CN112633315A - Electric power system disturbance classification method - Google Patents

Abstract

The invention provides a power system disturbance classification method, a feature extraction method based on a stacking denoising autoencoder can capture feature expression robust to lost data in disturbance data, and recognition of power system disturbance is realized by using a random forest classifier on the basis. The method can quickly and accurately classify the PMU disturbance data, has high identification accuracy for the PMU disturbance data containing the lost data, and has good anti-noise performance. Compared with the existing disturbance classification method, the method can quickly and accurately classify the PMU disturbance data containing the lost data, and realize the real-time monitoring of the dynamic behavior of the power system.

Description

Electric power system disturbance classification method

Technical Field

The invention relates to the technical field of power systems, in particular to a power system disturbance classification method, and particularly relates to a power system disturbance classification method considering PMU (phasor measurement Unit) lost data.

Background

With the continuous expansion of the scale of the power system and the access of a large number of power electronic devices, the complexity of the power grid structure is increased continuously, and the problem of power grid safety is highlighted gradually. In recent years, major power failure accidents frequently occur, and great influence is brought to the development of social economy and life of people. Researches find that a blackout accident usually starts from a single fault and finally causes the breakdown of a power grid through a series of chain reactions. Therefore, real-time monitoring and analysis of the disturbance of the power system play an important role in safe and stable operation of the power system. Synchronous Phasor Measurement Units (PMUs) can provide a data basis for system protection and closed-loop control due to the synchronism, rapidity and accuracy of the PMUs, so that real-time monitoring of power system disturbance becomes possible.

At present, the study of the disturbance classification of the power system by scholars at home and abroad is mainly divided into model-based and data-based methods. The model-based method needs to model the power grid through a system topological structure and parameters, and realizes disturbance type identification according to a disturbance triggering mechanism. However, for complex systems, the computational effort is large and may not even be resolvable. The data-based method analyzes historical data to obtain a nonlinear mapping relation between the data and a target, so as to realize disturbance type identification. With the increase of system complexity and the influx of massive power big data, the data-based method gradually becomes a more effective analysis method.

Most of the existing methods are researched on the assumption that PMU data is normal, and the influence of PMU data quality is ignored. However, about 10% to 17% of PMU data has different degrees of data quality problems, which severely restricts its application in power system disturbance classification.

Object of the Invention

The invention aims to provide a power system disturbance classification method based on a stack denoising autoencoder and a random forest classifier, aiming at the defects in the prior art.

Disclosure of Invention

The invention provides a power system disturbance classification method based on a stack denoising autoencoder and a random forest classifier, which comprises the following steps:

step 1: generating disturbance data of the power system by using an offline time domain simulation method;

step 2: carrying out standardization processing on the disturbance data obtained by the off-line simulation method in the step 1;

and step 3: constructing and training a deep neural network of the stacked denoising autoencoder, and training the stacked denoising autoencoder by taking the effective values of frequency and voltage within 0.5s after disturbance occurs as the input of the stacked denoising autoencoder;

and 4, step 4: extracting data features by using the stack denoising self-encoder trained in the step 3 to obtain high-level feature expression;

and 5: and (4) constructing and training a random forest classifier, and classifying the high-level features extracted in the step (4) through the trained random forest classifier to realize disturbance identification.

Further, the process of generating the disturbance data by using the offline time domain simulation method in the step 1 specifically includes: respectively selecting 6 disturbance types including three-phase short circuit 3-phi Flt, single-phase earth fault phi-g Flt, generator output reduction GL, load input, load shedding and three-phase line breaking LT for simulation, wherein the system is an IEEE 10 machine 39 system, simulation software is PSD-BPA, the simulation time is 30s, the simulation step length is set to be 0.02s, the disturbance is triggered after 5s, and the frequency and voltage effective values of each bus are output.

Still further, the process of performing the normalization process in step 2 is as follows: on the assumption of PMUThe frequency is 50Hz, the frequency and voltage within 0.5s are respectively denoted as f,

the frequency and voltage signals are normalized separately:

wherein the content of the first and second substances,

for normalized data, u and σ are the mean and standard deviation, respectively, of the Z corresponding variable.

Still further, the stacked denoising autoencoder SDAE in step 3 is a deep network model stacked by denoising autoencoders DAE, and the process of constructing and training the stacked denoising autoencoder deep neural network includes the following sub-steps:

s31: order to

For DAE input data, first, with a certain probability C pairs

To obtain the damaged disturbance data

The DAE then maps the corrupted data to a signature expression h ═ h of the hidden layer by an encoding operation₁,h₂,…,h_t]^TAnd then reconstruct the complete samples by decoding

The DAE encoding and decoding process is as follows:

wherein, W and W' are respectively an encoding matrix and a decoding matrix; b and b' are the coding offset vector and the offset bias vector, respectively; θ and θ' are parameters for encoding and decoding, respectively; f. of_θAnd g_θ' is an activation function, here the Sigmoid function is used:

s32: training the SDEA, and in the process, adjusting parameters by taking the minimum reconstruction error as a target:

wherein

Is the error of the reconstruction and is,

means that

The minimum corresponding parameters θ and θ'; for preprocessed perturbed data sets

N is the number of data, the reconstruction error

Expressed as:

wherein the content of the first and second substances,

for the ith preprocessed perturbation data,

the MSE is mean square error;

obtaining the optimal model parameters through the back propagation of errors and a gradient descent algorithm, wherein the parameter updating process comprises the following steps:

wherein eta is the learning rate;

in the training process, the SDAE optimizes the model through self-supervision learning, specifically, any two adjacent layers in the SDAE are regarded as one DAE, and the neural network is trained layer by layer with the aim of minimizing reconstruction errors.

Preferably, in step 4, the SDAE is propagated forward with the coding features of the previous DAE as input data for the next DAE.

Still further, the step 5 further comprises:

the random forest classifier is an integrated algorithm classifier taking a plurality of decision trees DT as weak classifiers, wherein a single DT is a classification regression tree CART, and for a given sample set D, the Gini coefficient is

Wherein, | C_kIs D_FThe number of the sample subsets belonging to the kth class, N is the number of samples, and K is the number of classes; gini: (D_F) Probability of a randomly selected sample in the representation being misclassified, Gini (D)_F) The smaller the size, the smaller the value of D_FThe lower the probability that the selected sample is misclassified, i.e. D_FThe higher the purity of (a);

sample feature set D_FAccording to the feature F ═ F₁,F₂,…F_kIn feature F_jIs divided into D₁And D₂Two parts are as follows:

D₁＝{(x,y)∈D_F|A(x)＝a},D₂＝D_F-D₁

then under the condition of feature a, the kini coefficient of set D is:

in the formula, | D₁I and I D₂Respectively representing the sets D₁And D₂The number of samples in (1); n is the number of samples;

Gini(D_F,F_ja) represents the set D_FAnd F_jUncertainty after a division; the larger the value of the kini index is, the larger the uncertainty of the sample set is;

generating n sub-data sets by using a bootstrap sampling method, and generating the n sub-data sets and corresponding n CARTs by using a Gini index as a segmentation criterion, thereby constructing a random forest classifier;

and training a random forest classifier by using the high-grade features extracted by the SDAE to realize disturbance identification and classification.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings needed to be used in the description of the embodiments are briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on the drawings without creative efforts.

FIG. 1 is a flowchart of a power system disturbance classification method based on a stacked denoising autoencoder and a random forest classifier according to an embodiment of the present invention;

FIG. 2 is a block diagram of a denoised self-encoder according to an embodiment of the invention;

FIG. 3 is a block diagram of a stacked denoising self-encoder according to an embodiment of the present invention;

FIG. 4 is a diagram illustrating corresponding reconstruction errors for different stacked denoising autocoders according to an embodiment of the present invention;

FIG. 5 is a diagram illustrating classification accuracy for different decision tree depths and counts according to an embodiment of the present invention;

FIG. 6 is a schematic diagram of two-dimensional visualization of features extracted at different depths of a denoised self-encoder neural network according to an embodiment of the present invention;

FIG. 7 is a schematic diagram of a disturbance identification confusion matrix of the disturbance classification method according to the embodiment of the invention;

FIG. 8 is a comparison of the recognition accuracy of the perturbation classification method according to the embodiment of the present invention and other methods under different data loss levels.

Detailed Description

The technical solutions in the embodiments of the present invention are clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it should be understood by those skilled in the art that the described embodiments are only used for illustrating the spirit and idea of the present invention and should not be construed as being limited by the scope of the present invention. All other variations and combinations of technical solutions which can be obtained by a person skilled in the art without making creative efforts based on the embodiments of the present invention fall within the protection scope of the present invention.

As shown in fig. 1, the method is a flowchart of a method for classifying disturbance of an electric power system based on a stack denoising autoencoder and a random forest classifier according to an embodiment of the present invention, and the method mainly includes the following steps:

step 1: generating disturbance data by using an offline time domain simulation method;

step 2: standardizing the off-line simulation data by utilizing standardization;

and step 3: constructing and training a stacked denoising self-encoder deep neural network. And training the effective values of the frequency and the voltage within 0.5s after the disturbance occurs as the input of the stacked denoising self-encoder.

And 4, step 4: and extracting the data features by utilizing the trained stacking denoising self-encoder to obtain high-level feature expression.

And 5: and constructing and training a random forest classifier, and classifying the extracted features through the trained random forest classifier to realize disturbance identification.

In step 1, the process of generating the disturbance data by using the offline time domain simulation method specifically includes:

according to the probability of disturbance occurrence and damage caused in the power system, 6 disturbance types of three-phase short circuit (3-phi Flt), single-phase earth fault (phi-g Flt), generator output reduction (GL), load switching on/off (L-on/off) and three-phase line breaking (LT) are selected for simulation. The system is an IEEE 10 machine 39 system, the simulation software is PSD-BPA, the simulation time is 30s, and the disturbance is triggered after 5s, so that the frequency and voltage effective value of each bus are output. The simulation step size is set to 0.02s, taking into account that the PMU upload rate is 50 Hz. To simulate the actual operating conditions of the power system, 60dB of white Gaussian noise was applied in the simulation.

TABLE 1 simulation method of different disturbances

In step 2, the process of standardizing the data of the off-line simulation by using standardization specifically includes: assuming a PMU up-conversion frequency of 50Hz, the frequency and voltage within 0.5s can be respectively denoted as f,

for the frequency and voltage signals, they are normalized separately:

wherein the content of the first and second substances,

for normalized data, u and σ are the mean and standard deviation, respectively, of the Z corresponding variable.

Before describing the specific process of step 3, a relevant description is first made for the stacked denoising self-encoder algorithm.

In step 3, the process of constructing and training the stacked denoising self-encoder deep neural network specifically comprises the following steps:

the network structure of a Denoise Autoencoder (DAE) is shown in fig. 2. Order to

For DAE input data, first pair with a certain probability C

To obtain the damaged disturbance data

The DAE then maps the corrupted data to a hidden layer's representation of features h ═ h through an encoding operation₁,h₂,…,h_t]^TAnd then reconstruct the complete samples by decoding

The encoding and decoding processes are as follows.

Wherein, W and W' are respectively an encoding matrix and a decoding matrix; b and b' are the coding offset vector and the offset bias vector, respectively; θ and θ' are parameters for encoding and decoding, respectively. f. of_θAnd g_θ'Is an activation function, here the Sigmoid function is used:

during the training process, the DAE adjusts the parameters with the aim of minimizing the reconstruction error:

wherein

Is the reconstruction error.

Means that

The minimum corresponding parameters theta and theta'. If the reconstruction error is small enough, it indicates that the hidden layer contains significant features that can characterize the original data.

For preprocessed perturbed data sets

And N is the number of data. The reconstruction error of the data set

Is represented as follows:

wherein the content of the first and second substances,

for the ith preprocessed perturbation data,

MSE is mean square error.

The optimal model parameters can be obtained through the back propagation of errors and a gradient descent algorithm. The parameter updating process is as follows:

wherein η is the learning rate.

A Stacked denoising auto-encoder (SDAE) is a deep network model Stacked by DAE, as shown in fig. 3. SDAE optimizes the model by self-supervised learning. First, any two adjacent layers in the SDAE are treated as one DAE, and the network is trained layer by layer with the goal of minimizing reconstruction errors. The feature extraction method based on SDAE is shown in an algorithm I:

in step 4, the specific process of extracting the data features by using the trained SDAE is as follows:

SDAE is forward propagated with the coding features of the previous DAE as input data for the next DAE. Therefore, for an SDAE with an implied layer number L, the feature extraction process is as follows:

in step 5, the specific process of constructing and training the random forest classifier is as follows:

random Forest (RF) is an integrated algorithm with multiple decision trees as weak classifiers. The final classification result is achieved by majority voting of multiple Decision Trees (DTs). The single DT selected in this study is a classification regression tree (CART) which uses the kini index as a criterion for selecting segmentation features, and the correlation formula is as follows:

for a given sample set D, the coefficient of kini is

Wherein, | C_kIs D_FThe number of the sample subsets belonging to the kth class, N is the number of samples, and K is the number of classes. Gini (D)_F) Probability of a randomly selected sample in the representation being misclassified, Gini (D)_F) The smaller the size, the smaller the value of D_FThe lower the probability that the selected sample is misclassified, i.e. D_FThe higher the purity of (c).

Sample feature set D_FAccording to the feature F ═ F₁,F₂,…F_kIn feature F_jIs divided into D₁And D₂Two parts are as follows:

D₁＝{(x,y)∈D_F|A(x)＝a},D₂＝D_F-D₁

then under the condition of feature a, the kini coefficient of set D is:

in the formula, | D₁I and|D₂respectively representing the sets D₁And D₂The number of samples in (1); n is the number of samples.

Gini(D_F,F_jA) represents the set D_FAnd F_jThe uncertainty after a division. The larger the value of the kini index, the greater the uncertainty of the sample set. Therefore, the feature with the minimum Gini index and the corresponding feature and feature value are selected as the optimal segmentation feature and segmentation point.

We generate n sub-datasets using bootstrap sampling method and generate n sub-datasets and corresponding n CART with the kini index as the segmentation criterion to construct the random forest classifier.

Random forests are used as classifiers to avoid the problem of low generalization performance of individual classifiers. Finally, disturbance identification is achieved by training the RF classifier with the high-level features extracted by SDAE. An event classification algorithm based on random forests is summarized in algorithm III, which describes the generation and class classification of RF.

Experiments also verify the effects of the above scheme of the embodiment of the present invention.

1. SDAE model structure and parameter setting

And taking the preprocessed data as SDAE input data, and setting the number of hidden layer neurons layer by layer. We first determine the optimal neuron number for the first layer and then fix the number of neurons in the first layer to determine the optimal neuron number for the second layer. This process continues until the reconstruction error MSE no longer decreases. Fig. 4 shows the effect of hidden layer number and hidden layer neuron number on MSE. For an IEEE-39 node system, when the data loss level is 50%, the SDAE identifies that the optimal number of hidden layers for power system disturbances is 4 layers, and the optimal number of hidden layer units per layer is 50,70,50, and 30.

And training the disturbance data with different loss levels by adopting the same parameter optimization method, and calculating the reconstruction error of the test set. The optimal network structure and test results at different data loss levels are shown in table 2, and when the loss level is greater than 50%, the reconstruction error increases rapidly, which indicates that the maximum loss level of the model capable of reconstructing the original data is 50%. Thus, for an IEEE-39 node system, the optimal number of hidden layers identified based on the perturbation of the SDAE neural network is 4 layers, and the optimal number of neurons per layer is [50,70,50,30 ]. The maximum data loss tolerance is 50%.

TABLE 2 optimal SDAE structure at different data loss levels

2. RF parameter setting

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

As shown in fig. 5, the classifier has the best performance when d is 6 and n is 40.

3. Feature extraction testing

To evaluate the performance of the method, raw data and SDAE extracted features were mapped into a two-dimensional space and the extracted features were visualized with a data loss rate of 50%. The visualization results are shown in fig. 6, where 1, 2, 3, 4, 5 and 6 correspond to three-phase short-circuit fault, single-phase earth fault, reduced engine output, load shedding and three-phase wire breakage, respectively.

As shown in fig. 6(a), there is too much overlap of samples for each class in the original data space to separate the classes. However, when using features extracted by SDAE, the categories are well separated. Fig. 6(b) - (d) show feature spaces for extracted features for different numbers of hidden layers in the SDAE. As the number of layers increases, the degree of overlap between the different classes decreases.

4. Testing at different levels of data loss

We assume that there is a data loss in the PMU data used for disturbance identification. Fig. 7 shows the confusion matrix of the proposed method, where the data loss level is set to 50%. The confusion matrix is a standard format for describing the accuracy of classification. It is a square matrix with rank K, where K is the number of classes. The rows of the confusion matrix are the disturbance classes obtained by the algorithm, and the columns are the actual disturbance classes. The values on the diagonal represent the probability of correctly identifying the class. The values on both sides of the diagonal indicate the probability that the jth class of disturbance was misidentified as an ith class of disturbance.

In the case shown in fig. 7, where the accuracy of the identification of the generator contribution reducing event is low. This is because the dynamic characteristics of the load shedding event and the generator derating event are similar, resulting in their misclassification. The overall accuracy of the test data is 98.73%, which shows that the method has good performance for identifying the disturbance of the power system with missing data.

In addition, under different data loss levels, different feature extraction methods and classifiers are combined to highlight the superiority of the method. Selected feature extraction methods include time-domain artificial features (MFT), frequency-domain artificial features (MFF), Stacked Automatic Encoder (SAE), and SDAE. 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 FIG. 8.

The following results can be obtained from the figure:

1) at all levels of data loss, the precision and Micro-F1 of all classifiers in four different feature spaces are on the rise, indicating that the deep neural network-based feature extraction capability is superior to the traditional methods (MFT and MFF).

2) With the increase of the data loss degree, the difference of the SDAE method and the traditional method in precision is larger and larger, which shows that the traditional method is very sensitive to the missing data, but the method has strong robustness to the missing data.

3) Under the same data loss level and the same feature space, the recognition rate of the integrated learning RF classifier can reach 98.73 percent at most.

5. Calculating a time estimate

Simulation experiments were performed on a computer with a CPU of i7-8700k (3.7GHz), GTX1080ti GPU and 16G memory. The calculation process includes feature extraction and classification. Table 2 shows the average calculated 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, and the result shows that the method is low in calculation complexity and high in real-time performance.

TABLE 2 calculation time of the proposed method

The above description is only for the preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope of the present invention are included in the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims (6)

1. A power system disturbance classification method based on a stack denoising autoencoder and a random forest classifier is characterized by comprising the following steps:

step 1: generating disturbance data of the power system by using an offline time domain simulation method;

step 2: carrying out standardization processing on the disturbance data obtained by the off-line simulation method in the step 1;

and step 3: constructing and training a deep neural network of the stacked denoising autoencoder, and training the stacked denoising autoencoder by taking the effective values of frequency and voltage within 0.5s after disturbance occurs as the input of the stacked denoising autoencoder;

and 4, step 4: extracting data features by using the stack denoising self-encoder trained in the step 3 to obtain high-level feature expression;

and 5: and (4) constructing and training a random forest classifier, and classifying the high-level features extracted in the step (4) through the trained random forest classifier to realize disturbance identification.

2. The method for classifying the disturbance of the power system according to claim 1, wherein the process of generating the disturbance data by using the offline time domain simulation method in the step 1 specifically comprises: respectively selecting 6 disturbance types including three-phase short circuit 3-phi Flt, single-phase earth fault phi-g Flt, generator output reduction GL, load input, load shedding and three-phase line breaking LT for simulation, wherein the system is an IEEE 10 machine 39 system, simulation software is PSD-BPA, the simulation time is 30s, the simulation step length is set to be 0.02s, the disturbance is triggered after 5s, and the frequency and voltage effective values of each bus are output.

3. The method for classifying disturbance of an electric power system according to claim 1, wherein the step 2 of performing the normalization process comprises: assuming a PMU up-conversion frequency of 50Hz, the frequency and voltage within 0.5s are respectively denoted as f,

the frequency and voltage signals are normalized separately:

wherein the content of the first and second substances,

for normalized data, u and σ are the mean and standard deviation, respectively, of the Z corresponding variable.

4. The method for classifying disturbance of electric power system as claimed in claim 1, wherein the stacked denoising autoencoder SDAE in step 3 is a deep network model stacked by denoising autoencoders DAE, and the process of constructing and training the stacked denoising autoencoder deep neural network includes the following sub-steps:

s31: order to

For DAE input data, first, with a certain probability C pairs

To obtain the damaged disturbance data

The DAE then maps the corrupted data to a signature expression h ═ h of the hidden layer by an encoding operation₁,h₂,…,h_t]^TAnd then reconstruct the complete samples by decoding

The DAE encoding and decoding process is as follows:

θ＝{W,b}

θ'＝{W',b'}

wherein, W and W' are respectively an encoding matrix and a decoding matrix; b and b' are the coding offset vector and the offset bias vector, respectively; θ and θ' are parameters for encoding and decoding, respectively; f. of_θAnd g_θ'Is an activation function, here the Sigmoid function is used:

s32: training the SDEA, and in the process, adjusting parameters by taking the minimum reconstruction error as a target:

wherein

Is the error of the reconstruction and is,

means that

The minimum corresponding parameters θ and θ'; for preprocessed perturbed data sets

N is the number of data, the reconstruction error

Expressed as:

wherein the content of the first and second substances,

for the ith preprocessed perturbation data,

is composed of

The MSE is mean square error;

obtaining the optimal model parameters through the back propagation of errors and a gradient descent algorithm, wherein the parameter updating process comprises the following steps:

wherein eta is the learning rate;

in the training process, the SDAE optimizes the model through self-supervision learning, specifically, any two adjacent layers in the SDAE are regarded as one DAE, and the neural network is trained layer by layer with the aim of minimizing reconstruction errors.

5. The method of claim 4, wherein in step 4, the SDAE is forward propagated by using the coding features of the previous DAE as input data of the next DAE.

6. The method according to claim 5, wherein the step 5 further comprises:

the random forest classifier is an integrated algorithm classifier taking a plurality of decision trees DT as weak classifiers, wherein a single DT is a classification regression tree CART, and for a given sample set D, the Gini coefficient is

Wherein, | C_kIs D_FThe number of the sample subsets belonging to the kth class, N is the number of samples, and K is the number of classes; gini (D)_F) Probability of a randomly selected sample in the representation being misclassified, Gini (D)_F) The smaller the size, the smaller the value of D_FThe lower the probability that the selected sample is misclassified, i.e. D_FThe higher the purity of (a);

sample feature set D_FAccording to the feature F ═ F₁,F₂,…F_kIn feature F_jIs divided into D₁And D₂Two parts are as follows:

D₁＝{(x,y)∈D_F|A(x)＝a},D₂＝D_F-D₁

then under the condition of feature a, the kini coefficient of set D is:

in the formula, | D₁I and I D₂Respectively representing the sets D₁And D₂The number of samples in (1); n is the number of samples;

Gini(D_F,F_ja) represents the set D_FAnd F_jUncertainty after a division; the larger the value of the kini index is, the larger the uncertainty of the sample set is;

generating n sub-data sets by using a bootstrap sampling method, and generating the n sub-data sets and corresponding n CARTs by using a Gini index as a segmentation criterion, thereby constructing a random forest classifier;

and training a random forest classifier by using the high-grade features extracted by the SDAE to realize disturbance identification and classification.

Images