Feature enhancement type adaptive transient stability assessment method

Abstract

The invention discloses a feature enhancement type adaptive transient stability evaluation method, which comprises the following steps of: acquiring operation data of a power system on line, extracting voltage features, inputting a trained PSC-HGRU transient stability evaluation model, and outputting a transient stability evaluation result of the power system, namely stability or instability of the power system. The PSC-HGRU transient stability evaluation model is constructed based on a probabilistic sparrow search algorithm PSA, a channel attention mechanism CAM and an enhanced gated circulation unit neural network HGRU; the HGRU comprehensively captures a complex dependency relationship before and after data in a transient process, and the capture capability of the model on transient stability features is enhanced; the CAM adaptively adjusts weights of different feature channels in the HGRU, interference of redundant features is reduced, and the accuracy of evaluation is further improved; the PSA dynamically adjusts the key parameters of the evaluation model, and improves the stability and convergence speed of the model.

Description

Technical Field

[0001] The invention relates to the technical field of power systems, and in particular to a feature-enhanced adaptive transient stability assessment method. Background Art

[0002] Transient stability of power systems refers to the ability of generators to maintain synchronous operation after the system is disturbed. Factors that lead to transient stability failure of power systems include short circuit faults, accidental disconnection of lines or generators, etc. Accurate and rapid transient stability assessment (TSA) is of great significance to ensure the safe operation of power systems. The development of smart grids has brought about the application of long-distance, large-capacity transmission and a high proportion of power electronic equipment. The introduction of these new technologies has brought new potential risks to the power system. Power supply interruptions and complex cascading failures further increase the complexity of power system stability analysis and control.

[0003] In traditional TSA, the direct method and the time domain simulation method are the two main analysis methods. The time domain simulation method involves solving the differential equations of the power system, which is a technically complex process. In contrast, the direct method provides a more computationally direct analysis approach. However, the direct method requires a series of simplifications and assumptions when constructing the transient energy function, which may affect the accuracy of the analysis results.

[0004] In recent years, a variety of machine learning methods including traditional machine learning and deep machine learning have been introduced in the field of TSA, such as decision trees, random forests, convolutional neural networks (CNNs), stacked autoencoders, gated recurrent unit neural networks (GRUs), and long short-term memory networks (LSTMs). Zhang et al. proposed a robust white-box model based on a decision tree model, which can maintain the accuracy of the evaluation under incomplete PMU data and ensure the interpretability and transparency of the model. Based on CNN, Yang et al. proposed a transient stability assessment method for AC / DC receiving-end power grid partitioning. James et al. proposed an intelligent time-varying transient stability assessment system based on LSTM networks to balance the trade-off between evaluation accuracy and response time.

[0005] Deep learning technology has achieved good results in the field of transient stability assessment, mainly due to its powerful feature extraction capabilities, but existing methods mainly process unidirectional data, which limits the model's ability to capture the front-to-back dependencies in sequence data. In addition, these deep learning methods have shortcomings in convergence speed, generalization ability, and stability, which seriously affect the performance of TSA assessment models. Summary of the invention

[0006] In view of the deficiencies in the prior art, the present invention provides a feature-enhanced adaptive transient stability assessment method.

[0007] The present invention achieves the above technical objectives through the following technical means.

[0008] A feature-enhanced adaptive transient stability assessment method:

[0009] Obtain the operation data of the power system online, extract the voltage characteristics, input the trained PSC-HGRU transient stability assessment model, and output the transient stability assessment results of the power system, that is, whether the power system is stable or unstable;

[0010] The PSC-HGRU transient stability assessment model is constructed based on a probabilistic sparrow search algorithm, a channel attention mechanism, and an enhanced gated recurrent unit neural network;

[0011] The enhanced gated recurrent unit neural network includes two independent GRU units: one processes the input sequence along the positive time axis, and the other processes the same input sequence along the negative time axis;

[0012] The channel attention mechanism is embedded in an enhanced gated recurrent unit neural network;

[0013] The probabilistic mahjong search algorithm optimizes the enhanced gated recurrent unit neural network.

[0014] In a further technical solution, the calculation formula of the enhanced gated recurrent unit neural network includes:

[0015] Final hidden state The calculation formula is:

[0016]

[0017] in, is the forward hidden state at time t is the reverse hidden state at time t, w1 and w2 are the forward hidden states and the reverse hidden state The weight of w3 is and The weight of the product, b is the bias term to be trained for the enhanced gated recurrent unit neural network.

[0018] A further technical solution is that the forward hidden state at time t The calculation formula is: x t is the input at time t.

[0019] A further technical solution is that the reverse hidden state at time t The calculation formula is: xt is the input at time t.

[0020] A further technical solution is that the channel attention mechanism is embedded in an enhanced gated recurrent unit neural network, including:

[0021] The final hidden state Aggregate each feature channel: Among them, T is the sequence time step, c is the number of channels, and F contains the global feature representation of each feature channel. express The feature representation of the cth channel;

[0022] Perform the excitation operation and use two fully connected layers to model F: F' = σ(W2δ(W1F)), where δ is the RELU function and W1 and W2 are the weights of the excitation operation;

[0023] Get the new weighted output:

[0024] Further technical solution, in the probabilistic sparrow search algorithm, the formula for updating the finder position is:

[0025]

[0026] in, is the information of the i-th sparrow in the j-dimension, t is the number of iterations, P is the selection probability, rd(0,1) is the Gaussian random distribution, R2 is the safety value, ST is the warning value, Q is the random step of the normal distribution, and α represents a random number between 0 and 1.

[0027] A further technical solution is that in the probabilistic sparrow search algorithm, the formula for updating the follower position is:

[0028]

[0029] in, is the current optimal solution, is the current worst solution, L represents a row vector whose elements are all 1, T m is the maximum number of iterations, n is the total population, and d is the number of sparrow individuals.

[0030] In a further technical solution, the selection probability

[0031] A further technical solution is that in the probabilistic sparrow search algorithm, the gap between the accuracy of the model on the training set and the perfect accuracy is defined as the fitness function. Each "sparrow" in the algorithm corresponds to a set of hyperparameters of the enhanced gated recurrent unit neural network, including training rounds, learning rate, and number of neural units.

[0032] A further technical solution is that the probabilistic sparrow search algorithm optimizes the enhanced gated recurrent unit neural network, specifically: in each training process, the probabilistic sparrow search algorithm adjusts the hyperparameters of the enhanced gated recurrent unit neural network according to the accuracy of the model on the training set, and the hyperparameters are used to guide the training of the enhanced gated recurrent unit neural network. The enhanced gated recurrent unit neural network uses the hyperparameters and the input transient stability data for learning, and performs performance evaluation on the test set. The evaluation results are fed back to the probabilistic sparrow search algorithm to update the positions of the discoverer, scout and follower, and calculate the corresponding fitness values ​​to guide the next update of the "sparrow" position until the set iteration upper limit is reached, and the probabilistic sparrow search algorithm converges to a set of optimal hyperparameters.

[0033] The beneficial effects of the present invention are:

[0034] (1) Based on the gated recurrent unit neural network (GRU), the present invention proposes an enhanced gated recurrent unit neural network (HGRU) to fully capture the complex before-and-after dependencies of data in the transient process and enhance the model's ability to capture transient stability features;

[0035] (2) The present invention designs a channel attention mechanism (CAM) that adaptively adjusts the weights of different feature channels in the HGRU network to enhance the extraction of important transient stability features, reduce the interference of redundant features, and further improve the accuracy of evaluation;

[0036] (3) The present invention proposes a probabilistic sparrow search algorithm (PSA), sets selection probabilities and uses random numbers to guide the search process, dynamically adjusts key parameters of the evaluation model (training rounds, learning rate, number of neural units), avoids the problem of traditional sparrow search algorithms easily falling into local optimality, improves the stability and convergence speed of the model, and reduces the workload of manual parameter adjustment. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0038] Figure 1 This is the GRU structure diagram of the present invention;

[0039] Figure 2 This is the HGRU structure diagram of the present invention;

[0040] Figure 3 This is a structural diagram of the PSC-HGRU transient stability assessment model of the present invention;

[0041] Figure 4 This is the PSC-HGRU transient stability assessment framework diagram of the present invention;

[0042] Figure 5 This is a comparison chart of transient stability assessment performance of the eight models described in the present invention. DETAILED DESCRIPTION

[0043] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0044] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0045] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0046] Example 1

[0047] Reference Figure 4 , which is the first embodiment of the present invention, provides a feature-enhanced adaptive transient stability assessment method, comprising:

[0048] (1) Based on the gated recurrent unit neural network (GRU), an enhanced gated recurrent unit neural network (HGRU) is proposed, which includes:

[0049] The transient stability data of power systems has significant time series characteristics. In the current research field, deep learning methods such as recurrent neural networks (RNN) and long short-term memory networks (LSTM) are widely used to deal with such problems. GRU effectively controls the information flow through a simple gating mechanism, allowing the model to capture long-distance dependencies and solve the problem of slow LSTM training. At the same time, GRU uses the connection between hidden layer nodes to make the network output at the previous moment become the input at the current moment. This structure effectively avoids the gradient explosion and vanishing problems of RNN.

[0050] In GRU, the reset gate and update gate are two important gating mechanisms. The update gate is responsible for adjusting the contribution of the hidden state of the previous moment to the hidden state of the current moment, while the reset gate determines the degree of association between the current input and the previous memory. This mechanism enables GRU to maintain the persistence of information when processing long sequence data, avoiding the problem of information decay or loss over time. The structure of GRU is as follows: Figure 1 shown.

[0051] Specifically, the reset gate (R t ) is calculated as:

[0052] R t =σ(W r x t +U r H t-1 +b r )

[0053] Where σ is the sigmoid function, W r and U r is the weight matrix to be trained, x t is the input at time t, b r is the bias term to be trained.

[0054] Update gate (Z t ) is calculated as:

[0055] Z t =σ(W z x t +U z H t-1 +b z )

[0056] Among them, W z and U z is the weight matrix to be trained, b z is the bias term to be trained.

[0057] Candidate hidden states at time t The calculation formula is:

[0058]

[0059] Where tanh(·) is the hyperbolic tangent function, W h and W h ' is the weight matrix to be trained, and · is the element-by-element multiplication operation.

[0060] The hidden state at time t (H t ) is calculated as:

[0061]

[0062] The HGRU model consists of two independent GRU units: one processes the input sequence along the positive time axis, and the other processes the same input sequence along the negative time axis. The specific network structure is as follows Figure 2 As shown. This bidirectional structure enables HGRU to simultaneously obtain the forward and backward feature information of the time series, thereby fully capturing the forward and backward dependencies in transient stability data. The calculation formula of HGRU is as follows:

[0063] The forward hidden state at time t The calculation formula is:

[0064]

[0065] The reverse hidden state at time t The calculation formula is:

[0066]

[0067] Final hidden state of HGRU The calculation formula is:

[0068]

[0069] Among them, w1 and w2 are the forward hidden states and the reverse hidden state The weight of w3 is and The weight of the product, b is the bias item of HGRU to be trained.

[0070] By merging the outputs of the forward and reverse GRUs, HGRU can capture transient stable features that GRU ignores. At the same time, by multiplying the corresponding elements of the feature vectors output by the forward and reverse GRUs, the model's sensitivity to transient features is enhanced, thereby improving the performance of the evaluation model.

[0071] (2) In order to further enhance the feature extraction capability of the transient stability assessment model, a channel attention mechanism (CAM) is designed, which includes:

[0072] The channel attention mechanism performs a dimensionality reduction operation called global average pooling on the hidden states generated by HGRU. Aggregate each feature channel of , the calculation formula of this operation is as follows:

[0073]

[0074] Among them, T is the sequence time step, c is the number of channels, and F contains the global feature representation of each feature channel. express The feature representation of the c-th channel of .

[0075] For each channel c, the average activation value at its time step is calculated. The global features in F are then used to calculate the attention weight of each channel in order to weight the importance of different feature channels in the subsequent processing of the model.

[0076] Next, we perform the excitation operation and use two fully connected layers to model F. The calculation formula for this operation is as follows:

[0077] F'=σ(W2δ(W1F))

[0078] Among them, δ is the RELU function, and W1 and W2 are the weights of the excitation operation.

[0079] Finally, the input of the original channel is multiplied by the obtained weight to obtain the new weighted output. The calculation formula for this operation is as follows:

[0080]

[0081] Through this channel attention mechanism, HGRU's attention to important features is effectively improved, redundant features are suppressed, and the accuracy of transient stability assessment is improved.

[0082] (3) A probabilistic sparrow search algorithm (PSA) is proposed. By setting the selection probability and using random numbers to guide the search process, the key parameters of the evaluation model (training rounds, learning rate, and number of neural units) are dynamically adjusted. This avoids the problem of traditional sparrow search algorithms easily falling into local optimality, improves the stability and convergence speed of the model, and reduces the workload of manual parameter adjustment, including:

[0083] In order to enhance the model's efficiency in parameter selection and accuracy in transient stability assessment, the present invention designs a probabilistic sparrow search algorithm to optimize HGRU. This algorithm is known for its excellent engineering practicality. It imitates the group behavior of sparrows and efficiently searches for the optimal parameter configuration in the parameter space of HGRU. Compared with traditional optimization techniques, the probabilistic sparrow search algorithm can achieve fast convergence while ensuring high accuracy of search results, which makes it perform well in optimizing HGRU.

[0084] The probabilistic sparrow search algorithm optimizes the solution to the problem by simulating the social behavior of sparrows. It is mainly divided into three key stages: discoverer position update, follower position update and scout position update. The following is an overview of the principles of these three stages:

[0085] a. Finder location update

[0086] The finder is responsible for finding food in the population and providing foraging areas and directions for the entire sparrow population. The position update formula is as follows:

[0087]

[0088] in, is the information of the i-th sparrow in the j-dimension, t is the number of iterations, T m is the maximum number of iterations, R2 is the safety value, ST is the warning value, Q is the random step of the normal distribution, and α represents a random number between 0 and 1. When R2 < ST, it means that the discoverer can perform extensive search operations. When R2 > ST, it means that the discoverer should adopt a more cautious search strategy instead of conducting large-scale exploration.

[0089] In the above formula, when R2<ST, when the sparrow individual directly migrates to the vicinity of the currently known optimal solution, although the algorithm can achieve rapid convergence, this approach may reduce the diversity of the population, thereby increasing the risk of the algorithm falling into the local optimum. In order to balance the relationship between rapid convergence and population diversity and avoid the problem of falling into the local optimal solution, the present invention updates the individual position by introducing selection probability and using random numbers. The specific discoverer position update formula is as follows:

[0090]

[0091] Among them, P is the selection probability and rd(0,1) is the Gaussian random distribution.

[0092] The calculation formula of P is:

[0093]

[0094] b. Follower position update

[0095] The follower uses the finder to obtain food, and its position update formula is as follows:

[0096]

[0097] in, is the current optimal solution, is the current worst solution, and L represents a row vector whose elements are all 1.

[0098] When i is less than half of the total population n This indicates that individuals with lower fitness rankings in the population (i.e., the i-th individual) have not yet obtained enough resources and therefore need to update their positions to find new food sources; They then perform random searches near their current locations in the hope of finding new food. It is noteworthy that the present invention uses random numbers rd(0,1) to ensure convergence and population diversity.

[0099] c. Scout location update

[0100] The scout is responsible for reconnaissance and will take corresponding actions when it finds a predator coming. The scout's position update strategy is as follows:

[0101]

[0102] Where β is the step size control parameter, K is a random number, τ is a small positive number (0.0001 in this embodiment), and f i is the fitness value of the i-th sparrow, f g is the optimal fitness value of the sparrow population, f w is the worst fitness value of the previous sparrow population, represents the optimal position in the sparrow population. i ≠f g When the sparrow is in a disadvantageous position; when f i =f g When the scout sparrows arrive, they get a signal of danger and start approaching other sparrows.

[0103] The three stages work together to enable the probabilistic sparrow search algorithm to demonstrate excellent performance in model parameter optimization. Its global search capability ensures extensive exploration of the parameter space, and its dynamic adaptability allows the algorithm to flexibly adjust the search strategy based on feedback from model training. The effective local search mechanism helps to quickly converge to the optimal or near-optimal parameter configuration, thereby improving the generalization ability and prediction accuracy of the model.

[0104] (4) Based on the probabilistic sparrow search algorithm, channel attention mechanism and enhanced gated recurrent unit neural network, a PSC-HGRU transient stability evaluation model is constructed, including:

[0105] The PSC-HGRU transient stability evaluation model proposed in the present invention includes: an input layer, an HGRU layer, a CAM layer, a fully connected layer, a Softmax layer and a PSA optimization module. The specific model structure is as follows: Figure 3 As shown in the figure, α1, α2, α3, α4 are the weights of each channel of HGRU, and h1, h2, h3, h4 are the outputs of HGRU.

[0106] The PSC-HGRU transient stability assessment framework proposed in this invention is shown in the figure Figure 4 shown.

[0107] The specific implementation steps are as follows:

[0108] 1) Transient stability data processing

[0109] Based on the dat and swi files of the power system, the present invention adopts BPA software to perform time domain simulation to generate transient stability simulation data; extracts voltage characteristics from the simulation results and performs preprocessing; then, divides them into a training set and a test set.

[0110] 2) Design and initialization of fitness function of PSA algorithm

[0111] The gap between the accuracy of the model on the training set and the perfect accuracy (100%) is defined as the fitness function. The lower the value of this function, the higher the accuracy of the model, which provides a clear optimization goal for model optimization. Subsequently, the probabilistic sparrow search algorithm (PSA) is started. Each "sparrow" in the algorithm corresponds to a set of hyperparameter configurations of the HGRU model, covering training rounds, learning rate, and number of neural units. The initial values ​​of these hyperparameters are randomly generated, laying the foundation for the algorithm to perform global optimization search.

[0112] 3) Overall training process and parameter update

[0113] During each training process, the PSA algorithm adjusts the hyperparameters of the HGRU model based on the accuracy of the model on the training set. Subsequently, these hyperparameters are used to guide the training of the HGRU model. The model uses these hyperparameters and the input transient stability data for learning, and the performance is evaluated on the test set. The evaluation results are then fed back to the PSA algorithm to update the positions of the discoverer, scout, and follower, and calculate the corresponding fitness values ​​to guide the next update of the "sparrow" position. This iterative cycle continues until the set iteration limit is reached. Ultimately, the PSA algorithm will converge to a set of optimal hyperparameter configurations, allowing the HGRU model to achieve the highest accuracy in the transient stability assessment task, completing the optimization process of the model parameters.

[0114] 4) Obtain the operation data of the power system online, extract the voltage characteristics, input the trained PSC-HGRU transient stability assessment model, and output the transient stability assessment result of the power system, that is, whether the power system is stable or unstable.

[0115] This embodiment also provides a computer device, which is suitable for the HGRU adaptive transient stability assessment method combining the probabilistic sparrow search algorithm and the attention mechanism, including a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute computer executable instructions to implement the HGRU adaptive transient stability assessment method combining the probabilistic sparrow search algorithm and the attention mechanism as proposed in the above embodiment.

[0116] The computer device may be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a key, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse, etc.

[0117] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, the HGRU adaptive transient stability evaluation method combining a probabilistic sparrow search algorithm and an attention mechanism as proposed in the above embodiment is implemented.

[0118] Example 2

[0119] Reference Figure 4 , which is the second embodiment of the present invention, and this embodiment provides a feature-enhanced adaptive transient stability assessment method. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through simulation experiments.

[0120] Exemplarily, the evaluation performance test and analysis of PSC-HGRU can be carried out through the following steps: In order to evaluate the superiority of PSC-HGRU for transient stability assessment tasks, it is compared with long short-term memory network (LSTM), gated recurrent unit neural network (GRU), recurrent neural network (RNN), convolutional neural network (CNN), decision tree (DT), support vector machine (SVM), and random forest (RF). Among them, DT adopts the C5.0 algorithm, SVM adopts the radial basis kernel function, and the number of base classifiers of RF is 20. LSTM, GRU, RNN and CNN are all configured to include two feature extraction layers and three fully connected layers, as well as a Softmax output layer. The learning rate of the model is set to 0.001, and the optimization process adopts the Adam optimization algorithm. Specifically, the evaluation results of the 8 models are as follows. Figure 5 shown.

[0121] Depend on Figure 5It can be seen that the PSC-HGRU model proposed in the present invention has an evaluation accuracy of 99.02%, which is significantly superior to deep learning models such as LSTM, GRU, RNN and CNN. However, DT, SVM, RF and other shallow machine learning models are difficult to deeply extract transient stability information, so their prediction performance is lower than that of deep learning models. In summary, the PSC-HGRU model has a more significant evaluation advantage.

[0122] Exemplarily, the ablation experiment can be performed through the following steps to verify the advantages of each module of the PSC-HGRU model.

[0123] Ablation model 1: does not contain the CAM layer and PSA optimization module, and only uses the HGRU layer, fully connected layer and Softmax layer to build the evaluation model. Other parameter settings are consistent with the PSC-HGRU model.

[0124] Ablation model 2: does not contain the CAM layer. Only the PSA optimization module and the HGRU layer, the fully connected layer, and the Softmax layer are combined to build the evaluation model. Other parameter settings are consistent with the PSC-HGRU model.

[0125] Ablation model 3: does not contain the PSA optimization module, and only combines the CAM layer, HGRU layer, fully connected layer and Softmax layer to build the evaluation model. Other parameter settings are consistent with the PSC-HGRU model.

[0126] The test results are shown in Table 1.

[0127] Table 1 Average results of each index of the four models

[0128]

[0129]

[0130] From the ablation experiment results, we can see that the PSC-HGRU model shows better performance than ablation models 1, 2, and 3. Removing the CAM layer or the PSA optimization module from the PSC-HGRU model will lead to a decrease in model performance.

[0131] The embodiments are preferred implementations of the present invention, but the present invention is not limited to the above-mentioned implementations. Any obvious improvements, substitutions or modifications that can be made by those skilled in the art without departing from the essential content of the present invention belong to the protection scope of the present invention.

Claims

1. A feature-enhanced adaptive transient stability assessment method, characterized in that: Obtain the operation data of the power system online, extract the voltage characteristics, input the trained PSC-HGRU transient stability assessment model, and output the transient stability assessment results of the power system, that is, whether the power system is stable or unstable; The PSC-HGRU transient stability assessment model is constructed based on a probabilistic sparrow search algorithm, a channel attention mechanism, and an enhanced gated recurrent unit neural network; The enhanced gated recurrent unit neural network includes two independent GRU units: one processes the input sequence along the positive time axis, and the other processes the same input sequence along the negative time axis; The channel attention mechanism is embedded in an enhanced gated recurrent unit neural network; The probabilistic mahjong search algorithm optimizes the enhanced gated recurrent unit neural network.

2. The feature-enhanced adaptive transient stability assessment method according to claim 1, characterized in that: The calculation formula of the enhanced gated recurrent unit neural network includes: Final hidden state The calculation formula is: in, is the forward hidden state at time t is the reverse hidden state at time t, w1 and w2 are the forward hidden states and the reverse hidden state The weight of w3 is and The weight of the product, b is the bias term to be trained for the enhanced gated recurrent unit neural network.

3. The feature-enhanced adaptive transient stability assessment method according to claim 2, characterized in that: The forward hidden state at time t The calculation formula is: x t is the input at time t.

4. The feature-enhanced adaptive transient stability assessment method according to claim 2, characterized in that: The reverse hidden state at time t The calculation formula is: x t is the input at time t.

5. The feature-enhanced adaptive transient stability assessment method according to claim 2, characterized in that: The channel attention mechanism is embedded in an enhanced gated recurrent unit neural network, including: The final hidden state Aggregate each feature channel: Among them, T is the sequence time step, c is the number of channels, and F contains the global feature representation of each feature channel. express The feature representation of the cth channel; Perform the excitation operation and use two fully connected layers to model F: F' = σ(W2δ(W1F)), where δ is the RELU function and W1 and W2 are the weights of the excitation operation; Get the new weighted output:

6. The feature-enhanced adaptive transient stability assessment method according to claim 1, characterized in that: In the probabilistic sparrow search algorithm, the formula for updating the finder's position is: in, is the information of the i-th sparrow in the j-dimension, t is the number of iterations, P is the selection probability, rd(0,1) is the Gaussian random distribution, R2 is the safety value, ST is the warning value, Q is the random step of the normal distribution, and α represents a random number between 0 and 1.

7. The feature-enhanced adaptive transient stability assessment method according to claim 6, characterized in that: In the probabilistic sparrow search algorithm, the formula for updating the follower position is: in, is the current optimal solution, is the current worst solution, L represents a row vector whose elements are all 1, T m is the maximum number of iterations, n is the total population, and d is the number of sparrow individuals.

8. The feature-enhanced adaptive transient stability assessment method according to claim 7, characterized in that: The selection probability 9. The feature-enhanced adaptive transient stability assessment method according to claim 1, characterized in that: In the probabilistic sparrow search algorithm, the gap between the accuracy of the model on the training set and the perfect accuracy is defined as the fitness function. Each "sparrow" in the algorithm corresponds to a set of hyperparameters of the enhanced gated recurrent unit neural network, including training rounds, learning rate, and number of neural units.

10. The feature-enhanced adaptive transient stability assessment method according to claim 9, characterized in that: The probabilistic sparrow search algorithm optimizes the enhanced gated recurrent unit neural network, specifically: in each training process, the probabilistic sparrow search algorithm adjusts the hyperparameters of the enhanced gated recurrent unit neural network according to the accuracy of the model on the training set, and the hyperparameters are used to guide the training of the enhanced gated recurrent unit neural network. The enhanced gated recurrent unit neural network uses the hyperparameters and input transient stability data for learning, and performs performance evaluation on the test set. The evaluation results are fed back to the probabilistic sparrow search algorithm to update the positions of the discoverer, scout and follower, and calculate the corresponding fitness values ​​to guide the next update of the "sparrow" position until the set iteration upper limit is reached, and the probabilistic sparrow search algorithm converges to a set of optimal hyperparameters.

Images