Power system abnormal event identification method and device based on deep learning

Abstract

The invention discloses a power system abnormal event identification method and device based on deep learning. The method comprises the steps of collecting PMU data of a power grid system; screening the PMU data according to a preset abnormal event detection index threshold value, and determining occurrence time of each abnormal event; performing abnormal node positioning on the abnormal events according to the occurrence time of each abnormal event and a preset energy operator, and determining a first voltage signal of each abnormal event; encoding the first voltage signal of each abnormal event to obtain a plurality of two-dimensional images; and according to a preset neural network model, identifying the plurality of two-dimensional images, and determining an event type corresponding to each abnormal event. According to the invention, the abnormal characteristics of the power grid operation condition can be captured timely and accurately, the event type is identified accordingly, and rapid detection and accurate identification of the abnormal event of the power system are realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of power system situation awareness, and relates to a method and device for identifying abnormal events in a power system based on deep learning. Background Art

[0002] In recent years, the power system has developed rapidly, and a large number of power electronic equipment have been connected while the scale has continued to expand. The system operation status has become more and more complex, and the safety and stability issues caused by abnormal operation of the power grid have become increasingly prominent. With the increasing degree of interconnection between regional power grids, if abnormal events that occur during the operation of the power grid cannot be discovered in time and corrective measures cannot be taken, it may cause damage to electrical equipment, voltage fluctuations, and even large-scale power outages, seriously affecting residential electricity consumption and industrial production activities, causing huge economic losses.

[0003] Traditional power system information monitoring equipment is mainly based on the Supervisory Control And Data Acquisition (SCADA) system, which has problems such as long sampling cycle, poor real-time performance, and inability to reflect the dynamic changes in the power system operation status. In addition, most of the existing abnormal event identification methods do not consider the occurrence time and signal selection issues before identifying the event, which affects the efficiency and performance of identification. Therefore, it is of great practical significance to study a real-time perception method for the power system operation status to detect and identify abnormal events. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present application provides a method and device for identifying abnormal events in a power system based on deep learning, so as to achieve rapid detection and accurate identification of abnormal events in the power system.

[0005] To achieve the above objectives, in a first aspect, the present invention provides a method for identifying abnormal events in a power system based on deep learning, comprising:

[0006] Collecting PMU data of the power grid system; wherein the PMU data includes branch voltage data of each PMU;

[0007] Filter the PMU data according to a preset abnormal event detection index threshold to determine the occurrence time of each abnormal event;

[0008] According to the occurrence time of each abnormal event and the preset energy operator, the abnormal node of the abnormal event is located respectively, and the first voltage signal of each abnormal event is determined; wherein the first voltage signal is the voltage signal of the node corresponding to the maximum energy peak value when the energy operator responds to the voltage signal when the abnormal event occurs;

[0009] respectively encoding the first voltage signal of each abnormal event to obtain a plurality of two-dimensional images;

[0010] According to a preset neural network model, the multiple two-dimensional images are respectively identified to determine the event type corresponding to each abnormal event.

[0011] Compared with the prior art, the embodiments of the present application have the following beneficial effects: by collecting PMU data of the power system and filtering the PMU data using thresholds, the power system can be monitored in real time, abnormal characteristics of the power grid operation conditions can be captured in a timely and accurate manner, and signals of abnormal events can be identified using energy operators. By image encoding and identifying the signals, the accuracy of identifying the types of abnormal events is improved.

[0012] In some embodiments of the first aspect of the present application, the PMU data is screened according to a preset abnormal event detection index threshold to determine the occurrence time of each abnormal event, including: determining the voltage coefficient corresponding to each branch based on the branch voltage data of each PMU in the PMU data, and taking the moment when the voltage coefficient is greater than the abnormal event detection index threshold as the occurrence time of each abnormal event.

[0013] Compared with the prior art, the above embodiment has the following beneficial effects: by using a preset threshold to screen the voltage coefficient, the occurrence of each abnormal event can be quickly determined by utilizing the characteristic of the voltages of the associated nodes of each branch changing in the same direction when an abnormal event occurs.

[0014] In some embodiments of the first aspect of the present application, the voltage coefficient corresponding to each branch is determined according to the branch voltage data of each PMU in the PMU data, including: setting a sliding time window in each branch, and the voltage coefficient r of the i-th branch at the last moment of the current window i for:

[0015]

[0016] Where T is the length of the time window, and the node a at the beginning and the node b at the end of each branch are the associated nodes of the branch. and are the voltage amplitudes of nodes a and b at time t, and They are and The average value in time T;

[0017] The average voltage coefficient of n branches is used As the final voltage coefficient:

[0018] Compared with the prior art, the above embodiment has the following beneficial effects: taking the average value of the voltage coefficients of multiple branches can avoid the failure of the method due to the failure of the characteristic of the same-direction change of the voltage of the associated node when an abnormal event occurs inside the associated node, and can also reduce randomness and errors.

[0019] In some embodiments of the first aspect of the present application, the time when the voltage coefficient is greater than the abnormal event detection index threshold is used as the occurrence time of each abnormal event, including: presetting the abnormal event detection index threshold as r th , will be The timestamp at h As the occurrence time of each abnormal event.

[0020] Compared with the prior art, the above embodiment has the following beneficial effects: when the voltage coefficient exceeds the threshold, it is considered that an abnormal event has occurred, and at the same time, the associated nodes corresponding to the voltage coefficient show a strong correlation, which is conducive to improving the accuracy of judging the occurrence of abnormal events and time positioning.

[0021] In some embodiments of the first aspect of the present application, the abnormal nodes of the abnormal events are located according to the occurrence time of each abnormal event and a preset energy operator, and the first voltage signal of each abnormal event is determined, including: at the occurrence time of each abnormal event, using the dTKEO energy operator to calculate the signal energy peak of each node of the power system, comparing the signal energy peak of each node, and taking the voltage signal of the node corresponding to the maximum value of the signal energy peak as the first voltage signal of each abnormal event.

[0022] Compared with the prior art, the above embodiment has the following beneficial effects: dTKEO has an extremely strong response to instantaneous changes in the signal, a signal energy peak will appear at the moment of the event, and the change is very small in the quasi-steady state or transient process after the disturbance. At the same time, its energy peak can also reflect the relative proximity of the PMU to the event location. By utilizing this characteristic of dTKEO, the accuracy of voltage signals for screening abnormal events is greatly improved.

[0023] In some embodiments of the first aspect of the present application, the first voltage signal of each abnormal event is encoded respectively to obtain multiple two-dimensional images, including: normalizing the first voltage signal data of each abnormal event to obtain standard data, converting the standard data from a Cartesian coordinate system into data in a polar coordinate system, calculating the data in the polar coordinate system to obtain a Gram matrix, and generating corresponding multiple GAF two-dimensional images.

[0024] Compared with the prior art, the above embodiment has the following beneficial effects: converting the original voltage signal data into a GAF two-dimensional image can highlight the differences between data points, simplify the data processing process, and facilitate subsequent feature extraction and event recognition of the image.

[0025] In some embodiments of the first aspect of the present application, the multiple two-dimensional images are identified separately according to a preset neural network model to determine the event type corresponding to each abnormal event, including: inputting the two-dimensional image into a convolutional neural network model, so that the convolutional neural network model outputs the event type corresponding to each abnormal event.

[0026] Compared with the prior art, the above embodiment has the following beneficial effects: the convolutional neural network model can automatically learn the feature extractor, has high flexibility and adaptability, and effectively ensures the accuracy of the image recognition results.

[0027] In a second aspect, an embodiment of the present application further provides a power system abnormal event identification device based on deep learning, comprising: an acquisition module, a screening module, a voltage signal determination module, an encoding module and an event type identification module;

[0028] Wherein, the acquisition module is used to acquire PMU data in the power system; wherein, the PMU data includes branch voltage data of each PMU;

[0029] The screening module is used to screen the PMU data according to a preset abnormal event detection index threshold to determine the occurrence time of each abnormal event;

[0030] The voltage signal determination module is used to locate abnormal nodes of the abnormal events according to the occurrence time of each abnormal event and the preset energy operator, and determine the first voltage signal of each abnormal event; wherein the first voltage signal is the voltage signal corresponding to the node with the maximum energy when the abnormal event occurs;

[0031] The encoding module is used to perform image encoding on the first voltage signal of each abnormal event to obtain multiple two-dimensional images;

[0032] The event type recognition module is used to recognize the multiple two-dimensional images respectively according to a preset neural network model to determine the event type corresponding to each abnormal event.

[0033] Compared with the prior art, the embodiments of the present application have the following beneficial effects: by collecting PMU data of the power system and filtering the PMU data using thresholds, the power system can be monitored in real time, abnormal characteristics of the power grid operation conditions can be captured in a timely and accurate manner, and signals of abnormal events can be identified using energy operators. By image encoding and identifying the signals, the accuracy of identifying the types of abnormal events is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 : A flow chart of a method for identifying abnormal events in a power system based on deep learning provided in some embodiments of the present invention.

[0035] Figure 2 : A voltage coefficient curve diagram of a test node before and after an abnormal event occurs in some embodiments of the present invention.

[0036] Figure 3 : A dTKEO response diagram of a test node after an abnormal event occurs in some embodiments of the present invention.

[0037] Figure 4 : A waveform diagram of a voltage signal of a test node and its corresponding GAF image provided in some embodiments of the present invention.

[0038] Figure 5 : A graph showing the test results of training a convolutional neural network provided in some embodiments of the present invention.

[0039] Figure 6 : A structural diagram of a power system abnormal event identification device based on deep learning provided in some embodiments of the present invention. DETAILED DESCRIPTION

[0040] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0041] Embodiment 1:

[0042] Please refer to Figure 1 , a method for identifying abnormal events in a power system based on deep learning is provided in an embodiment of the present invention, comprising steps S1 to S5:

[0043] Step S1: Collect PMU data of the power grid system.

[0044] In this embodiment, the PMU data includes branch voltage data of each PMU.

[0045] Step S2: Filter the PMU data according to a preset abnormal event detection index threshold to determine the occurrence time of each abnormal event.

[0046] In some embodiments of the present application, step S2 can be implemented by the following preferred implementations:

[0047] S21: Set a sliding time window in each branch, and the voltage coefficient r of the i-th branch at the last moment of the current window i for:

[0048] Where T is the length of the time window, and the node a at the beginning and the node b at the end of each branch are the associated nodes of the branch. and are the voltage amplitudes of nodes a and b at time t, and They are and The average value in time T;

[0049] In order to reduce errors and contingencies, the voltage coefficients of n branches can be averaged as the voltage coefficient for threshold screening. The calculation formula is as follows:

[0050] S22: Preset a threshold r th As an indicator for detecting abnormal events, The timestamp at h As the occurrence time of each abnormal event.

[0051] The threshold may be a Pearson coefficient, with a value of 0.6 used to verify whether there is a strong correlation between the voltage coefficients of the associated nodes, or different thresholds may be set according to actual conditions.

[0052] Compared with the prior art, the above embodiment has the following beneficial effects: when the voltage coefficient exceeds the threshold, it is considered that an abnormal event has occurred, and at the same time, the associated nodes corresponding to the voltage coefficient show a strong correlation, which is conducive to improving the accuracy of judging the occurrence of abnormal events and time positioning.

[0053] Step S3: locate abnormal nodes of the abnormal events according to the occurrence time of each abnormal event and the preset energy operator, and determine the first voltage signal of each abnormal event; wherein the first voltage signal is the voltage signal of the node corresponding to the maximum energy peak value when the energy operator responds to the voltage signal when the abnormal event occurs.

[0054] In some embodiments of the present application, step S3 can be implemented by the following preferred implementations:

[0055] S31: At the occurrence time of each abnormal event, using the dTKEO energy operator to calculate the signal energy peak value of each node of the power system;

[0056] The signal energy peak value Ψ d[u(t)] is calculated as follows:

[0057] ψ d [u(t)] = [u(t) - u(t-1)] 2 -[u(t-1)-u(t-2)][u(t+1)-u(t)]; where u(t) is the voltage of the node at time t.

[0058] S32: Compare the signal energy peak values ​​of each node, and use the voltage signal of the node corresponding to the maximum value of the signal energy peak value as the first voltage signal of each abnormal event.

[0059] Compared with the prior art, the above embodiment has the following beneficial effects: dTKEO has an extremely strong response to instantaneous changes in the signal, a signal energy peak will appear at the moment of the event, and the change is very small in the quasi-steady state or transient process after the disturbance. At the same time, its energy peak can also reflect the relative proximity of the PMU to the event location. By utilizing this characteristic of dTKEO, the accuracy of voltage signals for screening abnormal events is greatly improved.

[0060] Step S4: Encode the first voltage signal of each abnormal event respectively to obtain a plurality of two-dimensional images.

[0061] Preferably, in some embodiments of the present application, the first voltage signals are image-encoded to obtain a plurality of two-dimensional images, wherein the encoding method may adopt a Gram angular field algorithm, and the specific steps include S41 to S43, each step is as follows:

[0062] S41: First, the voltage signal data is normalized to obtain standard data. The normalization formula is as follows: Among them, u t is the data sequence of the first voltage signal, u i is the i-th element in the data sequence;

[0063] S42: Convert the standard data from the Cartesian coordinate system into data in the polar coordinate system, and the conversion formula is as follows: Where φ is the angle in polar coordinates, r is the radius in polar coordinates, and t i is the timestamp corresponding to the i-th element, and N is a constant factor;

[0064] S43: Calculate the data in the polar coordinate system to obtain a Gram matrix, and generate a corresponding plurality of GAF two-dimensional images, wherein the Gram matrix is ​​defined as follows:

[0065]

[0066] where φ nis the nth angle data in the data under the polar coordinate system, and n is a positive integer.

[0067] Compared with the prior art, the above embodiment has the following beneficial effects: converting the original voltage signal data into a GAF two-dimensional image can highlight the differences between data points, simplify the data processing process, and facilitate subsequent feature extraction and event recognition of the image.

[0068] Step S5: According to a preset neural network model, the multiple two-dimensional images are respectively identified to determine the event type corresponding to each abnormal event.

[0069] Preferably, in some embodiments of the present application, a convolutional neural network can be used for event recognition, and the two-dimensional image is input into a convolutional neural network model so that it outputs the event type corresponding to each abnormal event to complete event recognition.

[0070] Compared with the prior art, the above embodiment has the following beneficial effects: the convolutional neural network model can automatically learn the feature extractor, has high flexibility and adaptability, and effectively ensures the high accuracy of the image recognition results.

[0071] In summary, compared with the prior art, the embodiments of the present application have the following beneficial effects: by collecting PMU data of the power system and using thresholds to filter the PMU data, the power system can be monitored in real time, abnormal characteristics of the power grid operation conditions can be captured in a timely and accurate manner, and signals of abnormal events can be identified using energy operators. By performing image encoding and identification on the signals, the accuracy of identifying the types of abnormal events is improved.

[0072] In order to better illustrate the working principle and steps of this embodiment, the following example is used as an example:

[0073] This experimental example uses the IEEE39 node system for simulation verification. The algorithm is run in Jupyter Notebook. PMUs are installed at each node location to collect power data from each node and branch. A total of 30 abnormal events are simulated, including:

[0074] For short-circuit faults, nodes 1 to 10 are selected as the three-phase short-circuit fault locations, the fault occurrence time is set to 2s, and the fault clearing time is 2.2s;

[0075] For line tripping, line 1 to line 10 are selected as the tripping locations, and the occurrence time is set to 3s; for load mutation, load 1 to load 10 are selected as the event locations, the load mutation time is set to 4s, and the load mutation amplitude is ±1p.u.;

[0076] Taking into account that the actual measurement data has certain errors, a Gaussian noise of 40 dB is superimposed on the data during the event obtained by simulation, and the simulation time is uniformly set to 10 s and the simulation step is 0.01 s.

[0077] The simulation results are as follows:

[0078] (A) Time location of abnormal events:

[0079] The voltage coefficient is used to detect the above events, and the time window T = 50, such as Figure 2 As shown, Figure 2 (a) Figure 2 (b) and Figure 2 (c) The three sub-graphs are the voltage coefficient change curves of the corresponding nodes before and after the occurrence of the above three abnormal events. It can be seen from the figure that before the abnormal event occurs, the system is in normal operation, and the voltage coefficient at this time fluctuates around 0, indicating that there is no correlation between the voltages of the associated nodes of each branch. At the moment of the abnormal event, the voltage measurement data of all nodes suddenly changes, and the voltage coefficient quickly rises to a value close to 1. Finally, as the system transitions to a steady state, the voltage coefficient gradually drops. In addition, the voltage coefficient curves of the other 27 events are also consistent with Figure 2 similar.

[0080] (B) The first voltage signal determines:

[0081] for Figure 2 For the three events shown, the dTKEO response of the voltage signal at the time of the abnormal event is calculated. The results are as follows: Figure 3 As shown, Figure 3 (a) Figure 3 (b) and Figure 3 (c) The three sub-graphs are the result graphs of the corresponding nodes' responses to dTKEO when the above three abnormal events occur. It can be seen that a short circuit fault occurs at node 1 at 2.0s, and the dTKEO responses of all node voltages have peaks at this time, and the highest peak is node 1; similarly, line 2-3, that is, the line connecting node 2 and node 3, trips at 3.0s, and the signal energy peak of node 3 is the highest among all nodes. Finally, the voltage signals corresponding to the nodes with signal energy peaks are taken as the first voltage signals.

[0082] (C) Image coding: The three node voltage signals obtained by the first voltage signal determination step are subjected to image coding. The result is as follows: Figure 4 Indicates that the voltage waveform diagram is the voltage waveform of the three nodes within two seconds from the occurrence of the event, and the GAF image is the corresponding GAF image after the voltage waveform is encoded.

[0083] (D) Event recognition: A convolutional neural network model is used for event recognition. In order to verify the effectiveness of the CNN-based event recognition method, an event dataset is prepared to generate a model and evaluate the model performance. PSAT is used to simulate the IEEE39 node system, simulating three types of abnormal events and normal operation: three-phase symmetrical short circuit, line tripping, and load mutation. The occurrence time of the abnormal event is fixed at t o ,The simulation parameters are set as follows: the three-phase symmetrical short-circuit position is from node 1 to node 34, and the fault duration t con ~U(0.1,0.5); the tripping event location is line 1 to line 34; the load mutation event location is load 1 to load 17, of which the load sudden increase and load drop each account for half, and the load mutation amplitude is ±(0.5~2)pu. The simulation step size is set to 1ms, with t o The voltage data of nodes 1 to 30 are collected with a 2s window as the starting point, and the noise level is still 40dB. The data set contains a total of 20,400 samples, of which each type of event accounts for a quarter, and is divided into training and test sets in a ratio of 4:1.

[0084] Since Paddle has powerful GPU-accelerated tensor computing capabilities and is equipped with an automatic derivation system, we chose the Paddle deep learning framework to build the required CNN model. Since it is a classification task, the cross entropy loss function is used to measure the distance between the true value and the predicted value, and the Adam optimizer is used for batch stochastic gradient descent to reduce the value of the loss function and update the weight parameters of the network. The learning rate is 0.001, the batch size is 64, and the loss value and classification accuracy change as the training progresses. Figure 5 (a) and Figure 5 (b) As shown. By observation, it can be found that during the training process, the loss value and accuracy of the neural network gradually converge. The accuracy of the final saved model on the test set reached 98.63%, and its confusion matrix is ​​as follows Figure 5 (c) shows that SC refers to short circuit, LT refers to line tripping, LD refers to load mutation, and ST refers to normal operation. It can be seen that the classification effect is good, which verifies the feasibility and effectiveness of this scheme.

[0085] Embodiment 2:

[0086] Please refer to Figure 6 Based on the same inventive concept, an embodiment of the present invention discloses a device for identifying abnormal events in a power system based on deep learning, comprising: an acquisition module M1, a screening module M2, a voltage signal determination module M3, an encoding module M4 and an event type identification module M5.

[0087] Wherein, the acquisition module M1 is used to acquire PMU data in the power system; wherein, the PMU data includes branch voltage data of each PMU;

[0088] The screening module M2 is used to screen the PMU data according to a preset abnormal event detection index threshold to determine the occurrence time of each abnormal event;

[0089] The voltage signal determination module M3 is used to locate the abnormal nodes of the abnormal events according to the occurrence time of each abnormal event and the preset energy operator, and determine the first voltage signal of each abnormal event; wherein the first voltage signal is the voltage signal corresponding to the node with the maximum energy when the abnormal event occurs;

[0090] The encoding module M4 is used to perform image encoding on the first voltage signal of each abnormal event to obtain multiple two-dimensional images;

[0091] The event type identification module M5 is used to identify the multiple two-dimensional images respectively according to a preset neural network model to determine the event type corresponding to each abnormal event.

[0092] Wherein, the screening module M2 includes:

[0093] A voltage coefficient calculation unit M21 is used to calculate the branch voltage data of each PMU in the PMU data and determine the voltage coefficient corresponding to each branch;

[0094] The time positioning unit M22 is used to locate the moment when the voltage coefficient is greater than the abnormal event detection index threshold as the occurrence time of each abnormal event.

[0095] The voltage coefficient calculation unit M21 is used to set a sliding time window in each branch and calculate the voltage coefficient r of the i-th branch at the last moment of the current window. i for:

[0096]

[0097] Where T is the length of the time window, the node a at the beginning of the branch and the node b at the end are the associated nodes of the branch, and are the voltage amplitudes of nodes a and b at time t, and are the voltage amplitudes at the beginning and end nodes of the branch at time t, and They are and The average value in time T;

[0098] The average voltage coefficient of n branches is used As the final voltage coefficient:

[0099]

[0100] The time positioning unit M22 is used to preset the abnormal event detection index threshold value as r th , will be The timestamp at h As the occurrence time of each abnormal event.

[0101] The voltage signal determination module M3 includes:

[0102] An energy operator calculation unit M31 is used to calculate the signal energy peak value of each node of the power system using the dTKEO energy operator at the time of occurrence of each abnormal event;

[0103] The first signal determination unit M32 is used to compare the signal energy peak values ​​of each node, and use the voltage signal of the node corresponding to the maximum value of the signal energy peak value as the first voltage signal of each abnormal event.

[0104] The energy operator calculation unit M31 is used to calculate the corresponding signal energy peak value Ψ in the dTKEO discrete domain using the voltage signal of each node at the time t of occurrence of each abnormal event d [u(t)]:

[0105] ψ d [u(t)] = [u(t) - u(t-1)] 2 -[u(t-1)-u(t-2)][u(t+1)-u(t)]; where u(t) is the voltage of the node at time t.

[0106] The encoding module M4 includes:

[0107] The normalization unit M41 is used to perform normalization processing on the first voltage signal data of each abnormal event. The processing formula of the normalization unit is:

[0108] Among them, u t is the data sequence of the first voltage signal, u i is the i-th element in the data sequence;

[0109] Coordinate conversion unit M42: used to convert the normalized data from the Cartesian coordinate system into data in the polar coordinate system. The conversion formula of the coordinate conversion unit is:

[0110]

[0111] Where φ is the angle in polar coordinates, r is the radius in polar coordinates, and t i is the timestamp corresponding to the i-th element, and N is a constant factor;

[0112] Image generation unit M43: used to calculate the data in the polar coordinate system to obtain a Gram matrix, and generate a corresponding plurality of GAF two-dimensional images. The Gram matrix used by the image generation unit is defined as:

[0113] where φ n is the nth angle data in the data under the polar coordinate system, and n is a positive integer.

[0114] The event type identification module M5 includes: inputting the two-dimensional image into a convolutional neural network model, so that the convolutional neural network model outputs the event type corresponding to each abnormal event.

[0115] Compared with the prior art, the embodiments of the present application have the following beneficial effects: by collecting PMU data of the power system and filtering the PMU data using thresholds, the power system can be monitored in real time, abnormal characteristics of the power grid operation conditions can be captured in a timely and accurate manner, and signals of abnormal events can be identified using energy operators. By image encoding and identifying the signals, the accuracy of identifying the types of abnormal events is improved.

[0116] The specific working process of each module described above can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here. The module division is only a logical function division, and there may be other division methods in actual implementation, such as multiple modules can be combined or integrated into another system.

[0117] The specific embodiments described above further illustrate the purpose, technical solutions and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. It is particularly pointed out that for those skilled in the art, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for identifying abnormal events in a power system based on deep learning, characterized in that: include: Collecting PMU data of the power grid system; wherein the PMU data includes branch voltage data of each PMU; Filter the PMU data according to a preset abnormal event detection index threshold to determine the occurrence time of each abnormal event; According to the occurrence time of each abnormal event and the preset energy operator, the abnormal node of the abnormal event is located respectively, and the first voltage signal of each abnormal event is determined; wherein the first voltage signal is the voltage signal of the node corresponding to the maximum energy peak value when the energy operator responds to the voltage signal when the abnormal event occurs; respectively encoding the first voltage signal of each abnormal event to obtain a plurality of two-dimensional images; According to a preset neural network model, the multiple two-dimensional images are respectively identified to determine the event type corresponding to each abnormal event.

2. The method for identifying abnormal events in a power system based on deep learning according to claim 1, characterized in that: The screening of the PMU data according to a preset abnormal event detection index threshold to determine the occurrence time of each abnormal event includes: Determine the voltage coefficient corresponding to each branch according to the branch voltage data of each PMU in the PMU data; The time when the voltage coefficient is greater than the abnormal event detection index threshold is taken as the occurrence time of each abnormal event.

3. The method for identifying abnormal events in a power system based on deep learning according to claim 2, characterized in that: The step of determining the voltage coefficient corresponding to each branch according to the branch voltage data of each PMU in the PMU data includes: A sliding time window is set for each branch, and the voltage coefficient r of the i-th branch at the last moment of the current window is i for: Where T is the length of the time window, and the node a at the beginning and the node b at the end of each branch are the associated nodes of the branch. and are the voltage amplitudes of nodes a and b at time t, and They are and The average value in time T; The average voltage coefficient of n branches is used As the final voltage coefficient:

4. The method for identifying abnormal events in a power system based on deep learning according to claim 3, characterized in that: The time when the voltage coefficient is greater than the abnormal event detection index threshold is taken as the occurrence time of each abnormal event, including: The threshold value of the abnormal event detection index is preset to r th , will be The timestamp at t h As the occurrence time of each abnormal event.

5. The method for identifying abnormal events in a power system based on deep learning according to claim 4, characterized in that: The method of locating abnormal nodes of the abnormal events according to the occurrence time of each abnormal event and a preset energy operator, and determining the first voltage signal of each abnormal event, includes: At the occurrence time of each abnormal event, using the dTKEO energy operator to calculate the signal energy peak value of each node of the power system; The signal energy peak values ​​of each node are compared, and the voltage signal of the node corresponding to the maximum value of the signal energy peak value is used as the first voltage signal of each abnormal event.

6. The method for identifying abnormal events in a power system based on deep learning according to claim 5, characterized in that: At the occurrence time of each abnormal event, using the dTKEO energy operator to calculate the signal energy peak value of each node of the power system includes: Using the voltage signals of the nodes at the time t of each abnormal event, the corresponding signal energy peak value Ψ in the dTKEO discrete domain is calculated. d [u(t)]: ψ d [u(t)]=[u(t)-u(t-1)] 2 -[u(t-1)-u(t-2)][u(t+1)-u(t)]。 7. The method for identifying abnormal events in a power system based on deep learning according to claim 1, characterized in that: The first voltage signal of each abnormal event is encoded respectively to obtain a plurality of two-dimensional images, including: Normalizing the first voltage signal data of each abnormal event to obtain standard data; Converting the standard data from a Cartesian coordinate system into data in a polar coordinate system; The data in the polar coordinate system is calculated to obtain a Gram matrix, and a corresponding plurality of GAF two-dimensional images are generated.

8. The method for identifying abnormal events in a power system based on deep learning according to claim 7, characterized in that: The first voltage signal of each abnormal event is encoded respectively to obtain a plurality of two-dimensional images, including: The formula for the normalization process is: Among them, u t is the data sequence of the first voltage signal, u i is the i-th element in the data sequence; The formula for the coordinate system transformation is: Where φ is the angle in polar coordinates, r is the radius in polar coordinates, and t i is the timestamp corresponding to the i-th element, and N is a constant factor; The Gram matrix is ​​defined as: where φ n is the nth angle data in the data under the polar coordinate system, and n is a positive integer.

9. The method for identifying abnormal events in a power system based on deep learning according to claim 1, characterized in that: According to a preset neural network model, the plurality of two-dimensional images are respectively identified to determine the event type corresponding to each abnormal event, including: The two-dimensional image is input into a convolutional neural network model, so that the convolutional neural network model outputs the event type corresponding to each abnormal event.

10. A power system abnormal event identification device based on deep learning, characterized in that: include: an acquisition module, a screening module, a voltage signal determination module, an encoding module and an event type identification module; Wherein, the acquisition module is used to acquire PMU data in the power system; wherein, the PMU data includes branch voltage data of each PMU; The screening module is used to screen the PMU data according to a preset abnormal event detection index threshold to determine the occurrence time of each abnormal event; The voltage signal determination module is used to locate abnormal nodes of the abnormal events according to the occurrence time of each abnormal event and the preset energy operator, and determine the first voltage signal of each abnormal event; wherein the first voltage signal is the voltage signal corresponding to the node with the maximum energy when the abnormal event occurs; The encoding module is used to perform image encoding on the first voltage signal of each abnormal event to obtain multiple two-dimensional images; The event type recognition module is used to recognize the multiple two-dimensional images respectively according to a preset neural network model to determine the event type corresponding to each abnormal event.

Images