Transient voltage evaluation method and system based on data enhancement and bidirectional gating circulation

Abstract

The invention discloses a transient voltage evaluation method and system based on data enhancement and bidirectional gating circulation. Relates to the technical field of power system transient evaluation. The method level and the network structure level are improved on the basis of a traditional power system transient stability evaluation technology, data enhancement is carried out on an original data set based on an improved deep learning network model, and the diversity of samples is enriched; meanwhile, in order to cope with inherent characteristics of multi-characteristic time sequence sample data, potential information is extracted in multiple directions by utilizing a transient voltage evaluation model of a bidirectional gating circulation unit network based on an attention mechanism, and the attention mechanism improves the characteristic learning ability by automatically distributing attention according to the importance of input information; and finally, based on an improved focus loss function, the problems of tendency and misjudgment caused by sample imbalance in the training process of the transient voltage stability evaluation model are relieved to a certain extent through a weight coefficient and a modulation factor.

Description

Technical Field

[0001] The present invention relates to the technical field of power system transient assessment, and in particular to a transient voltage assessment method and system based on data enhancement and bidirectional gating cycle. Background Art

[0002] With the active construction of new power systems based on new energy, a high proportion of new energy such as photovoltaic and wind power has gradually replaced traditional thermal power. At the same time, this also poses a serious threat to the voltage stability of modern power systems. Short-term voltage stability assessment has always been considered a key task to ensure the safe and stable operation of power systems. The present invention aims to accurately and timely detect the short-term voltage stability state of power systems.

[0003] The transient stability assessment method of power system based on one-dimensional convolutional neural network provides the one-dimensional convolutional neural network with all the time series information of the transient process by constructing the initial features of the full-stage time series, and classifies the transient stability. Then, the model is trained by generating training samples through time domain simulation. Finally, an example analysis is carried out on the New England 10-machine 39-node system to verify the effect of accurately characterizing the transient stability state of the system.

[0004] The transient stability assessment method of power system based on bidirectional long short-term memory network establishes a nonlinear mapping relationship between the underlying measurement data and the transient stability category of the power system through the Bi-LSTM network, and uses the accuracy, F1 index and FPR index to comprehensively evaluate the performance of the Bi-LSTM network model. On this basis, the t-distributed random neighbor embedding (t-SNE) dimension reduction method and k-nearest neighbor (KNN) classifier are used to further improve the accuracy of transient stability assessment. The assessment model is analyzed by visualization method and network prediction score. The results show that the Bi-LSTM network model has a strong ability to extract transient process features of the power system and is suitable for the assessment of transient stability of the power system.

[0005] However, the transient stability assessment method of power system based on one-dimensional convolutional neural network does not consider the influence of noise and missing values ​​in the measurement data of phasor measurement unit on the generalization ability of the model, and its sample coverage is limited and cannot well verify the robustness of the assessment under complex disturbances of the power grid. The transient stability assessment method of power system based on bidirectional long short-term memory network takes into account the factors before and after the sequence. Therefore, while improving the accuracy, the time cost of model training will also increase. In addition, a large amount of transient data is required when training the network model, and the process of collecting these data is extremely cumbersome and inefficient; in the power system, voltage instability is usually rare, which leads to a natural imbalance between voltage instability samples and voltage stability samples. If effective measures are not taken, the model may be biased towards stable samples during training. When instability occurs, if the model mistakenly judges it as stable, it may cause significant economic losses to the power grid. Summary of the invention

[0006] The technical problem to be solved by the present invention is that the evaluation model of the traditional power system transient stability evaluation method is greatly affected by data noise and missing values, and is highly dependent on transient data. There is a natural imbalance between voltage instability samples and voltage stability samples, which leads to defects in model accuracy. The purpose of the present invention is to provide a transient voltage evaluation method and system based on data enhancement and bidirectional gated loop, to improve the traditional power system transient stability evaluation technology at the method level and network structure level, to enhance the original data set based on the improved deep learning network model, and to enrich the diversity of samples. At the same time, in order to cope with the inherent characteristics of multi-feature time series sample data, a transient voltage evaluation model of a bidirectional gated recurrent unit network based on an attention mechanism is used to extract potential information in multiple directions, and the attention mechanism improves the feature learning ability by automatically allocating attention according to the importance of the input information.

[0007] The present invention is achieved through the following technical solutions:

[0008] This solution provides a transient voltage assessment method based on data enhancement and bidirectional gating cycles, including:

[0009] Based on historical data, a time domain simulation of the power grid to be evaluated is performed to generate an original data set;

[0010] Performing data enhancement processing on the original data set;

[0011] A transient voltage assessment model is trained based on the original data set and the original data set after data enhancement processing; the transient voltage assessment model is a bidirectional gated recurrent unit network based on an attention mechanism;

[0012] The real-time data of the power grid to be evaluated is collected, and the real-time data is subjected to data enhancement processing and then input into a trained transient voltage evaluation model for transient voltage evaluation.

[0013] Working principle of this scheme: The evaluation model of the traditional power system transient stability assessment method is greatly affected by data noise and missing values, and is highly dependent on transient data. There is a natural imbalance between voltage instability samples and voltage stability samples, which leads to defects in model accuracy. The purpose of the present invention is to provide a transient voltage assessment method based on data enhancement and bidirectional gated loop, to improve the traditional power system transient stability assessment technology at the method level and network structure level, to enhance the original data set based on the improved deep learning network model, and to enrich the diversity of samples. At the same time, in order to cope with the inherent characteristics of multi-feature time series sample data, a transient voltage assessment model based on a bidirectional gated recurrent unit network of an attention mechanism is used to extract potential information in multiple directions. The attention mechanism improves the feature learning ability by automatically allocating attention according to the importance of the input information. Finally, based on the improved focal loss function, the weight coefficient and the modulation factor are used to alleviate the tendency and misjudgment problems due to sample imbalance in the training process of the transient voltage stability assessment model to a certain extent.

[0014] A further optimization scheme is that the time domain simulation of the power grid to be evaluated based on historical data generates an original data set, including the following method:

[0015] Collecting three-stage electrical characteristic quantities of the power grid to be evaluated; the three-stage electrical characteristic quantities represent electrical characteristic quantities before, during and after the fault is removed, and the electrical characteristic quantities include: node voltage amplitude, phase angle, branch active power, and branch reactive power;

[0016] The original data set S is obtained by performing time domain transient stability simulation on the power grid to be evaluated:

[0017] S={S1,S2,...,S f ,...S i ,...S c ,...,S T}

[0018] i=1,2,…,T;for the original data S at time i i The set is represented as:

[0019] S i = {U 1i ,U 2i ,...U ni ,θ 1i ,θ 2i ,...θ ni ,

[0020] P 1i ,P 2i ,....P li ,Q 1i ,Q 2i ,....Q li}

[0021] Among them, S c , S f Represent the fault start and fault removal time respectively; T represents the total number of samples; U ni ,θ ni , P ni and Q ni They respectively represent the voltage amplitude, phase angle, branch active power and branch reactive power of the grid node n to be evaluated at time i.

[0022] A further optimization scheme is that the data enhancement processing is performed on the original data set, including the method:

[0023] The original data set is normalized based on the following formula;

[0024] The normalized original data set is input into the constructed generative deep learning network model for data enhancement; during the construction process of the generative deep learning network model, the maximum mean difference and FID similarity are used as evaluation indicators.

[0025] A further optimization scheme is that the original data set is normalized based on the following formula, including a method:

[0026] The original data set is normalized according to the following formula:

[0027]

[0028] Among them, x represents the transient time series data that needs to be standardized; μ represents the mean value of the time series data; and σ represents the standard deviation of the time series data.

[0029] A further optimization scheme is that the bidirectional gated recurrent unit network based on the attention mechanism includes:

[0030] The input layer is used to construct the input feature vector based on the original data set after data augmentation;

[0031] A BiGRU layer, for extracting latent features and hidden features from an input feature vector; the hidden state of the BiGRU layer is affected by both a forward hidden state and a reverse hidden state;

[0032] An attention layer, used to increase the weight coefficient of the hidden feature;

[0033] The fully connected layer is used for transient voltage stability judgment, and outputs the transient voltage stability judgment result as a probability value of 0 to 1 based on the output of the softmax activation function.

[0034] A further optimization scheme is that the method for determining transient voltage stability includes:

[0035] The transient voltage stability is determined according to the following formula:

[0036]

[0037] Among them U min Indicates the minimum voltage value during the transient period; U s Indicates the preset voltage threshold; T u Indicates the time during which the voltage during the transient period is lower than the set voltage threshold; T s Indicates that the voltage is allowed to be lower than U s The longest time; U STVSA Indicates the transient voltage stability state, 1 indicates stability, and 0 indicates instability.

[0038] A further optimization scheme is that the transient voltage assessment model is trained based on the original data set and the original data set after data enhancement processing, including the method:

[0039] The transient voltage assessment model is trained with the original data set after data enhancement;

[0040] The original data set is input into the trained transient voltage assessment model, and the parameters of the transient voltage assessment model are adjusted according to the preset evaluation indicators.

[0041] A further optimization scheme is that the preset evaluation indicators include: precision Pre, recall rate Rec and F1 value;

[0042]

[0043] Where: Acc represents the accuracy; TP represents the number of truly stable and predicted stable results; FN represents the number of truly unstable and predicted stable results; FP represents the number of truly stable and predicted unstable results; TN represents the number of truly unstable and predicted unstable results.

[0044] A further optimization scheme is to use an improved focus loss function as the loss function during the training of the transient voltage assessment model; the improved focus loss function L fl It is expressed as:

[0045]

[0046] Where: α represents the weight factor; γ represents the modulation factor; y′ represents the predicted value of the sample; y represents the label of the real sample; ln represents the logarithmic function.

[0047] The present solution also provides a transient voltage evaluation system based on data enhancement and bidirectional gating cycle, which is used to implement the above-mentioned transient voltage evaluation method based on data enhancement and bidirectional gating cycle; the system comprises:

[0048] A data generation module, used for generating original data sets based on time domain simulation;

[0049] A data enhancement module, used to perform data enhancement processing on the original data set;

[0050] A training module, used for inputting the original data set and the original data set after data enhancement processing into a transient voltage assessment model for training, wherein the transient voltage assessment model is a bidirectional gated recurrent unit network based on an attention mechanism;

[0051] The evaluation module is used to perform online transient voltage evaluation based on the trained transient voltage evaluation model.

[0052] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0053] The present invention provides a transient voltage assessment method and system based on data enhancement and bidirectional gated loop, improves the traditional power system transient stability assessment technology at the method level and network structure level, and performs data enhancement on the original data set based on the improved deep learning network model to enrich the diversity of samples; at the same time, in order to cope with the inherent characteristics of multi-feature time series sample data, a transient voltage assessment model of a bidirectional gated recurrent unit network based on an attention mechanism is used to extract potential information in multiple directions, and the attention mechanism improves the feature learning ability by automatically allocating attention according to the importance of the input information; finally, based on the improved focus loss function, the tendency and misjudgment problems due to sample imbalance in the training process of the transient voltage stability assessment model are alleviated to a certain extent through the weight coefficient and the modulation factor. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other relevant drawings can be obtained based on these drawings without creative work. In the drawings:

[0055] Figure 1 It is a flow chart of a transient voltage assessment method based on data enhancement and bidirectional gating cycle;

[0056] Figure 2 Schematic diagram of transient voltage evaluation principle based on data enhancement and bidirectional gating cycle;

[0057] Figure 3 It is a schematic diagram of the transient voltage assessment model architecture;

[0058] Figure 4 This is a schematic diagram of the improved New England 39-node system results;

[0059] Figure 5 It is a schematic diagram of the original data;

[0060] Figure 6 This is a schematic diagram of the original data distribution after data enhancement processing;

[0061] Figure 7 Schematic diagram of the output data distribution of the BiGRU layer;

[0062] Figure 8 Schematic diagram of FC layer output data distribution. DETAILED DESCRIPTION

[0063] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.

[0064] The evaluation model of the traditional power system transient stability evaluation method is greatly affected by data noise and missing values, and is highly dependent on transient data. There is a natural imbalance between voltage instability samples and voltage stability samples, which leads to defects in model accuracy.

[0065] In view of this, this solution provides the following embodiments to solve the above technical problems:

[0066] Example 1

[0067] This embodiment provides a transient voltage evaluation method based on data enhancement and bidirectional gating cycle, such as Figure 1 and Figure 2 As shown, including:

[0068] Step 1: Based on historical data, a time domain simulation is performed on the power grid to be evaluated to generate an original data set. This step specifically includes the following methods:

[0069] S11, collecting three-segment electrical characteristic quantities of the power grid to be evaluated; the three-segment electrical characteristic quantities represent the electrical characteristic quantities before, during and after the fault is removed, and the electrical characteristic quantities include: node voltage amplitude, phase angle, branch active power, branch reactive power; in order to train the evaluation model of the full link of transient voltage evaluation, the electrical characteristics obtained by time domain simulation are used as the input of the model, and the final evaluation effect depends to a large extent on the selection of input characteristics. Therefore, the input characteristics selected in this scheme need to be able to accurately reflect the transient characteristics of the power system so as to accurately judge the transient process; the electrical characteristic quantities that accurately reflect the transient voltage are used as input characteristics.

[0070] S12, perform time domain transient stability simulation on the power grid to be evaluated to obtain the original data set S:

[0071] S={S1,S2,...,S f ,...S i ,...S c ,...,S T}

[0072] i=1,2,…,T;for the original data S at time i i The set is represented as:

[0073] S i = {U 1i ,U 2i ,...U ni ,θ 1i ,θ 2i ,...θ ni ,

[0074] P 1i ,P 2i ,....P li ,Q 1i ,Q 2i ,....Q li}

[0075] Among them, S c , S f Represent the fault start and fault removal time respectively; T represents the total number of samples; U ni ,θ ni , P ni and Q ni They respectively represent the voltage amplitude, phase angle, branch active power and branch reactive power of the grid node n to be evaluated at time i.

[0076] In order to unify the dimensions of various electrical quantities, the collected data needs to be standardized.

[0077] Step 2: Perform data enhancement processing on the original data set; this step specifically includes the following method:

[0078] The original data set is normalized based on the following formula:

[0079]

[0080] Among them, x represents the transient time series data that needs to be standardized; μ represents the mean value of the time series data; and σ represents the standard deviation of the time series data.

[0081] Mapping data to the [0,1] interval can improve the accuracy and convergence speed of the model.

[0082] The normalized original data set is input into the constructed generative deep learning network model for data enhancement; during the construction process of the generative deep learning network model, the maximum mean difference and FID similarity are used as evaluation indicators.

[0083] In this embodiment, the deep learning network model is used for data enhancement, and it plays an important role as a link between the original data set and the deep learning technology in the proposed transient voltage stability assessment process. Specifically, data enhancement uses the original data set obtained by time domain simulation as input and outputs a large amount of high-quality labeled data, which provides the necessary prerequisite for training deep learning classification models in terms of data resources. The large training data set obtained by data enhancement is used as the input of the evaluation model of the bidirectional gated recurrent unit network based on the attention mechanism.

[0084] The deep learning network model structure used in this embodiment is shown in Table 1 below:

[0085] Table 1 Deep learning network model structure

[0086]

[0087] This scheme uses maximum mean difference (MMD) and FID similarity as quantitative evaluation indicators to verify the diversity and quality of generated samples obtained by data augmentation based on deep learning network models.

[0088] The calculation formula of the maximum mean difference MMD is as follows:

[0089]

[0090] Where: P data represents the distribution of the original data; P g Represents the distribution of generated data; Represents the distribution P data The expected value of the kernel function of the samples x and x′ extracted from ; Represents the distribution P data The sample x is drawn from the distribution P g The expected value of the kernel function of the sample y extracted from ; From the distribution P g The expected value of the kernel function of the samples y and y′ drawn from . If P data and P g The closer the two distributions are, the lower the MMD value is, indicating that the quality of the generated samples is better.

[0091] The calculation of the maximum mean difference (MMD) involves the mean and covariance of the features, and the formula is:

[0092]

[0093] Where: Represents the characteristic mean of the original data; represents the characteristic mean of the generated data; Σ represents the covariance matrix; Tr represents the trace of the matrix.

[0094] Since the transient time series data constructed by this solution is similar to a two-dimensional grayscale image, it can be processed as an image. Therefore, FID can be used to judge the quality of the generated data. The lower the FID value, the better the quality and diversity of the generated data. At the same time, FID can also effectively reflect the model collapse problem.

[0095] Step 3: training a transient voltage assessment model based on the original data set and the original data set after data enhancement processing; the transient voltage assessment model is a bidirectional gated recurrent unit network based on an attention mechanism; Figure 3 As shown in Figure 1, the bidirectional gated recurrent unit network based on the attention mechanism includes:

[0096] The input layer is used to construct the input feature vector based on the original data set after data enhancement. In the input layer, x t It represents the value of all electrical quantities of the system at time t for a sample, and it is an m-dimensional feature vector.

[0097] The BiGRU layer is used to extract the forward hidden features and the reverse hidden features from the input feature vector; the hidden state of the BiGRU layer is affected by both the forward hidden state and the reverse hidden state; the BiGRU layer is designed to extract the potential time dependency from the input features. Different from the original GRU, the hidden state h of the BiGRU layer in this scheme t It receives the influence of both the forward hidden state and the reverse hidden state at the same time. This design enables BiGRU to capture dependencies more comprehensively.

[0098] The attention layer is used to increase the weight coefficients of the forward hidden features and the reverse hidden features; the attention mechanism automatically gives high weight coefficients to those hidden features that are closely related to the target. By highlighting the information that is critical to the stable state evaluation performance, the attention layer improves the performance of the model.

[0099] The fully connected layer is used for transient voltage stability judgment, and outputs the transient voltage stability judgment result as a probability value of 0 to 1 based on the output of the softmax activation function.

[0100] Defining a classification threshold can convert the probability value of the model output into a label that is stable at 1 or unstable at 0:

[0101]

[0102] Where: output is the model output result, by setting the classification threshold δ, output pred In order to finally output stable and unstable labels, the classification threshold δ defaults to 0.5.

[0103] The method for determining transient voltage stability comprises:

[0104] Transient voltage stability determination is a binary classification problem. Stable sample voltage is marked as 1, and unstable sample voltage is marked as 0. Transient voltage stability determination is performed according to the following formula:

[0105]

[0106] Among them U min Indicates the minimum voltage value during the transient period; U s Indicates the preset voltage threshold; T u Indicates the time during which the voltage during the transient period is lower than the set voltage threshold; T s Indicates that the voltage is allowed to be lower than U s The longest time; U STVSA Indicates the transient voltage stable state, 1 indicates stability, and 0 indicates instability. This solution sets T according to the engineering practical criterion. s Indicates 10s, U s is 0.8pu, and as long as the voltage of one node meets the instability criterion, the system is considered to have voltage instability.

[0107] The method of training a transient voltage assessment model based on the original data set and the original data set after data enhancement processing includes:

[0108] S31, training a transient voltage assessment model using the original data set after data enhancement processing;

[0109] S32, inputting the original data set into the trained transient voltage assessment model, and adjusting the parameters of the transient voltage assessment model according to the preset evaluation index.

[0110] In the transient voltage stability assessment and analysis, "misjudgment of stability" and "missed judgment of stability" have different impacts on the final power system. The former will have a huge impact on the system, while the latter will basically not cause any harm to the system. Taking the above into consideration, it is difficult to accurately evaluate the quality of the transient voltage stability assessment model using only conventional accuracy. Therefore, this solution also introduces indicators that can reflect unbalanced samples and the cost of misjudgment. The corresponding preset evaluation indicators include: precision Pre, recall rate Rec, and F1 value;

[0111]

[0112] Where: Acc represents the accuracy; TP represents the number of true stable and predicted stable results; FN represents the number of true unstable and predicted stable results; FP represents the number of true stable and predicted unstable results; TN represents the number of true unstable and predicted unstable results; the corresponding confusion matrix is ​​shown in Table 2:

[0113] Table 2 Confusion matrix

[0114]

[0115] In the binary classification problem, the commonly used loss function is the cross entropy loss function. However, in the transient voltage stability assessment, there are problems such as sample imbalance causing serious consequences of stability misclassification. Therefore, in the process of training the transient voltage assessment model, this scheme uses the improved focus loss function as the loss function; the improved focus loss function introduces a weight factor on the basis of the original one, which can effectively solve the problem of sample imbalance. The improved focus loss function L fl It is expressed as:

[0116]

[0117] In the formula: α represents the weight factor; γ represents the modulation factor; y′ represents the predicted value of the sample; y represents the label of the real sample; ln represents the logarithmic function; the weight of the minority samples is increased by changing the value of the weight factor α; the modulation factor γ is used to increase the weight of the difficult classification.

[0118] Step 4: collect real-time data of the power grid to be evaluated, perform data enhancement processing on the real-time data, and then input the data into a trained transient voltage evaluation model to perform transient voltage evaluation.

[0119] The real-time data of the power grid to be evaluated is collected in real time by the phase measurement unit (PMU). The collected real-time data is processed into the data format required by the model, and then input into the trained transient voltage evaluation model to determine the transient voltage results of the system. If the evaluation results show that the system cannot maintain a stable state, remedial control measures must be taken to prevent the system from immediately experiencing a voltage collapse; otherwise, the evaluation model will continue to monitor the system's stable state. Note that in the online application stage, the trained model can be updated regularly to improve the adaptability of the proposed intelligent system to emergencies under various operating conditions.

[0120] Example 2

[0121] In order to verify the effectiveness of the deep learning network model of the embodiment, this embodiment compares it with the generative adversarial network model GAN ​​under the same conditions. In addition, in order to ensure the reliability of the experiment, the present invention uses different random seeds and conducts multiple experiments, and the average values ​​are shown in Table 3 below.

[0122] Compared with the generative adversarial network model GAN, it can be seen that the MMD value of the deep learning network model (WGAN) is reduced by 79.7%, and the FID value is reduced by 65.9%, showing better performance in terms of affinity, quality, and diversity. In other words, the deep learning network model (WGAN) can generate synthetic data with high-quality data, showing its outstanding advantages in data enhancement.

[0123] Table 3 Comparison of MMD and FID between WGAN and GAN

[0124]

[0125] In order to verify the superiority of data enhancement and the transient voltage stability assessment model proposed in the present invention, one-dimensional convolutional neural network (CNN-1d), two-dimensional convolutional network (CNN-2d), and LSTM are selected for comparative testing. Among them, the one-dimensional convolution adopts three layers of one-dimensional convolution, and each layer of convolution is connected to the maximum pooling layer and the Dropout layer. Finally, there are two layers of fully connected layers. The two-dimensional convolution process is similar, and the original data is processed into a single-channel data format similar to a grayscale image. The number of LSTM layers and the number of neurons are consistent with the model proposed in this article; among them, BIGRU-ATTENTION1 means that data enhancement is not used; BIGRU-ATTENTION2 represents the model after data enhancement by the deep learning network model (WGAN). The present invention builds a model based on Pytorch, and the simulation hardware uses Intel (R) Xeon (R).

[0126] It can be seen from Table 4 that except for the relatively poor performance of CNN-2d, the accuracy of other models has reached more than 96%.

[0127] Table 4 Evaluation performance indicators of each model

[0128]

[0129] This is because the two-dimensional convolution does not consider the time series characteristics when processing the data of this solution, but simply performs transient voltage evaluation according to the feature extraction similar to image processing, so it performs the worst among the above models. The BIGRU-ATTENTION1 proposed in this solution with bidirectional transmission and attention mechanism is higher than LSTM and CNN-1d in all indicators. Finally, after data enhancement, BIGRU-ATTENTION2 has improved the model accuracy, precision, recall rate, and F1 by 0.0149, 0.0106, 0.0071, and 0.0089 compared with BIGRU-ATTENTION1, which is the highest among all models.

[0130] Since the phase measurement unit (PMU) is inevitably affected by measurement noise during the actual sampling process, in order to evaluate the robustness of the model under noise conditions, Gaussian white noise with different signal-to-noise ratios (SNR) is included in the original data set. According to the IEEE C37.118 standard, the measurement error of the phase measurement unit should be less than 1%, according to the signal-to-noise ratio calculation formula:

[0131]

[0132] Where: P S Indicates the effective power of the signal; P N Indicates the effective power of noise, 1% noise is converted to a signal-to-noise ratio of 20dB. The indicators of the model under different signal-to-noise ratios are shown in Table 5 below.

[0133] Table 5 Performance comparison under different signal-to-noise ratios

[0134]

[0135]

[0136] It can be seen from Table 5 that after data enhancement, although the evaluation accuracy will decrease slightly with the increase of background noise level, it is still above 97% in various noise environments. Similarly, the other three indicators Pre, F1-score and Rec also maintain a high level, which fully verifies that the scheme has good robustness to noise.

[0137] Example 3

[0138] This embodiment provides a transient voltage evaluation system based on data enhancement and bidirectional gating cycle, which is used to implement the transient voltage evaluation method based on data enhancement and bidirectional gating cycle described in Example 1; the system includes:

[0139] A data generation module, used for generating original data sets based on time domain simulation;

[0140] A data enhancement module, used to perform data enhancement processing on the original data set;

[0141] A training module, used for inputting the original data set and the original data set after data enhancement processing into a transient voltage assessment model for training, wherein the transient voltage assessment model is a bidirectional gated recurrent unit network based on an attention mechanism;

[0142] The evaluation module is used to perform online transient voltage evaluation based on the trained transient voltage evaluation model.

[0143] Example 4

[0144] This embodiment selects the improved New England 39-node system (such as Figure 4 As shown in the figure, 34 nodes are replaced by wind turbines) for transient simulation. The total simulation time is set to 10 seconds, and sampling is performed every 0.01 seconds. Taking the load of the original network as the benchmark load, five different load levels of 85%, 95%, 100%, 105% and 110% are set, and three different wind power penetration rates of 5%, 10% and 15% are considered, and the output of each generator is adjusted accordingly to ensure that the power flow converges and the bus voltage remains within the normal range. The most serious short-circuit fault (three-phase short-circuit fault) is set for the 33 AC lines in the system, and the fault is set at 10%, 30%, 50%, 70% and 90% of the line respectively. The fault duration is 8 to 10 cycles, and the system frequency is 60Hz. A total of 4455 samples are generated in the PSASP software, of which 3442 are stable samples and 1013 are unstable samples. The samples generated by the simulation are divided into training set and test set according to 7:3. The training set is fed into the deep learning network model, and finally 10,326 training samples are generated.

[0145] In order to more intuitively understand the distribution of original data and the processing results of each layer of the model on the data, this paper adopts the t-distribution random neighbor embedding algorithm to compress the original data and each layer of the model into a two-dimensional plane. The data distribution of each layer in the two-dimensional space is as follows: Figure 5-Figure 8 As shown. Figure 5 It can be seen that after the original data is compressed into two-dimensional space through multi-dimensional features, the samples are mixed with each other, the distribution is discrete, and it is difficult to identify stable boundaries. Figure 6It shows that the generated data not only maintains the discreteness of the original data, but also significantly increases the diversity of samples. Figure 7 and Figure 8 The following figure shows the visualization results of sample features extracted by the GRU layer and processed by the fully connected layer. In these two figures, the output results show obvious stable boundaries. This shows that GRU can effectively extract high-dimensional time series features, convert them into low-dimensional features, and establish a good mapping relationship with the original labels, thereby significantly improving the evaluation performance.

[0146] For complex power networks with high-dimensional nonlinearity, it is difficult for transient voltage assessment models to achieve 100% assessment accuracy, so there will be cases of false judgment of stability. However, this paper enhances the number of overall samples through data enhancement, so that the model can extract more features of unstable samples and reduce the number of false judgments of stability. The confusion matrix with and without data enhancement is shown in the following table.

[0147] Table 6 Confusion matrix before and after data augmentation

[0148]

[0149] As can be seen from Table 6, after using data enhancement, the number of false positives dropped from 20 to 8, reducing the false positive rate by 60%. At the same time, the model is easier to grasp the instability situation, and the number of accurate instability judgments increased from 190 to 202, an increase of 6.32%.

[0150] In actual power systems, the probability of transient voltage instability is very small. Therefore, it is necessary to consider the evaluation ability of the model when the data is extremely unbalanced. In the original data set, the ratio of stable and unstable samples was selected as 1:5, 1:7, and 1:9 for testing. The research results are shown in Table 7 below:

[0151] In Table 7, model 1 is BiGRU-Attention using cross entropy as the loss function, and model 2 is BiGRU-Attention using the improved focus function. It is not difficult to see from the table that when the ratio of positive samples to negative samples in the original data set increases, the use of traditional loss functions will cause the model performance to deteriorate seriously. The improved focus loss function proposed in this paper improves the tendency of model training by adjusting the weight factor and reduces model misjudgment by adjusting the factor, which can well ensure the robust adaptability of the model to different ratios between positive and negative samples.

[0152] Table 7 Model evaluation indicators under imbalanced conditions

[0153]

[0154] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A transient voltage assessment method based on data enhancement and bidirectional gating cycle, characterized in that: include: Based on historical data, a time domain simulation of the power grid to be evaluated is performed to generate an original data set; Performing data enhancement processing on the original data set; Training a transient voltage assessment model based on the original data set and the original data set after data enhancement processing; The transient voltage assessment model is a bidirectional gated recurrent unit network based on an attention mechanism; The real-time data of the power grid to be evaluated is collected, and the real-time data is subjected to data enhancement processing and then input into a trained transient voltage evaluation model for transient voltage evaluation.

2. The transient voltage assessment method based on data enhancement and bidirectional gating cycle according to claim 1 is characterized in that: The method of generating an original data set by performing time domain simulation on the power grid to be evaluated based on historical data includes: Collecting three-stage electrical characteristic quantities of the power grid to be evaluated; the three-stage electrical characteristic quantities represent electrical characteristic quantities before, during and after the fault is removed, and the electrical characteristic quantities include: node voltage amplitude, phase angle, branch active power, and branch reactive power; The original data set S is obtained by performing time domain transient stability simulation on the power grid to be evaluated: S={S1,S2,...,S f ,...S i ,...S c ,...,S T } i=1,2,…,T;for the original data S at time i i The set is represented as: S i ={U 1i ,U 2i ,....U ni ,i 1i ,i 2i ,...i ni , P 1i ,P 2i ,....P li ,Q 1i ,Q 2i ,....Q li } Among them, S c , S f Represent the fault start and fault removal time respectively; T represents the total number of samples; U ni ,θ ni , P ni and Q ni They respectively represent the voltage amplitude, phase angle, branch active power and branch reactive power of the grid node n to be evaluated at time i.

3. The transient voltage assessment method based on data enhancement and bidirectional gating cycle according to claim 1 is characterized in that: The data enhancement processing of the original data set includes a method: The original data set is normalized based on the following formula; The normalized original data set is input into the constructed deep learning network model for data enhancement; during the construction process of the generated deep learning network model, the maximum mean difference and FID similarity are used as evaluation indicators.

4. The transient voltage assessment method based on data enhancement and bidirectional gating cycle according to claim 3 is characterized in that: The normalization process of the original data set based on the following formula includes a method: The original data set is normalized according to the following formula: Among them, x represents the transient time series data that needs to be standardized; μ represents the mean value of the time series data; and σ represents the standard deviation of the time series data.

5. The transient voltage assessment method based on data enhancement and bidirectional gating cycle according to claim 1 is characterized in that: The bidirectional gated recurrent unit network based on the attention mechanism includes: The input layer is used to construct the input feature vector based on the original data set after data augmentation; A BiGRU layer, for extracting latent features and hidden features from an input feature vector; the hidden state of the BiGRU layer is affected by both a forward hidden state and a reverse hidden state; An attention layer, used to increase the weight coefficient of the hidden feature; The fully connected layer is used for transient voltage stability judgment, and outputs the transient voltage stability judgment result as a probability value of 0 to 1 based on the output of the softmax activation function.

6. The transient voltage assessment method based on data enhancement and bidirectional gating cycle according to claim 5, characterized in that: The method for determining transient voltage stability comprises: The transient voltage stability is determined according to the following formula: Among them U min Indicates the minimum voltage value during the transient period; U s Indicates the preset voltage threshold; T u Indicates the time during which the voltage during the transient period is lower than the set voltage threshold; T s Indicates that the voltage is allowed to be lower than U s The longest time; U STVSA Indicates the transient voltage stability state, 1 indicates stability, and 0 indicates instability.

7. The transient voltage assessment method based on data enhancement and bidirectional gating cycle according to claim 1 is characterized in that: The method of training a transient voltage assessment model based on the original data set and the original data set after data enhancement processing includes: The transient voltage assessment model is trained with the original data set after data enhancement; The original data set is input into the trained transient voltage assessment model, and the parameters of the transient voltage assessment model are adjusted according to the preset evaluation indicators.

8. The transient voltage assessment method based on data enhancement and bidirectional gating cycle according to claim 7 is characterized in that: The preset evaluation indicators include: precision Pre, recall Rec and F1 value; Where: Acc represents the accuracy; TP represents the number of truly stable and predicted stable results; FN represents the number of truly unstable and predicted stable results; FP represents the number of truly stable and predicted unstable results; TN represents the number of truly unstable and predicted unstable results.

9. The transient voltage assessment method based on data enhancement and bidirectional gating cycle according to claim 1, characterized in that: In the process of training the transient voltage assessment model, an improved focus loss function is used as the loss function; the improved focus loss function L fl It is expressed as: Where: α represents the weight factor; γ represents the modulation factor; y′ represents the predicted value of the sample; y represents the label of the real sample; ln represents the logarithmic function.

10. A transient voltage assessment system based on data enhancement and bidirectional gating cycle, characterized in that: A transient voltage assessment method based on data enhancement and bidirectional gating cycle for implementing any one of claims 1 to 9; the system comprises: A data generation module, used for generating original data sets based on time domain simulation; A data enhancement module, used to perform data enhancement processing on the original data set; A training module, used for inputting the original data set and the original data set after data enhancement processing into a transient voltage assessment model for training, wherein the transient voltage assessment model is a bidirectional gated recurrent unit network based on an attention mechanism; The evaluation module is used to perform online transient voltage evaluation based on the trained transient voltage evaluation model.

Images