Electric power system transient voltage stability evaluation method based on convolutional nerve and bidirectional long-short term memory neural network

Abstract

The invention discloses an electric power system transient voltage stability evaluation method based on a convolutional neural network and a bidirectional long-short term memory neural network, and belongs to the technical field of electric power system universality transient voltage stability evaluation. The method comprises the following steps: carrying out feature set coarse screening by adopting a Pearson's rule; carrying out shortest time window search based on a bisection method; selecting an optimal feature subset for transient voltage stability evaluation; the obtained optimal feature subset is input into a CNN-BiLSTM model for training; carrying out transient voltage stability evaluation based on a deep learning CNN-BiLSTM model; original features are constructed from the time sequence dimension and the space dimension, and an obtained feature subset serves as input of a machine learning model to conduct online evaluation on transient voltage stability of the power system. According to the method, transient voltage stability evaluation can be carried out only by obtaining the optimal model through large-scale model training, the dimensionality of the feature subset is reduced to a great extent, the evaluation accuracy is improved, and the time for online evaluation calculation is shortened.

Description

Technical Field

[0001] The present invention relates to the technical field of transient voltage stability assessment of power systems, and in particular to a method for transient voltage stability assessment of power systems. Background Art

[0002] The new grid operation characteristics such as the continuous growth in the scale of grid-connected new energy power sources such as wind and solar power, the "hollowing out" of load centers caused by the withdrawal of traditional power generating units, and the high proportion of power electronic equipment investment will make the transient voltage stability problem more prominent, and the existing transient voltage stability assessment methods are no longer applicable.

[0003] For machine learning methods, the selection of initial features has a great impact on the final evaluation performance of the model. At present, there is no unified method for constructing initial features. Generally, "three-segment" features and post-disturbance trajectory features are commonly used. The former selects electrical parameters or their statistical indicators before, during and after the fault is removed as initial features; the latter uses the electrical parameter trajectory (time series) after the fault is removed as the initial feature. Although both initial features contain time dimension information, the machine learning model does not have the ability to mine time dimension information, resulting in the time series features being submerged in a large number of feature points, which restricts the improvement of evaluation performance. When the feature dimension is high, in order to overcome the shortcomings of the limited expression ability of the machine learning model, feature extraction technology is also needed to reduce the dimensionality of the input data to improve the model performance. This results in the need for experts to design new feature extraction algorithms when the system operation mode changes, which increases the cost of using the model.

[0004] Deep learning achieves its goal by customizing deep neural network structures for different tasks and then using back-propagation algorithms to reveal complex associations in large data sets. Convolutional neural networks (CNNs), an important branch of deep learning, have good sequence feature extraction. Initially, they mainly used two-dimensional convolutional neural networks for image recognition and language generation. After that, many scholars and experts found that they have good performance in sequence data processing and feature extraction. When evaluating transient voltage stability problems, it is necessary to explore the complex high-dimensional nonlinear mapping relationship between input electrical parameters and output stable state or stability, and it is also necessary to do so under the constraint of fast processing of high-dimensional data. Compared with other deep learning models, CNN can more comprehensively capture the high-dimensional mapping relationship between input and output. In addition, another important branch of deep learning, long short-term memory neural networks (LSTMs), improves the shortcomings of gradient vanishing and gradient exploding in recurrent neural networks to capture information from both directions of the sequence, which can better utilize the correlation between the front and back of the input sequence information, and can enable the model to make more accurate predictions.

[0005] The rapid development of artificial intelligence has provided a new perspective for transient voltage stability assessment methods, and combined with deep neural network methods, it is expected to better cope with the above challenges. How to obtain the spatiotemporal key response features suitable for transient voltage stability assessment of new power systems from ultra-high dimensional features, and how to improve the spatial and temporal dimensions of the input feature quantities to find a more suitable subset of key response features for transient voltage stability assessment are issues that need to be urgently addressed. Summary of the invention

[0006] The purpose of the present invention is to propose a method for evaluating transient voltage stability of a power system based on a convolutional neural network and a bidirectional long short-term memory neural network, the method comprising the following steps:

[0007] Step 1: obtain simulation results under different operating modes and different fault conditions through simulation; perform Pearson correlation coefficient dimensionality reduction on the obtained feature quantities to obtain feature quantities that are strongly correlated with transient voltage stability; perform homogenization on the obtained feature quantities to obtain the required feature set samples.

[0008] Step 2: Input the feature set samples obtained in step 1 into the CNN-BiLSTM model for training;

[0009] Step 3: Perform transient voltage stability assessment based on the trained CNN-BiLSTM model; input the real-time data measured online as the initial feature into the trained transient stability assessment model to predict the stability of the system, and then determine the stability output continuity label: Substitute the output label into the formula to obtain the stability label and complete the stability margin division.

[0010] The step 2 specifically includes:

[0011] Step 21: partition the feature set into stability margins using a single binary table practical criterion;

[0012] Step 22: Input the feature set samples into the CNN model to minimize the loss function, realize time series data processing and automatic feature extraction, and reduce the multi-dimensional input to one-dimensional input;

[0013] Step 23: Import the extracted input into the BiLSTM model to capture information from both directions of the sequence, and use the correlation between the front and back ends of the input sequence information to enable the model to make more accurate predictions;

[0014] Step 24: After the preliminary prediction is completed, the model performance is evaluated by accuracy, missed detection rate and false positive rate; and the shortest time window is searched through binary search; repeated iterations are performed to ensure that the model meets the accuracy requirements.

[0015] The step 22 specifically includes:

[0016] Step 221: Input feature is X = [x1, x2, x3, ... x n ] T is a one-dimensional array as the input layer of the model. t×d , t is the length of the time series, and d is the number of input feature arrays. The obtained one-dimensional input signal is mapped to the convolution layer through a one-dimensional convolution operation, which is expressed as:

[0017]

[0018] f r (z)=max(z,0) (13)

[0019] Where: is the convolution operator; Represented by the convolution kernel The calculated j-th feature map, j∈[1,n c ], where n c is the number of convolution kernels; is a weight matrix, m is the convolution kernel size, which determines the length of the local time window corresponding to the convolution kernel when extracting features; b is the bias term.

[0020] Step 222: The pooling layer can reduce the size of the feature space and the number of network parameters through downsampling operations, thereby reducing the computational complexity and helping to prevent overfitting. In the maximum pooling, the maximum value in the window is selected as the output value, which is expressed as

[0021]

[0022] in To select the maximum value of two consecutive points in the sequence, the sequence length is reduced by half. The above formula is the global max-pooling, which is the last operation of the pooling layer. The function of this formula is to extract the most useful global time series information and reduce the sequence length to 1.

[0023] Step 223: The fully connected layer further abstracts and combines the global temporal features by performing matrix concatenation on the feature quantities extracted by the convolutional layer and the pooling layer, and is expressed as:

[0024] a fc =f r (a p,last W fc +b) (16)

[0025] In the above formula: W fc ∈R nfc is the weight matrix in the fully connected layer; n fc is the hidden layer. The obtained feature quantity is passed through the fully connected layer and the obtained classification output is:

[0026]

[0027] Where f a (z) is the activation function sigmoid function, and the output is That is, the probability distribution of different voltage stability conditions, and the transient voltage stability is judged according to the obtained probabilities of different voltage stability conditions.

[0028] The output label of step 3 is substituted into the formula to obtain the stability label, that is,

[0029]

[0030] Where: V min is the minimum node voltage after the fault occurs; V th is the per-unit value of the transient voltage threshold, which is set to 0.75 as required; T th is the duration of operation below the transient voltage threshold, which is set to 1s as required; Ts is the time the bus voltage is below the threshold. k is the critical voltage offset time factor. Where η represents the transient voltage stability margin. When η is smaller, it means the system stability margin is poor.

[0031] Based on the value of η, four stability conditions are constructed: super stable system, stable system, critical stable system, and unstable system. A stability margin function is constructed to classify systems with different stability conditions and normalize P so that P∈[0,3]. That is,

[0032] P=me -η +n (21)

[0033]

[0034] Where: m and n are unknown coefficients, which can be calculated through the critical conditions of the hyperstable system and the unstable system.

[0035] P th1 With P th2 are the thresholds of stable state and critical stable state respectively. Samples with P = 0 are super stable samples, and samples with P = 1 are super stable samples.

[0036] The sample with P=2 is a stable sample, and the probability of system instability is small. The sample with P=2 is a critical state sample with a greater risk of instability, and the sample with P=3 is an unstable sample.

[0037] The beneficial effects of the present invention are:

[0038] 1. The method of the present invention can greatly reduce the dimension of feature subsets, improve evaluation accuracy and reduce the time required for online evaluation calculations.

[0039] 2. The method of the present invention improves the original optimization algorithm's high initial fitness value, slow convergence speed, and falling into the local optimal solution, and effectively improves the optimization efficiency and fitness value. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a flow chart of the method for evaluating transient voltage stability of a power system based on convolutional neural network and bidirectional long short-term memory neural network of the present invention;

[0041] Figure 2 Flowchart for searching the shortest time window for the binary search;

[0042] Figure 3 It is a schematic diagram of the algorithm structure of convolutional neural network and bidirectional long short-term memory neural network;

[0043] Figure 4 This is the wiring diagram of an AC / DC hybrid power grid with a high proportion of new energy access;

[0044] Figure 5 China Electric Power Research Institute's 10,000-node standard calculation example

[0045] Figure 6 is the model training curve. DETAILED DESCRIPTION

[0046] The present invention proposes a method for evaluating transient voltage stability of a power system based on a convolutional neural network and a bidirectional long short-term memory neural network. The present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0047] Figure 1 The flowchart of the method for evaluating transient voltage stability of a power system based on convolutional neural network and bidirectional long short-term memory neural network of the present invention is as follows:

[0048] 1. Initial feature set construction

[0049] 1) The Pearson correlation coefficient calculates the correlation between data set X and data set Y to achieve feature dimension reduction. The specific formula is as follows:

[0050]

[0051] In the formula, are the average values ​​of data set X and data set Y respectively. The coefficient ρ is the Pearson correlation coefficient, and the value range of ρ is -1≤ρ≤1. It is stipulated that the larger the |ρ|, the greater the correlation between the two.

[0052] 2) Since the loads and currents in different systems have numerical differences, averaging is used to eliminate the numerical differences of the same characteristic quantities in different systems:

[0053]

[0054] Among them, μ i It is the average value of each variable. m is the number of features.

[0055] 3) Stability margin division based on transient voltage stability assessment. The bus voltage drops continuously below 0.75 (per unit value) for no more than 1 second during the transient process after the fault, which is a practical criterion for transient voltage stability. The specific formula for describing the transient voltage stability through a single binary table practical criterion is as follows:

[0056]

[0057] Where: V min is the minimum node voltage after the fault occurs; V th is the per-unit value of the transient voltage threshold, which is set to 0.75 as required; T th The duration of operation below the transient voltage threshold is set to 1s as required; T s

[0058] is the time when the bus voltage is lower than the threshold. k is the critical voltage deviation time factor. Where η represents the transient voltage stability margin. When η is smaller, it means that the system stability margin is poor.

[0059] In order to intuitively show the transient stability margin of the system, this paper constructs four stability situations based on the value of η: super stable system, stable system, critical stable system, and unstable system. The systems in different stability situations are classified and normalized with P so that P∈[0,3].

[0060] P=me -η +n (26)

[0061]

[0062] Where: m and n are unknown coefficients, which can be calculated through the critical conditions of the hyperstable system and the unstable system.

[0063] P th1 With P th2 are the thresholds of stable state and critical stable state respectively. Samples with P = 0 are super stable samples, and samples with P = 1 are super stable samples.

[0064] The sample with P=2 is a stable sample, and the probability of system instability is small. The sample with P=2 is a critical state sample with a greater risk of instability, and the sample with P=3 is an unstable sample.

[0065] 2. Training based on CNN-BiLSTM model

[0066] Model training is performed in two steps:

[0067] 1) Use Adam optimization algorithm to adjust the learning rate of CNN-BiLSTM model.

[0068] Adam algorithm: The core idea of ​​the Adam algorithm is to estimate the first-order moment of the gradient (momentum) by calculating

[0069] And the second-order moment estimate (the weighted average of the square of the gradient) to adjust the learning rate of each parameter. Specifically, Adam's update rule for each parameter is as follows:

[0070] m t =β1m t-1 +(1-β1)g t (28)

[0071] v t =β2v t-1 +(1-β2)gt2 (29)

[0072] Among them, g t is the gradient at time step t, m t is the first-order moment estimate, v t is the second-order moment estimate, β1 and β2 are the exponential decay rates used to compute the weighted average of momentum and squared gradient, respectively.

[0073] Due to m t and v t The data are calculated using exponentially weighted moving averages, so in the initial stages, they will be biased. To correct this bias, the Adam algorithm introduces bias correction:

[0074]

[0075] Here, θt is the parameter at the tth step, η is the learning rate, and ε is a small constant to prevent the denominator from being zero.

[0076] 2) Shortest time window search: Use binary search to search for the shortest time window.

[0077] The binary search for the shortest time window is divided into three steps:

[0078] The first step is to initialize the obtained data, and the right boundary of the search range is 5 seconds after the model disturbance fault ends, denoted as T right0 , the fault occurrence time is T f .

[0079] The second step is to import the obtained data into the model for training. The input time window is T0 = T right0 -T f . Since the accuracy rate P obtained by CNN model training AccThere is randomness, so this paper will train the data of the same time window 100 times and get The average accuracy P is obtained by averaging Acc .

[0080] Step 3: If the obtained If it is less than 98%, it is considered that the import time window is too short and needs to be extended. If the obtained P Acc If greater than 98%, The input time window is T (k+1) =T right(k+1) -T f . Similarly, the difference between the current and the next two input time windows ΔT = |T k+1 -T k |Stop searching when the required accuracy δ is met.

[0081] The overall process and optimization process of the proposed binary search for the shortest time window are as follows: Figure 2 shown.

[0082] 3. Application of CNN-BiLSTM model in transient voltage stability assessment

[0083] The application of CNN-BiLSTM model in transient voltage stability assessment is mainly divided into three steps: initial feature set construction, CNN-BiLSTM model training and model transferability verification.

[0084] 1) Original feature construction: construct the original feature set.

[0085] The feature set is studied from three aspects: voltage, load, and phase angle. Considering the collection of more comprehensive types of electrical quantities, and fully considering the impact of new energy and DC access on the transient voltage stability of the system, including: thermal power unit output and power angle, new energy output, 500kV main grid transmission line flow and load power; 500kV bus voltage, phase; DC voltage, DC power. The feature dimension is 5646 in total.

[0086] The Pearson filtering algorithm is used to preliminarily screen the original feature dimensions to obtain the initial subset, and the original feature set of each sample contains 1168 dimensions of features.

[0087] 2) CNN-BiLSTM model training:

[0088] Based on the obtained initial feature set, input it into the CNN-BiLSTM model training.

[0089] 3) Online evaluation: Use step 2) to obtain the CNN-BiLSTM model for transient voltage stability evaluation.

[0090] Based on the Missing Alarm Rate (MAR) and the False Alarm Rate (FAR), the composite screening framework model evaluation indicators, the Missing Alarm Rate and the False Alarm Rate, as well as the Accuracy (ACC), are defined as follows:

[0091]

[0092] Case analysis:

[0093] Figure 4 The figure shows the connection diagram of a hybrid AC / DC power grid with a high proportion of new energy access, with a total of 100 nodes, including 32 nodes of the 500kV main grid. The installed capacity ratio is new energy (2.4 million kilowatts): conventional power (6.3 million kilowatts) = 1:2.62. The basic operation mode is 1.8 million kilowatts of new energy power generation (1.2 million kilowatts of wind power and 600,000 kilowatts of photovoltaic power generation) and 3.86 million kilowatts of conventional units. There is one DC in the system, and the DC input power is 800,000 kilowatts.

[0094] Figure 5 The 10,000-node standard calculation example of China Electric Power Research Institute is a complex AC / DC hybrid power grid with a high proportion of new energy access. The system is divided into the sending side and the receiving side. The sending side system has a total of 3 areas, denoted as S areas, namely the high proportion of new energy wind and thermal bundled transmission area, the conventional load center area and the high proportion of hydropower transmission area; the receiving side system has a total of 4 areas, denoted as R areas, namely the receiving area, the high proportion of new energy area, the AC / DC hybrid receiving area and the pure power base. The voltage of the sending and receiving sides is 500KV, including 10,575 three-phase nodes, about 550 synchronous power sources, and 1,031 wind / solar renewable energy stations with a total installed capacity of 289.58 million kilowatts, 178.96 million kilowatts of conventional power installed capacity, and 110.62 million kilowatts of new energy installed capacity (accounting for 38.2% of the total installed capacity); the receiving side of the system is connected to the sending side through 3 8 million DC lines, of which the sending side is connected to other power grids through 1 line, and the receiving side is connected to 2 lines. The load adopts a comprehensive load model. The calculation example is built based on the BPA simulation platform of the Institute of Electrical Engineering, and data processing and model training are jointly carried out in MATLAB 2020 and Python 3.9.

[0095] Dimensionality reduction effectiveness evaluation:

[0096] As shown in Table 1, only 1162 feature nodes with high importance weights are retained after Pearson dimensionality reduction of the original feature set, and more than 79.3% of the features are considered to have no significant impact on the learning process. In terms of evaluation accuracy, the evaluation model can give the lowest missed rate with the minimum feature subset dimension and improve the overall evaluation accuracy. Since the feature subset dimension is greatly reduced, it only takes 0.041s to get the stability judgment result during online evaluation, which can meet the needs of online real-time applications. The effectiveness of the proposed feature selection method in transient voltage stability evaluation applications is verified.

[0097] Table 1 Comparison of feature selection evaluation indicators

[0098]

[0099] CNN-BiLSTM model effectiveness evaluation:

[0100] Some traditional transient voltage stability assessment methods often used in the packaging method, such as traditional artificial neural network (ANN), support vector machine (SVM), decision tree (DT), and multi-binary table (MBT), are compared. The average performance parameters of 20 tests are shown in Table 2.

[0101] Table 2 Comparison of transient voltage stability evaluation performance of different models

[0102]

[0103] By comparing the advantages and disadvantages of the evaluation indicators of each method, the following conclusions can be drawn:

[0104] 1) The accuracy of CNN-BiLSTM is as high as 99.03%, which is the highest compared with the other four models, and has the lowest missed alarm rate and false alarm rate.

[0105] 2) CNN achieved a good score of 96.51%, and the BiLSTM algorithm also achieved a high accuracy of 95.26% during the evaluation process, but it was still lower than CNN-BiLSTM. The time consumed was longer than that of the CNN-BiLSTM model, reaching 0.136s and 0.88s, and the evaluation effect was not good.

[0106] Evaluation of generalization ability of feature combination

[0107] In terms of the combination of characteristic quantities, the following are set as the basic reference group (Group 1) that only includes the bus node voltage amplitude and the new energy unit terminal voltage amplitude as characteristic quantities, the bus node voltage amplitude, the new energy unit terminal voltage amplitude and the node active load (Group 2), the bus node voltage amplitude, the new energy unit terminal voltage amplitude and the node reactive load (Group 3), the bus node voltage amplitude, the new energy unit terminal voltage amplitude and the node power factor (Group 4), and the bus node voltage amplitude and power angle and the new energy unit terminal voltage amplitude (Group 5). The nodes used in the above characteristic quantities are the same, and the input time window sampling time is from the 5th cycle to the 15th cycle. The other control parameters remain unchanged, and the evaluation nodes are shown in Table 3.

[0108] Table 3 Influence of different spatial characteristics

[0109]

[0110] By comparing the advantages and disadvantages of the evaluation indicators of each method, the following conclusions can be drawn:

[0111] When the line topology is greatly changed, the evaluation effect of using only the voltage amplitude is relatively general; if only the voltage amplitude and the node input active power or reactive power or the voltage amplitude and power factor are used, the model accuracy will decrease, indicating that the power amplitude alone cannot well reflect the stability of the unknown system, and to a certain extent, it is used as a fuzzy label or an error label in the model training, indicating that the feature is weakly correlated with the system stability feature, which easily leads to the model to make an incorrect prediction. Using voltage amplitude and power angle as feature quantities has the best evaluation accuracy. Compared with the power of the node, the power angle can significantly reflect certain faults or abnormal conditions in the power system, which helps the model capture key features related to the stability of the power system, or extract more implicit information to improve the accuracy of model evaluation.

[0112] Sampling window length comparison

[0113] Perform the shortest time window search for samples input into different models. Set the fault trigger time T error At the end of 1s; right boundary T right0 =2s; Confidence Acc = 95%; Accuracy δ = 0.02. The intercepted data window is brought into the model, and it is iterated repeatedly until the difference between the two data window lengths meets the accuracy requirements. The comparison of transient voltage stability assessment under the shortest sampling window length of different models is shown in Table 4.

[0114] Table 4 Comparison of sampling window lengths of different models

[0115]

[0116] By comparing the length of the time windows of each model, the following conclusions can be drawn:

[0117] It can be seen that when the sampling window length increases and the data of the input feature set increases, the accuracy of model evaluation is improved. In summary of the above five models, the CNN-BiLSTM model can reach a confidence level of 95 with only a time window of 0.38s. Compared with the CNN and BiLSTM models, the CNN-BiLSTM model combines the strengths of both, that is, it can effectively extract local features of data and process time series in both directions, so it can capture more time series information in the same length window, and obtain the best accuracy and the shortest time window. In addition, compared with traditional machine learning algorithms DT and SVM, when the time window reaches 0.68s, the confidence of the DT and SVM models only reaches 95%, while the CNN-BiLSTM model has a stability judgment accuracy of 98.64%. This shows that the model has a strong ability to extract time series features and can meet the requirements of transient voltage stability assessment.

[0118] Model transferability verification

[0119] The generalization performance of the constructed CNN-BiLSTM model in large-scale systems was verified by using the China Electric Power Research Institute Voltage Collapse Example (CEPRI-VC) test system. The system includes a total of 100 nodes. The voltage level of the main grid nodes is 500kV; the system's installed capacity of new energy is 2.4 million kilowatts, and the installed capacity of traditional energy is 6.3 million kilowatts. Among them, new energy power generation includes wind turbines with an output of 1.2 million kilowatts, photovoltaic units with an output of 600,000 kilowatts, and traditional energy power generation is all thermal units, generating 3.86 million kilowatts: the system contains a single-circuit DC line with a DC transmission power of 800,000 kilowatts. Similarly, the unit output is adjusted to simulate the situation where the peak and valley changes of the sending and receiving end power grid are large, and 10 sets of operation modes are set based on the three conditions of peak, valley, and flat valley.

[0120] Table 5 CNN-BiLSTM model generalization test

[0121]

[0122] The following conclusions can be drawn from the above table:

[0123] The CNN-BiLSTM model trained by homogenized data still performs well in transient voltage stability judgment for complex and large-scale power systems. For the above 10 different operating conditions, the accuracy of all evaluations can reach 97.8% or above, while the missed alarm rate is less than 1.5% and the false alarm rate is less than 0.1%. The CNN-BiLSTM model also has good evaluation capabilities for models with unknown topological changes; and when the output load of new energy units changes, it can still achieve a high evaluation accuracy. It can be seen that the CNN-BiLSTM model, which uses the characteristic quantity of the combination of voltage and power angle, has good generalization and can judge the stability of systems in new scenarios and models with easily changing operating conditions.

[0124] The present invention proposes a new CNN-BiLSTM model and applies it to the transient voltage stability assessment of a new power system. The Pearson rule is used to perform a rough screening of the feature set, and the shortest time window search is performed based on the dichotomy method to select the optimal feature subset for transient voltage stability assessment, which can greatly reduce the dimension of the feature subset, improve the assessment accuracy and reduce the online assessment calculation time. In addition, the generalization of the model is improved, and the transient voltage stability judgment can be performed in another model with a completely different topology without training, and with good accuracy.

Claims

1. A method for evaluating transient voltage stability of a power system based on convolutional neural network and bidirectional long short-term memory neural network, the method comprising the following steps: Step 1: obtain simulation results under different operation modes and different fault conditions through standard example simulation; perform Pearson correlation coefficient dimensionality reduction on the obtained feature quantities to obtain feature quantities such as node voltage and phase angle that are strongly correlated with transient voltage stability; perform homogenization on the obtained feature quantities to obtain the required feature set samples; Step 2: Input the feature set samples obtained in step 1 into the CNN-BiLSTM model for training; and use the binary search method to find the shortest time window required by the model; Step 3: When applying this method, there is no need to train the transient stability assessment model trained by using the real-time data measured online as the initial feature input. The transient voltage stability of the system can be obtained by predicting the stability of the system.

2. According to the method for evaluating transient voltage stability of a power system based on convolutional neural network and bidirectional long short-term memory neural network according to claim 1, the step 2 specifically comprises: Step 21: partition the feature set into stability margins using a single binary table practical criterion; Step 22: Input the feature set samples into the CNN model to minimize the loss function, realize time series data processing and automatic feature extraction, and reduce the multi-dimensional input to one-dimensional input; Step 23: Import the extracted input into the BiLSTM model to capture information from both directions of the sequence, and use the correlation between the front and back ends of the input sequence information to enable the model to make more accurate predictions; Step 24: After the initial prediction is completed, the model performance is evaluated by accuracy, missed detection rate and false detection rate; The shortest time window is searched through binary search; repeated iterations are performed to ensure that the model meets the accuracy requirements.

3. According to the method for evaluating transient voltage stability of a power system based on convolutional neural network and bidirectional long short-term memory neural network according to claim 2, the step 22 specifically comprises: Step 221: Input feature is X = [x1, x2, x3, ... x n ] T is a one-dimensional array as the input layer of the model. t×d , t is the length of the time series, and d is the number of input feature arrays. The obtained one-dimensional input signal is mapped to the convolution layer through a one-dimensional convolution operation, which is expressed as: f r (z)=max(z,0) (2) Where: is the convolution operator; Represented by the convolution kernel The calculated j-th feature map, j∈[1,n c ], where n c is the number of convolution kernels; is a weight matrix, m is the convolution kernel size, which determines the length of the local time window corresponding to the convolution kernel when extracting features; b is the bias term; Step 222: The pooling layer can reduce the size of the feature space and the number of network parameters through downsampling operations, thereby reducing the computational complexity and helping to prevent overfitting. In the maximum pooling, the maximum value in the window is selected as the output value, which is expressed as in To select the maximum value of two consecutive points in the sequence, the sequence length is reduced by half. The above formula is global max-pooling, which is the last operation of the pooling layer. The function of this formula is to extract the most useful global time series information and reduce the sequence length to 1; Step 223: The fully connected layer further abstracts and combines the global temporal features by performing matrix concatenation on the feature quantities extracted by the convolutional layer and the pooling layer, and is expressed as: a fc =f r (a p,last W fc +b) (5) In the above formula: W fc ∈R nfc is the weight matrix in the fully connected layer; n fc As the hidden layer, the obtained feature quantity is passed through the fully connected layer to output the obtained classification: Where f a (z) is the activation function sigmoid function, and the output is That is, the probability distribution of different voltage stability conditions, and the transient voltage stability is judged according to the obtained probabilities of different voltage stability conditions.

4. The method for evaluating transient voltage stability of a power system based on convolutional neural network and bidirectional long short-term memory neural network according to claim 2 is characterized in that In step 23, the input quantity is extracted and imported into the BiLSTM model to capture information from two directions of the sequence, namely In the formula, x t is the input of the model at time t, h t is the output of the model at time t; where i t 、f t , o t By inputting x t and the hidden layer output h of the previous step t-1 Perform linear transformation, and then obtain the output of the input gate, forget gate, and output gate through the activation function; W i , W f , W o , W c is the weight matrix of the input gate, forget gate, and output gate; U i , U f , U o , U c is the coupling matrix, b f 、b i 、b o 、b c It is the memory bias of input gate, forget gate and output gate.

5. According to the method for evaluating transient voltage stability of a power system based on convolutional neural network and bidirectional long short-term memory neural network in claim 1, the label outputted from step 3 is substituted into the formula to obtain the stability label, that is, Where: V min is the minimum node voltage after the fault occurs; V th is the per-unit value of the transient voltage threshold, which is set to 0.75 as required; T th The duration of operation below the transient voltage threshold is set to 1s as required; T s is the time when the bus voltage is lower than the threshold; k is the critical voltage deviation time factor; where η represents the transient voltage stability margin, and the smaller η is, the worse the system stability margin is. Based on the value of η, four stability conditions are constructed: super stable system, stable system, critical stable system, and unstable system; a stability margin function is constructed to classify systems in different stability conditions and normalize them with P so that P∈[0,3]; that is, P=me -η +n (10) Where: m and n are unknown coefficients, which can be calculated through the critical conditions of the super-stable system and the unstable system; P th1 With P th2 are the thresholds of stable state and critical stable state respectively; P=0 sample is an ultra-stable sample, P=1 sample is a stable sample with a small probability of system instability, P=2 sample is a critical state sample with a large risk of instability, and P=3 sample is an unstable sample.

Images