A method for classifying power system disturbances

Through the combination method of stacked denoising autoencoder and random forest classifier, the problem of disturbance classification of power system under the influence of PMU data loss is solved, real-time monitoring and accurate identification of dynamic behavior of power system is realized, and the robustness of the system is improved and identification accuracy.

CN112633315BActive Publication Date: 2024-11-26GUANGZHOU POWER SUPPLY BUREAU GUANGDONG POWER GRID CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202011132259.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2020-10-21
Publication Date
2024-11-26
Estimated Expiration
2040-10-21

AI Technical Summary

Technical Problem

The existing power system disturbance classification method is poor in handling PMU data loss, resulting in difficulty in monitoring the dynamic behavior of the power system. Especially in complex large systems and massive data environments, data quality problems seriously restrict their application.

Method used

Using a method based on stacked denoising autoencoder and random forest classifier, the disturbance data is generated through offline time domain simulation, and the high-level features are extracted using stacked denoising autoencoder, and the random forest classifier performs disturbance identification. Implement fast and accurate classification of PMU perturbation data containing lost data.

Benefits of technology

Real-time monitoring of the dynamic behavior of the power system is realized, the accuracy and robustness of disturbance recognition are improved, and the recognition accuracy can be maintained at high data loss levels, the calculation complexity is low, and real-time is available.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN112633315B_ABST
    Figure CN112633315B_ABST
Patent Text Reader

Abstract

The present invention provides a method for classifying disturbances in a power system. The feature extraction method based on the stacked denoising autoencoder can capture the feature expression of the disturbance data that is robust to the missing data. On this basis, the random forest classifier is used to realize the recognition of the disturbance in the power system. The method can quickly and accurately classify the PMU disturbance data, and still has a high recognition accuracy rate and good anti-noise performance for the PMU disturbance data containing the missing data. Compared with the existing disturbance classification method, the method can quickly and accurately classify the PMU disturbance data containing the missing data, and realize the real-time monitoring of the dynamic behavior of the power system.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of power systems, and in particular to a power system disturbance classification method, and in particular to a power system disturbance classification method considering PMU lost data. Background Art

[0002] With the continuous expansion of the scale of power systems and the access of a large number of power electronic equipment, the complexity of the power grid structure has continued to increase, and the safety issues of the power grid have become increasingly prominent. In recent years, major power outages have occurred frequently, which has had a huge impact on social and economic development and people's lives. Studies have found that major power outages usually start with a single fault, and after a series of chain reactions, they eventually lead to the collapse of the power grid. Therefore, real-time monitoring and analysis of power system disturbances are important for the safe and stable operation of power systems. Synchronous phasor measurement units (PMUs) can provide a data basis for system protection and closed-loop control due to their synchronization, rapidity and accuracy, making real-time monitoring of power system disturbances possible.

[0003] At present, the research on power system disturbance classification by domestic and foreign scholars is mainly divided into model-based and data-based methods. The model-based method needs to model the power grid through the system topology and parameters, and identify the disturbance type according to the disturbance triggering mechanism. However, for complex and large systems, the amount of calculation is very large, and it may even be impossible to analyze. The data-based method analyzes historical data to obtain a nonlinear mapping relationship between data and targets to identify the disturbance type. With the increasing complexity of the system and the influx of massive power big data, data-based methods have gradually become a more effective analysis method.

[0004] Most existing methods are based on the assumption that PMU data is normal, ignoring the impact of PMU data quality. However, about 10% to 17% of PMU data have data quality problems to varying degrees, which seriously restricts its application in power system disturbance classification.

[0005] Purpose of the Invention

[0006] The purpose of the present invention is to address the deficiencies in the above-mentioned prior art and to provide a power system disturbance classification method based on stacked denoising autoencoders and random forest classifiers. Compared with other disturbance classification methods, this method can quickly and accurately classify PMU disturbance data containing lost data, thereby realizing real-time monitoring of the dynamic behavior of the power system. Summary of the invention

[0007] The present invention provides a method for classifying power system disturbances based on a stacked denoising autoencoder and a random forest classifier, comprising the following steps:

[0008] Step 1: Generate disturbance data of the power system using an offline time-domain simulation method;

[0009] Step 2: Standardize the disturbance data obtained by the offline simulation method in step 1;

[0010] Step 3: Construct and train a stacked denoising autoencoder deep neural network, using the frequency and voltage RMS values ​​within 0.5s after the disturbance occurs as the input of the stacked denoising autoencoder to train the stacked denoising autoencoder;

[0011] Step 4: Use the stacked denoising autoencoder trained in step 3 to extract data features and obtain high-level feature expressions;

[0012] Step 5: Construct and train a random forest classifier, and use the trained random forest classifier to classify the high-level features extracted in step 4 to achieve disturbance identification.

[0013] Furthermore, the process of generating disturbance data using the offline time domain simulation method in step 1 is specifically as follows: 6 types of disturbances, including three-phase short circuit 3-φFlt, single-phase grounding fault φ-g Flt, generator output reduction GL, load addition, load shedding, and three-phase line break LT, are selected for simulation respectively. The system example is the IEEE 10 machine 39 system, the simulation software is PSD-BPA, the simulation time is 30s, the simulation step is set to 0.02s, the disturbance is triggered after 5s, and the effective values ​​of the frequency and voltage of each bus are output.

[0014] Furthermore, the process of standardization described in step 2 is as follows: Assuming that the PMU uplink frequency is 50 Hz, the frequency and voltage within 0.5 s are expressed as f, The frequency and voltage signals are standardized separately:

[0015]

[0016] in, is the standardized data, u and σ are the mean and standard deviation of the variable corresponding to Z respectively.

[0017] Furthermore, the stacked denoising autoencoder SDAE in step 3 is a deep network model formed by stacking denoising autoencoders DAE, and the process of constructing and training the stacked denoising autoencoder deep neural network includes the following sub-steps:

[0018] S31: Order As the input data of DAE, first, with a certain probability C The data in is randomly set to zero to obtain the damaged perturbation data Then, DAE maps the corrupted data into the feature representation of the hidden layer h = [h1,h2,…,h t ] T , and then reconstruct the complete sample by decoding The DAE encoding and decoding process is as follows:

[0019]

[0020]

[0021] Where W and W' are the encoding matrix and decoding matrix respectively; b and b' are the encoding bias vector and bias deviation vector respectively; θ and θ' are the parameters used for encoding and decoding respectively; f θ and g θ ' is the activation function, here we use the Sigmoid function:

[0022]

[0023] S32: Train SDEA. In this process, adjust the parameters with the goal of minimizing the reconstruction error:

[0024]

[0025] in is the reconstruction error, means The parameters θ and θ' corresponding to the minimum; for the preprocessed perturbation data set N is the number of data, and its reconstruction error It is expressed as:

[0026]

[0027] in, is the i-th preprocessed disturbance data, The reconstructed data, MSE is the mean square error;

[0028] The optimal model parameters are obtained through the back propagation of errors and the gradient descent algorithm. The parameter update process is as follows:

[0029]

[0030] Where η is the learning rate;

[0031] During the training process, SDAE optimizes the model through self-supervised learning. Specifically, any two adjacent layers in SDAE are regarded as a DAE, and the neural network is trained layer by layer with the goal of minimizing the reconstruction error.

[0032] Preferably, in step 4, SDAE performs forward propagation using the encoding features of the previous DAE as input data of the next DAE.

[0033] Furthermore, the step 5 further comprises:

[0034] The random forest classifier is an integrated algorithm classifier that uses multiple decision trees DT as weak classifiers, where a single DT is a classification and regression tree CART. For a given sample set D, its Gini coefficient is

[0035]

[0036] Among them, |C k |It is D F The number of sample subsets belonging to the kth class in Gini(D F ) represents the probability that a randomly selected sample is misclassified, Gini(D F ) is smaller, indicating that D F The lower the probability that the sample selected in is misclassified, that is, D F The higher the purity;

[0037] Sample feature set D F According to the feature F={F1,F2,…F k}Feature F j A specific value a is divided into two parts D1 and D2:

[0038] D1={(x,y)∈D F |A(x)=a},D2=D F -D1

[0039] Then under the condition of feature A, the Gini coefficient of set D is:

[0040]

[0041] Where |D1| and |D2| represent the number of samples in sets D1 and D2 respectively; N is the number of samples;

[0042] Gini (D F ,F j =a) represents the set D F and F j = uncertainty after a division; the larger the value of the Gini index, the greater the uncertainty of the sample set;

[0043] Use bootstrap sampling method to generate n sub-datasets, and use Gini index as segmentation criterion to generate n sub-datasets and corresponding n CARTs, so as to construct random forest classifier;

[0044] By training a random forest classifier with high-level features extracted by SDAE, perturbation recognition and classification are achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0046] Figure 1 A flowchart of a method for classifying power system disturbances based on a stacked denoising autoencoder and a random forest classifier according to an embodiment of the present invention;

[0047] Figure 2 is a structural diagram of a denoising autoencoder according to an embodiment of the present invention;

[0048] Figure 3 A structural diagram of a stacked denoising autoencoder according to an embodiment of the present invention;

[0049] Figure 4 A schematic diagram of reconstruction errors corresponding to different stacked denoising autoencoder structures according to an embodiment of the present invention;

[0050] Figure 5 A schematic diagram of classification accuracy rates corresponding to different decision tree depths and numbers according to an embodiment of the present invention;

[0051] Figure 6 A two-dimensional visualization diagram of features extracted at different denoising autoencoder neural network depths according to an embodiment of the present invention;

[0052] Figure 7 A schematic diagram of a disturbance identification confusion matrix of a disturbance classification method according to an embodiment of the present invention;

[0053] Figure 8 Schematic diagram showing the comparison of recognition accuracy between the disturbance classification method according to an embodiment of the present invention and other methods at different data loss levels. DETAILED DESCRIPTION

[0054] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Those skilled in the art should understand that the described embodiments are only used to illustrate the subject matter and ideas of the present invention and should not be regarded as limiting the scope of the present invention. Based on the embodiments of the present invention, all other variants and combined technical solutions obtained by ordinary technicians in this field without creative work fall within the scope of protection of the present invention.

[0055] like Figure 1 As shown, it is a flow chart of a method for classifying disturbances in a power system based on a stacked denoising autoencoder and a random forest classifier according to an embodiment of the present invention. The disturbance classification method of the present invention mainly includes the following steps:

[0056] Step 1: Generate disturbance data using an offline time-domain simulation method;

[0057] Step 2: Standardize the offline simulation data by using standardization;

[0058] Step 3: Construct and train the stacked denoising autoencoder deep neural network. The frequency and voltage RMS values ​​within 0.5s after the disturbance occur are used as the input of the stacked denoising autoencoder for training.

[0059] Step 4: Use the trained stacked denoising autoencoder to extract data features and obtain high-level feature expression.

[0060] Step 5: Construct and train a random forest classifier, and classify the extracted features through the trained random forest classifier to achieve disturbance identification.

[0061] In step 1, the process of generating disturbance data using the offline time domain simulation method is as follows:

[0062] Combined with the probability of disturbances in the power system and the harm caused, six types of disturbances are selected for simulation: three-phase short circuit (3-φFlt), single-phase ground fault (φ-g Flt), generator output reduction (GL), load switching (L-on / off), and three-phase line break (LT). The system example is the IEEE 10-machine 39 system, the simulation software is PSD-BPA, the simulation time is 30s, the disturbance is triggered after 5s, and the effective values ​​of the frequency and voltage of each bus are output. Considering that the PMU upload rate is 50Hz, the simulation step is set to 0.02s. In order to simulate the actual operation state of the power system, 60dB Gaussian white noise is applied in the simulation.

[0063] Table 1. Simulation methods for different disturbances

[0064]

[0065] In step 2, the process of standardizing the offline simulation data by using standardization is as follows: Assuming that the PMU uploading frequency is 50Hz, the frequency and voltage within 0.5s can be expressed as f, For frequency and voltage signals, they are standardized respectively:

[0066]

[0067] in, is the standardized data, u and σ are the mean and standard deviation of the variable corresponding to Z respectively.

[0068] Before introducing the specific process of step 3, we first introduce the stacked denoising autoencoder algorithm.

[0069] In step 3, the process of constructing and training the stacked denoising autoencoder deep neural network is as follows:

[0070] The network structure of the denoising autoencoder (DAE) is as follows Figure 2 As shown. As the input data of DAE, firstly The data in is randomly set to zero to obtain the damaged perturbation data Then DAE maps the damaged data into the feature representation of the hidden layer h = [h1,h2,…,h t ] T , and then reconstruct the complete sample by decoding The encoding and decoding process is shown below.

[0071]

[0072]

[0073] Where W and W' are the encoding matrix and decoding matrix respectively; b and b' are the encoding bias vector and bias deviation vector respectively; θ and θ' are the parameters used for encoding and decoding respectively. θ and g θ' is the activation function, here we use the Sigmoid function:

[0074]

[0075] During the training process, DAE adjusts the parameters with the goal of minimizing the reconstruction error:

[0076]

[0077] in is the reconstruction error. means The corresponding parameters θ and θ' are the minimum. If the reconstruction error is small enough, it means that the hidden layer contains significant features that can characterize the original data.

[0078] For the preprocessed perturbation dataset N is the number of data. Then the reconstruction error of the data set is It is expressed as follows:

[0079]

[0080] in, is the i-th preprocessed disturbance data, The reconstructed data is , and MSE is the mean square error.

[0081] The optimal model parameters can be obtained through error back propagation and gradient descent algorithm. The parameter update process is as follows:

[0082]

[0083] Among them, η is the learning rate.

[0084] Stacked denoising autoencoders (SDAE) are deep network models stacked by DAEs, such as Figure 3 As shown in Figure 1, SDAE optimizes the model through self-supervised learning. First, any two adjacent layers in SDAE are regarded as a DAE, and the network is trained layer by layer with the goal of minimizing the reconstruction error. The feature extraction method based on SDAE is shown in Algorithm I:

[0085]

[0086]

[0087] In step 4, the specific process of extracting data features using the trained SDAE is as follows:

[0088] SDAE uses the encoded features of the previous DAE as the input data of the next DAE for forward propagation. Therefore, for SDAE with L hidden layers, the feature extraction process is as follows:

[0089]

[0090]

[0091] In step 5, the specific process of constructing and training the random forest classifier is:

[0092] Random Forest (RF) is an integrated algorithm that uses multiple decision trees as weak classifiers. The final classification result is achieved through majority voting of multiple decision trees (DT). The single DT selected in this study is the classification and regression tree (CART), which uses the Gini index as the criterion for selecting segmentation features. The relevant formula is as follows:

[0093] For a given sample set D, its Gini coefficient is

[0094]

[0095] Among them, |C k |It is D F The number of sample subsets belonging to the kth class in Gini(D F ) represents the probability that a randomly selected sample is misclassified, Gini(D F ) is smaller, indicating that D F The lower the probability that the sample selected in is misclassified, that is, D F The higher the purity.

[0096] Sample feature set D F According to the feature F={F1,F2,…F k}Feature F j A specific value a is divided into two parts D1 and D2:

[0097] D1={(x,y)∈D F |A(x)=a},D2=D F -D1

[0098] Then under the condition of feature A, the Gini coefficient of set D is:

[0099]

[0100] Where |D1| and |D2| represent the number of samples in sets D1 and D2 respectively; N is the number of samples.

[0101] Gini (D F ,F j =a) represents the set D F and F j = uncertainty after a division. The larger the value of the Gini index, the greater the uncertainty of the sample set. Therefore, the feature with the smallest Gini index and its corresponding feature and feature value are selected as the best segmentation feature and segmentation point.

[0102] We use the bootstrap sampling method to generate n sub-datasets, and use the Gini index as the segmentation criterion to generate n sub-datasets and corresponding n CARTs to construct a random forest classifier.

[0103] Random forest is used as a classifier to avoid the problem of low generalization performance of a single classifier. Finally, perturbation identification is achieved by training the RF classifier with high-level features extracted by SDAE. The event classification algorithm based on random forest is summarized in Algorithm III, which describes the generation of RF and category classification.

[0104]

[0105]

[0106] In order to illustrate the effect of the above solution of the embodiment of the present invention, it is verified through experiments.

[0107] 1. SDAE model structure and parameter setting

[0108] The preprocessed data is used as the SDAE input data, and the number of hidden layer neurons is set layer by layer. We first determine the optimal number of neurons in the first layer, and then fix the number of neurons in the first layer to determine the optimal number of neurons in the second layer. This process continues until the reconstruction error MSE no longer decreases. Figure 4 The influence of the number of hidden layers and the number of hidden layer neurons on MSE is shown. For the IEEE-39 node system, when the data loss level is 50%, the optimal number of hidden layers for SDAE to identify power system disturbances is 4, and the optimal number of hidden units in each layer is 50, 70, 50 and 30.

[0109] The same parameter optimization method is used to train the perturbation data with different loss levels, and the reconstruction error of the test set is calculated. The optimal network structure and test results under different data loss levels are shown in Table 2. When the loss level is greater than 50%, the reconstruction error increases rapidly, indicating that the maximum loss level at which the model can reconstruct the original data is 50%. Therefore, for the IEEE-39 node system, the optimal number of hidden layers for perturbation recognition based on the SDAE neural network is 4, and the optimal number of neurons in each layer is [50, 70, 50, 30]. The maximum data loss tolerance is 50%.

[0110] Table 2. Best SDAE structures under different data loss levels

[0111]

[0112] 2. RF parameter settings

[0113] The optimal parameters of RF are selected based on the features extracted by SDAE on the perturbation data. First, the optimal depth d of DT is determined, and at the optimal depth, the number of optimal DTs n is determined. The optimal depth d and n are determined by comparing the average accuracy of 10 cross-validations on the validation set.

[0114] like Figure 5 As shown, when d=6 and n=40, the classifier has the best performance.

[0115] 3. Feature extraction test

[0116] In order to evaluate the performance of the method, the original data and the features extracted by SDAE are mapped to a two-dimensional space, and the extracted features are visualized under the condition of 50% data loss rate. Figure 6 As shown, 1, 2, 3, 4, 5 and 6 correspond to three-phase short circuit fault, single-phase grounding fault, engine output reduction, load shedding, load addition and three-phase disconnection respectively.

[0117] like Figure 6 As shown in (a), the samples of each category in the original data space have too much overlap and the categories cannot be separated. However, when using the features extracted by SDAE, the categories are well separated. Figure 6 (b)-(d) show the feature space of extracted features with different numbers of hidden layers in SDAE. As the number of layers increases, the overlap between different classes decreases.

[0118] 4. Testing at different data loss levels

[0119] We assume that there is data missing in the PMU data used for perturbation identification. Figure 7 The confusion matrix of the proposed method is shown, where the data loss level is set to 50%. The confusion matrix is ​​a standard format for describing classification accuracy. It is a square matrix of rank K, where K is the number of classes. The rows of the confusion matrix are the perturbation categories obtained by the algorithm, and the columns are the actual perturbation categories. The values ​​on the diagonal represent the probability of correctly identifying the category. The values ​​on both sides of the diagonal represent the probability that the j-th perturbation is mistakenly identified as the i-th perturbation.

[0120] exist Figure 7 In the case shown, the recognition accuracy for the generator output reduction event is low. This is because the dynamic characteristics of the load event and the generator output reduction event are similar, which leads to their misclassification. The overall accuracy of the test data is 98.73%, indicating that the method has good performance for power system disturbance identification with missing data.

[0121] In addition, we combine different feature extraction methods and classifiers to highlight the superiority of the proposed method under different data loss levels. The selected feature extraction methods include time domain artificial features (MFT), frequency domain artificial features (MFF), stacked autoencoder (SAE) and SDAE. The classifiers include Softmax, extreme learning machine (ELM), linear support vector machine (LSVM), Gaussian support vector machine (GSVM), DT and RF. The results are shown in Figure 2. Figure 8 shown.

[0122] The following results can be obtained from the figure:

[0123] 1) Under all data loss levels, the accuracy and Micro-F1 of all classifiers in four different feature spaces show an upward trend, indicating that the feature extraction ability based on deep neural networks is better than that of traditional methods (MFT and MFF).

[0124] 2) As the degree of data loss increases, the difference in accuracy between the SDAE method and the traditional method becomes larger and larger, indicating that the traditional method is very sensitive to missing data, but the proposed method is very robust to missing data.

[0125] 3) Under the same data loss level and the same feature space, the recognition rate of the ensemble-learned RF classifier can reach up to 98.73%.

[0126] 5. Computation time evaluation

[0127] The simulation experiment was carried out on a computer with a CPU of i7-8700k (3.7GHz), a GTX1080ti GPU and 16G memory. The calculation process includes feature extraction and classification. Table 2 shows the average calculation time for each sample in the test set. The calculation time is 3.507ms, the feature extraction time is 3.482ms, and the classification time is 0.025ms. The results show that the method has low computational complexity and high real-time performance.

[0128] Table 2. Computation time of the proposed method

[0129]

[0130] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for classifying power system disturbances based on stacked denoising autoencoders and random forest classifiers, characterized in that: The following steps are involved: Step 1: Generate disturbance data of the power system using an offline time-domain simulation method; Step 2: Standardize the disturbance data obtained by the offline simulation method in step 1; Step 3: Construct and train a stacked denoising autoencoder deep neural network, using the frequency and voltage RMS values ​​within 0.5s after the disturbance occurs as the input of the stacked denoising autoencoder to train the stacked denoising autoencoder; The stacked denoising autoencoder SDAE in step 3 is a deep network model formed by stacking denoising autoencoders DAE. The process of constructing and training the stacked denoising autoencoder deep neural network includes the following sub-steps: S31: Order As the input data of DAE, first, with a certain probability C The data in is randomly set to zero to obtain the damaged perturbation data Then, DAE maps the corrupted data into the feature representation of the hidden layer h = [h1,h2,…,h t ] T , and then reconstruct the complete sample by decoding The DAE encoding and decoding process is as follows: θ={W,b} θ'={W',b'} Where W and W' are the encoding matrix and decoding matrix respectively; b and b' are the encoding bias vector and bias deviation vector respectively; θ and θ' are the parameters used for encoding and decoding respectively; f θ and g θ' is the activation function, here we use the Sigmoid function: S32: Train SDAE. During this process, adjust the parameters with the goal of minimizing the reconstruction error: in is the reconstruction error, means The parameters θ and θ' corresponding to the minimum; for the preprocessed perturbation data set N is the number of data, and its reconstruction error It is expressed as: in, is the i-th preprocessed disturbance data, for The reconstructed data, MSE is the mean square error; The optimal model parameters are obtained through the back propagation of errors and the gradient descent algorithm. The parameter update process is as follows: Where η is the learning rate; During the training process, SDAE optimizes the model through self-supervised learning. Specifically, any two adjacent layers in SDAE are regarded as a DAE, and the neural network is trained layer by layer with the goal of minimizing the reconstruction error. Step 4: Use the stacked denoising autoencoder trained in step 3 to extract data features and obtain high-level feature expressions; Step 5: Construct and train a random forest classifier, and use the trained random forest classifier to classify the high-level features extracted in step 4 to achieve disturbance identification.

2. The method for classifying disturbances in a power system according to claim 1, characterized in that: The process of generating disturbance data using the offline time domain simulation method in step 1 is as follows: select six disturbance types, including three-phase short circuit 3-φFlt, single-phase ground fault φ-gFlt, generator output reduction GL, load addition, load shedding, and three-phase line break LT, for simulation. The system example is the IEEE 10 machine 39 system, the simulation software is PSD-BPA, the simulation time is 30s, the simulation step is set to 0.02s, the disturbance is triggered after 5s, and the effective values ​​of each bus frequency and voltage are output.

3. The method for classifying disturbances in a power system according to claim 1, characterized in that: The process of standardization described in step 2 is as follows: Assuming that the PMU uplink frequency is 50 Hz, the frequency and voltage within 0.5 s are represented as f, The frequency and voltage signals are standardized separately: in, is the standardized data, u and σ are the mean and standard deviation of the variable corresponding to Z respectively.

4. The method for classifying disturbances in a power system according to claim 1, characterized in that: In step 4, SDAE uses the encoded features of the previous DAE as the input data of the next DAE for forward propagation.

5. The method for classifying disturbances in a power system according to claim 4, characterized in that: The step 5 further comprises: The random forest classifier is an integrated algorithm classifier that uses multiple decision trees DT as weak classifiers, where a single DT is a classification and regression tree CART. For a given sample set D, its Gini coefficient is Among them, C k Yes D F The number of sample subsets belonging to the kth class in Gini(D F ) represents the probability that a randomly selected sample is misclassified, Gini(D F ) is smaller, indicating that D F The lower the probability that the sample selected in is misclassified, that is, D F The higher the purity; Sample feature set D F According to the feature F=(F1,F2,…,F k ) in feature F j A specific value a is divided into two parts D1 and D2: D1={(x,y)∈D F |A(x)=a},D2=D F -D1 Then under the condition of feature A, the Gini coefficient of set D is: Where |D1| and |D2| represent the number of samples in sets D1 and D2 respectively; N is the number of samples; Gini (D F ,F j =a) represents the set D F and F j = uncertainty after a division; the larger the value of the Gini index, the greater the uncertainty of the sample set; Use bootstrap sampling method to generate n sub-datasets, and use Gini index as segmentation criterion to generate n sub-datasets and corresponding n CARTs, so as to construct random forest classifiers; By training a random forest classifier with high-level features extracted by SDAE, perturbation recognition and classification are achieved.