Transient stability prevention and control strategy generation method and system under multiple accident scenes

Abstract

The invention belongs to the technical field of transient stability prevention and control of a power system, and discloses a transient stability prevention and control strategy generation method and system under a multi-accident scene, and the method comprises the steps: obtaining the stability probability of each operation point under N different anticipated accidents based on a transient stability evaluation neural network model; calculating the operation safety degree and the fault importance degree of each round of iteration for operation point search; if the operation safety degree of the input operation point in the current round is smaller than a set value M, the anticipated accident with the high fault importance degree is preferentially considered when the output operation point in the current round is searched, otherwise, alternative operation points are searched in the safety domain of the instability accident, and the operation point with the maximum operation safety degree value serves as the output operation point; and circulating the operation point search until the output operation point recovers transient stability or the operation safety degree is reduced, and taking the output operation point as a final control strategy. By coordinating the safety operation domains of different accidents, the problems of difficult solution and low efficiency of prevention and control optimization in a multi-accident scene are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of transient stability preventive control of power systems, and more specifically, relates to a method and system for generating a transient stability preventive control strategy under multiple accident scenarios. Background Art

[0002] Transient stability problems are one of the common causes of large power grid safety accidents. Rapid and effective transient stability prevention and control are of great significance to ensuring the safe and stable operation of power systems online. The actual power system is not only flexible and changeable in operation points, but also has the possibility of failures at any part. When conducting online safety analysis, it is necessary to conduct transient stability assessment and preventive control for some of the most risky anticipated accidents. In this multi-accident scenario, the safe operation domains corresponding to different accidents are different, and the transient stability constraints they represent are also different. Using traditional mathematical optimization methods for preventive control requires taking into account many transient stability constraints. Not only is it difficult to balance accuracy and simplicity in the modeling of transient stability constraints, but it is also difficult to satisfy multiple stability constraints at the same time, making the solution efficiency low or even impossible to solve. At this point, traditional methods can no longer guarantee the timeliness of online safety analysis.

[0003] Therefore, the existing technology proposes to use artificial intelligence technology to generate preventive control strategies. However, there are many kinds of accidents that may occur during the online operation of the actual power system, and it is difficult to cover the unknown accidents that may occur in the entire network with a few expected accident scenarios. The existing preventive control strategy generation methods are often targeted at a single expected accident or a few expected accident scenarios. Not only that, they also fail to take into account the mutual exclusion of different accident safety operation domains that may occur in the power system, making it impossible to extend it to the actual power system online analysis application. Summary of the invention

[0004] In view of the above defects or improvement needs of the prior art, the present invention provides a method and system for generating transient stability preventive control strategies under multiple accident scenarios, which aims to solve the problems of difficulty and low efficiency in optimizing and solving preventive control strategies under multiple anticipated accident scenarios.

[0005] To achieve the above object, according to a first aspect of the present invention, a method for generating a transient stability prevention control strategy in a multi-accident scenario is provided, comprising:

[0006] The steady-state power flow information reflecting the status of each operating point of the power grid is input into the trained transient stability assessment neural network model to obtain the stability probability of each operating point under N different expected accidents;

[0007] For the current operating point, the operating safety and fault importance of the operating point are calculated based on the stability probability; wherein the operating safety is the total number of expected accidents with a stability probability greater than a preset safety and stability threshold, and the fault importance is used to characterize the importance of different expected accidents to the operating point;

[0008] When the operating safety Q of the current operating point is less than the set value M, the expected accidents ranked first S in fault importance are taken to obtain the characteristic importance vector of the active output of the power grid generator under the S expected accidents at the current operating point; based on the characteristic importance vector, the active power of the generator with the smallest characteristic importance is transferred to the generator with the largest characteristic importance, and S candidate operating points are obtained; among the S candidate operating points, the one that increases the transient stability probability of all expected accidents the most is determined as the current new operating point; repeat this step until Q ≥ M, and the updated current operating point is obtained;

[0009] For the remaining D instability accidents at the current operating point, D=NQ, the current operating point is adjusted to the safety domain of each instability accident respectively, and among the D candidate operating points obtained, the one with the largest operating safety is updated as the current operating point; repeat this step until the operating safety of the current operating point is N, or the operating safety of the current operating point is reduced, and the current operating point is used as the power grid control strategy.

[0010] Furthermore, the fault importance is:

[0011]

[0012] In the formula, I i is the fault importance of instability anticipated accident i among N different anticipated accidents, y i is the stability probability of a given operating point under the anticipated accident i; G is the set of all anticipated instability accidents; p i is the probability of transient instability of the power system after the expected accident i occurs; α is the weight coefficient.

[0013] Furthermore, the feature importance vector is calculated as follows:

[0014] LIME is used to construct a local linear proxy model K(x) to locally interpret the transient stability assessment neural network model to obtain the feature importance vector; wherein K(x) satisfies:

[0015] K(x)=w0+w1x1+…+w n x n

[0016] Where K(x) is the local linear proxy model of the transient stability assessment neural network model, x represents the input of the transient stability assessment neural network model, which is a vector containing the steady-state power flow information reflecting the current operating point state and the current expected accident information, and x i is the active output characteristic of the i-th generator in the model, i = 1, 2, ..., n, n is the total number of generators in the power grid; w i is the feature importance corresponding to the active output feature of the i-th generator, w1, w2, …, w n It constitutes the characteristic importance vector of the active output of the power grid generator at the current operating point under the current expected accident.

[0017] Furthermore, for the remaining D instability accidents at the current operating point, D=NQ, the current operating point is adjusted to the safety region of each instability accident, including:

[0018] Under a given instability accident, based on the characteristic importance vector of the current operating point, the active power of the generator with the smallest characteristic importance is transferred as a fixed power P to the generator with the largest characteristic importance to adjust the current operating point. This process is executed repeatedly until the adjusted current operating point is within the safety domain of the given instability accident, and D candidate operating points are obtained.

[0019] Furthermore, the transient stability assessment neural network model is obtained by using knowledge distillation training, which specifically includes:

[0020] A transient stability assessment neural network model actually used in a power grid is used as a teacher model, and the knowledge of the teacher model is transferred to a student model through knowledge distillation; wherein the student model is the transient stability assessment neural network model.

[0021] Furthermore, during the knowledge distillation training process, each training sample input into the student model and the teacher model includes steady-state power flow information reflecting the current operating point of the power grid and the current expected accident information, and the label is the corresponding transient stability probability; the loss function used is:

[0022] Loss = λL hard +(1-λ)T 2 L soft

[0023] Where, L hard is the classification loss, L soft is the distillation loss, T is the temperature coefficient of the softmax activation function of the teacher model, and T>1, λ is the weight coefficient;

[0024] Classification loss L hardis the cross entropy loss between the sample label prediction value output by the student model and the sample label, wherein the temperature coefficient of the activation function of the student model is 1;

[0025] Distillation loss L soft is the KL divergence between the soft label prediction value and the soft label, wherein the soft label prediction value is the output of the student model when the temperature coefficient of the softmax activation function of the student model is set to T, and the soft label is the output of the teacher model.

[0026] Furthermore, S is calculated as:

[0027] S = [(N-M+1) / 2]

[0028] Where [·] is the rounding function.

[0029] According to a second aspect of the present invention, there is provided a system for generating transient stability prevention control strategies under multiple accident scenarios, comprising a computer-readable storage medium and a processor;

[0030] The computer-readable storage medium is used to store executable instructions;

[0031] The processor is used to read the executable instructions stored in the computer-readable storage medium to execute the method for generating a transient stability prevention control strategy under multiple accident scenarios as described in any one of the first aspects.

[0032] According to a third aspect of the present invention, there is provided a computer-readable storage medium having a computer program stored thereon, wherein when the program is executed by a processor, the method for generating a transient stability prevention control strategy under multiple accident scenarios as described in any one of the first aspects is implemented.

[0033] According to a fourth aspect of the present invention, there is provided a computer program product, comprising a computer program, which, when executed on a computer, enables the computer to execute the method for generating a transient stability prevention control strategy under multiple accident scenarios as described in any one of the first aspects.

[0034] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects:

[0035] (1) The method for generating a transient stability prevention control strategy under multiple accident scenarios of the present invention combines an artificial intelligence transient stability assessment model, designs an operation point search strategy based on the operation safety and fault importance of a given operation point. Specifically, in the process of preventing and controlling the strategy search, when the operation safety of the current operation point is less than the set value M, the operation point search is performed with priority given to the expected accidents with high fault importance when searching the output operation point of this round (the current new operation point), which can ensure that the operation point is safe and stable under most accidents with high fault importance, and improve the search efficiency. After the operation safety Q of the current operation point is greater than or equal to the set value M, the alternative operation point is searched in the safety domain of the instability accident. The present invention selects a suitable operation point adjustment direction based on the operation safety and fault importance of the operation point, thereby gradually adjusting the operation point to the intersection of the transient safety domains of multiple expected accidents, and is suitable for preventing and controlling the strategy generation under multiple expected accident scenarios; and, compared with the existing solution method based on the optimization algorithm, the control strategy generation method of the present invention consumes less time and resources, improves the efficiency of control strategy generation, and is more adaptable to the online operation of the power grid.

[0036] (2) Furthermore, the fault importance provided by the present invention describes the severity of a specific accident occurring at a given operating point under the occurrence of a specific accident. The greater the fault importance of the specific accident, the higher the risk of transient instability of the power grid caused by the accident occurring at the current operating point. Conversely, the smaller the fault importance of the specific accident, the lower the risk of transient instability of the power grid caused by the accident occurring at the current operating point.

[0037] (3) As a preferred method, compared with the evaluation model obtained by directly using the training data to train the neural network classifier using the data set, the transient stability evaluation neural network model with high classification accuracy actually used in the power grid is used as the teacher model, and the knowledge of the teacher model is transferred to the lightweight student model through "distillation" by using knowledge distillation. The student model is the required transient stability evaluation neural network model, which can improve the classification accuracy of the transient stability evaluation neural network model.

[0038] (4) Furthermore, in the process of knowledge distillation training, the loss functions used include distillation loss and classification loss. When designing the distillation loss, the temperature coefficient of the activation function T>1, which can reduce the difference between the predicted positive and negative labels and further improve the classification accuracy of the student model obtained after knowledge distillation.

[0039] The operating point adjustment strategy of the present invention fully exploits the information provided by the transient stability assessment model of multiple anticipated accidents. Based on the priorities of different faults and operating points, it greatly ensures that the obtained operating point exists in the transient safe operation domain of different anticipated accidents, thereby realizing a fast and accurate search for transient stability prevention and control strategies for multiple accidents. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 A schematic diagram of a method for generating a transient stability prevention control strategy in a multi-accident scenario provided by an embodiment of the present invention;

[0041] Figure 2 A schematic diagram of a transient stability assessment model training process based on knowledge distillation provided in an embodiment of the present invention;

[0042] Figure 3 Schematic diagram of the TabNet classifier structure provided in an embodiment of the present invention, (a) is a schematic diagram of the overall structure of the classifier, (b) is a schematic diagram of the structure of the Feature transformer module, and (c) is a schematic diagram of the structure of the Attentive transformer module;

[0043] Figure 4 A schematic diagram of a multi-anticipated accident prevention control strategy search process provided by an embodiment of the present invention;

[0044] Figure 5 A schematic diagram of the change in active output of a generator before and after preventive control is performed at a given operating point by the method in an embodiment of the present invention;

[0045] Figure 6 It is a schematic diagram of a time domain simulation curve of a generator rotor angle variation before and after preventive control is performed at a given operating point by the method in an embodiment of the present invention. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0047] Example 1

[0048] like Figure 1 As shown, an embodiment of the present invention provides a method for generating a transient stability prevention control strategy in a multi-accident scenario, which mainly includes:

[0049] S1. Input the steady-state power flow information reflecting the state of each operating point of the power grid into the trained transient stability assessment neural network model to obtain the stability probability of each operating point of the power grid under N different expected accidents. In the embodiment of the present invention, the operating point with a stability probability of 0.5 under a specific expected accident constitutes a transient safety boundary.

[0050] S2. For the current operating point, the operating safety and fault importance of the operating point are calculated based on the stability probability; wherein the total number of expected accidents with a stability probability greater than a preset safety and stability threshold is counted as the operating safety Q of the operating point, and the fault importance is used to characterize the importance of different expected accidents to the operating point. In the embodiment of the present invention, the preset safety and stability threshold is 0.5.

[0051] S3. When the operation safety degree Q of the current operation point is less than the set value M, take the expected accidents ranked first S in fault importance, and obtain the characteristic importance vector of the active output of the power grid generator at the current operation point under these S expected accidents; wherein each element in the characteristic importance vector represents the characteristic importance of the active output of each generator at the current operation point under the current accident; based on the characteristic importance vector, transfer the active power P of the generator with the smallest characteristic importance to the generator with the largest characteristic importance, and obtain S candidate operation points; among the S candidate operation points, the one that increases the transient stability probability of all expected accidents the most is determined as the current new operation point; repeat this step until the operation safety degree Q of the current operation point is greater than or equal to the set value M, and obtain the current operation point with an updated operation safety degree of Q;

[0052] S4. For the remaining D instability accidents (D=NQ) at the current operating point, adjust the current operating point to the safety domain of each instability accident to obtain D candidate operating points; among the obtained D candidate operating points, the one with the largest operating safety is used as the new current operating point; repeat this step until the operating safety of the current operating point is N, or the operating safety of the current operating point is reduced, and the current operating point is used as the final control strategy to make the power grid operate at the current operating point.

[0053] As a preferred implementation, in S1, knowledge distillation is used to obtain a trained transient stability assessment neural network model to improve the classification accuracy of the transient stability assessment neural network model. Knowledge distillation transfers the knowledge of the trained teacher model to a lightweight student model through "distillation". The student model is the required transient stability assessment neural network model.

[0054] In order to evaluate the transient stability of the power grid under different anticipated accidents, the input of the transient stability evaluation neural network model (evaluation model) integrates the anticipated accident information reflecting different anticipated accidents and the steady-state flow information reflecting the state of the power grid operating point (characterizing the power grid operating point). That is, in the present invention, different anticipated accidents are one-hot encoded as the anticipated accident information of the evaluation model in the present invention. The steady-state flow information of the evaluation model in the present invention only needs to include the state quantity of the steady-state flow, and the output of the evaluation model is the stability probability of a given operating point under each given anticipated accident.

[0055] The training data used to train the evaluation model comes from the actual power system operation data (including steady-state power flow information reflecting the status of the power grid operation point) or the simulated power system operation data obtained through time domain simulation according to the power grid operation parameters; the power system operation data and the given expected accident information constitute a piece of data, each piece of data corresponds to an operation point of the actual power system, and each piece of data under different expected accident scenarios is simulated in the time domain to obtain a transient stability label (if the system is stable after the fault, it is marked as 1, otherwise it is marked as 0), that is, a sample of the training data contains the state information of the steady-state power flow of an operation point and specific expected accident information, and the label is the corresponding transient stability probability.

[0056] Transient stability refers to the situation in which the power grid can transition from the original operating state to a new operating state without losing synchronization and operate stably in the new operating state when it is subjected to a large disturbance. In the online safety analysis of power systems, an expected accident set is often set based on historical operating data, and the occurrence of any accident in the expected accident set is used to simulate large disturbances.

[0057] Figure 2 The training process of transient stability assessment model based on knowledge distillation is demonstrated. Knowledge distillation transfers the knowledge of the trained teacher model to the lightweight student model through "distillation". The student model is the required transient stability assessment neural network model. The teacher model can be a transient stability assessment neural network model with high classification accuracy that is actually used in the power grid. Compared with the evaluation model obtained by directly training the neural network classifier using the training data set, the teacher model has higher classification accuracy. In the embodiment of the present invention, the teacher model and the student model use the same neural network architecture. The difference is that the teacher model input not only includes the expected accident information and the steady-state flow state quantity information, but also includes the steady-state flow non-state quantity information, which has higher classification accuracy. However, the complex and coupled input features of the teacher model are difficult to apply to preventive control search. In this case, the teacher model has more input information to provide more accurate transient stability assessment results, while the input features of the student model are independent of each other and can evaluate the transient stability of the adjusted operating point.

[0058] Specifically, the steady-state power flow information containing expected accident information and reflecting the state of the power grid operation point is input into the teacher model, the temperature coefficient T of the softmax function of the output layer (fully connected layer) of the teacher model is set to be greater than 1, and the probability output by the teacher model is used as the soft label. The steady-state power flow information containing expected accident information and reflecting the state of the power grid operation point is input into the student model, the temperature coefficient T of the softmax function of the output layer of the student model is set to be greater than 1, and the probability output by the student model is used as the soft label prediction value. The steady-state power flow information containing expected accident information and reflecting the state of the power grid operation point is input into the student model, the temperature coefficient T of the softmax function of the output layer of the student model is set to be equal to 1, and the student model outputs the corresponding stable probability (sample label prediction value).

[0059] During the student model training process, the cross entropy loss between the sample label prediction value output by the student model and the sample label is used as the classification loss, and the KL divergence between the soft label prediction value output by the student model and the soft label output by the teacher model is used as the distillation loss. The student model is trained to obtain the trained student model, which is used as the required transient stability evaluation neural network model.

[0060] Specifically, the loss function when training the student model is divided into distillation loss and classification loss, and the calculation is obtained by the following formula:

[0061]

[0062] Loss = λL hard +(1-λ)T 2 L soft

[0063] Where z is the output of the teacher model or the student model fully connected layer when T>1, T(T>1) is the temperature coefficient of the softmax function (the activation function of the fully connected layer), and the softmax function is used to convert the features into probability outputs by nonlinear mapping; q' is the soft label output by the teacher model (T>1) or the predicted value of the soft label output by the student model (T>1); when T>1, the difference between the predicted positive and negative labels can be reduced; λ is the weight coefficient, which takes values ​​between 0 and 1; L hard is the classification loss of the student model, calculated by the cross entropy of the sample label obtained by time domain simulation and the sample label prediction value (the temperature coefficient of the student model is 1), L soft The distillation loss is calculated as the KL divergence between the sample soft label (the output of the teacher model when T>1) and the soft label prediction value (the output of the student model when T>1).

[0064] In other embodiments, the data set may be directly used to train a neural network classifier to obtain the desired transient stability assessment neural network model.

[0065] As a preferred implementation, the present invention uses the existing neural network TabNet as the classification architecture of the teacher model and the student model, which shows strong performance in the classification of tabular data and is suitable for steady-state flow data. Figure 3 shown.

[0066] As a preferred implementation method, in S2, for a given operating point, different anticipated accidents have different stability probabilities in the aforementioned transient stability assessment model. In this embodiment of the present invention, the total number of anticipated accidents with a stability probability greater than a safety stability threshold of 0.5 is counted as the operating safety value of the operating point.

[0067] For a given operating point, the risk and severity of different anticipated accidents are different, that is, different anticipated accidents have different importance to the operating point, which is calculated by the following formula:

[0068]

[0069] In the formula, I i is the fault importance of instability anticipated accident i in the anticipated accident set (including N different anticipated accidents), y i is the stability probability of a given operating point under an anticipated accident i, that is, the output of the evaluation model under a given operating point and an anticipated accident i; G is the set of all unstable anticipated accidents, which is a subset of the anticipated accident set; p i In order to predict the probability of transient instability of the power system after the occurrence of accident i, the upper and lower limits of the operation space are set to make the operation point fluctuate randomly between the upper and lower limits, and then the p is obtained through time domain simulation. i ; α is the weight coefficient.

[0070] Operational safety measures the risk of transient instability at a given operating point when an accident in the expected accident concentration occurs. The greater the operational safety of the operating point, the fewer expected accidents that cause instability at the operating point, and the safer the operating point; conversely, the smaller the operational safety of the operating point, the more expected accidents that cause instability at the operating point, and the less safe the operating point. Fault importance describes the severity of a specific accident at a given operating point under a specific accident. The greater the fault importance of a specific accident, the higher the risk of transient instability of the power grid caused by the accident at the current operating point, and the more attention should be paid to the accident; conversely, the smaller the fault importance of a specific accident, the lower the risk of transient instability of the power grid caused by the accident at the current operating point. In actual operation, the value of the weight coefficient α determines whether the fault importance pays more attention to the sample as a whole or a single sample. The closer α is to 0, the more the fault importance calculation depends on the transient safety risk of the operating space of the operating point under a given accident; if α is closer to 1, the fault importance calculation depends more on the transient safety risk of a given operating point under a given accident by the evaluation model. More often, when both factors need to be considered, α can be taken as 0.5.

[0071] Specifically, in S3 and S4, the operation point of transient stability of the power grid is searched iteratively. If the operation safety of the input operation point of this round (that is, the current operation point mentioned above) is less than the set value M, the expected accidents with high fault importance are given priority when searching for the output operation point of this round (the current new operation point). If the total number of expected accidents is N, the expected accidents with the top S fault importance ranking are taken to obtain the characteristic importance vectors of the active output of the power grid generator under these S expected accidents, and a total of S characteristic importance vectors.

[0072] For each group of feature importance vectors, according to the feature importance corresponding to each generator in the feature importance vector, the active power of the grid generator is transferred from the generator with the smallest feature importance to the generator with the largest feature importance. The transferred power is set to a fixed value P, and a candidate operating point corresponding to the group of feature importance vectors is obtained, and a total of S candidate operating points are obtained.

[0073] Among the S candidate operating points, the one that maximizes the transient stability probability of all anticipated accidents is determined as the new round of operating points, that is, the output operating point of this round (the current new operating point); among them, among the S candidate operating points, the transient stability probability of all anticipated accidents refers to the sum of the stability probabilities of each candidate operating point under N different anticipated accidents, and the candidate operating point corresponding to the maximum sum of stability probabilities is taken as the current new operating point.

[0074] Based on the current new operating point, the above process is repeated until the operating safety degree of the current operating point is greater than or equal to the set value M, and the updated current operating point is obtained.

[0075] After the current operating point is given, that is, the input of the evaluation model is given, the evaluation model performs transient stability evaluation. At this time, different input features have different importance to the output evaluation results, that is, the feature importance is different. In the embodiment of the present invention, an artificial intelligence interpretability method such as LIME is used to locally interpret the transient evaluation model to obtain a feature importance vector. The expression of LIME to obtain feature importance is as follows:

[0076] F(x)≈K(x)=w0+w1x1+…+w n x n

[0077] Where F is the transient stability assessment neural network model, K is the local linear proxy model of F, x is the input vector containing the expected accident information and the steady-state power flow information state quantity (characterizing the current operating point), and x i (i=1,2,…,n) is the active output characteristic of the generator (the i-th element of the corresponding generator active output in the input vector x), n is the total number of active output characteristics of the generator in the input vector, which is consistent with the number of generators in the power grid, and w i (i=1,2,...,n) is the corresponding feature weight, that is, the feature importance corresponding to the active output feature of the i-th generator, w1, w2,..., w n It constitutes the characteristic importance vector of the active output of the power grid generator at the current operating point under the current expected accident.

[0078] As a preferred implementation, S is calculated as follows:

[0079] S = [(N-M+1) / 2]

[0080] Where [·] is the rounding function.

[0081] In the embodiment of the present invention, N=6 anticipated accidents are taken as an example to illustrate the operation point adjustment process of giving priority to anticipated accidents with high fault importance. If the initial operation point is x0, M is set to 2, and the operation mode is changed to x1 after one round of search, the anticipated accidents that still cause system instability are F1, F2, and F3, that is, the operation safety is 3, then the second round of search process is as follows:

[0082] Calculate the fault importance of the instability anticipated accidents F1, F2, and F3 under the operation mode x1 and sort them from large to small (assuming the fault importance is: F1>F2>F3); obtain the characteristic importance vector of the generator set under the first S=2 anticipated accidents (F1, F2) of x1 respectively; determine the additional and reduced generating units according to the characteristic importance and adjust them to obtain 2 sets of candidate operating points Using transient stability assessment model Perform a stability assessment and calculate the sum of the stability probabilities under N contingency cases; compare and select the operating point with the largest sum of stability probabilities as the current new operating point x2.

[0083] Conduct the next round of search based on the current new operating point x2. That is, at the current new operating point x2, calculate the corresponding operating safety degree. When the operating safety degree at the current operating point is less than the set value M, give priority to considering the contingency cases with high fault importance when searching for the output operating point of this round (the current new operating point) until the operating safety degree of the current new operating point is greater than or equal to the set value M.

[0084] At this time, search for alternative operating points within the safety domain of the instability accident, and select the one with the largest operating safety degree value as the output operating point. Specifically, under a given instability accident, cyclically adjust the operating point according to the characteristic importance vector of the active power output of the generators at the operating point (transfer the active power from the generator with the smallest characteristic importance to the generator with the largest characteristic importance). The transferred active power for each adjustment is P, so that the operating point is adjusted into the safety domain of the given accident. Adjust the operating point into the safety domain of each of the D (D < M) instability accidents to obtain D candidate operating points; among the D candidate operating points, the one with the largest operating safety degree is determined as the new round of operating point.

[0085] In the embodiment of the present invention, take N = 6 contingency cases as an example to illustrate the adjustment process. If M = 2 and the operating safety degree of the operating point x k after k rounds of search is 4, that is, there are still 2 contingency cases that cause the operating point x k to become unstable, denoted as F1 and F2, then the (k + 1)-th round of search process is as follows:

[0086] For the instability accidents Fj with j = 1, 2 j , obtain the characteristic importance vector of the generator set at the operating point x k , determine the generator sets with increased power output and decreased power output according to the characteristic importance and conduct adjustments, and adjust x k into the corresponding safety domain of F j to obtain 2 candidate operating modes Calculate the operating safety degree of k+1 (the current new operating point) and select the one with the largest operating safety degree as the new round of operating mode x

[0087] Based on the current new operating point x k+1 , calculate the current new operating point x k+1When the operation safety degree of the current operating point does not meet the search termination condition that the operation safety degree of the current operating point is N, or the operation safety degree of the current operating point is reduced, the above process is repeated, that is, for the remaining D instability accidents at the current operating point, the current operating point is adjusted to the safety domain of each instability accident to obtain D candidate operating points. Among the obtained D candidate operating points, the one with the largest operation safety degree is used as the new current operating point until the search termination condition is met, and the current operating point is output as the final control strategy to make the power grid operate at the current operating point.

[0088] During the operation point adjustment process, the operation safety degree decreases, which means that there are mutual exclusions in different accident safety domains. Continuing the adjustment will increase the number of instability accidents. At this time, the operation point with the largest operation safety degree should be used as the final control strategy to prompt the dispatcher to formulate further emergency control measures for accidents that still have the risk of instability; if the operation point can maintain transient stability under all expected accidents, it means that the preventive control search target is achieved, and the operation point is output as the final control strategy. The entire preventive control strategy search process is as follows: Figure 4 shown.

[0089] In order to further illustrate the method for generating transient stability prevention control strategies under multiple accident scenarios provided by an embodiment of the present invention, in a specific embodiment, the New England 10-machine 39-node system is selected as the test system, and a total of 33 anticipated accidents are selected to simulate different operation scenarios to generate original operation data. These anticipated accidents are all three-phase short-circuit faults occurring at the end of the line, and are cleared 0.25s after the occurrence of the fault. The specific accident lines are shown in Table 1. Under 33 anticipated accidents, the simulation randomly generated a total of 118,826 valid samples. Subsequently, the sample set was randomly divided into a training set, a validation set, and a test set at a ratio of 0.7:0.15:0.15.

[0090] Table 1 Predicted accident routes

[0091]

[0092] The multi-layer perceptron MLP, stacked autoencoder SAE and TabNet are used to train the transient stability assessment model of multiple expected accidents respectively. The training method is direct training without using knowledge distillation technology. At the same time, the transient stability assessment neural network model proposed in the embodiment of the present invention is obtained by training TabNet using knowledge distillation. The distillation parameters during training are T=3, λ=0.5. The performance results of each evaluation model are shown in Table 2.

[0093] Table 2 Performance of different transient stability assessment models

[0094]

[0095] As can be seen from Table 2, the multi-contingency transient stability assessment model obtained by training TabNet using knowledge distillation technology proposed in the present invention has obvious advantages in the accurate prediction of transient stability compared with other models, and the comprehensive index G-mean is improved by 1% compared with other models.

[0096] An unstable operation point A is randomly selected from the test set as the original operation condition for transient stability preventive control of multiple expected accidents. After the evaluation model prediction, the operation point A experienced transient instability when 28 expected accidents occurred. The operation mode obtained after the preventive control adjustment using the proposed method can maintain safe and stable operation under all expected accidents, and the total calculation time is 4.73s. The stability probability of the operation mode under each expected accident is shown in Table 3. The comparison of the active output of the generator set before and after the adjustment is shown in Figure 5 , Figure 5 The active output is mapped to the upper and lower limits of the output, that is, the lower limit of the active output of the generator corresponds to 0, and the upper limit corresponds to 1. The generator sets that adjust the active output are G33, G35, G38 and G39. Except for G39, which is a positive adjustment unit, the rest are negative adjustment units. At the same time, the expected accident 33 is selected for time domain simulation before and after the adjustment to obtain the generator rotor angle change curve, such as Figure 6 As shown. It can be seen from the curve that the evaluation model captures the instability state of the unit well. After the generator output is adjusted in a targeted manner through the strategy proposed by the present invention, the system has restored transient safety. After adjustment, the active output of each generator has not reached its power limit, and there is still a certain adjustment space to adapt to other emergencies.

[0097] Table 3 Stability probability of operating point A under various anticipated accidents before and after preventive control

[0098]

[0099] 500 groups of instability samples were randomly selected from the test set, and single-expected accident prevention control and multi-expected accident prevention control were carried out according to this scheme. The average time consumption was calculated, and the results of the methods proposed in other literature on the same test system were compared, as shown in Table 4. At the same time, the traditional transient stability prevention control generation method is through multiple alternating iterations of time domain simulation and optimal power flow calculation, and the time consumption is usually several times the sum of the time consumption of time domain simulation and optimal power flow calculation, which is particularly obvious in the multi-expected accident scenario. Taking the New England 10-machine 39-node system as an example, the single-accident time domain simulation time consumption is 9 seconds, and the optimal power flow time consumption is 2.7 seconds. Taking operating point A as an example, preventive control is carried out under 33 kinds of expected accidents. Even if only one round of time domain simulation and optimal power flow solution is completed, the time consumption is more than 5 minutes, and there is also the risk of non-convergence of solution. Therefore, compared with the traditional transient stability prevention control, the present invention greatly improves the speed of multi-expected accident prevention control and the feasibility of online application.

[0100] Table 4 Time consumption of different preventive control methods

[0101]

[0102] In the embodiment of the present invention, based on the knowledge distillation technology, the transient stability assessment knowledge in the teacher model is transferred to the student model to assist the training of the transient stability assessment model for multiple expected accidents, fully exploit the information provided by the transient stability assessment model for multiple expected accidents, and improve the accuracy of the stability assessment. Fault importance and operation safety indicators are proposed for the evaluation of operating points under different expected accidents. In the process of searching for unstable accident prevention and control problems, when the operation safety Q of the current operating point is less than the set value M, the expected accidents with high fault importance are given priority for operating point search, and the search for operating points is decomposed into multiple single expected accident prevention and control problems, and the obtained operating points are selected preferentially. At the same time, for the single expected accident prevention and control problem, the evaluation model is used to identify the characteristics and equipment of the prevention and control adjustment based on the feature importance of different features, and targeted adjustments are performed. The method proposed in the present invention can effectively serve the online operation of the power grid and the prevention and control scenario of multiple expected accidents, showing high solution efficiency and method feasibility.

[0103] The more anticipated accidents considered during stability assessment and preventive control, the wider the coverage, the more capable of identifying and resolving potential risks of the power grid. The present invention does not limit the number of anticipated accidents. Experimental simulations have also proved that the method of the present invention can be applied to the generation of transient stability preventive control strategies in multiple accident scenarios (covering at least 33 anticipated accidents). Compared with the existing preventive control strategy generation methods that only consider one or several anticipated accidents, it has obvious advantages and can achieve accurate and rapid strategy generation, which is suitable for online operation of power grids.

[0104] The present invention fully exploits the intrinsic information of the artificial intelligence transient stability assessment model, accurately divides the transient stability boundaries of different anticipated accidents and determines the characteristic importance of the input, and quickly adjusts the operating point to the intersection of multiple accident safety domains according to the two indicators of fault importance and operation safety, thereby solving the problems of difficulty and low efficiency in optimizing the preventive control strategy under multiple anticipated accident scenarios.

[0105] Example 2

[0106] An embodiment of the present invention provides a system for generating a transient stability prevention control strategy under multiple accident scenarios, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method in the above-mentioned embodiment 1 when executing the computer program.

[0107] The relevant technical solutions are the same as above and will not be described in detail here.

[0108] Example 3

[0109] An embodiment of the present invention provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the method in the above-mentioned embodiment 1 are implemented.

[0110] Specifically, the memory may include a high-speed random access memory, and may also include a non-volatile memory, such as a hard disk, an internal memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0111] The relevant technical solutions are the same as above and will not be described in detail here.

[0112] Example 4

[0113] An embodiment of the present application provides a computer program product, including a computer program. When the computer program is run on a computer, the computer executes the steps of the method in the above-mentioned embodiment 1.

[0114] The relevant technical solutions are the same as above and will not be described in detail here.

[0115] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for generating a transient stability prevention control strategy under multiple accident scenarios, characterized in that: include: The steady-state power flow information reflecting the status of each operating point of the power grid is input into the trained transient stability assessment neural network model to obtain the stability probability of each operating point under N different expected accidents; For the current operating point, the operating safety and fault importance of the operating point are calculated based on the stability probability; wherein the operating safety is the total number of expected accidents with a stability probability greater than a preset safety and stability threshold, and the fault importance is used to characterize the importance of different expected accidents to the operating point; When the operating safety Q of the current operating point is less than the set value M, the expected accidents ranked first S in fault importance are taken to obtain the characteristic importance vector of the active output of the power grid generator under the S expected accidents at the current operating point; based on the characteristic importance vector, the active power of the generator with the smallest characteristic importance is transferred to the generator with the largest characteristic importance, and S candidate operating points are obtained; among the S candidate operating points, the one that increases the transient stability probability of all expected accidents the most is determined as the current new operating point; repeat this step until Q ≥ M, and the updated current operating point is obtained; For the remaining D instability accidents at the current operating point, D=NQ, the current operating point is adjusted to the safety domain of each instability accident respectively, and among the D candidate operating points obtained, the one with the largest operating safety is updated as the current operating point; repeat this step until the operating safety of the current operating point is N, or the operating safety of the current operating point is reduced, and the current operating point is used as the power grid control strategy.

2. The method for generating transient stability prevention control strategies under multiple accident scenarios according to claim 1 is characterized in that: The fault importance is: In the formula, I i is the fault importance of instability anticipated accident i among N different anticipated accidents, y i is the stability probability of a given operating point under the anticipated accident i; G is the set of all anticipated instability accidents; p i is the probability of transient instability of the power system after the expected accident i occurs; α is the weight coefficient.

3. The method for generating transient stability prevention control strategy under multiple accident scenarios according to claim 1 or 2, characterized in that: The feature importance vector is calculated as follows: LIME is used to construct a local linear proxy model K(x) to locally interpret the transient stability assessment neural network model to obtain the feature importance vector; wherein K(x) satisfies: K(x)=w0+w1x1+…+w n x n Where K(x) is the local linear proxy model of the transient stability assessment neural network model, x represents the input of the transient stability assessment neural network model, which is a vector containing the steady-state power flow information reflecting the current operating point state and the current expected accident information, and x i is the active output characteristic of the i-th generator in the model, i = 1, 2, ..., n, n is the total number of generators in the power grid; w i is the feature importance corresponding to the active output feature of the i-th generator, w1, w2, …, w n It constitutes the characteristic importance vector of the active output of the power grid generator at the current operating point under the current expected accident.

4. The method for generating transient stability prevention control strategies under multiple accident scenarios according to claim 3 is characterized in that: For the remaining D instability accidents at the current operating point, D = NQ, the current operating point is adjusted to the safety region of each instability accident, including: Under a given instability accident, based on the characteristic importance vector of the current operating point, the active power of the generator with the smallest characteristic importance is transferred as a fixed power P to the generator with the largest characteristic importance to adjust the current operating point. This process is executed repeatedly until the adjusted current operating point is within the safety domain of the given instability accident, and D candidate operating points are obtained.

5. The method for generating transient stability prevention control strategies under multiple accident scenarios according to claim 1, characterized in that: The transient stability assessment neural network model is obtained by using knowledge distillation training, which specifically includes: A transient stability assessment neural network model actually used in a power grid is used as a teacher model, and the knowledge of the teacher model is transferred to a student model through knowledge distillation; wherein the student model is the transient stability assessment neural network model.

6. The method for generating transient stability prevention control strategies under multiple accident scenarios according to claim 5 is characterized in that: During the knowledge distillation training process, each training sample input into the student model and the teacher model includes steady-state power flow information reflecting the current operating point of the power grid and the current expected accident information, and the label is the corresponding transient stability probability; the loss function used is: Loss=λL hard +(1-λ)T 2 L soft Where, L hard is the classification loss, L soft is the distillation loss, T is the temperature coefficient of the softmax activation function of the teacher model, and T>1, λ is the weight coefficient; Classification loss L hard is the cross entropy loss between the sample label prediction value output by the student model and the sample label, wherein the temperature coefficient of the activation function of the student model is 1; Distillation loss L soft is the KL divergence between the soft label prediction value and the soft label, wherein the soft label prediction value is the output of the student model when the temperature coefficient of the softmax activation function of the student model is set to T, and the soft label is the output of the teacher model.

7. The method for generating transient stability prevention control strategies under multiple accident scenarios according to claim 1, characterized in that: S is calculated as: S = [(N-M+1) / 2] Where [·] is the rounding function.

8. A transient stability prevention control strategy generation system under multiple accident scenarios, characterized in that: comprising a computer readable storage medium and a processor; The computer-readable storage medium is used to store executable instructions; The processor is used to read the executable instructions stored in the computer-readable storage medium to execute the method for generating a transient stability prevention control strategy under multiple accident scenarios as described in any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method for generating a transient stability prevention control strategy under a multi-accident scenario as described in any one of claims 1 to 7 is implemented.

10. A computer program product, characterized in that It includes a computer program, which, when running on a computer, enables the computer to execute the method for generating a transient stability prevention control strategy under a multi-accident scenario as described in any one of claims 1 to 7.

Images