss. The training is performed by minimizing the comprehensive training loss, and finally a trained transient stability evaluation model is obtained.

[0009] Furthermore, the step 1 comprises:

[0010] A key curve selection method is adopted to ensure the consistency of input features of power systems across scales, and the training data are divided into multiple categories according to the labels. In each category, the training samples are divided into multiple groups through clustering for subsequent intra-class distance minimization;

[0011] Through the transient time series curve characteristics of each hidden layer, the distribution distance between groups of different categories is increased and the distribution distance between groups of the same category is reduced to enhance the characteristic representation of the transient time series curve;

[0012] A spatiotemporal importance evaluation algorithm is used to quantify the importance of each neuron in each hidden layer of the temporal neural network for temporary stability evaluation, and a random upsampling method is used to ensure the balance of each group of samples.

[0013] Furthermore, the key curve selection method is adopted to ensure the consistency of input characteristics of power systems across scales, including:

[0014] The transient voltage curves are sorted in descending order of variance, and the transient generator power angle curves are sorted in descending order of maximum absolute value. The first q curves are selected from each sorted set as the key curve set, which is used as the input feature vector of the data-driven transient stability assessment model. The transient voltage curves and transient generator power angle curves are clustered using their inherent characteristics to unify the dimensions of the key curve set.

[0015] Furthermore, in each round of the training process in the initial stage, random upsampling is performed so that the group with smaller sample size randomly selects additional batches according to the maximum batch number among all groups to ensure that the number of batches among the groups is consistent.

[0016] Furthermore, the spatiotemporal importance evaluation algorithm is used to quantify the importance of each neuron in each hidden layer of the temporal neural network for temporary stability evaluation, including:

[0017] At the beginning, each time state is considered equally important, and The values ​​of are initialized to the same value 1 / V. During the training process, The value of is adaptively adjusted according to the distribution distance of each round; represents the weights of the i-th group of samples and the j-th group of samples in the neuron t state, which aims to learn the relative importance of V hidden layer neurons in the temporal neural network.

[0018] Further, The adaptive adjustment formula is:

[0019]

[0020] In the formula, represents the weights of the i-th group of samples and the j-th group of samples in the neuron t state in round n+1, P is the update function for calculating the temporal distribution matching loss at different learning stages, is the distribution distance of time step t in round n, and σ(·) is the sigmoid function.

[0021] Further, The expression is:

[0022]

[0023] in, For input x si,1 The feature representation of the h-th layer neuron at state t, For input x si,2 The feature of the h-th layer neuron in state t, x si,1 and x si,2 are different groups in the same category. For input x sj,1 The feature representation of the h-th layer neuron at state t, For input x sj,2 The feature of the h-th layer neuron in state t, x sj,1 and x sj,2 are different groups in the same category. for and The splicing form, For input and The concatenation form of , ε is a very small hyperparameter, and d is the distance.

[0024] Furthermore, from equations (8) and (9), we can see that if the distribution distance of round n is greater than the distribution distance of round n-1, then to expand its effect of enhancing feature representation.

[0025] Furthermore, the expression of the timing distribution matching loss is:

[0026]

[0027] Among them, L tdm is the time series distribution matching loss, β h It represents the weight of the hth hidden layer, which represents the spatial importance between hidden layers, and exp is the natural exponential function.

[0028] Furthermore, β h The empirical formula is:

[0029]

[0030] Among them, l is the number of hidden layers.

[0031] Due to the adoption of the above technical solution, the present invention has the following advantages:

[0032] 1. The present invention proposes a simple and effective key curve selection method, which is in line with the engineering experience of temporary stability judgment in actual engineering, and does not require pre-definition of complex screening rules. It can effectively solve the problem of inconsistent data dimensions in cross-system temporary stability assessment.

[0033] 2. The present invention takes into account the timing characteristics of transient curves during the migration process, which can effectively enhance the feature representation of transient timing curves, improve the inter-class distance of transient curves and reduce their intra-class differences. Without the need for target domain samples, the generalization performance of the data-driven model for transient stability assessment of power systems of different scales can be enhanced.

[0034] 3. The present invention can be widely used in artificial intelligence-based power system transient stability judgment, and is particularly suitable for transient stability assessment under different system scales. It can also be further extended to other time series classification problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the embodiments of the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0036] Figure 1 The present invention is a flowchart of a method for acquiring a robust transient stability assessment model based on time series transfer learning according to an embodiment of the present invention. DETAILED DESCRIPTION

[0037] The present invention is further described in conjunction with the accompanying drawings and embodiments, and the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by ordinary technicians in this field should fall within the scope of protection of the embodiments of the present invention.

[0038] See also Figure 1 The present invention provides an embodiment of a method for acquiring a robust transient stability assessment model based on time series transfer learning, which includes:

[0039] Step 1: In the initial stage, a set of key curves is selected from the training samples of the original scene and used as the training samples of the data-driven transient stability assessment model; the z-score normalization method is applied to the training samples, and the normalized training samples are input into the time series neural network to pre-train the data-driven transient stability assessment model by minimizing the classification loss;

[0040] Step 2: In the second stage, the training samples are first divided into multiple categories according to the labels. In each category, the training samples are divided into multiple groups by clustering. The classification loss and the time series distribution matching loss are combined to obtain the comprehensive training loss. The training is performed by minimizing the comprehensive training loss, and finally the trained transient stability evaluation model is obtained. The classification loss corresponds to formula (4).

[0041] In one embodiment, step 1 includes:

[0042] A key curve selection method is adopted to ensure the consistency of input features of power systems across scales, and the training data are divided into multiple categories according to the labels. In each category, the training samples are divided into multiple groups through clustering for subsequent intra-class distance minimization;

[0043] Through the transient time series curve characteristics of each hidden layer, the distribution distance between groups of different categories is increased and the distribution distance between groups of the same category is reduced to enhance the characteristic representation of the transient time series curve;

[0044] A spatiotemporal importance evaluation algorithm is used to quantify the importance of each neuron in each hidden layer of the temporal neural network for temporary stability evaluation, and a random upsampling method is used to ensure the balance of each group of samples.

[0045] In this embodiment, a simple key curve selection method is proposed to ensure the consistency of input features of power systems across scales, and the training data is divided into two categories according to the labels. In each category, the training samples are divided into two groups by the k-means method for subsequent intra-class distance minimization; secondly, the feature representation of the transient time series curve is enhanced by increasing the distribution distance of different categories and reducing the distance of the same category distribution for the transient time series curve features of each hidden layer.

[0046] In addition, a spatiotemporal importance evaluation algorithm is proposed to quantify the importance of each neuron in each hidden layer of the temporal neural network for temporary stability evaluation, and a random upsampling method is further used to ensure the balance of each group of samples. Finally, the effectiveness of the proposed method is verified by simulation using the CEPRI-36 node system and its modified system. The specific steps of the method are as follows:

[0047] (1) Training sample feature division:

[0048] In practical applications, changes in the scale of power systems will lead to changes in the number of transient voltage and generator power angle curves. Usually, the key transient voltage and generator power angle curves are focused on, which show fast response and significant fluctuation amplitude, and transient stability assessment is performed based on key curves. A method for selecting key curves is proposed. The curves that are given priority correspond to the curves with the largest fluctuation amplitude.

[0049] The transient voltage curves are sorted in descending order by variance, and the transient generator power angle curves are sorted in descending order by the maximum absolute value. The first q curves are selected from each sorted set as the key curve set, as shown in formula (1). This key curve set is used as the input feature vector Xin-rank = {U rank ,θ rank The method uses the inherent characteristics of the curves for clustering, thereby unifying the dimensions of the key curve set. This eliminates the need to predefine fixed filtering rules and solves the dimensionality inconsistency problem of data-driven models migrating between power systems of different sizes.

[0050] U rank = {U rank,1 ,L,U rank,q}

[0051] θ rank ={θ rank,1 ,L,θ rank,q} (1)

[0052] Among them, U rank,1 , U rank,q The first and qth curves representing the transient voltage variance in descending order, θ rank,1 ,θ rank,q The first and qth power angle curves are sorted in descending order according to the maximum absolute value.

[0053] To ensure that the training samples are divided into four different groups with significant differences, the samples are initially classified according to the sample labels as shown in Equation (2). To enhance the classification ability of the model, the distributions of the two types of samples should be as different as possible. Since stable and unstable modes are often similar in cross-scale power systems, the resulting prediction model is expected to show excellent generalization ability.

[0054] X si ={x s |y s =0}

[0055] X sj ={x s |y s =1} (2)

[0056] Among them, X si and X sj are different categories, x s is the sample, y s is the sample label.

[0057] In each category, the training samples are further clustered into two different groups using the k-means algorithm. The k value in the k-means algorithm is fixed to 2. The k-means algorithm is chosen because it is simple and easy to implement. Therefore, the training data can be expressed as shown in Equation (3). By making the two different groups within each category closely combined, the performance of the data-driven TAS model can be enhanced.

[0058] X si ={X si,1 ,X si,2}

[0059] X sj ={X sj,1 ,X sj,2} (3)

[0060] Among them, X si and X sj The k-means algorithm is clustered into X si,1 ,X si,2 and X sj,1 ,X sj,2 .

[0061] To address the problem of unbalanced sample sizes between different groups, a random upsampling technique was used. During each epoch of the training process, the group with a smaller sample size randomly selected additional batches based on the maximum batch number among all groups to ensure that the number of batches across groups was consistent.

[0062] (2) Alignment of transient curve timing distribution:

[0063] Given the sample groups, the transient curve time series distribution alignment module aims to enhance their feature representation by matching the distributions of different groups. Therefore, the learned model M is expected to achieve good generalization in unknown scenarios compared to methods that only rely on local or statistical information.

[0064] TDM is used for classification loss L pred , referred to as classification loss, can be expressed as formula (4):

[0065]

[0066] Among them, loss(.,.) is the loss function, such as MSE loss, are learnable model parameters.

[0067] However, minimizing (4) only promotes the extraction of predictive knowledge from each group. It cannot simultaneously enhance the distribution diversity between different categories and improve the sample density within a single category. Referring to existing domain adaptation work, a simple solution is to maximize the inter-class discreteness between the four groups and minimize the intra-class discreteness as a regularization term. However, existing work usually matches the distribution on the deep representation. h Represents the V hidden states of the hth hidden layer of the time series neural network. Several commonly used distribution matching distances include cosine distance, maximum mean discrepancy (MMD) and adversarial distance. Existing work usually aligns the distribution of the last hidden layer features of the time series neural network, which can be formulated as:

[0068]

[0069] Where l is the number of hidden layers; to avoid numerical problems, ε is a very small hyperparameter to avoid division by zero, and is generally set to 1×10-8.

[0070] Since each hidden neuron only contains part of the distribution information of the input transient voltage or generator power angle sequence, the above regularization term cannot capture the information of each hidden layer neuron, that is, it cannot effectively characterize the distribution information of the transient timing curve. Therefore, attention should be paid to each hidden layer neuron of the timing neural network. In addition, in order to quantify the relative importance of the Vth hidden layer neuron in different layers, the present invention proposes a spatiotemporal importance evaluation algorithm. Finally, the proposed timing distribution matching loss Ltdm is expressed as:

[0071]

[0072] Among them, L tdm is the time series distribution matching loss, β h represents the weight of the hth hidden layer, representing the spatial importance between hidden layers, exp is the natural exponential function, For input x si,1 The feature representation of the h-th layer neuron at state t, For input x si,2 The feature of the h-th layer neuron in state t, x si,1 and x si,2 are different groups in the same category. For input x sj,1 The feature representation of the h-th layer neuron at state t, For input x sj,2 The feature of the h-th layer neuron in state t, x sj,2 and x sj,2 are different groups in the same category. for and The splicing form, For input and The concatenation form of , ε is a very small hyperparameter, and d is the distance.

[0073] Generally speaking, deeper feature representations have a greater impact on the results. h The empirical formula is:

[0074]

[0075] in, It means that the sum of all spatial weights is 1, and l is the number of hidden layers.

[0076] Each time state is considered equally important at the beginning, and The values ​​of are initialized to the same value 1 / V. During the training process, The value of can be adaptively adjusted according to the distribution distance of each round; represents the weights of the i-th group of samples and the j-th group of samples in the neuron t state, which aims to learn the relative importance of V hidden layer neurons in the temporal neural network.

[0077] The adaptive adjustment formula is:

[0078]

[0079] In the formula, represents the weights of the i-th group of samples and the j-th group of samples in the neuron t state in round n+1, P is the update function for calculating the temporal distribution matching loss at different learning stages, is the distribution distance of time step t in round n, and σ(·) is the sigmoid function.

[0080]

[0081] It can be seen from equations (8) and (9) that if the distribution distance of round n is greater than the distribution distance of round n-1, then to expand its effect of enhancing feature representation.

[0082] In summary, the training process of the data-driven temporary stability intelligent evaluation model based on the proposed method is summarized as follows: In the initial stage, the training samples of the original scene are selected by key curves through formula (1). In order to solve the numerical problem, the z-score normalization method is applied to the samples. This method only requires the mean and standard deviation of historical statistical data, which is very suitable for unknown scenes. Compared with other preprocessing techniques (such as minimization-maximization methods), the z-score method better preserves the shape of the input distribution. Then, the normalized samples are input into the temporal neural network, including RNN, LSTM, GRU, TCN and Transformer. The data-driven model M is pre-trained by minimizing formula (4). In the second stage, the normalized samples are first divided into four groups by formulas (2)-(3). Combining formulas (4) and (6), the comprehensive training loss is obtained:

[0083] L(θ,α)=L pred (θ)+λ1L tmd (θ,α) (10)

[0084] Where λ1 is a hyperparameter. The data-driven model M is trained by equation (10). The trained model M can improve the computational performance on the new scene while maintaining the performance level on the source scene.

[0085] For ease of understanding, the present invention provides a more specific embodiment:

[0086] (1) Simulation information and comparison method

[0087] In this embodiment, the CEPRI-36 node system is used as the source domain system, and the PSASP production samples are used. The fault type, occurrence time and location, power flow level, and load motor component ratio setting parameters are shown in Reference 1 (Zhou Rui, Yang Yan, Yu Juan, et al. Data-driven power system transient dominant instability mode identification method adapted to topology changes [J]. Proceedings of the CSEE, published online, 2024.), and a total of 7800 samples are obtained. The target domain includes adding a bus between BUS21 and BUS23 of the CEPRI-36 node, copying the original generator BUS3 and transformer T2W3, connecting them to the newly added bus, named CEPRI-37 node system, and connecting the 37 node system to form a new system, named CEPRI-74 node system.

[0088] The accuracy of a classifier is usually measured by the following indicators: Accuracy, Precision, F1-score and AUC (Area Under Curve, AUC).

[0089]

[0090] Among them, TP is a true positive example, TN is a true negative example, FP is a false positive example, and FN is a false negative example.

[0091]

[0092] Among them, Precision is the precision rate, Recall is the recall rate,

[0093] AUC is the area under the ROC curve. The ROC curve is a curve with the false positive rate (FPR) as the horizontal axis and the true positive rate (TPR, i.e. recall rate) as the vertical axis.

[0094] The present invention verifies the effectiveness of the proposed method on five common time series neural networks (RNN, GRU, LSTM, TCN, Transformer), and uses the proposed feature migration method, namely RNN+TTLF, GRU+TTLF, LSTM+TTLF, TCN+TTLF, Transformer+TTLF, to compare the test results in different scenarios after training with only source domain samples. Each time series neural network has 4 hidden layers, each hidden layer has 64 neurons, the number of training batches is 32, the number of pre-training rounds is 50, the total number of training times does not exceed 550, the number of rounds of the early stopping method is 30, and the model is trained using the adam algorithm.

[0095] (2) Verification of the effectiveness of the proposed method

[0096]

[0097]

[0098] This experiment evaluated the performance of different models in the source domain system (36 nodes) and the target domain system (37 nodes and 74 nodes). In the source domain system, the present invention maintains the performance of the traditional model, and has excellent performance in terms of accuracy, precision, F1 value and AUC. In the target domain system with more complex scenarios, the present invention shows better performance than the traditional model. In 37 nodes, the present invention is 12% higher than RNN in accuracy, and more than 10% higher in F1 value and AUC. Compared with Transformer, there is an improvement of about 5% in accuracy, F1 value and AUC. In the 74-node system, the present invention is significantly improved compared with RNN, with an accuracy improvement of about 14% and an F1 value improvement of more than 20%. It also maintains a good improvement in other traditional models. The experimental results show that the present invention performs well in the process of transfer learning, maintaining high accuracy and good performance in scenarios of different scales.

[0099] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A robust transient stability assessment model acquisition method based on time series transfer learning, It is characterized in that include: Step 1: In the initial stage, a set of key curves is selected from the training samples of the original scenario and used as the training samples of the data-driven transient stability assessment model; The z-score normalization method is applied to the training samples, and the normalized training samples are input into the time series neural network to pre-train the data-driven transient stability assessment model by minimizing the classification loss. Step 2: In the second stage, the training samples are divided into multiple categories according to the labels. In each category, the training samples are divided into multiple groups by clustering. The comprehensive classification loss and time series distribution matching loss are combined to obtain the comprehensive training loss. The training is performed by minimizing the comprehensive training loss, and finally a trained transient stability evaluation model is obtained.

2. The method according to claim 1, It is characterized in that The step 1 comprises: A key curve selection method is adopted to ensure the consistency of input features of power systems across scales, and the training data are divided into multiple categories according to the labels. In each category, the training samples are divided into multiple groups through clustering for subsequent intra-class distance minimization; Through the transient time series curve characteristics of each hidden layer, the distribution distance between groups of different categories is increased and the distribution distance between groups of the same category is reduced to enhance the characteristic representation of the transient time series curve; A spatiotemporal importance evaluation algorithm is used to quantify the importance of each neuron in each hidden layer of the temporal neural network for temporary stability evaluation, and a random upsampling method is used to ensure the balance of each group of samples.

3. The method according to claim 2, It is characterized in that The critical curve selection method is described to ensure consistency of input characteristics across power systems of different scales, including: The transient voltage curves are sorted in descending order of variance, and the transient generator power angle curves are sorted in descending order of maximum absolute value. The first q curves are selected from each sorted set as the key curve set, which is used as the input feature vector of the data-driven transient stability assessment model. The transient voltage curves and transient generator power angle curves are clustered using their inherent characteristics to unify the dimensions of the key curve set.

4. The method according to claim 1, It is characterized in that In each round of the initial training process, random upsampling is performed so that the group with smaller sample size randomly selects additional batches according to the maximum batch number among all groups to ensure consistent batch numbers across groups.

5. The method according to claim 2, It is characterized in that The spatiotemporal importance evaluation algorithm is used to quantify the importance of each neuron in each hidden layer of the temporal neural network for temporary stability evaluation, including: At the beginning, each time state is considered equally important, and The values ​​of are initialized to the same value 1 / V. During the training process, The value of is adaptively adjusted according to the distribution distance of each round; represents the weights of the i-th group of samples and the j-th group of samples in the neuron t state, which aims to learn the relative importance of V hidden layer neurons in the temporal neural network.

6. The method according to claim 5, It is characterized in that The adaptive adjustment formula is: In the formula, represents the weights of the i-th group of samples and the j-th group of samples in the neuron t state in round n+1, P is the update function for calculating the temporal distribution matching loss at different learning stages, is the distribution distance of time step t in round n, and σ(·) is the sigmoid function.

7. The method according to claim 6, It is characterized in that The expression is: in, For input x si,1 The feature representation of the h-th layer neuron at state t, For input x si,2 The feature of the h-th layer neuron in state t, x si,1 and x si,2 are different groups in the same category. For input x sj,1 The feature representation of the h-th layer neuron at state t, For input x sj,2 The feature of the h-th layer neuron in state t, x sj,1 and x sj,2 are different groups in the same category. for and The splicing form, For input and The concatenation form of , ε is a very small hyperparameter, and d is the distance.

8. The method according to claim 6, It is characterized in that From equations (8) and (9), we can see that if the distribution distance of round n is greater than the distribution distance of round n-1, then to expand its effect of enhancing feature representation.

9. The method according to claim 6, It is characterized in that The expression of the timing distribution matching loss is: Among them, L tdm is the time series distribution matching loss, β h It represents the weight of the hth hidden layer, which represents the spatial importance between hidden layers, and exp is the natural exponential function.

10. The method according to claim 9, It is characterized in that β h The empirical formula is: Among them, l is the number of hidden layers.

Citation Information

Patent Citations

  • Behavior recognition method by extracting coordinates of human joint points

    CN110826453A

  • Power system transient stability adaptive evaluation method based on transfer learning

    CN110879917A

  • Remote sensing image cross-domain classification method based on interpretable deep learning

    CN113537317A

  • Unsupervised bearing fault diagnosis method based on adaptive residual adversarial network

    CN115127814A

  • Parameter identification method and system of load model, computer equipment and storage medium

    CN115422869A

Cited By

  • Power system multi-scene adaptive transient stability assessment method based on active transfer learning

    CN120851390A