A method and system for power system transient stability assessment

Abstract

The application discloses a power system transient stability evaluation method and system, and belongs to the technical field of power systems. The power system transient stability evaluation method comprises the following steps: iteratively training a first deep neural network according to an initial sample set to obtain an initial evaluation model; determining an importance index according to the weight of the first deep neural network in each iteration training process, the prediction result of sample operation data and the label of the sample operation data; determining a stability parameter based on the initial evaluation model; determining the stability score of sample data according to the importance index and the stability parameter; removing sample data with a stability score greater than a maximum threshold value or less than a minimum threshold value in the initial sample set, and iteratively training a second deep neural network; and determining a transient stability state based on real-time operation data of the power system according to an optimal evaluation model. The application saves training time, improves the generalization ability of the evaluation model and the accuracy of transient stability evaluation.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power systems, in particular to a power system transient stability evaluation method and system based on a deep learning importance index. BACKGROUND

[0002] With the expansion of the scale of the power system and the access of new power electronic devices, the complexity of the power system is continuously improved. From the perspective of system characteristics or operating data characteristics, studying the transient stability of the power system is one of the key problems to ensure the safe and stable operation of the power system.

[0003] In recent years, artificial intelligence algorithms represented by artificial neural networks have realized the transient stability evaluation of complex nonlinear power systems by establishing the mapping relationship between system operating characteristic quantities and transient stability, overcoming the limitations of numerical simulation methods based on simulation calculation and direct methods based on Lyapunov stability analysis for transient stability evaluation of complex systems. However, compared with the direct method based on the construction of Lyapunov function for stability analysis, the current neural network can only represent the nonlinear relationship between operating data and transient stability, the available information is relatively single, lacks the support of power system related theories, and the interpretability is poor. Therefore, the stability judgment performance is easily affected by the quality of the training sample, and the generalization performance is low. At the same time, the construction of the data set required by deep learning training will consume a large amount of computing resources and time, which limits the wide application of the transient stability evaluation algorithm based on deep learning. SUMMARY

[0004] The purpose of the present application is to provide a power system transient stability evaluation method and system, which can improve the accuracy of transient stability evaluation and the generalization performance of the model.

[0005] To achieve the above purpose, the present application provides the following solutions:

[0006] A power system transient stability evaluation method, comprising:

[0007] An initial sample set is obtained; the initial sample set includes a plurality of sample data, and each sample data includes sample operating data after a transient fault occurs in the power system and a label of the sample operating data;

[0008] According to the initial sample set, a first deep neural network is iteratively trained to obtain an initial evaluation model, and the weight of the first deep neural network in each iteration training process and the prediction result of the first deep neural network for each sample operating data in each iteration training process are recorded;

[0009] For each sample data, according to the weight of the first deep neural network in each iteration training process, the prediction result of the first deep neural network in each iteration training process on the sample running data in the sample data, and the label of the sample running data, an importance index of the sample data is determined;

[0010] Based on the initial evaluation model, a stability parameter of the sample data is determined; the stability parameter is a prediction result of the initial evaluation model on the sample data;

[0011] According to the importance index and the stability parameter of the sample data, a stability score of the sample data is determined;

[0012] The sample data with a stability score greater than a maximum threshold value or less than a minimum threshold value in the initial sample set is removed, and an optimal sample set is obtained;

[0013] According to the optimal sample set, a second deep neural network is iteratively trained, and an optimal evaluation model is obtained;

[0014] Based on the optimal evaluation model, a transient stability state of the power system at a corresponding time is determined according to real-time running data of the power system.

[0015] Optionally, the sample running data includes active power, reactive power, voltage amplitude and voltage phase angle after the transient fault of the power system occurs.

[0016] Optionally, the initial sample set is obtained, specifically including:

[0017] Based on the simulation platform, a plurality of sample running data after the transient fault of the power system occurs is simulated, and an initial label corresponding to each sample running data is determined; the initial label is 0 or 1, and the initial label of 0 indicates that the corresponding sample running data is unstable, and the initial label of 1 indicates that the corresponding sample running data is stable;

[0018] The initial label of each sample running data is one-hot encoded to obtain a final label of each sample running data.

[0019] Optionally, the following formula is used to calculate the importance index of the jth sample data:

[0020]

[0021] wherein, SSI j is the importance index of the jth sample data, w t is the weight of the first deep neural network in the tth iteration process, x j is the sample running data in the jth sample data, y j is the label of the sample running data in the jth sample data, and p(wt x j is a prediction result of the first deep neural network on the sample running data x j in the tth iteration process, denotes expectation, and || ||2 denotes two-norm.

[0022] Optionally, the stability score of the sample data is determined according to the importance indicator and the stability parameter of the sample data, and specifically includes:

[0023] The importance indicator of the sample data is normalized to obtain a normalized importance degree.

[0024] The stability score of the sample data is determined according to the normalized importance degree and the stability parameter.

[0025] Optionally, the importance indicator of the jth sample data is normalized by using the following formula:

[0026]

[0027] wherein, SSI_norm j is the normalized importance degree of the jth sample data, SSI j is the importance indicator of the jth sample data.

[0028] Optionally, the stability score of the jth sample data is calculated by using the following formula:

[0029]

[0030] wherein, SS j is the stability score of the jth sample data, a j is the stability parameter of the jth sample data, SSI_norm j is the normalized importance degree of the jth sample data.

[0031] To achieve the above object, the present application further provides the following schemes:

[0032] A power system transient stability evaluation system, comprising:

[0033] An initial sample acquisition unit is configured to acquire an initial sample set, wherein the initial sample set comprises a plurality of sample data, and each sample data comprises sample running data after a transient fault of a power system occurs and a label of the sample running data.

[0034] The first training unit is connected with the initial sample obtaining unit, and is configured to perform iterative training on the first deep neural network according to the initial sample set, to obtain an initial evaluation model, and to record the weight of the first deep neural network in each iteration training process and the prediction result of the first deep neural network on the sample running data in each iteration training process;

[0035] The importance determining unit is connected with the initial sample obtaining unit and the first training unit respectively, and is configured to determine an importance index of sample data according to the weight of the first deep neural network in each iteration training process, the prediction result of the first deep neural network on sample running data in the sample data in each iteration training process, and the label of the sample running data.

[0036] The stable parameter determining unit is connected with the first training unit, and is configured to determine a stable parameter of the sample data based on the initial evaluation model; the stable parameter is the prediction result of the initial evaluation model on the sample data.

[0037] The stable score determining unit is connected with the importance determining unit and the stable parameter determining unit respectively, and is configured to determine a stability score of the sample data according to the importance index and the stable parameter of the sample data.

[0038] The optimal sample determining unit is connected with the initial sample obtaining unit and the stable score determining unit respectively, and is configured to remove sample data with a stability score greater than a maximum threshold value or less than a minimum threshold value from the initial sample set, to obtain an optimal sample set.

[0039] The second training unit is connected with the optimal sample determining unit, and is configured to perform iterative training on the second deep neural network according to the optimal sample set, to obtain an optimal evaluation model.

[0040] The prediction unit is connected with the second training unit, and is configured to determine the transient stability state of the power system at a corresponding moment according to real-time running data of the power system based on the optimal evaluation model.

[0041] According to the specific embodiments of the present application, the following technical effects are disclosed: the first deep neural network is iteratively trained according to the initial sample set, an initial evaluation model is obtained, and the weight of the first deep neural network and the prediction result of the running data of each sample in the iterative training process are recorded, and then the importance index and the stability parameter of each sample data are determined, the stability score corresponding to each sample data is determined according to the importance index and the stability parameter of each sample data, the sample data with a stability score greater than a maximum threshold or less than a minimum threshold in the initial sample set is removed, and an optimal sample set is obtained, the second deep neural network is iteratively trained according to the optimal sample set, and an optimal evaluation model is obtained, the model is trained using the pruned training set, the training time is saved, and the generalization ability of the evaluation model is improved, and based on the optimal evaluation model, the transient stability state of the power system at the corresponding time is determined according to the real-time running data of the power system, and the accuracy of the transient stability evaluation of the power system is improved. BRIEF DESCRIPTION OF DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. 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.

[0043] Figure 1 Flowchart of the power system transient stability evaluation method of the present application;

[0044] Figure 2 Module schematic diagram of the power system transient stability evaluation system of the present application.

[0045] Symbol explanation:

[0046] Initial sample acquisition unit-1, first training unit-2, importance determination unit-3, stability parameter determination unit-4, stability score determination unit-5, optimal sample determination unit-6, second training unit-7, prediction unit-8. DETAILED DESCRIPTION

[0047] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0048] The application aims at the problem that the transient stability degree, as one of the system operation characteristics in the deep learning process of power system transient stability evaluation, is difficult to obtain. From the data perspective, the dynamic influence degree of the sample on the deep learning training and the boundary relationship with the transient stability domain of the power system are considered. Based on the sample importance index, the stability score of the sample data is calculated, the sample data with extremely high or low stability score is eliminated, and then the deep neural network is trained, thereby improving the accuracy of transient stability evaluation and the generalization performance of the evaluation model.

[0049] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below in combination with the drawings and specific embodiments.

[0050] Embodiment one

[0051] As Figure 1 shown, the power system transient stability evaluation method provided by the embodiment includes:

[0052] S1: obtaining an initial sample set The initial sample set includes a plurality of sample data (x i ,y i ), each sample data includes sample operation data x i after the transient fault of the power system occurs and a label y i of the sample operation data x i , y i ∈{0,1} 2 , and N is the total number of sample data.

[0053] Specifically, first, a power system model is constructed based on a simulation platform, a plurality of sets of sample operation data after the transient fault of the power system occurs are simulated, and initial labels corresponding to the sample operation data are determined. The initial label is 0 or 1, the initial label of 0 indicates that the corresponding sample operation data is unstable, and the initial label of 1 indicates that the corresponding sample operation data is stable. In the embodiment, the sample operation data includes active power, reactive power, voltage amplitude and voltage phase angle after the transient fault of the power system occurs.

[0054] Then, the initial labels of the sample operation data are one-hot encoded to obtain the final labels of the sample operation data. That is, for K-class data labels, the original class label n is numbered from 0 to K-1, and after one-hot processing, each label corresponds to a 1xK array, where the K-n bit is 1 and the remaining positions are 0. Generally, the labels of the transient stability data set are divided into stable and unstable two categories, that is, K=2, the initial labels are 1 and 0, and after one-hot processing, the label of the stable sample operation data is {0,1}, and the label of the unstable sample operation data is {1,0}.

[0055] S2: iteratively train the first deep neural network according to the initial sample set to obtain an initial evaluation model, and record the weight value of the first deep neural network in each iteration training process and the prediction result of the first deep neural network on the running data of each sample in each iteration training process.

[0056] Specifically, the initial sample set is used to iteratively train the first deep neural network, and the hyperparameters (including learning rate, iteration number, number of hidden layers, etc.) are continuously adjusted to improve the accuracy of the transient stability judgment, and the neural network parameters are recorded.

[0057] In this embodiment, the initial sample set is input into the first deep neural network, and the output of the first deep neural network is calculated using wherein w represents the weight in the first deep neural network, is a real number set. Based on the output result of the first deep neural network and the label of the sample data, the loss value is calculated using the following formula:

[0058]

[0059] wherein l(p,y) is the loss value, y (k) represents the kth value (0 or 1) in the label of the sample running data, p (k) represents the kth value (0 or 1) in the output result of the first deep neural network.

[0060] The w0, w1, w2,..., w T represent the weight value of each iteration in the stochastic gradient descent process, and T is the maximum iteration number. The weight value w t of the tth iteration is:

[0061]

[0062] wherein η is the learning rate, S is the initial sample set, and (x,y) is the sample data in the initial sample set, is the gradient operator symbol, is the gradient of the loss function with respect to the weight w t-1 , l(p(w t-1 , x) represents the cross-entropy loss function value of the t-1th iteration. Therefore, g t-1 (x,y) is the gradient of the loss function with respect to the weight, and has

[0063] Further based on the back propagation of the error, the weight of each neuron is adjusted until the accuracy or the iteration number reaches the set standard, and the change of the neuron parameters and the network parameters in the iteration process is recorded.

[0064] S3: For each sample data, according to the weight of the first deep neural network in each iteration of the training process, the prediction result of the first deep neural network in each iteration of the training process on the sample data, and the label of the sample data, determine the importance index of the sample data.

[0065] S4: Based on the initial evaluation model, determine the stability parameter of the sample data. The stability parameter is the prediction result of the initial evaluation model on the sample data.

[0066] Specifically, the sample data is input into the initial evaluation model, and the network output result of the sample data is judged. The stability is consistent with the one-hot encoding principle, that is, for the output result (output2, output1) (where output1+output2=1), if output1 is greater than output2, the sample data is in a stable state, otherwise it is in an unstable state. Based on the stability result, the stability parameter a of the sample data is assigned a value, which is 1 if it is stable, otherwise it is 0.

[0067] S5: According to the importance index and the stability parameter of the sample data, determine the stability score of the sample data.

[0068] Specifically, first, the importance index of the sample data is normalized to obtain the normalized importance. Then, according to the normalized importance and the stability parameter, the stability score of the sample data is determined. The higher the stability score, the more stable the system state represented by the sample data.

[0069] In this embodiment, the following formula is used to normalize the importance index of the jth sample data:

[0070]

[0071] where SSI_norm j is the normalized importance of the jth sample data, SSI j is the importance index of the jth sample data.

[0072] The following formula is used to calculate the stability score of the jth sample data:

[0073]

[0074] where SS j is the stability score of the jth sample data, a j is the stability parameter of the jth sample data, SSI_norm j is the normalized importance of the jth sample data.

[0075] S6: Discard sample data with stability score greater than a maximum threshold value or less than a minimum threshold value in the initial sample set to obtain an optimal sample set.

[0076] The sample stability evaluation method based on the importance index of deep learning is adopted, the connection between the sample importance index and the transient stability degree is constructed from the data perspective, the stability score realizes the intuitive representation of the sample stability degree, can become an important index for the sample transient stability feature analysis, further guides the construction of the training set based on the deep learning transient stability evaluation, and improves the generalization ability of the deep learning transient stability evaluation model.

[0077] S7: Iteratively training a second deep neural network according to the optimal sample set to obtain an optimal evaluation model.

[0078] S8: Determining the transient stability state of the power system at a corresponding moment based on the optimal evaluation model and real-time operation data of the power system.

[0079] Further, in step S3, the importance index of each sample data is derived to calculate the importance index of each sample data in the first deep neural network according to the degree of influence of removing any sample data from the initial sample set on the training process of the first deep neural network.

[0080] To simplify the analysis, the discrete training iteration process is approximated as a continuous dynamic process, and therefore, the change of the loss value with the iteration process is represented as:

[0081]

[0082] where Δ t ((x,y),S) is the change of the loss value in the tth iteration process, w t is the weight value of the first deep neural network in the tth iteration process, (x,y) is the sample data, x is the sample operation data, y is the label, l(p(w t , x), y) is the loss value in the tth iteration process.

[0083] Based on the chain rule, the above formula can be further represented as:

[0084]

[0085] Therefore, the influence of removing any sample (x j ,y j ) from the initial sample set S on the loss value at the tth iteration can be obtained:

[0086]

[0087] where S -jTo remove samples (x) from the initial sample set S j ,y j The sample set after (x) is given, where l is the loss value during training using the initial sample set S, and l' is the loss value between the initial sample set and the sample removed from it (x). j ,y j The loss value during subsequent training.

[0088] Since c is a constant independent of the sample data, any sample (x) can be removed from the initial sample set S. j ,y j The impact of ||g|| on the loss value is mainly determined by ||g|| t (x j ,y j )||Decision.

[0089] Will Substituting into the chain rule, we get:

[0090]

[0091] in, and All are gradient operator symbols. Let f be the gradient of the loss function. For function pair w t The gradient value.

[0092] Let the output of the penultimate layer of the first deep neural network be... When the last layer of the first deep neural network is a Softmax function, the output of the first deep neural network is:

[0093]

[0094] Where, p (k) (w,x) represents the output of the first deep neural network, f (k) (w,x) represents the output of the penultimate layer of the first deep neural network.

[0095] Substituting the above equation into... In the middle, we get:

[0096]

[0097] And thus obtain

[0098] when When the approximation condition is met, the degree of influence of removing the j-th sample data on the neural network training process, i.e., the importance index of the j-th sample data, is:

[0099]

[0100] wherein, SSI j is the importance index of the jth sample data, w t is the weight of the first deep neural network in the tth iteration process, x j is the sample operation data in the jth sample data, y j is the label of the sample operation data in the jth sample data, p(w t , x j ) is the prediction result of the first deep neural network on the sample operation data x j in the tth iteration process, represents the expectation, and || ||2 represents the two-norm.

[0101] Generally, the sample stability degree is represented by the difference between the fault limit cutting time and the actual cutting time, but the limit cutting time of a complex system is difficult to obtain. In power system operation, the fault limit cutting time actually represents the boundary of the system stability domain, and the samples with extremely strong and extremely weak stability will be far away from this boundary, and the probability of misjudgment of such data in transient stability assessment is extremely low. This feature can be represented by the sample importance index.

[0102] The transient stability assessment of the power system based on deep learning is essentially a classification problem, that is, to distinguish between stable (white) samples and unstable (black) samples, and the gray samples on the boundary between stability and instability are very important for the training of the model. According to related research experiments, the samples with extremely high or extremely low stability have less influence on the learning of the transient stability boundary of the model in the training samples of the transient stability assessment based on deep learning. The training set generated after removing such samples (data with extremely high and extremely low stability) can improve the generalization ability of the discrimination model to a certain extent and save the training time.

[0103] Therefore, the sample stability degree, as one of the important features, can be used as an important index for distinguishing samples under the same conditions, and is a basic condition for mining the transient operation characteristics of the power system in the deep learning training process. The basic operation data can be used as auxiliary features to input the network.

[0104] Embodiment Two

[0105] In order to perform the method corresponding to the above-mentioned embodiment one, to realize the corresponding functions and technical effects, the following provides a power system transient stability assessment system.

[0106] As Figure 2 shown, the power system transient stability assessment system provided by the embodiment includes: an initial sample acquisition unit 1, a first training unit 2, an importance determination unit 3, a stability parameter determination unit 4, a stability score determination unit 5, an optimal sample determination unit 6, a second training unit 7, and a prediction unit 8.

[0107] The initial sample acquisition unit 1 is configured to acquire an initial sample set. The initial sample set includes a plurality of sample data, and each sample data includes sample operation data after a transient fault occurs in a power system and a label of the sample operation data.

[0108] The first training unit 2 is connected with the initial sample acquisition unit 1. The first training unit 2 is configured to perform iterative training on a first deep neural network according to the initial sample set, to obtain an initial evaluation model, and to record a weight of the first deep neural network in each iteration process and a prediction result of the first deep neural network for each sample operation data in each iteration process.

[0109] The importance determination unit 3 is connected with the initial sample acquisition unit 1 and the first training unit 2 respectively. The importance determination unit 3 is configured to determine an importance index of each sample data according to the weight of the first deep neural network in each iteration process, the prediction result of the first deep neural network for sample operation data in the sample data in each iteration process, and the label of the sample operation data.

[0110] The stability parameter determination unit 4 is connected with the first training unit 2. The stability parameter determination unit 4 is configured to determine a stability parameter of the sample data based on the initial evaluation model. The stability parameter is a prediction result of the initial evaluation model for the sample data.

[0111] The stability score determination unit 5 is connected with the importance determination unit 3 and the stability parameter determination unit 4 respectively. The stability score determination unit 5 is configured to determine a stability score of the sample data according to the importance index and the stability parameter of the sample data.

[0112] The optimal sample determination unit 6 is connected with the initial sample acquisition unit 1 and the stability score determination unit 5 respectively. The optimal sample determination unit 6 is configured to remove sample data with a stability score greater than a maximum threshold value or less than a minimum threshold value from the initial sample set, to obtain an optimal sample set.

[0113] The second training unit 7 is connected with the optimal sample determination unit 6. The second training unit 7 is configured to perform iterative training on a second deep neural network according to the optimal sample set, to obtain an optimal evaluation model.

[0114] The prediction unit 8 is connected with the second training unit 7. The prediction unit 8 is configured to determine a transient stability state of the power system at a corresponding time according to real-time operation data of the power system based on the optimal evaluation model.

[0115] The various embodiments described in this specification are presented for the purpose of illustrating the principles of the present application and its best mode of operation. Each of the embodiments described in this specification has been provided for the purpose of illustration and is not intended to limit the application.

[0116] The principles and implementations of the present application have been described in the above embodiments, which are only used to help understand the method of the present application and its core idea. Meanwhile, for those skilled in the art, the specific implementation and application range of the present application can be changed according to the idea of the present application. In summary, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A method of power system transient stability assessment, characterized by, The power system transient stability evaluation method comprises: An initial sample set is obtained; the initial sample set comprises a plurality of sample data, and each sample data comprises sample operation data after a power system transient fault occurs and a label of the sample operation data; An initial evaluation model is obtained by iteratively training a first deep neural network according to the initial sample set, and the weight of the first deep neural network in each iteration training process and the prediction result of the first deep neural network for each sample operation data in each iteration training process are recorded; For any sample data, the importance index of the sample data is determined according to the weight of the first deep neural network in each iteration training process, the prediction result of the first deep neural network for the sample operation data in the sample data in each iteration training process and the label of the sample operation data; The stability parameter of the sample data is determined based on the initial evaluation model; the stability parameter is the prediction result of the initial evaluation model for the sample data; The stability score of the sample data is determined according to the importance index and the stability parameter of the sample data; Sample data with a stability score greater than a maximum threshold value or less than a minimum threshold value in the initial sample set is removed to obtain an optimal sample set; An optimal evaluation model is obtained by iteratively training a second deep neural network according to the optimal sample set; The transient stability state of the power system at a corresponding time is determined based on the real-time operation data of the power system according to the optimal evaluation model.

2. The power system transient stability assessment method according to claim 1, characterized in that, The sample operation data comprises active power, reactive power, voltage amplitude and voltage phase angle after a power system transient fault occurs.

3. The power system transient stability assessment method according to claim 1, wherein, The initial sample set is obtained in the following manner: A plurality of sets of sample operation data after power system transient faults occur are simulated based on a simulation platform, and initial labels corresponding to the sample operation data are determined; the initial label is 0 or 1, and the initial label 0 indicates that the corresponding sample operation data is unstable, and the initial label 1 indicates that the corresponding sample operation data is stable; The initial labels of the sample operation data are one-hot encoded to obtain final labels of the sample operation data.

4. The power system transient stability assessment method according to claim 1, characterized in that, The importance index of the jth sample data is calculated by the following formula: wherein SSI j is an importance indicator of the jth sample data, w t is a weight of the first deep neural network in the tth iteration process, x j is sample running data in the jth sample data, y j is a label of the sample running data in the jth sample data, p(w t ,x j ) is a prediction result of the first deep neural network on the sample running data x j in the tth iteration process, denotes expectation, and || ||2 denotes two-norm.

5. The power system transient stability assessment method according to claim 1, wherein, The stability score of the sample data is determined according to the importance index and the stability parameter of the sample data in the following manner: The importance index of the sample data is normalized to obtain a normalized importance degree; The stability score of the sample data is determined according to the normalized importance degree and the stability parameter.

6. The power system transient stability assessment method according to claim 5, wherein, The importance index of the jth sample data is normalized by the following formula: wherein SSI_norm j is the normalized importance degree of the jth sample data, SSI j is the importance index of the jth sample data.

7. The power system transient stability assessment method according to claim 5, wherein, The stability score of the jth sample data is calculated by the following formula: wherein SS j is the stability score of the jth sample data, a j is the stability parameter of the jth sample data, SSI_norm j is the normalized importance of the jth sample data.

8. A power system transient stability assessment system applied to the power system transient stability assessment method of any one of claims 1-7, characterized in that, The power system transient stability evaluation system comprises: An initial sample obtaining unit is configured to obtain an initial sample set; the initial sample set comprises a plurality of sample data, and each sample data comprises sample operation data after a power system transient fault occurs and a label of the sample operation data; The first training unit is connected with the initial sample obtaining unit, configured to perform iterative training on the first deep neural network according to the initial sample set, obtain an initial evaluation model, and record the weight of the first deep neural network in each iteration training process and the prediction result of the first deep neural network on the sample running data in each iteration training process; The importance determining unit is connected with the initial sample obtaining unit and the first training unit respectively, configured to determine the importance index of the sample data according to the weight of the first deep neural network in each iteration training process, the prediction result of the first deep neural network on the sample running data in the sample data in each iteration training process and the label of the sample running data; The stable parameter determining unit is connected with the first training unit, configured to determine the stable parameter of the sample data based on the initial evaluation model; the stable parameter is the prediction result of the initial evaluation model on the sample data; The stable score determining unit is connected with the importance determining unit and the stable parameter determining unit respectively, configured to determine the stability score of the sample data according to the importance index and the stable parameter of the sample data; The optimal sample determining unit is connected with the initial sample obtaining unit and the stable score determining unit respectively, configured to remove the sample data with the stability score greater than the maximum threshold value or less than the minimum threshold value in the initial sample set, and obtain an optimal sample set; The second training unit is connected with the optimal sample determining unit, configured to perform iterative training on the second deep neural network according to the optimal sample set, and obtain an optimal evaluation model; The prediction unit is connected with the second training unit, configured to determine the transient stability state of the power system at the corresponding moment based on the real-time running data of the power system according to the optimal evaluation model.

Images