Fine adjustment method, transient evaluation method and system for transient stability evaluation model of power system

Abstract

The invention discloses a fine tuning method, a transient evaluation method and a transient evaluation system for a transient stability evaluation model of a power system, and belongs to the technical field of transient stability evaluation of the power system. In a deep learning model comprising a convolutional layer, a pooling layer and a full connection layer, the convolutional layer extracts basic features of working condition data, the features are high in universality among different topologies, and when topology changes are small, the features can be effectively migrated to a transient stability evaluation model under a new topology. Therefore, the topology change state of the power system is accurately determined according to the difference between the current working condition characteristics of the power system and the working condition characteristics of the power system in the basic topology state, and a corresponding model fine tuning scheme is designed; when the power system is in a small topology change state, the model is finely adjusted by fixing all convolutional layers and retraining parameters of a full connection layer, and compatibility of accuracy and model response time can be achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of transient stability assessment of power systems, and more specifically, relates to a fine-tuning method for a transient stability assessment model of a power system, a transient assessment method and a system. Background Art

[0002] In recent years, with the rapid development of renewable energy and the large-scale access of power electronic equipment to the power grid, the complexity and interconnection of power systems have continued to deepen. At the same time, there are random fluctuations in load demand and power supply, which requires the power system to achieve energy supply and demand balance, and its stability characteristics and control mechanisms are becoming increasingly complex. The emergence of these challenges has put the security and stability of the power grid under severe test. In order to ensure the transient stability of the power system, first of all, a variety of emergency plans can be taken to protect the safety of the power system when an accident occurs, such as relay protection to quickly cut off the fault, interlocking cutting, etc.; secondly, it is necessary to evaluate the transient stability of the power system to make full preparations when the fault does not occur. It can be seen that transient stability assessment (TSA) is one of the main measures to ensure the safe and stable operation of modern power systems.

[0003] Time domain simulation method, direct method and artificial intelligence method are the three main transient stability assessment methods at present. The simulation analysis and transient stability assessment by time domain simulation method has high model dimension and long calculation time, that is, time domain simulation method is difficult to meet the requirements of online real-time analysis and decision-making. The direct method is a method to judge transient stability by comparing the transient energy function value and critical value at the end of the disturbance. Its disadvantage is that the Lyapunov function is not easy to construct, it is not suitable for large systems or situations subject to a series of disturbances, and the calculation results are conservative, which is difficult to be applied in modern large power grids that require fast and accurate response.

[0004] Artificial intelligence technology has low dependence on physical models and can learn and mine nonlinear complex relationships between large amounts of data, providing new solutions for dynamic safety analysis and decision-making of complex power grids. As a major field of machine learning, deep learning methods have developed rapidly in recent years. Compared with shallow machine learning, deep learning has stronger data mining and autonomous learning capabilities. Combining the network structure of deep learning with the complexity of power systems can provide a new research direction for transient stability assessment problems.

[0005] However, the existing transient stability assessment methods based on artificial intelligence technology are usually aimed at the situation where all generators and lines of the power system are put into operation. At this time, the power system is in a basic topology state. The method of training a transient stability assessment model for the power system to perform transient stability analysis of the power grid has achieved remarkable results. However, when the topology of the power system changes, such as a system failure causing individual lines to be disconnected or generators to be put into or withdrawn, the data characteristics of each node and line in the system will also change accordingly. At this time, the existing transient stability assessment model cannot accurately assess the transient stability in the current scenario. It is often necessary to rebuild the transient stability assessment model framework for the current scenario and perform a series of tasks such as flow analysis, data processing, and model training. It takes a long time and cannot meet the model response time requirements in actual applications, posing a challenge to the safe and stable operation of the power grid. Summary of the invention

[0006] In response to the above defects or improvement needs of the prior art, the present invention provides a fine-tuning method for a transient stability assessment model of an electric power system, a transient assessment method and a system, so as to solve the technical problem that the prior art is not compatible with the accuracy of transient stability assessment and the model response time when the topology structure of the electric power system changes.

[0007] In order to achieve the above-mentioned object, in a first aspect, the present invention provides a fine-tuning method for a transient stability assessment model of a power system, comprising:

[0008] Obtain the pre-trained transient stability assessment model M when the power system is in the basic topology state tsa ; The basic topological state is the topological state when all generators and lines in the power system are put into operation; M tsa A deep learning model consisting of convolutional layers, pooling layers, and fully connected layers;

[0009] Determine the current topology change state of the power system. When the power system is in a small topology change state, fix M tsa The parameters of the other parts except the fully connected layer in M tsa In the example, the goal is to minimize the difference loss between the evaluation results of the transient stable state of each training sample in the first training set and the corresponding true label. tsa Fine-tune the parameters of the fully connected layer in the proposed method to obtain a transient stability assessment model under small topological changes.

[0010] Each training sample in the first training set includes: the fault line number and corresponding operating condition data of the faulty line when the power system is in a state of small topological change;

[0011] The above-mentioned determination of the current topology change state of the power system includes:

[0012] When the difference D between the current operating characteristics of the power system and the operating characteristics of the power system in the basic topology state is less than the preset threshold D 0 When , the power system is in a state of small topological change; otherwise, the power system is in a state of large topological change;

[0013] The operating condition characteristics include the following category characteristics: node admittance matrix, node active output matrix, node reactive output matrix, bus voltage amplitude matrix and bus voltage phase angle matrix, all of which are N×N in size; N is the total number of nodes in the power system;

[0014] The elements in the i-th row and i-th column of the node admittance matrix are the mutual admittance between the i-th node and the i-th node of the power system, and the elements in the i-th row and i-th column are the self-admittance between the i-th node and the i-th node of the power system; the node active output matrix, node reactive output matrix, bus voltage amplitude matrix and bus voltage phase angle matrix are all diagonal matrices, and the elements in the i-th row and i-th column are the active output, reactive output, voltage amplitude and voltage phase angle of the i-th node of the power system respectively; i=1,2,…,N,j=1,2,…,N,i≠j.

[0015] Further preferably, when the power system is in a state of large topological change, the second training set is input into M tsa In the second training set, the goal is to minimize the difference between the evaluation results of the transient stable state of each training sample and the corresponding true label. tsa By fine-tuning the parameters in , a transient stability assessment model under large topological change conditions is obtained;

[0016] Each training sample in the second training set includes: a fault line number and corresponding operating condition data of a fault occurring when the power system is in a state of large topological changes.

[0017] Further preferably, the pre-trained transient stability assessment model M tsa It is obtained through the following pre-training process:

[0018] The third training set is input into the transient stability assessment model. The parameters in the transient stability assessment model are trained with the goal of minimizing the difference loss between the transient stability state assessment results of each training sample in the third training set and the corresponding true label to obtain the pre-trained transient stability assessment model M. tsa ;

[0019] Each training sample in the third training set includes: a fault line number and corresponding operating condition data of a fault occurring when the power system is in a basic topology state.

[0020] More preferably, Among them, K is the number of feature categories contained in the working condition characteristics; w k is the weight of the feature of the kth category in the working condition feature; A k is the feature of the kth category in the current working condition feature; A 0k is the feature of the kth category in the operating characteristics of the power system in the basic topology state; ||·|| F is the F-norm.

[0021] Further preferably, the operating condition data include: steady-state generator output, steady-state bus voltage and steady-state line flow before the fault occurs.

[0022] Further preferably, the operating condition data includes: transient generator output, transient bus voltage and transient line power flow after a fault occurs.

[0023] Further preferably, the above fault is: a symmetrical three-phase short circuit fault, a two-phase short circuit to ground fault or a single-phase short circuit to ground fault.

[0024] Further preferably, the faulty line number is a preprocessed faulty line number; wherein the preprocessing includes: executing One-Hot and Embedding encoding in sequence.

[0025] In a second aspect, the present invention provides a method for transient assessment of a power system, comprising: determining a current topology change state of the power system, and when the power system is in a state of small topology change, collecting a fault line number of a faulty current power system and corresponding operating condition data, and inputting the data into a transient stability assessment model in a state of small topology change, to obtain an assessment result of the transient stability state;

[0026] The transient stability assessment model under a small topology change state is obtained by fine-tuning the method for fine-tuning the transient stability assessment model of the power system when the power system is in a small topology change state provided by the first aspect of the present invention;

[0027] The method for determining the current topology change state of the power system adopts the method for determining the current topology change state of the power system in the fine-tuning method for the power system transient stability assessment model provided in the first aspect of the present invention.

[0028] Further preferably, the above-mentioned power system transient assessment method further includes: when the power system is in a state of large topology change, collecting the fault line number and corresponding operating condition data of the current power system fault, and inputting them into the transient stability assessment model under the state of large topology change to obtain the transient stability state assessment result;

[0029] The transient stability assessment model under a large topology change state is obtained by fine-tuning the method for fine-tuning the transient stability assessment model of a power system when the power system is in a large topology change state provided by the first aspect of the present invention.

[0030] In a fourth aspect, the present invention provides a power system transient assessment system, comprising: a memory and a processor, wherein the memory stores a computer program, and the processor executes the method provided in the first aspect, the second aspect or the third aspect of the present invention when executing the computer program.

[0031] In a fifth aspect, the present invention further provides a computer-readable storage medium, which includes a stored computer program, wherein when the computer program is executed by a processor, the device where the storage medium is located is controlled to execute the method provided in the first aspect, the second aspect or the third aspect of the present invention.

[0032] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects:

[0033] 1. The present invention provides a method for fine-tuning a transient stability assessment model for a power system. Considering that in a deep learning model including a convolutional layer, a pooling layer, and a fully connected layer, the convolutional layer extracts the basic features of the operating condition data, and these features have high universality between different topologies. When the topology changes are small, they can be effectively migrated to the transient stability assessment model under the new topology. Therefore, the present invention designs a corresponding model fine-tuning scheme according to the topological change state of the power system; by constructing operating condition characteristics including a node admittance matrix, a node active output matrix, a node reactive output matrix, a bus voltage amplitude matrix, and a bus voltage phase angle matrix, and based on the difference between the current operating condition characteristics of the power system and the operating condition characteristics of the power system in the basic topological state, the topological change state of the power system is accurately determined. When small topological changes occur, all convolutional layers are fixed. Considering that the fully connected layer is a further abstraction and mapping of the features extracted by the convolutional layer, it is necessary to correctly map it to the new label space when facing new scenarios. Therefore, when the power system is in a state of small topological changes, by fixing all convolutional layers and retraining the parameters of the fully connected layer to fine-tune the model, it is possible to achieve compatibility between accuracy and model response time.

[0034] 2. The present invention first evaluates the degree of difference in model topology changes, and then adopts different model fine-tuning methods. In this way, more appropriate fine-tuning methods can be used for topology changes of different scales, which can effectively improve the generalization performance, computational efficiency and prediction accuracy of the model, while reducing human intervention and optimizing resource utilization, so that the system can better maintain performance and meet the needs of practical applications when facing complex and dynamically changing environments.

[0035] 3. Furthermore, in the fine-tuning method of the power system transient stability assessment model provided by the present invention, for the fault line number in the training sample, considering that it is discrete category data, One-Hot and Embedding encoding are performed on it and converted into low-dimensional continuous data for subsequent training. This can avoid the model from mistakenly processing discrete category data as continuous data, thereby improving the assessment accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A flowchart of a method for fine-tuning a transient stability assessment model for a power system provided by an embodiment of the present invention;

[0037] Figure 2 A schematic diagram of the structure of a convolutional neural network constructed according to an embodiment of the present invention;

[0038] Figure 3 A diagram showing the pre-training effect of the transient stability assessment model provided in an embodiment of the present invention;

[0039] Figure 4 A schematic diagram of two scenario settings in the IEEE 39-node standard test system provided by an embodiment of the present invention;

[0040] Figure 5 Schematic diagram of four fine-tuning solutions provided by embodiments of the present invention;

[0041] Figure 6 Effect diagrams of four fine-tuning solutions provided in the embodiments of the present invention in two scenarios. DETAILED DESCRIPTION

[0042] 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.

[0043] In order to achieve the above objectives, firstly, Figure 1 As shown, the present invention provides a fine-tuning method for a transient stability assessment model of a power system, comprising:

[0044] Obtain the pre-trained transient stability assessment model M when the power system is in the basic topology state tsa ; The basic topological state is the topological state when all generators and lines in the power system are put into operation; M tsaA deep learning model including a convolutional layer, a pooling layer, and a fully connected layer, which may be CNN, EfficientNet, Xception, etc., is not limited here, and CNN is preferably used in the present invention;

[0045] Determine the current topology change state of the power system. When the power system is in a small topology change state, fix M tsa The parameters of the other parts except the fully connected layer in M tsa In the example, the goal is to minimize the difference loss between the evaluation results of the transient stable state of each training sample in the first training set and the corresponding true label. tsa Fine-tune the parameters of the fully connected layer in the proposed method to obtain a transient stability assessment model under small topological changes.

[0046] Each training sample in the first training set includes: the fault line number and corresponding operating condition data of the faulty line when the power system is in a state of small topological change;

[0047] The above-mentioned determination of the current topology change state of the power system includes:

[0048] When the difference D between the current operating characteristics of the power system and the operating characteristics of the power system in the basic topology state is less than the preset threshold D 0 When , the power system is in a state of small topological change; otherwise, the power system is in a state of large topological change;

[0049] Among them, the operating condition characteristics include the following category characteristics: node admittance matrix, node active output matrix, node reactive output matrix, bus voltage amplitude matrix and bus voltage phase angle matrix, all of which are N×N in size; N is the total number of nodes in the power system;

[0050] The elements in the i-th row and i-th column of the node admittance matrix are the mutual admittance between the i-th node and the i-th node of the power system, and the elements in the i-th row and i-th column are the self-admittance between the i-th node and the i-th node of the power system; the node active output matrix, node reactive output matrix, bus voltage amplitude matrix and bus voltage phase angle matrix are all diagonal matrices, and the elements in the i-th row and i-th column are the active output, reactive output, voltage amplitude and voltage phase angle of the i-th node of the power system respectively; i=1,2,…,N,j=1,2,…,N,i≠j.

[0051] It should be noted that when a small topology change occurs, it will take too long to retrain or fine-tune the entire network. In this case, it is necessary to consider fixing the parameters of some layers and only train the remaining layers. In the deep learning model including convolutional layers, pooling layers, and fully connected layers, the convolutional layers extract the basic features of the operating data. These features are highly universal between different topologies and can be effectively migrated to the transient stability assessment model under the new topology. The weights of the fully connected layers need to be adjusted according to the transient stability assessment task under the new topology, because they are further abstracted and mapped to the features extracted by the convolutional layers. If the fully connected layers are not retrained, they may not be correctly mapped to the new label space when faced with new data, resulting in low accuracy after migration. Therefore, when a small topology change occurs, all convolutional layers are fixed and the parameters of the fully connected layers are retrained to fine-tune the model to be compatible with accuracy and model response time.

[0052] In an alternative embodiment, Among them, K is the number of feature categories contained in the working condition characteristics; w k is the weight of the feature of the kth category in the working condition feature; A k is the feature of the kth category in the current working condition feature; A 0k is the feature of the kth category in the operating characteristics of the power system in the basic topology state; ||·|| F is the F-norm.

[0053] It should be noted that the F-norm metric is only a measure of A k With A 0k One way to measure the difference between them is to use Euclidean distance, cosine similarity (converting the matrix into a vector for measurement), etc., which is not limited here.

[0054] In an optional implementation, when the power system is in a state of large topological change, the second training set is input into M tsa In the second training set, the goal is to minimize the difference between the evaluation results of the transient stable state of each training sample and the corresponding true label. tsa By fine-tuning the parameters in , a transient stability assessment model under large topological change conditions is obtained;

[0055] Each training sample in the second training set includes: a fault line number and corresponding operating condition data of a fault occurring when the power system is in a state of large topological changes.

[0056] When a large topology change occurs, the new topology is too different from the original topology. If the convolution layer parameters are fixed, the model evaluation effect will be poor. Therefore, it is necessary to fine-tune the overall transient stability evaluation model parameters based on the pre-trained model to ensure accuracy.

[0057] In an optional implementation, the pre-trained transient stability assessment model M tsa It is obtained through the following pre-training process:

[0058] The third training set is input into the transient stability assessment model. The parameters in the transient stability assessment model are trained with the goal of minimizing the difference loss between the transient stability state assessment results of each training sample in the third training set and the corresponding true label to obtain the pre-trained transient stability assessment model M. tsa ;

[0059] Each training sample in the third training set includes: a fault line number and corresponding operating condition data of a fault occurring when the power system is in a basic topology state.

[0060] It should be noted that the fault type in the present invention may be a symmetrical three-phase short circuit fault, a two-phase short circuit to ground fault, a single-phase short circuit to ground fault, etc. The fault types designed in the first training sample, the second training sample and the third training sample are all the same.

[0061] In an optional implementation manner, the operating condition data includes: steady-state generator output, steady-state bus voltage and steady-state line flow before the fault occurs; wherein the steady-state generator output includes steady-state generator active output and / or reactive output.

[0062] In an optional implementation manner, the operating condition data includes: transient generator output, transient bus voltage and transient line flow after a fault occurs; wherein the steady-state generator output includes steady-state generator active output and / or reactive output.

[0063] In an optional implementation manner, the fault line number is a preprocessed fault line number; wherein the preprocessing includes: executing One-Hot and Embedding encoding in sequence.

[0064] In order to further illustrate the fine-tuning method of the power system transient stability assessment model provided by the present invention, a specific embodiment is described in detail below:

[0065] Construct the training set:

[0066] This embodiment selects the steady-state generator active output, steady-state bus voltage and steady-state line flow before the fault occurs as the operating condition data;

[0067] For each line in the power system, expected faults are set in the 0-50% segment and 50%-100% segment of the line respectively, and a training sample including the fault line number of the power system fault and the corresponding operating condition data is constructed.

[0068] Calculate the steady-state stability index In the formula, |Δδ max | is the absolute value of the maximum power angle difference between any two generators during the simulation time. The TSI index can be used to characterize the degree of disturbance of the target system, and thus to characterize the transient stability. When TSI>0, it means that the system is stable; otherwise, it means that the system is unstable, and the corresponding sample label is obtained.

[0069] By randomly discarding the generated stable samples, the ratio of stable to unstable samples is controlled to 2:1 to prevent the imbalance of the binary classification samples from causing poor training results.

[0070] For the fault line number in the training sample, considering that it is discrete category data, One-Hot and Embedding encoding are performed on it and converted into low-dimensional continuous data for subsequent training, which can further ensure the training effect.

[0071] By adopting the above method, corresponding training sets are obtained when the power system is in a basic topology state, a small topology change state and a large topology change state, respectively, to obtain a third training set, a first training set and a second training set.

[0072] It should be noted that the above training set can be obtained in a real scene, can be generated by time domain simulation, or can be obtained by both methods, which is not limited here.

[0073] Pre-training of the model:

[0074] Obtain the pre-trained transient stability assessment model M when the power system is in the basic topology state tsa ; The basic topological state is the topological state when all generators and lines in the power system are put into operation; M tsa A deep learning model consisting of convolutional layers, pooling layers, and fully connected layers;

[0075] Pre-trained transient stability assessment model M tsa It is obtained through the following pre-training process:

[0076] The third training set is input into the transient stability assessment model. The parameters in the transient stability assessment model are trained with the goal of minimizing the difference loss between the transient stability state assessment results of each training sample in the third training set and the corresponding true label to obtain the pre-trained transient stability assessment model M. tsa ;

[0077] Each training sample in the third training set includes: a fault line number and corresponding operating condition data of a fault occurring when the power system is in a basic topology state.

[0078] In this embodiment, CNN is selected as the transient stability assessment model. In order to avoid the disadvantage of the traditional time-domain simulation method that the training time is too long, a convolutional neural network including two cascaded convolutional layers + pooling layers and a fully connected layer is constructed; wherein the kernel size of the convolutional layer and the pooling layer is both 2.

[0079] The learning rate exponential decay strategy is adopted. The learning rate decays exponentially from the initial value of 0.001, thereby ensuring a faster convergence speed in the early stage of training and the training results are not limited to the local optimal solution.

[0080] Accuracy, F1 value, missed detection rate and training time are selected as evaluation indicators of model training effect.

[0081] Regarding the structure of the convolutional neural network constructed in this embodiment, illustratively, Figure 2 As shown. Convolutional neural networks mainly consist of four structures: convolution layer, activation layer, pooling layer and fully connected layer. Convolution is the basis of CNN, which extracts features by sliding a small convolution kernel over the input data and calculating the weighted sum of the local area. Specifically, given an input image I and a filter (convolution kernel) K, the convolution operation is expressed as follows:

[0082]

[0083] Among them, S is the output feature map, m and n are the height and width of the filter.

[0084] The features generated by the convolution operation are transformed nonlinearly through the ReLU activation function. The ReLU activation function is expressed as follows:

[0085] f(x) = max(0, x)

[0086] The pooling layer is used to reduce the size of the feature map, thereby reducing the amount of computation and the risk of overfitting. The maximum pooling representation is as follows:

[0087]

[0088] At the end of the CNN, a fully connected layer is set to flatten the feature maps and map them to output categories.

[0089] By combining convolution operations, activation functions, pooling layers, and fully connected layers, convolutional neural networks can effectively extract hierarchical features from raw data, thereby achieving classification and recognition of complex data.

[0090] The exponential decay strategy of learning rate is explained as follows: Learning rate is a key parameter in deep learning training. A learning rate that is too large may make it difficult for the model to converge. A learning rate that is too small may cause the model to converge too slowly and the program to run for too long. Only when the learning rate is appropriate can the model achieve a better overall effect. In order to optimize the training effect of CNN and shorten the training time, the present invention adopts an exponential decay strategy for learning rate. In the early stage of training, in order to make the model converge quickly to a better solution first, a larger learning rate is adopted; in the later stage of training, in order to improve the stability of model training, a smaller learning rate is adopted. With the iteration of training rounds, the learning rate decreases exponentially. The exponential decay rate is 0.98, and the exponential decay formula is shown as follows:

[0091] decayed lr=0.001×0.98 ( global_step / decay_step ) .

[0092] In a specific embodiment, 80% of the data is used as a training set, 20% of the data is used as a test set, and the IEEE 39 node system is selected as a test system for example testing. Figure 3 The pre-training effect of the transient stability assessment model in this embodiment is shown. It can be seen that with the increase in the number of training rounds, the accuracy and F1 value of the transient stability assessment model constructed by the present invention show an upward trend, while the FPR and training loss show a downward trend. Finally, the accuracy stabilized at 99.75%, the F1 value stabilized at 99.63%, and the FPR stabilized at 0.56%. The training effect meets the requirements of the transient stability performance of the system in actual operation.

[0093] In a specific embodiment, 80% of the data is used as a training set and 20% of the data is used as a test set. The IEEE39 node system is selected as the test system, and support vector machine (SVM), k-nearest neighbor (KNN), and graph neural network (GNN) are used as comparison algorithms for example testing.

[0094] It should be noted that the transient stability assessment of the power system belongs to an unbalanced binary classification problem, and its imbalance is mainly manifested in that the number of stable samples is much larger than the number of unstable samples. For unbalanced binary classification problems, accuracy (ACC) and F1 value are usually used to describe the effectiveness of the model. Since the losses caused by missed judgments are often much greater than the losses caused by misjudgments, in order to ensure the safe and stable operation of the power system, the present invention will focus on the proportion of negative samples that are missed as positive samples, that is, the missed judgment rate (FPR). In addition, the time required to evaluate each piece of data (evaluation time) is also an important indicator. Therefore, the present invention selects the four indicators of ACC, F1 value, FPR and evaluation time as evaluation indicators of the CNN model.

[0095] Table 1 shows the effect comparison when different models are selected as transient stability assessment models. Compared with the other three methods, the CNN method has the most outstanding ACC and F1 values, the lowest FPR, and the assessment speed meets the requirements of actual operation. Although SVM has a superior assessment speed, its ACC and F1 values ​​are less than 99%, and its FPR is larger than that of the CNN method. It may cause greater losses than CNN in actual power system fault conditions, so it is not suitable for selection.

[0096] It can be seen that the CNN method selected by the present invention has achieved excellent results in the three indicators of ACC, F1 value and FPR while ensuring timeliness, and has certain advantages.

[0097] Table 1

[0098] Selected Model ACC(%) F1(％) FPR(%) Evaluation time SVM 98.80 98.15 0.75 ～4.76e-5 KNN 93.60 95.17 7.10 ～1.31e-3 GNN 92.26 94.24 6.61 ～8.50e-1 CNN 99.75 99.63 0.56 ～2.07e-3

[0099] Fine-tuning of the model:

[0100] Determine the current topology change state of the power system. When the difference D between the current operating characteristics of the power system and the operating characteristics of the power system in the basic topology state is less than the preset threshold D 0 When , the power system is in a state of small topological change; otherwise, the power system is in a state of large topological change;

[0101] Among them, the operating condition characteristics include the following category characteristics: node admittance matrix, node active output matrix, node reactive output matrix, bus voltage amplitude matrix and bus voltage phase angle matrix, all of which are N×N in size; N is the total number of nodes in the power system;

[0102] The elements in the i-th row and j-th column of the node admittance matrix are the mutual admittance between the i-th node and the j-th node of the power system, and the elements in the i-th row and i-th column are the self-admittance between the i-th node and the i-th node of the power system; the node active output matrix is ​​a diagonal matrix, and the elements in its i-th row and i-th column are the active output of the i-th node of the power system; the node reactive output matrix is ​​a diagonal matrix, and the elements in its i-th row and i-th column are the reactive output of the i-th node of the power system; the bus voltage amplitude matrix is ​​a diagonal matrix, and the elements in its i-th row and i-th column are the voltage amplitudes of the i-th node of the power system; the bus voltage phase angle matrix is ​​a diagonal matrix, and the elements in its i-th row and i-th column are the voltage phase angles of the i-th node of the power system; i=1,2,…,N,j=1,2,…,N,i≠j。

[0103] In this embodiment, Wherein, K is the number of feature categories contained in the working condition feature, and in this embodiment, the value is 5; k is the weight of the feature of the kth category in the working condition feature; A kis the feature of the kth category in the current working condition feature; A 0k is the feature of the kth category in the operating characteristics of the power system in the basic topology state; ||·|| F is the F norm. In this embodiment, the preset threshold D 0 The value is 400. 1 、w 2 、w 3 、w 4 and w 5 The values ​​are 0.6, 0.1, 0.1, 0.1 and 0.1 respectively.

[0104] Specifically, the first category of features in the current operating condition features is the node admittance matrix of the power system in the current topology change state:

[0105] The second category of features in the current operating condition features is the node active output matrix A of the power system in the current topology change state. 2 =P; The third category of the current operating condition characteristics is the node reactive power matrix A of the power system in the current topology change state 3 =Q. The node active output matrix P is a diagonal matrix. For the generator node i put into operation, P ii is a positive value, defined as the active power output by the generator; for load node j, P jj is a negative value, which is defined as the active power absorbed by the load; for other nodes, it is defined as zero. The node reactive power matrix Q is similarly defined as the reactive power emitted and absorbed at each node. Specifically:

[0106]

[0107] The fourth category of the current operating condition characteristics is the bus voltage amplitude matrix A when the power system is in the current topology change state. 4 =V; The fifth category of the current operating condition characteristics is the bus voltage phase angle matrix A when the power system is in the current topology change state 5 =θ; specifically:

[0108]

[0109] When the power system is in a state of small topological change, the fixed M tsa The parameters of the other parts except the fully connected layer in M tsa In the example, the goal is to minimize the difference loss between the evaluation results of the transient stable state of each training sample in the first training set and the corresponding true label. tsaFine-tune the parameters of the fully connected layer in the proposed method to obtain a transient stability assessment model under small topological changes.

[0110] Each training sample in the first training set includes: a fault line number and corresponding operating condition data of a fault occurring when the power system is in a state of small topology change.

[0111] When the power system is in a state of large topological change, the second training set is input into M tsa In the second training set, the goal is to minimize the difference between the evaluation results of the transient stable state of each training sample and the corresponding true label. tsa By fine-tuning the parameters in , a transient stability assessment model under large topological change conditions is obtained;

[0112] Each training sample in the second training set includes: a fault line number and corresponding operating condition data of a fault occurring when the power system is in a state of large topological changes.

[0113] In order to further illustrate the accuracy and model response time of the fine-tuning method of the power system transient stability assessment model provided by the present invention, the present invention sets up two scenarios with different degrees of topology change on the IEEE 39-node standard test system to study the learning effect of parameter transfer when the degree of topology change is different. The two scenario settings can be described as follows, as shown in the figure. Figure 4 Then, four different fine-tuning schemes were set up in two scenarios: small topology change and large topology change. The migration effects of several fine-tuning schemes in the two scenarios were compared and verified by the IEEE 39-node standard test system example. Among them, the four fine-tuning schemes can be described as follows, as shown in the figure. Figure 5 shown.

[0114] Solution 1: Use a randomly initialized power system transient stability assessment model and retrain it.

[0115] Solution 2: Using the pre-trained transient stability evaluation model M tsa The parameters are used as initial values ​​for training.

[0116] Solution 3: Using the pre-trained transient stability evaluation model M tsa The parameters of are used as initial values, the parameters of the first convolutional layer are fixed, and the parameters of the second convolutional layer and the fully connected layer are trained.

[0117] Solution 4: Using the pre-trained transient stability evaluation model M tsa The parameters of are used as initial values, the parameters of all two convolutional layers are fixed, and only the parameters of the fully connected layer are trained.

[0118] The IEEE 39-bus system is selected as the test system. The training results of the four schemes in scenario 1 and scenario 2 are shown in the figure. Figure 6 As shown. Figure 6 It can be seen that in scenario 1 where the system topology has only slightly changed, the model accuracy of schemes 3 and 4 is slightly lower than that of schemes 1 and 2, but both are higher than 98.5%. At the same time, the training time required for the model is significantly reduced from scheme 1 to scheme 4. This is because when the topology changes slightly, the new topology is highly similar to the original topology, and only the fully connected layer needs to be fine-tuned to achieve a good evaluation effect. This shows that when the topology changes slightly, scheme 4, which fixes all convolutional layers and only trains the parameters of the fully connected layer, can significantly reduce the time cost of model training while ensuring a high evaluation accuracy.

[0119] from Figure 6 It can also be seen that for scenario 2 with a large topology change, the model training time is also gradually decreasing from solution 1 to solution 4. However, since the system topology is quite different from the original topology at this time, it is not possible to fine-tune the model directly by fixing the convolution layer and training the fully connected layer, and a new model needs to be reconstructed. Therefore, the solution 2 with the original model parameters as the initial value is selected as the model fine-tuning solution under large topology changes. Under the condition of making full use of the network parameters of the original model, the new model can complete the transient stability assessment work quickly and well.

[0120] In a second aspect, the present invention provides a method for transient assessment of a power system, comprising: when the power system is in a state of small topology change, collecting the fault line number and corresponding operating condition data of the current power system, inputting them into a transient stability assessment model under a state of small topology change, and obtaining an assessment result of the transient stability state;

[0121] The transient stability assessment model under a small topology change state is obtained by fine-tuning the method for fine-tuning the transient stability assessment model of the power system when the power system is in a small topology change state provided by the first aspect of the present invention;

[0122] The method for determining the current topology change state of the power system adopts the method for determining the current topology change state of the power system in the fine-tuning method for the power system transient stability assessment model provided in the first aspect of the present invention.

[0123] The related technical solution is the same as the fine-tuning method of the power system transient stability assessment model provided by the first aspect of the present invention, and will not be described in detail here.

[0124] Further preferably, the above-mentioned power system transient assessment method further includes: when the power system is in a state of large topology change, collecting the fault line number and corresponding operating condition data of the current power system fault, and inputting them into the transient stability assessment model under the state of large topology change to obtain the transient stability state assessment result;

[0125] The transient stability assessment model under a large topology change state is obtained by fine-tuning the method for fine-tuning the transient stability assessment model of a power system when the power system is in a large topology change state provided by the first aspect of the present invention.

[0126] The related technical solution is the same as the fine-tuning method of the power system transient stability assessment model provided by the first aspect of the present invention, and will not be described in detail here.

[0127] In a fourth aspect, the present invention provides a power system transient assessment system, comprising: a memory and a processor, wherein the memory stores a computer program, and the processor executes the method provided in the first aspect, the second aspect or the third aspect of the present invention when executing the computer program.

[0128] The related technical solutions are the same as the methods provided in the first, second and third aspects of the present invention, and will not be described in detail here.

[0129] In a fifth aspect, the present invention further provides a computer-readable storage medium, which includes a stored computer program, wherein when the computer program is executed by a processor, the device where the storage medium is located is controlled to execute the method provided in the first aspect, the second aspect or the third aspect of the present invention.

[0130] The related technical solutions are the same as the methods provided in the first, second and third aspects of the present invention, and will not be described in detail here.

[0131] 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 fine-tuning method for a transient stability assessment model of a power system, characterized in that: include: Get the pre-trained transient stability assessment model M when the power system is in the basic topology state tsa ; The basic topological state is the topological state when all generators and lines in the power system are put into operation; the M tsa A deep learning model consisting of convolutional layers, pooling layers, and fully connected layers; Determine the current topology change state of the power system. When the power system is in a small topology change state, fix the M tsa The parameters of the other parts except the fully connected layer in M tsa In the example, the M is trained with the goal of minimizing the difference loss between the evaluation result of the transient stable state of each training sample in the first training set and the corresponding true label. tsa Fine-tune the parameters of the fully connected layer in the proposed method to obtain a transient stability assessment model under small topological changes. Wherein, each training sample in the first training set includes: the fault line number and corresponding operating condition data of the fault occurring when the power system is in a state of small topology change; The determining of the current topology change state of the power system includes: When the difference D between the current operating characteristics of the power system and the operating characteristics of the power system in the basic topology state is less than the preset threshold D0, the power system is in a small topology change state; otherwise, the power system is in a large topology change state; The operating condition characteristics include the following category characteristics: node admittance matrix, node active output matrix, node reactive output matrix, bus voltage amplitude matrix and bus voltage phase angle matrix, all of which are N×N in size; N is the total number of nodes in the power system; The elements of the i-th row and i-th column of the node admittance matrix are the mutual admittances between the i-th node and the i-th node of the power system, and the elements of the i-th row and i-th column are the self-admittances between the i-th node and the i-th node of the power system; the node active output matrix, the node reactive output matrix, the bus voltage amplitude matrix and the bus voltage phase angle matrix are all diagonal matrices, and the elements of the i-th row and i-th column are the active output, reactive output, voltage amplitude and voltage phase angle of the i-th node of the power system respectively; i=1, 2,…, N, j=1, 2,…, N, i≠j.

2. The fine-tuning method for the transient stability assessment model of a power system according to claim 1, characterized in that: When the power system is in a state of large topological change, the second training set is input into the M tsa In the example, the M is trained with the goal of minimizing the difference loss between the evaluation result of the transient stable state of each training sample in the second training set and the corresponding true label. tsa By fine-tuning the parameters in , a transient stability assessment model under large topological change conditions is obtained; Wherein, each training sample in the second training set includes: a fault line number and corresponding operating condition data of a fault occurring when the power system is in a state of large topology change.

3. The fine-tuning method for the transient stability assessment model of a power system according to claim 1, characterized in that: The M tsa It is obtained through the following pre-training process: The third training set is input into the transient stability assessment model, and the parameters in the transient stability assessment model are trained with the goal of minimizing the difference loss between the transient stability state assessment results of each training sample in the third training set and the corresponding true label, so as to obtain the M tsa ; Wherein, each training sample in the third training set includes: a fault line number and corresponding operating condition data of a fault occurring when the power system is in a basic topology state.

4. The fine-tuning method for the transient stability assessment model of a power system according to any one of claims 1 to 3, characterized in that: Among them, K is the number of feature categories contained in the working condition characteristics; w k is the weight of the feature of the kth category in the working condition feature; A k is the feature of the kth category in the current working condition feature; A 0k is the feature of the kth category in the operating characteristics of the power system in the basic topology state; ||·|| F is the F-norm.

5. The fine-tuning method for the transient stability assessment model of a power system according to any one of claims 1 to 3, characterized in that: The operating condition data include: steady-state generator output, steady-state bus voltage and steady-state line flow before the fault occurs; Alternatively, the operating condition data include: transient generator output, transient bus voltage and transient line flow after a fault occurs.

6. The fine-tuning method for the transient stability assessment model of a power system according to any one of claims 1 to 3, characterized in that: The fault line number is a pre-processed fault line number; wherein the pre-processing includes: executing One-Hot and Embedding encoding in sequence.

7. A method for transient assessment of a power system, characterized in that: include: Determine the current topology change state of the power system. When the power system is in a small topology change state, collect the fault line number and corresponding operating condition data of the current power system, input them into the transient stability assessment model under the small topology change state, and obtain the transient stability state assessment result; Wherein, the transient stability assessment model under the state of small topology change is obtained by fine-tuning the transient stability assessment model of the power system according to any one of claims 1 and 3-6; The method for determining the current topology change state of the power system adopts the method for determining the current topology change state of the power system in the fine-tuning method of the power system transient stability assessment model described in any one of claims 1-6.

8. The method for transient assessment of a power system according to claim 7, characterized in that: Also includes: When the power system is in a state of large topological change, the fault line number and corresponding operating condition data of the current power system fault are collected and input into the transient stability assessment model under the state of large topological change to obtain the assessment result of the transient stability state; The transient stability assessment model under a large topology change state is obtained by fine-tuning the power system transient stability assessment model according to any one of claims 2 to 6.

9. A power system transient assessment system, characterized in that: include: A memory and a processor, wherein the memory stores a computer program, and the processor executes the method according to any one of claims 1 to 8 when executing the computer program.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored computer program, wherein when the computer program is executed by a processor, the device where the storage medium is located is controlled to execute the method according to any one of claims 1 to 8.

Images