An unsupervised power system disturbance identification method based on synchronous phasor measurement
By combining the LST-TimeGAN model with the LightGBM classifier, the problem of unsupervised disturbance identification in power systems is solved, enabling accurate disturbance identification and power quality monitoring in situations with no or few labels, thereby improving the safety and stability of power systems.
Patent Information
- Application Number
- CN202310491917.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-05-04
- Publication Date
- 2025-12-09
- Estimated Expiration
- 2043-05-04
AI Technical Summary
Existing technologies struggle to effectively identify unsupervised disturbances in power systems with a high proportion of renewable energy and power electronic equipment, especially in cases with few or no labels. Traditional methods are insufficient in identifying unpredictable disturbances, and data-driven methods require a large number of labels and are highly complex, resulting in low practicality.
An unsupervised power system disturbance identification method based on synchronous phasor measurement is adopted. By constructing a long short-time time series generative adversarial network (LST-TimeGAN) model for feature extraction, and combining it with a LightGBM classifier for clustering identification, the method can accurately identify power system disturbances and provide early warning of power quality problems.
It enables accurate identification of power system disturbances in the absence of or with few tags, monitors power quality issues and provides early warning of potential faults, reduces tag requirements and complexity, and improves the practicality and accuracy of identification.
Smart Images
Figure CN116502126B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of synchronous phasor measurement application, and in particular to a method for identifying power system disturbance based on synchronous phasor measurement. BACKGROUND
[0002] Under the background of new power system development, the characteristics of high proportion of renewable energy and high proportion of power electronic equipment access will make the dynamic behavior of power system more complex and variable, and the safety of power grid is facing challenges. Identifying disturbance is crucial for quickly judging the type and propagation behavior of disturbance to ensure the safe operation of power grid. Wide area measurement system refers to a new generation of power grid dynamic monitoring and control system composed of synchronous phasor measurement units, and synchronous phasor measurement units continuously provide real-time data basis for accurate disturbance identification and stability control. Such devices are collectively referred to as synchronous phasor measurement devices (SMD).
[0003] With the increase of system scale and complexity, data-driven disturbance identification methods have shown significant advantages compared to model-driven methods. At present, data-driven power system disturbance identification based on synchronous phasor measurement has been widely studied. Feature extraction is an important step for data-driven disturbance identification methods. Existing technologies are all supervised feature extraction methods, which need to construct artificial features based on experience or deep learning based on labels. Such supervised learning methods require a large amount of experience data, and the required labels are difficult to obtain due to high complexity and large amount of manual work, which leads to low practicality of the above research in the initial stage of data-driven disturbance identification development, and few mature applications. At the same time, under the background of "double high" power system, the types of power system disturbances differ from traditional types, and the infrequency, non-planning and strong unpredictability of disturbance events make it difficult to know which specific features need to be identified. Traditional methods are specific to certain events, but have poor recognition ability for unpredictable events. There are few data-driven unsupervised disturbance identification schemes without experience in the prior art. SUMMARY
[0004] The purpose of the present application is to provide a method for identifying power system disturbance based on synchronous phasor measurement, which can be used for disturbance identification in no-label or few-label situations, and can accurately identify power system disturbance, so as to monitor power quality problems of power system and give early warning of potential distribution network faults.
[0005] The purpose of the present application is achieved by the following technical solutions:
[0006] A method for identifying power system disturbance based on synchronous phasor measurement, the method comprising:
[0007] Step 1, first acquire the synchronous phasor measurement data of the power system to be processed, including frequency, voltage amplitude and voltage phase angle, and normalize the frequency, voltage amplitude and voltage phase angle as the input of the subsequent model;
[0008] Step 2, construct a long short time sequence generation adversarial network LST-TimeGAN model, and use the constructed model to extract non-event features of long time, short time and ultra-short time windows;
[0009] Step 3, according to the output of the model in step 2, combine the electrical quantity of the event and the time window feature to perform disturbance pre-classification; wherein, it specifically includes four traditional disturbance events of generator tripping, load shedding, short circuit and open circuit, as well as new events of unknown types and power quality problems;
[0010] Step 4, using the LightGBM classifier to cluster and re-identify the pre-classified event categories, to realize the identification and classification of the disturbance.
[0011] It can be seen from the above technical solution provided by the present application that the above method can be used for disturbance identification in no-label or few-label situations, and can accurately identify the power system disturbance, so as to monitor the power quality problem of the power system and prewarn the potential distribution network fault. BRIEF DESCRIPTION OF DRAWINGS
[0012] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.
[0013] Figure 1 The unsupervised power system disturbance identification method based on synchronous phasor measurement provided by the embodiment of the present application is shown in the flowchart;
[0014] Figure 2 The architecture diagram of the LST-TimeGAN model described in the embodiment of the present application is shown in the flowchart;
[0015] Figure 3 The feature extraction process diagram of the disturbance event based on LST-TimeGAN provided by the embodiment of the present application is shown in the flowchart;
[0016] Figure 4 The SMD-L field data diagram of the frequency and voltage phase angle, amplitude of each node measurement when the disturbance occurs in the first case of the embodiment of the present application is shown in the flowchart;
[0017] Figure 5SMD-L field data schematic diagram of frequency and voltage phase angle, amplitude of each node measurement when the disturbance occurs in the second case of the embodiment of the present application;
[0018] Figure 6 SMD-L field data schematic diagram of frequency and voltage phase angle, amplitude of each node measurement in the third case of the embodiment of the present application;
[0019] Figure 7 SMD-L field data schematic diagram of frequency and voltage phase angle, amplitude of each node measurement when the disturbance occurs in the fourth case of the embodiment of the present application;
[0020] Figure 8 SMD-L field measurement voltage amplitude data schematic diagram of the power quality problem in the embodiment of the present application. DETAILED DESCRIPTION
[0021] The technical solutions in the embodiments of the present application will be described clearly and completely in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, which does not constitute a limitation of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the present application.
[0022] As Figure 1 shown is a flowchart of an unsupervised power system disturbance identification method based on synchronous phasor measurement provided by the embodiment of the present application, and the method comprises the following steps:
[0023] Step 1, first, synchronous phasor measurement data of a power system to be processed is acquired, including frequency, voltage amplitude and voltage phase angle, and the frequency, voltage amplitude and voltage phase angle are normalized for processing, as input of a subsequent model;
[0024] In this step, the acquired frequency f and voltage amplitude U are normalized and preprocessed to obtain input After the voltage phase angle φ is offset with a unified reference to obtain a relative phase angle, the relative phase angle is normalized to obtain input
[0025] Let frequency f, voltage amplitude U and voltage phase angle φ before normalization be X, and the normalization formula is represented as:
[0026]
[0027] wherein, is the normalized value of the i th sample data; X max and X min respectively represent the maximum value and the minimum value in the sample data;
[0028] Considering each group of synchronized phasor measurement data In the steady-state (quasi-steady-state) condition of the power system, two types of characteristics are included, namely static characteristics and dynamic characteristics. The static characteristics refer to the fixed properties of space, such as the static influence of the measurement point position, environment, etc. of a single PMU on the measurement data, etc. The dynamic characteristics refer to the properties that change with time, such as the dynamic influence of seasons, morning and evening peak electricity consumption, etc. on the measurement data, etc. Therefore, the definitions of and are respectively the vector spaces representing the static and dynamic characteristics of the input , and and represent the static and dynamic vectors in the vector spaces, respectively.
[0029] Let (s, x 1：T ) obey the joint distribution P in the vector space, where T represents the length of the time series; if s n and x n represent the elements in the vectors s and x, and n∈{1,…,N}, the training set is represented as
[0030] Step 2, a long-short time time series generative adversarial network (LST-TimeGAN) model is constructed, and the constructed model is used for non-event feature extraction of long, short and ultra-short time windows;
[0031] In this step, as shown in Figure 2 , the architecture schematic diagram of the LST-TimeGAN model described in the embodiment of the application is shown, the constructed long-short time time series generative adversarial network LST-TimeGAN model includes an encoder, a decoder, a generator and a discriminator, all of which are constructed by a neural network containing 3 hidden layers, the number of nodes of the hidden layers is 24, and 128 samples are processed per round; the last layer of the generator is a new generation of recurrent neural network GRU (Gated Recurrent Unit); the last layer of the discriminator is a bidirectional LSTM (Bi-directional Long Short-Term Memory, BiLSTM); both of the above are variants of LSTM, wherein the GRU combines the forget gate and the input gate of the LSTM into a single update gate while mixing the cell state and the hidden state, and the model is simpler and faster in calculation than the standard LSTM model, and the BiLSTM is composed of two separate LSTM in opposite directions, which makes the output vector of the discriminator not only related to the previous synchronized phasor measurement time sequence state, but also related to the subsequent time sequence state;
[0032] Except that the last layer of the discriminator is a direct mapping based on the least square loss function, the activation function used in the LST-TimeGAN model is LeakyReLU, which is expressed as:
[0033]
[0034] where x represents the input of the LeakyReLU function; a is the slope of the LeakyReLU function, which is 0.01, and the function amplifies the value range while adjusting the zero gradient problem of negative values;
[0035] The process of using the constructed model to extract non-event features in long, short and ultra-short time windows is as follows:
[0036] (1) Optimizing the spatial feature extraction by using the scaled dot-product attention mechanism
[0037] Each real input matrix is taken as a key sequence, denoted by K; the random input matrix of the generator is taken as a query sequence, denoted by Q; F is the mapping of each element in Q to each element in K; and the attention of each generator random input is described as the mapping of the key sequence on the query. Since the length of the query and the key is the same, the similarity between K and Q is evaluated by using the scaled dot-product model to reduce the computational cost:
[0038]
[0039] where S represents the attention score based on the similarity; d k represents the dimension of K; this formula can be regarded as a matrix that calculates the degree of correlation between the data distribution in the real synchronous phasor measurement matrix and the input matrix of the existing generator;
[0040] The softmax function is used to do multi-class normalization on the attention score S, and the attention aggregation formula is:
[0041]
[0042] where a represents the attention weight coefficient matrix; the attention weight coefficient matrix a and the random input matrix Q are weighted as the attention output, which becomes the new generator random input sequence Q a , which is defined as:
[0043] Αttention(Q a )=αQ
[0044] The weight coefficient obtained by multiplying the scaled dot-product in the random input matrix Q is used to evaluate the The contribution of the input to the attention matrix data difference can distinguish the contribution of the information of different measuring devices in the random input matrix Q to the data difference of Attention(Q a ) so as to make the random generation of the generator focused, thereby realizing feature extraction of the multi-point measuring device data;
[0045] (2) The decision loss function based on least squares contains distance features in the output and can represent the degree of data anomaly
[0046] The gradient feedback of the decision loss function to the generator neural network and the output of the discriminator score are crucial, and the power system disturbance identification method needs the decision loss function to reflect the abnormal degree of each disturbance event, thereby providing a data basis for unsupervised feature extraction.
[0047] The log loss based on the least squares loss function is improved as follows:
[0048]
[0049] Wherein, the first term expects to represent the least squares distance of y s , y t and 1; y s , y t is the classification result of the static and dynamic parts of the real data by the discriminator; the second term expects to represent and the least squares distance of 0; and is the classification result of the static and dynamic data generated by the generator by the discriminator; the formula measures the difference between the generated model and the real sequence by the least squares loss;
[0050] In addition, when L′ U is used as the loss function, the last layer of the discriminator will no longer use the sigmoid activation layer, but will directly use linear mapping to map the output to [0, 1];
[0051] In the specific implementation, compared with the sigmoid cross-entropy loss function which only focuses on classification, the least squares loss function can distinguish between normal and abnormal, and the output of the loss function can also reflect the distance of the abnormal data from the normal condition, so that the output vector can accurately reflect the degree of abnormality. The value can be used as a feature to provide a basis for accurate classification. In the improvement of the decision loss function based on least squares and the improvement of the random input of the generator based on the scaled dot product attention.
[0052] (3) Construct long, short, and ultra-short time window structures, and use the sensitivity of different types of disturbances to different length time windows to extract unsupervised features of the disturbances
[0053] The long-time TimeGAN extracts long-term time sequence features of multi-dimensional synchronous measurement data with a 24h cycle, and sets the time window for disturbance feature extraction to 5min;
[0054] The short-time TimeGAN learns normal fluctuation characteristics of previous synchronous phasor measurement data with a 60min cycle, and sets the time window for disturbance feature extraction to 1min;
[0055] The ultra-short-time TimeGAN judges whether a non-normal impulse signal appears in the current time window according to normal impulse characteristics of synchronous phasor measurement data with a 1min cycle, and sets the ultra-short-time disturbance feature extraction time window to 1s;
[0056] As Figure 3 Fig. 1 shows a feature extraction process of a disturbance event based on LST-TimeGAN according to an embodiment of the present application, and a total of 9 TimeGAN models, i.e., [long-time, short-time, ultra-short-time]*[frequency f, voltage amplitude U, and voltage phase angle φ] models are trained, and 9 groups of discriminator score vectors are output;
[0057] If the score of any score vector or multiple score vectors is not close enough to the global optimum, it is judged that a disturbance event occurs in the window, and the output value of the decision loss function is taken as the score of data anomaly; then a group of discriminator score vectors is obtained by rolling scoring on the event window, and the score vector represents the change of the window abnormality degree with time; the score vector from the beginning of the event to the end of the event is extracted as the data basis for subsequent event classification.
[0058] Step 3, according to the output of the model in step 2, combining the electrical quantity and time window characteristics of the event to pre-classify the disturbance; wherein, it specifically includes four traditional disturbance events of machine tripping, load shedding, short circuit and open circuit, as well as new events of unknown types and power quality problems;
[0059] In this step, it is specifically to preliminarily distinguish the four traditional disturbance events including machine tripping, load shedding, short circuit and open circuit according to the time scale characteristics and slope characteristics of frequency and phase angle changes; wherein, when no disturbance occurs in the power transmission network, the identified abnormality is considered as a local disturbance which has no significant impact on the power transmission network and only has an impact on the power quality of the distribution network;
[0060] Firstly, the abnormality of the frequency quantity in different time scale windows can distinguish slow disturbance and fast disturbance; wherein, the slow disturbance includes machine tripping and load shedding; the fast disturbance includes short circuit and open circuit;
[0061] The two kinds of slow disturbances can be distinguished by the fact that the frequency variation slope of the cutting machine is greater than 0 and the frequency variation slope of the cutting load is less than 0;
[0062] The two kinds of fast disturbances can be distinguished by the fact that the voltage phase angle abnormality time window characteristics of the short circuit and tripping events are different, wherein:
[0063] The tripping event is usually caused by the line being cut off immediately after the event occurs, and the recovery is fast, and the influence on the voltage phase angle is less than 20s, and only the abnormality in the ultra-short time characteristic is abnormal; the influence of the short circuit event on the voltage phase angle is long, and in addition to the abnormality in the ultra-short time characteristic, the abnormality will also be detected in the short time characteristic;
[0064] In addition to the above events, the frequency and phase angle abnormalities are recorded as new events of unknown type of the power transmission network, which will be identified again in the future;
[0065] If the voltage amplitude is abnormal under the condition that the power transmission network is not abnormal, it is judged as a power quality problem of the local distribution network.
[0066] The abnormal characteristic of each major disturbance is shown in the following table 1:
[0067] Table 1: Abnormal characteristic of each major disturbance
[0068]
[0069]
[0070] In table 1, ES, S and L represent ultra-short time, short time and long time characteristic quantities respectively; "abnormal" represents that the corresponding characteristic is abnormal; "normal" represents that no abnormality is identified for the characteristic; " / " represents a necessary characteristic that is not classified, and can be any value.
[0071] Step 4: using the LightGBM classifier to cluster and re-identify the pre-classified event categories, to realize the identification and classification of the disturbance.
[0072] In this step, after the classification framework preliminarily confirms the major category to which the event belongs, the event characteristics and the correct event information of the major category are further utilized to improve the classification accuracy. Specifically, the discriminator score vector in the constructed model is used as a feature vector, and the LightGBM classifier and the existing major category information are used to re-cluster the events that have been preliminarily classified. The LightGBM classifier completes the labeling of the data based on the distribution properties of the data itself; wherein, since the feature vector can reflect the severity of the event in each time window scale, the LightGBM classifier classifies based on this, and divides the events with similar severity and electrical quantity change characteristics in the existing major category information into the same subcategory.
[0073] In the specific implementation, the automatic clustering of small categories under various large categories can effectively reduce the workload of experience-based tagging, and the classification method is flexible, and it is even impossible to predict before automatic classification that the classifier will divide events into several categories and will be classified by what features. The clustering that can automatically increase the group can cluster events with similar characteristics in the same small category, which is convenient for subsequent research and analysis.
[0074] It is worth noting that the contents not described in detail in the embodiments of the present application belong to the prior art known to those skilled in the art.
[0075] To illustrate the effectiveness of the method proposed in the embodiments of the present application, the following simulation test is performed with specific examples:
[0076] In this example, the DigSILENT simulation software is used to simulate the normal operation and the occurrence of disturbance events of the IEEE 39-node system, and a data set containing normal operation state and six types of power grid disturbance is constructed. The data set is constructed in accordance with the normal data and disturbance proportion of the real power system, and the time sequence synchrophasor measurement data is constructed synchronously, and 60dB of environmental noise is applied. The training set and the test set are constructed in a ratio of 10:1, the simulation step is set to 0.02s according to the sampling rate of the synchronous measurement device, and 600 disturbance events are generated by simulation, 100 for each event. The simulation method of each disturbance is shown in the following table 2:
[0077] Table 2 Simulation method of different disturbances
[0078]
[0079] To prove that the proposed method can meet the real-time requirements of the identification algorithm, the disturbance identification is performed on the computer, and the average time used by the method for identification is 40.67ms.
[0080] To verify the discrimination degree of the discriminator score vector obtained by the proposed TimeGAN model under the LightGBM classifier for each event, the features of the discriminator score vector extracted from each disturbance event are projected onto the same two-dimensional plane. The discriminator score vector of each type of disturbance extracted by the proposed method has discrimination degree, and the cut load event can be accurately divided from other disturbance events. The features of the cut load and the cut machine events are similar because both of them have a similar process of changing the frequency of the power system due to power shortage; the three-phase short circuit, single-phase short circuit and tripping events all belong to fast dynamic events, that is, the features of the disturbance events are represented in a short time window, and the extracted features have certain similarity. In the unlabeled feature extraction scene, the events with the above-mentioned intra-class similarity can be distinguished by features, and the expected disturbance identification effect of the model is achieved.
[0081] According to the method, 12 traditional power grid disturbances and 1 unknown event are identified. Among them, 1 is a generator trip event, 2 are load shedding events, 1 is a short circuit event, 7 are trip events, and 1 is an unknown event. According to the experience of experts, the unknown event is a frequency oscillation event. In addition, 124 local power quality events in the distribution network are identified. The identification results have been verified, and the event statistics are shown in Table 3:
[0082] Table 3: Classification results of identified events
[0083]
[0084]
[0085] Case 1: Take a generator trip event in North China Power Grid on a certain day as an example. The total power decreases by 700,000 kW. The SMD-L field data of frequency and voltage phase angle, amplitude of each node measurement at the time of disturbance is shown in Figure 4 .
[0086] For this event, the mean of the LST-GAN discriminator score vector of the present application is shown in Table 4 (the score is obtained by averaging the discriminator score vector containing complete event data, the score is between 0 and 1, and the score greater than 0.5 is judged to be abnormal)
[0087] Table 4: LST-GAN discriminator score vector mean abnormality detection results of case 1
[0088]
[0089] According to the pre-classification method of the proposed disturbance identification, the event belongs to the generator trip event.
[0090] Case 2: Take a load shedding event in North China Power Grid on a certain day as an example. The SMD-L field data of frequency and voltage phase angle, amplitude of each node measurement at the time of disturbance is shown in Figure 5 .
[0091] The mean of the LST-GAN discriminator score vector of the event is shown in Table 5.
[0092] Table 5: LST-GAN discriminator score vector mean abnormality detection results of case 2
[0093]
[0094] According to the pre-classification framework, the event belongs to the load shedding event; and it is confirmed that the event characteristics of the load shedding event and another load shedding event have very high characteristic similarity, so it is confirmed that the two events belong to one small class, and belong to the load shedding event class.
[0095] Case three: Take the circuit breaker tripping event of North China Power Grid on a certain day as an example, the SMD-L field data of the frequency and voltage phase angle, amplitude measured by each node are shown in Figure 6 Table 6, and the mean of the LST-GAN discriminator score vector of this event is shown in Table 6.
[0096] Table 6 LST-GAN discriminator score vector mean abnormality detection results of case three
[0097]
[0098] According to the pre-classification framework, it is judged that the event belongs to the tripping event.
[0099] Case four: Take a frequency oscillation event of North China Power Grid on a certain day as an example, the SMD-L field data of the frequency and voltage phase angle, amplitude measured by each node are shown in Figure 7 Table 7, and the mean of the LST-GAN discriminator score vector of this event is shown in Table 7.
[0100] Table 7 LST-GAN discriminator score vector mean abnormality detection results of case four
[0101]
[0102] The short-time frequency, phase angle and amplitude time series vector of this event all appear abnormal characteristics, and the frequency does not rise sharply. According to the pre-classification framework algorithm, it is judged that the event belongs to the non-traditional power grid event.
[0103] The above cases come from the SMD-L of the power quality problem case of the distribution network identified by the method, although only according to the sparse point SMD-L data, the specific event details cannot be confirmed, but the method can monitor the power quality problem of the distribution network where SMD-L is located to a certain extent, and can warn potential distribution network faults.
[0104] The identified power quality problems are automatically clustered by the LightGBM classifier and divided into 7 small categories, namely voltage surge, voltage sag, voltage temporary rise, voltage temporary drop, voltage interruption, voltage oscillation and unknown type problem. Among them, the SMD-L field measured voltage amplitude data of the first 6 categories of power quality problems are shown in the cases of Figure 8 (a), (b), (c), (d), (e) and (f) of Table 8.
[0105] Table 8 Statistics of identified distribution event classification results
[0106]
[0107]
[0108] Those skilled in the art can understand that all or part of the steps in the methods of the above embodiments can be instructed by programs to the relevant hardware to complete, and the corresponding programs can be stored in a computer readable storage medium. The storage medium mentioned above can be a read-only memory, a magnetic disk or an optical disk, etc.
[0109] The above description is merely preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto. Any changes or replacements easily conceived by those skilled in the art within the technical scope disclosed by the present application should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims. The information disclosed in the background section of the present application is merely intended to deepen the understanding of the overall background of the present application, and should not be regarded as acknowledging or implying in any form that the information constitutes the prior art known by those skilled in the art.
Claims
1. A method for unsupervised power system disturbance identification based on synchronized phasor measurements, characterized in that, The method comprises: Step 1, first, synchronous phasor measurement data of a power system to be processed is acquired, including frequency, voltage amplitude and voltage phase angle, and the frequency, voltage amplitude and voltage phase angle are normalized as inputs of a subsequent model; Step 2, a long short-term time sequence generation adversarial network LST-TimeGAN model is constructed, and the constructed model is used for non-event feature extraction of long-time, short-time and ultra-short-time time windows; In step 2, the constructed long short-term time sequence generation adversarial network LST-TimeGAN model comprises an encoder, a decoder, a generator and a discriminator, all of which are constructed by a neural network comprising three hidden layers, the number of nodes of the hidden layers is 24, and 128 samples are processed per round; the last layer of the generator is a new generation of recurrent neural network GRU; the last layer of the discriminator is a bidirectional LSTM; In addition to the last layer of the discriminator being a direct mapping based on a least square loss function, the activation functions used in the LST-TimeGAN model are all LeakyReLU, which is expressed as: f(x) = a * x, x < 0 a * x, x >= 0 wherein x represents the input of the LeakyReLU function; a is the slope of the LeakyReLU function, and is 0.01; The process of using the constructed model for non-event feature extraction of long-time, short-time and ultra-short-time time windows is as follows: (1) optimizing time and space feature extraction by scaled dot-product attention mechanism Each real input matrix is taken as a key sequence, denoted as K; a random input matrix of the generator is taken as a query sequence, denoted as Q; F is the mapping of each element in Q to each element in K; the attention of each random input of the generator is described as the mapping of the key sequence on the query, and since the lengths of the query and the key are the same, the similarity of K and Q is evaluated by using a scaled dot-product model to reduce the calculation cost, and the formula is as follows: The attention score S is normalized by using a softmax function to obtain the attention aggregation formula as follows: where S represents a similarity-based attention score; d k represents the dimension of K; (2) the decision loss function based on least square enables the output to contain distance features and be able to represent the degree of data abnormality Wherein, a represents the attention weight coefficient matrix; the attention weight coefficient matrix a and the random input matrix Q are weighted as the attention output, and become the new generator random input sequence Q a defined as: Attention(Q a ) = αQ The logarithmic loss is improved based on the least square loss function to obtain the following formula: where the first term expectation represents y s , y t is the least square distance to 1 ; y s , y t is the classification result of the discriminator on the real data static and dynamic parts; the second term expectation represents and is the least square distance to 0; and is the classification result of the discriminator on the static and dynamic data generated by the generator; the formula measures the difference between the generated model and the real sequence in terms of the least square loss; Furthermore, with L' = L + 1 U As loss function, the last layer of the discriminator will no longer use a sigmoid activation layer, but a linear mapping directly to [0,1] will be used. (3) constructing a long-time, short-time and ultra-short-time three-window architecture, and using the sensitivity of different kinds of perturbations to different lengths of time windows to extract unsupervised features of the perturbations The long-time TimeGAN extracts long-term time sequence features of multi-dimensional synchronous measurement data with a 24h cycle, and the time window for perturbation feature extraction is set to 5min; The short-time TimeGAN learns the normal fluctuation characteristics of the previous synchronous phasor measurement data with a 60min cycle, and the time window for perturbation feature extraction is set to 1min; The ultra-short-time TimeGAN uses the normal impulse features of the synchronous phasor measurement data with a 1min cycle to judge whether a non-normal impulse signal appears in the current time window, and the ultra-short-time perturbation feature extraction time window is set to 1s; A total of 9 TimeGAN models, i.e. [long-time, short-time, ultra-short-time] *[frequency f, voltage amplitude U and voltage phase angle φ] models are trained, and 9 groups of discriminator score vectors are output. If the score of any score vector or multiple score vectors is not close enough to the global optimum, it is determined that a disturbance event occurs in the window, and the output value of the decision loss function is taken as the score of data anomaly; a group of discriminator score vectors representing the change of window anomaly degree over time is obtained by scoring the rolling window containing the event window; the score vector from the start of the event to the end of the event is extracted as the data basis for subsequent event classification Step 3: According to the output of the model in step 2, combined with the electrical quantity and time window characteristics of the event, pre-classification of the disturbance is carried out; wherein, it specifically includes four traditional disturbance events of machine tripping, load shedding, short circuit and open circuit, and new events of unknown types and power quality problems; Specifically, the four traditional disturbance events of machine tripping, load shedding, short circuit and open circuit are preliminarily distinguished according to the time scale characteristics and slope characteristics of frequency and phase angle changes; In addition to the above events, the frequency and phase angle anomalies are recorded as new events of unknown types in the power transmission network, which will be identified again in the subsequent steps; Step 4: Using the LightGBM classifier to cluster and identify the pre-classified event categories, the identification and classification of the disturbance are realized.
2. The unsupervised power system disturbance identification method based on synchronized phasor measurements according to claim 1, characterized in that, In step 1, the acquired frequency f, voltage amplitude U are normalized and pre-processed to obtain input After the voltage phase angle φ is offset to a unified reference to obtain a relative phase angle, normalization processing is performed to obtain input Let the frequency f, voltage amplitude U and voltage phase angle φ before normalization be X, and the normalization formula be: wherein, Xi is the normalized value of the i-th sample data; X max and X min respectively represent the maximum value and the minimum value in the sample data; Definitions and χ are vector spaces representing static and dynamic characteristics of inputs and x e χ represent static and dynamic vectors within said vector spaces, respectively and x e χ represent static and dynamic vectors within said vector spaces, respectively Let (s, x 1：T ) be subject to a joint distribution P over a vector space, where T represents the length of the time series; if s n and x n represent elements in the vectors s and x, and n e {1,..., N}, then the training set is represented as 3. The unsupervised power system disturbance identification method based on synchronized phasor measurements according to claim 1, characterized in that, In step 3, first, the frequency quantity in different time scale windows can be used to distinguish slow and fast disturbances; wherein, the slow disturbance includes machine tripping and load shedding; the fast disturbance includes short circuit and open circuit; By the fact that the frequency change slope of machine tripping is greater than 0 and the frequency change slope of load shedding is less than 0, the two slow disturbances can be distinguished; By the different voltage phase angle anomaly time window characteristics of short circuit and trip-out events, the two fast disturbances can be distinguished, wherein: The trip-out event is usually caused by the immediate removal of the line after the event, and the recovery is fast, so the influence on the voltage phase angle is less than 20s, and only the anomaly in the ultra-short time characteristic is shown; the influence time of the short circuit event on the voltage phase angle is longer, and in addition to the anomaly in the ultra-short time characteristic, the anomaly in the short time characteristic is also detected; If the voltage amplitude is abnormal under the condition that the power transmission network is not abnormal, it is determined to be a power quality problem of the local distribution network.
4. The unsupervised power system disturbance identification method based on synchronized phasor measurements according to claim 1, characterized in that, In step 4, the discriminator score vector in the constructed model is taken as the feature vector, and the LightGBM classifier and the existing category information are used to re-cluster the events that have been preliminarily classified; The LightGBM classifier completes the labeling of data based on the similarity measurement and the distribution properties of data itself; wherein, since the feature vector can reflect the severity of the event in each time window scale, the LightGBM classifier classifies based on this, and events with similar severity and electrical quantity change characteristics in the existing category information are divided into the same subcategory.