Power system instability state identification model training method and identification method

Abstract

This invention discloses a training method and identification method for a power system instability state identification model, belonging to the field of power system stability analysis technology. The training method includes noise identification of the original sample set D0 carrying instability state labels before training the instability state identification model. Noise identification includes: training model M0 with sample set D0, and dividing sample set D0 into simple sample sets D based on the magnitude of the loss. s and complex sample set D H / N , sample set D s The unstable state labels of some samples in the dataset are changed to form an artificially labeled noisy sample set D. a ; with sample set D a Train model M1, and adjust the sample set D according to the loss value. a Divided into simple sample sets D a，s and complex sample set D a，H / N According to sample set D a，H / N Training model M a Through model M a Construct a mapping relationship between the loss value sequence and the label type; then based on model M a Identify sample set D H / N The above process allows for the identification of noise in the original sample set, enabling more effective training of the power system instability identification model.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of power system stability analysis, and more particularly relates to a power system instability state identification model training method and identification method. BACKGROUND

[0002] The safe and stable operation of a power system is of great importance to energy security and economic and social development, and its safety and stability are problems that system planning, operation and protection workers focus on. After a large disturbance is applied to a power system, it is necessary to quickly and accurately determine whether the system is stable and the specific instability mode, which will provide a basis for taking emergency control measures and effectively ensure the safety and stability of the system.

[0003] Time-domain simulation, energy function, bifurcation analysis and the like are commonly used methods for instability mode discrimination, but there are many difficulties in applying them to actual large power grids. In recent years, some research has applied deep learning to instability state identification, and deep learning methods can learn the mapping relationship between simulation data or measured data and stability and dominant instability mode without complex physical mechanisms, and can make judgments directly from data, which has broad application prospects in dominant instability mode discrimination problems.

[0004] However, in practical applications, the success of deep learning is largely dependent on obtaining a large number of accurately labeled training samples. Because different instability modes can have similar appearances and coupling properties, it is difficult to accurately determine the instability mode by means of mechanism, and there is no reliable automatic program or simple rule that can be directly applied. Therefore, the instability mode sample set relies on manual labeling by power system experts.

[0005] However, even for power system experts, sample labeling for instability state identification is a quite complex and challenging task, and errors are inevitable in labeling. Incorrect labeling (label noise) can have a serious negative impact on deep learning model training. If not handled, the deep learning model will overfit the incorrect labels and learn incorrect mapping relationships, which will seriously affect correct stability analysis and dominant instability state identification.

[0006] Therefore, how to deal with labeling errors in the sample set and minimize the impact of labeling errors on deep learning models is a technical problem that needs to be solved. SUMMARY

[0007] In view of the above defects or improvement needs of the prior art, the present application provides a power system instability state identification model training method and identification method, which aims to reduce the requirement for sample labeling accuracy, and can train an accurate instability state identification model based on a sample set containing label noise, and improve the model implementation efficiency.

[0008] To achieve the above object, according to one aspect of the present application, a training method of a power system instability state recognition model is provided, comprising:

[0009] Label noise recognition is performed on an original sample set D0 carrying an instability state label; the original sample includes power grid operation data, and the instability state includes whether instability and instability mode;

[0010] A target neural network is trained with the power grid operation data as input and the instability state as output, to obtain a power system instability state recognition model; wherein,

[0011] The process of label noise recognition comprises:

[0012] The model M0 is trained with the power grid operation data of the original sample set D0 as input and the instability state as output, and the original sample set D0 is divided into a simple sample set D s with smaller loss and a complex sample set D H / N with larger loss according to the final loss value, and the loss value sequence L H / N of each complex sample in the complex sample set D 0，H / N is recorded in the whole training process;

[0013] The instability state label of part of the samples in the simple sample set D S is changed to an error label, to form an artificial label noise sample set D a ;

[0014] The model M1 is trained with the power grid operation data of the artificial label noise sample set D a as input and the instability state as output, and the artificial label noise sample set D a is divided into a simple sample set D a，S with smaller loss and a complex sample set D a，H / N with larger loss according to the final loss value, and the loss value sequence L a，H / N and the label type η a，H / N of each complex sample in the complex sample set D a，H / N are recorded in the whole training process, and the label type η a，H / N records whether the label of the corresponding complex sample is an error label;

[0015] The auxiliary model M a，H / N is trained with the loss value sequence L a，H / N of the complex sample set D a，H / N as input and the label type η a as output;

[0016] The loss value sequence L H / N of the complex sample set D 0，H / N is input into the auxiliary model M aa complex sample set D H / N a noise sample set D N with incorrect labels, and a difficult sample set D H .

[0017] In one embodiment, the training sample set used to train the target neural network is obtained by removing the noise sample set D N from the original sample set D0, and the target neural network is directly trained using the training sample set.

[0018] In one embodiment, the training sample set used to train the target neural network includes an annotated sample set and an unannotated sample set, wherein the annotated sample set is obtained by removing the noise sample set D N from the original sample set D0, and contains simple samples D S and difficult samples D H , and the unannotated sample set is obtained by removing the labels in the noise sample set D N .

[0019] Training the target neural network includes performing virtual adversarial training using the training sample set, wherein the total loss function of the virtual adversarial training includes a regular classification loss function and a virtual adversarial loss function, the regular classification loss function evaluates the deviation between the actual labels and the predicted labels of the annotated sample set, and the virtual adversarial loss function evaluates the difference between the predicted labels of the original training sample set and the predicted labels of the training sample set after adding perturbations.

[0020] In one embodiment, the difference between the predicted labels of the original training sample set and the predicted labels of the training sample set after adding perturbations is evaluated using KL divergence.

[0021] In one embodiment, when the unstable labels of part of the samples in the simple sample set D S are changed to incorrect labels, the proportion of the changed simple samples in the simple sample set is the same as the proportion of the noise in the original sample set.

[0022] In one embodiment, the original sample set D0 is divided into a simple sample set D S with smaller loss and a complex sample set D H / N with larger loss according to the final loss value, including dividing according to a loss threshold, wherein the loss threshold is 20% to 40% of the maximum loss value finally obtained during training.

[0023] In one embodiment, the artificial label noise sample set D a is divided into a simple sample set D a，S with smaller loss and a complex sample set D a，H / N, including: dividing according to a loss threshold, the loss threshold being 20%-40% of the maximum loss value finally obtained in training.

[0024] In one of the embodiments, the auxiliary model M a The long short-term memory neural network is adopted.

[0025] In one of the embodiments, the grid operation data includes voltage data and power angle data, and the instability mode includes power angle instability and voltage instability.

[0026] According to another aspect of the present application, a power system instability state identification method is provided, comprising:

[0027] The grid operation data is input into the power system instability state identification model to obtain the instability state, wherein the power system instability state identification model is obtained by training using the training method of the power system instability state identification model.

[0028] Overall, compared with the prior art, the above technical solutions conceived by the present application can achieve the following beneficial effects:

[0029] The present application firstly trains the model M0 using the original sample set D0, although the original sample set D0 contains noise, but the correct label samples still account for a large proportion, therefore, the model M0 trained has a certain accuracy for the classification of samples, the greater the loss function, the greater the possibility of the input sample being noise. Based on this, the original sample set D0 is preliminarily divided into a simple sample set D s and a complex sample set D H / N . The loss of the simple sample set D s is small, and the probability of containing label noise is small, and the present application defaults that the simple sample has no noise label; the loss of the complex sample set D H / N is large, and the probability of containing error label is large, but it may also contain correct label samples, and the complex sample set D H / N needs to be further divided. Therefore, the present application takes the simple sample set D S as the correct label sample set, randomly selects part of the samples and modifies the labels thereof to error labels, and the simple sample set D S is modified to an artificial label noise sample set D a , in which the noise is known. Then, the model M1 is trained using the artificial label noise sample set D a , and the artificial label noise sample set D a is divided into a simple sample set D a，S and a complex sample set D a，H / N in the same way. At this time, the loss value sequence in the training process of the complex sample set D a，H / N and whether it is noise are known, and then the complex sample set Da，H / N training the auxiliary model M a , using the auxiliary model M a learning the mapping relationship between the loss value sequence of the complex sample set and whether the label contains noise. Finally, using the auxiliary model M a , based on the loss value sequence of the complex sample set D H / N further classification into a noise sample set D N with incorrect labels and a difficult sample set D H with correct labels. Through the above process, the noise in the original sample set can be identified, i.e. the noise sample set D N is noise, the difficult sample set D H and the simple sample set D S are correctly labeled samples, based on which the power system instability state recognition model can be more effectively trained.

[0030] Further, after identifying the noise samples, only the labels of the noise samples are deleted while the input features are retained, and through virtual adversarial training, all sample information is used to obtain a smooth model to counteract the non-smoothness caused by missing label noise, further improving the performance of the dominant instability state recognition model. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 is a step flowchart of a training method of a power system instability state recognition model of an embodiment;

[0032] Figure 2 is a step schematic diagram of noise recognition of an embodiment;

[0033] Figure 3 is a schematic diagram of combining noise recognition and virtual adversarial training of an embodiment;

[0034] Figure 4 is a step schematic diagram of virtual adversarial training of an embodiment;

[0035] Figure 5 is a single-line diagram of an 8-machine 36-node test system of an embodiment;

[0036] Figure 6 is an effect diagram of four methods on different types and intensities of label noise of an embodiment, wherein (a) is an effect diagram of the four methods trained on a non-uniformly distributed noise sample set and (b) is an effect diagram of the four methods trained on a uniformly distributed noise sample set. DETAILED DESCRIPTION

[0037] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and should not be used to limit the present application. In addition, the technical features involved in the various embodiments of the present application described below can be combined with each other as long as they do not conflict with each other.

[0038] As Figure 1 shown is a step flow chart of a training method of a power system instability state recognition model in an embodiment, which mainly includes the following steps:

[0039] Step S100: Noise label recognition is performed on the original sample set D0 carrying instability state labels.

[0040] The original sample includes power grid operation data, and the instability state includes whether instability and specific instability mode.

[0041] Specifically, the original sample is power grid operation data such as voltage and power angle data obtained by setting different power grid operation conditions and faults and performing transient stability simulation. The instability mode can be voltage instability and power angle instability, and the instability state label can be divided into stable, voltage instability and power angle instability. For example, the instability label is specifically 0, 1 and 2, wherein 0 represents stable, 1 represents power angle instability, and 2 represents voltage instability. The added label in the original sample may have noise, which needs to be recognized to obtain training samples available for training.

[0042] In combination with Figure 2 shown, the noise label recognition process includes:

[0043] Step S110: training a model M0 with the power grid operation data of the original sample set D0 as input and the instability state as output, dividing the original sample set D0 into a simple sample set D s with smaller loss and a complex sample set D H / N with larger loss according to the final loss value, and recording the loss value sequence L 0，H / N of each complex sample in the complex sample set D H / N in the entire training process.

[0044] Specifically, the basic dominant instability state recognition model M0 is trained on the original sample set D0 containing noise, and the loss value L0 of each sample at each training round is recorded to form a training loss sequence for each sample. At this time, although the original sample set D0 contains noise, the correct label samples still account for a large proportion, and therefore the model M0 trained has a certain accuracy for sample classification. The larger the loss value, the greater the possibility that the input sample is noise. According to the final loss value L0, D0 is divided into two parts, a simple sample set D S(losses are smaller) and complex sample set D H / N (losses are larger). L0is correspondingly divided into L 0，S and L 0，H / N , where L 0，S is the loss value sequence of each sample in simple sample set D S , and L 0，H / N is the loss value sequence of each sample in complex sample set D H / N .

[0045] In an embodiment, the original sample set D0is divided into two parts, and a small-loss criterion is used to preliminarily screen a conservative (small) loss threshold. The final loss of a sample less than the loss threshold is a simple sample, and otherwise is a complex sample.

[0046] The screened simple sample has a small loss and by default does not contain noise, i.e., by default all samples are correctly labeled.

[0047] The screened complex sample has a large loss and may contain noise, which needs to be further distinguished.

[0048] It can be understood that the setting of the loss threshold in this step can be flexibly selected according to actual conditions. The smaller the loss threshold, the fewer the simple samples obtained, and the smaller the number of artificial label noise sample set D a obtained, and the effect of the auxiliary model M a trained may be poor. The larger the loss threshold, the more simple samples obtained, but the greater the probability of noise. In this embodiment, the loss threshold is 20% to 40% of the maximum loss value finally obtained by training. Preferably, the loss threshold of the preliminary screening is selected to be 30% of the maximum loss value.

[0049] Step S120: The loss state label of part of the samples in simple sample set D s is changed to an error label to form an artificial label noise sample set D a .

[0050] Specifically, a ratio r S of samples in simple sample set D a is randomly selected to add label noise, i.e., part of the correct labels are changed to error labels to form an artificial label noise sample set D a , and whether the label of each simple sample is correct (i.e., whether it is modified) is recorded as a vector η a = {η1, …, η n}, where n is the number of simple samples, and η i indicates whether the label of the i-th simple sample is correct. For example, η i = 0 indicates that the label of the i-th simple sample is correct, and η i= 1 indicates that the label of the i-th simple sample is wrong.

[0051] In an embodiment, the ratio r S of adding label noise in the simple sample set D a is the same as the noise ratio in the original sample, and the noise ratio in the original sample can be obtained by evaluating the noise probability by random sampling.

[0052] Further, the type of adding label noise in the simple sample set D s is also the same as the noise type in the original sample, so that the obtained auxiliary model M a can be better applied to the original sample set, and the noise type of the original sample set can be obtained by cross-validation.

[0053] For example, the label noise can be divided into two categories: non-uniformly distributed label noise and uniformly distributed label noise. The non-uniformly distributed label noise is caused by the similar characteristics of power angle instability and voltage instability interlaced, and the label noise is not uniformly distributed in different categories, which is caused by insufficient cognitive experience of the annotator or the sample itself has misleading. To simulate this kind of noise, a proportion of r a of the simple samples that occur simultaneously with power angle instability and voltage instability are randomly selected to modify their labels, i.e., voltage instability is changed to power angle instability, and power angle instability is changed to voltage instability. The uniformly distributed label noise is a random misjudgment between any different types, and the distribution of the label noise is uniformly distributed in different categories, which is caused by accidental operation errors of the annotator. To simulate this kind of noise, a proportion of r a of the simple samples in each category are randomly selected to modify their labels, i.e., voltage instability is changed to power angle instability, and power angle instability is changed to voltage instability.

[0054] Step S130: training the model M1 with the power grid operation data of the artificial label noise sample set D a as input and the instability state as output, and dividing the artificial label noise sample set D a into a simple sample set D a，S with smaller loss and a complex sample set D a，H / N with larger loss according to the final loss value, recording the loss value sequence L a，H / N and the label type η a，H / N of each complex sample in the complex sample set D a，H / N during the entire training process.

[0055] Among them, the label type η a，H / N records whether the label of the corresponding complex sample is a wrong label.

[0056] Specifically, in the artificial label noise sample set D aThe training model M1 is trained again, and a training loss sequence L is recorded again a , and D is divided into two parts, a simple part D a and a complex part D a，S , η a，H / N and L a of the complex part are denoted as η a and L a，H / N , respectively. a，H / N

[0057] Step S140: training an auxiliary model M a，H / N with the loss value sequence L a，H / N of the complex sample set D a，H / N as input and the label type η a as output.

[0058] The complex sample set D a，H / N screened out through step S130, the loss value of each complex sample and whether it has an incorrect label are known, and a mapping relationship between the loss value sequence and whether the label is incorrect can be established through training of the auxiliary model M a .

[0059] Specifically, the auxiliary model M a adopts a long short-term memory neural network, and the auxiliary model M a can identify whether the complex sample under each loss value sequence is noise.

[0060] Step S150: inputting the loss value sequence L H / N of the complex sample set D 0，H / N into the auxiliary model M a to classify the complex sample set D H / N into a noise sample set D N with incorrect labels and a difficult sample set D H with correct labels.

[0061] After the auxiliary model M a is established through step S140, the auxiliary model M a can identify whether the complex sample under each loss value sequence is noise. Therefore, by using the auxiliary model M a , further identification can be performed on the complex samples divided in step S110, that is, inputting the loss value sequence L H / N of the complex sample set D 0，H / N into the auxiliary model M a , the labels of the corresponding complex samples can be directly identified, so that the complex sample set D H / N is classified into a noise sample set D N with incorrect labels and a difficult sample set D H with correct labels. ​

[0062] Through analysis, the difficult samples generally have greater confusion, and are easily classified as noise only by the size of the single loss function value. Through the above method, noise samples can be more accurately screened from the original samples according to the dynamic change of the loss function in the entire training process, and more non-noise samples are retained.

[0063] Through the above process, the noise sample set is D N , and the non-noise sample set is D S +D H After distinguishing noise and non-noise samples, appropriate strategies can be taken to train the target neural network to obtain the power system instability state recognition model, that is, after step S100, the following steps are performed:

[0064] Step S200: training the target neural network with power grid operation data as input and instability state as output to obtain the power system instability state recognition model.

[0065] In an embodiment, the network architecture of the target neural network, the model M0 and the model M1 is the same, and the same network architecture is used, which can exclude the influence of model architecture and hyperparameters on the loss value sequence.

[0066] In an embodiment, the training sample set used for training the target neural network is obtained after removing the noise sample set D N from the original sample set D0, and the target neural network is directly trained using the training sample.

[0067] In another embodiment, as shown in Figure 3 , the training sample set used for training the target neural network includes a labeled sample set and an unlabeled sample set, wherein the labeled sample set is obtained after removing the noise sample set D N from the original sample set D0, and the unlabeled sample set is obtained by deleting the label in the noise sample set D N At this time, the training sample set can be used for virtual adversarial training.

[0068] Specifically, the simple sample set D S and the difficult sample set D H are combined into a labeled sample set D l , the label of the sample in D N is deleted, and is used as an unlabeled sample set D ul . D l and D ul are used for virtual adversarial training to obtain the final dominant instability state recognition model M.

[0069] When conducting adversarial training, the total loss function for virtual adversarial training is set to include both the regular classification loss function and the virtual adversarial loss function. The regular classification loss function evaluates the deviation between the actual labels and the predicted labels of the labeled sample set, while the virtual adversarial loss function evaluates the difference between the predicted labels of the original training sample set and the predicted labels of the training sample set after perturbation.

[0070] Specifically, such as Figure 4 As shown, let the training sample be x, and the added perturbation be r. vadv Add perturbation r to training sample x vadv Later became a perturbation sample The training sample x and the corresponding perturbation sample Inputting the data into a neural network (deep learning model) yields the predicted value p(y|x,θ) for the training sample x, and the corresponding perturbation sample. Predicted value θ represents the parameters of the current neural network. By setting a virtual adversarial loss function, perturbation samples can be made more efficient. Predicted value The predicted value p(y|x,θ) of the labeled sample x is approximated by a standard classification loss function. The deviation between the actual and predicted labels of the labeled sample set is calculated to make the predicted values ​​of the labeled sample set closer to the actual labels. In this embodiment, only the labels of noisy samples are removed while retaining their input features. Through virtual adversarial training, a smooth model is obtained using all sample information to counteract the non-smoothness caused by missing label noise, further improving the performance of the dominant unstable state recognition model.

[0071] Specifically, virtual adversarial training involves adding a virtual adversarial loss function to the objective function of conventional neural network training. The total loss function is:

[0072] L(D l D ul ,θ)=L clf (D l ,θ)+αL VAT (D l D ul ,θ)

[0073] Among them, L clf (D l Let θ) be the conventional classification loss function for the labeled sample set, and its expression is:

[0074]

[0075] Where, N l To specify the number of samples in the labeled sample set, p(y|x,θ) represents the deviation between the predicted value p(y|x,θ) and the actual value y for a single sample x.

[0076] L VAT (D l , D ul , theta) is a virtual adversarial loss function calculated based on all samples, and the calculation formula is:

[0077]

[0078] Wherein, N l and N ul are the number of samples of the labeled sample set and the unlabeled sample set respectively, The virtual adversarial loss of a single sample x, and the calculation formula is:

[0079]

[0080] Wherein,

[0081] Wherein, D[p, p'] is the KL divergence of the difference between the metric distribution p and p'.

[0082] The above-mentioned weakly supervised learning method of the present application can train an accurate dominant instability state recognition model on a sample set containing label noise, thereby reducing the dependence on expert accurate labeling, reducing the cost of intelligent dominant instability state recognition, and improving the practicability.

[0083] Correspondingly, the present application also relates to a power system instability state recognition method, which inputs the power grid operation data into the power system instability state recognition model trained by the above-mentioned method to obtain the instability state. In an embodiment, the power grid operation data is the power angle data and voltage data obtained based on the simulation of the actual power grid operation condition.

[0084] Next, the above-mentioned method will be specifically described by taking an 8-machine 36-node test system as an example, and the single-line diagram of the system is as Figure 5 In the sample set generation stage, different power flow and fault conditions are set to perform transient stability simulation on the PSASP software, and finally 7800 samples are obtained. The sample set is randomly divided into three parts, namely the training set, the validation set and the test set.

[0085] The effects of four methods under different types and different intensities of label noise are tested respectively, and the four methods are:

[0086] Method one: no label noise learning, that is, conventional deep learning;

[0087] Method two: after identifying and removing noise by the noise identification method in the present application, conventional deep learning is performed, that is, stage one;

[0088] Method three: without identifying noise, directly based on the original sample, using the virtual adversarial learning method, that is, stage two;

[0089] Method four: through the noise identification method in the application, the noise is identified and removed, and then the virtual adversarial learning method is used, that is, stages one and two.

[0090] The results are shown in (a) and (b) of FIG. 6. Figure 6

[0091] It can be seen that method one does not take any measures, and the existence of label noise will greatly reduce the accuracy of identifying the dominant unstable state. The methods (method two and method four) of the application can improve the recognition accuracy under two types of label noise. Moreover, method two, method three and method four are all superior to method one, and method two and method four are generally superior to method three, wherein method four is the best.

[0092] Those skilled in the art can easily understand that the above is only a preferred embodiment of the application and is not intended to limit the application. Any modification, equivalent replacement and improvement made within the spirit and principle of the application shall be included in the protection scope of the application.​

Claims

1. A method for training a power system instability state identification model, characterized in that, The method comprises the following steps: Label noise identification is performed on an original sample set carrying instability state labels; The original sample set comprises power grid operation data, and the instability state comprises whether instability occurs and an instability mode; A target neural network is trained by taking the power grid operation data as input and the instability state as output to obtain a power system instability state identification model; wherein The label noise identification process comprises the following steps: The model is trained by taking the power grid operation data of the original sample set as input and the instability state as output, and the original sample set is divided into a simple sample set with smaller loss and a complex sample set with larger loss according to the final loss value, and the loss value sequence of each complex sample in the complex sample set in the entire training process is recorded; The instability state labels of part of the samples in the simple sample set are changed to error labels to form an artificial label noise sample set; The model M1 is trained by taking the power grid operation data of the artificial label noise sample set as input and the instability state as output, and the artificial label noise sample set is divided into a simple sample set with smaller loss and a complex sample set with larger loss according to the final loss value, and the loss value sequence of each complex sample in the complex sample set in the entire training process and the label type are recorded, and the label type records whether the label of the corresponding complex sample is an error label; An auxiliary model is trained by taking the loss value sequence of the complex sample set as input and the label type as output; The loss value sequence of the complex sample set is input into the auxiliary model to classify the complex sample set into a noise sample set with error labels and a difficult sample set with correct labels. 2.The method of claim 1, wherein, The training sample set used for training the target neural network is obtained by removing the noise sample set in the original sample set, and the target neural network is directly trained by using the training sample. 3.The method of claim 1, wherein, The training sample set used for training the target neural network comprises an annotated sample set and an unannotated sample set, wherein the annotated sample set is obtained by removing the noise sample set in the original sample set and contains simple samples and difficult samples, and the unannotated sample set is obtained by deleting the labels in the noise sample set. The training of the target neural network comprises virtual adversarial training by using the training sample set, and the total loss function of the virtual adversarial training comprises a conventional classification loss function and a virtual adversarial loss function, the conventional classification loss function evaluates the deviation between the actual label and the predicted label of the annotated sample set, and the virtual adversarial loss function evaluates the difference between the predicted label of the original training sample set and the predicted label of the training sample set after adding perturbation. 4.The method of claim 3, wherein, The difference between the predicted label of the original training sample set and the predicted label of the training sample set after adding perturbation is evaluated by using the KL divergence. 5.The method of claim 1, wherein, When the instability labels of part of the samples in the simple sample set are changed to error labels, the proportion of the changed simple samples in the simple sample set is the same as the proportion of the noise in the original sample set. 6.The method of claim 1, wherein, The original sample set is divided into a simple sample set with smaller loss and a complex sample set with larger loss according to the final loss value, which comprises dividing according to a loss threshold, and the loss threshold is 20% to 40% of the maximum loss value finally obtained during training. 7.The method of claim 6, wherein, The artificial label noise sample set is divided into a simple sample set with smaller loss and a complex sample set with larger loss, including: dividing according to a loss threshold, the loss threshold being 20% to 40% of the maximum loss value finally obtained in training. 8.The method of claim 1, wherein, The auxiliary model adopts a long short-term memory neural network. 9.The method of claim 1, wherein, The grid operation data include voltage data and power angle data, and the instability mode includes power angle instability and voltage instability.

10. A method for identifying an unstable state of a power system, characterized by, The method comprises the steps of: The grid operation data are input into the power system instability state identification model to obtain an instability state, wherein the power system instability state identification model is obtained by training the training method of the power system instability state identification model according to any one of claims 1 to 9.

Images