Power system disturbance identification method considering unknown disturbance types based on deep learning

Abstract

The present invention discloses a method for identifying power system disturbances based on deep learning and taking into account unknown disturbance types, comprising the following steps: S1, according to power system parameters, using power grid transient simulation data and power grid historical operation data, obtaining time series data when various disturbances occur, and generating a time series disturbance data set; S2, using filtering technology to denoise the time series disturbance data set; 3, performing feature extraction on the time series disturbance data set according to feature indicators, generating a disturbance identification feature sample set, and dividing it into a training set and a test set; S4, constructing a composite model including a deep neural network and a heuristic judgment layer, and training the deep neural network using the training set and its corresponding disturbance type labels; S5, inputting the test set into the composite model, the deep neural network outputting a pre-classification result, and judging the similarity between the pre-classification result and the test set and the training set, the heuristic judgment layer outputting a final disturbance identification result taking into account the unknown disturbance.

Description

Technical Field

[0001] The present invention belongs to the field of power system power quality analysis, and in particular, relates to a power system disturbance identification method based on deep learning and considering unknown disturbance types. Background Art

[0002] Traditionally, the field of power grid disturbance identification has primarily focused on model-driven, offline methods. These methods rely on power system topology and parameters to model the grid and identify disturbance types based on their characteristics and triggering mechanisms. For example, traditional disturbance identification methods often rely on real-time synchronized phase measurement unit (PMU) data and power system power flow calculations. This requires accurate knowledge of the power system topology and operating conditions, and places certain requirements on information interoperability between various components of the power system. Some techniques formulate constrained optimization problems based on the power balance model, behavioral characteristics, and specific state information of the power grid, serving as supplementary checks within power grid early warning systems or larger multi-model detection systems. Other research reformulates the DC power flow model as a sparse overcomplete extension and utilizes compressed sampling and variable selection to identify power system disconnection faults. Some techniques use bus voltage to predict bus voltage amplitudes using an impedance matrix and the maximum available fault current. The error function is estimated by taking the difference between the measured and predicted bus voltage amplitudes, thereby enabling disturbance identification. Other technologies use the busbar voltage phase angle difference observed by synchronous phase measurement devices before and after system events to construct and solve optimization problems to detect the occurrence of system events. Researchers have developed a large-scale power grid security analysis, early warning, and control system capable of preliminary rehearsal, analysis, and early warning of various power grid faults and accidents.

[0003] However, with the continuous development of high intelligence and automation of power systems, on the other hand, the accumulation of a large number of measurement methods and multi-time scale data has also brought new challenges to the operation analysis and evaluation of power grids. The traditional "model-driven" power grid disturbance identification technology has the following problems: ① The scale of power systems is getting larger and larger, and large-scale regional interconnection, AC / DC hybrid and other phenomena have led to increasingly complex grid forms, and the grid has gradually transformed into a high-dimensional dynamic system. The disturbance identification method based on the accurate model of the power system is becoming more and more difficult to establish and calculate; ② The access of a high proportion of new energy has led to increasingly complex operation of the power grid, and the volatility and uncertainty of the power grid have increased. Model-driven methods are no longer suitable for this purpose. The online update method is difficult to adapt to the complex and changeable operation mode of the current power grid; ③ The coupling relationship between the power system and other energy systems is becoming increasingly close. In the future, a comprehensive energy system with electricity as the core will be formed. Traditional methods such as model construction and parameter optimization may be difficult to adapt to the rapid development of intelligent power systems and meet the power system's requirements for accuracy or efficiency; ④ Due to the limitation of computing power, the "model-driven" method may have high requirements for computing efficiency and it is difficult to achieve real-time calculation, thus losing the rapidity of disturbance identification; ⑤ The model-driven method focuses more on the analysis of the disturbance mechanism, but for unknown types of disturbances, it is difficult to judge them because of the inability to grasp their changing characteristics.

[0004] Therefore, there is an urgent need to carry out research on data generation and preprocessing technologies related to data-driven power system disturbance identification in the cyber-physical environment of smart grids, and disturbance classification considering unknown disturbance types. Summary of the Invention

[0005] The purpose of the present invention is to overcome the defect of the existing technology that it is difficult to perform real-time online identification of disturbances in new power grids, and to provide a power system disturbance identification method based on deep learning considering unknown disturbance types, so as to realize data-driven power grid disturbance discrimination in the context of new smart grids.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] A method for identifying power system disturbances based on deep learning and considering unknown disturbance types includes the following steps:

[0008] S1. Based on the power system parameters, using the power grid transient simulation data and the power grid historical operation data, obtain the time series data of the positive sequence voltage, negative sequence voltage, and zero sequence voltage when various disturbances occur, and generate the time series disturbance data set of the power system;

[0009] S2, using filtering technology to denoise the time series disturbance dataset of the power system;

[0010] S3. Extract features from the time series disturbance dataset based on the feature index to obtain disturbance feature data and use the disturbance type corresponding to the disturbance feature data as a label to generate a disturbance identification feature sample set, which is divided into a training set and a test set.

[0011] S4. Construct a composite model for power system disturbance identification based on a deep neural network, which includes a deep neural network and a heuristic judgment layer, and train the deep neural network using the training set data and its corresponding disturbance type labels;

[0012] S5. Input the test set data into the power system disturbance identification composite model based on deep neural network. The deep neural network outputs the pre-classification result. Based on the pre-classification result and the similarity between the test set data and the training set data, the heuristic judgment layer outputs the final disturbance identification result taking into account the unknown disturbance type.

[0013] Furthermore, the power system parameters include the grid structure of the power system, the physical parameters of the electrical equipment contained in the power system, and the output and load demand data of the generator sets under different operating modes;

[0014] The power grid transient simulation data is generated by the power grid through transient calculation simulation using various analysis software that can perform power system transient simulation calculations;

[0015] The historical operation data of the power grid is the historical operation data of the bus voltage in the time series recorded when disturbances occur during the actual operation of the power grid;

[0016] The time series data of the positive-sequence voltage, negative-sequence voltage, and zero-sequence voltage when various disturbances occur are the time series data of the positive-sequence voltage amplitude, negative-sequence voltage amplitude, zero-sequence voltage amplitude, and phase of the busbar when a single-phase short circuit to grounding, two-phase short circuit, two-phase short circuit to grounding, three-phase short circuit, induction motor self-starting, or other different types of disturbances occur.

[0017] Furthermore, the filtering technology is sliding average filtering, and the filtering formula is:

[0018]

[0019] Where k is the filter window length, is the data before filtering, {x i ,i=1,2,...,N-k+1} is the filtered data.

[0020] Furthermore, the characteristic indicators include kurtosis factor, total time series variation, C3 coefficient, FFT spectrum center, power spectrum density and standard deviation.

[0021] Furthermore, for a time series perturbation dataset {x i,i=1,2,...,n}, the kurtosis factor K measures the smoothness of the disturbance waveform. The formula of the kurtosis factor K is:

[0022]

[0023] Where x i is the i-th data in the time series perturbation data set, n is the length of a time series perturbation data set, is the average value of the time series perturbation data set;

[0024] Total time series change C abs Measure the change amplitude of the disturbance waveform, the total change in time series C abs The formula is:

[0025]

[0026] Where x i+1 is the i+1th data in the time series perturbation dataset;

[0027] The C3 coefficient measures the nonlinearity of the disturbance waveform. The formula for the C3 coefficient is:

[0028]

[0029] Where x i+2lag is the i+2lag data in the time series perturbation data set, x i+lag is the i+lag data in the time series disturbance data set, where lag is an integer set artificially to represent the phase lag;

[0030] FFT spectrum center C FFT Measure the frequency domain distribution of the disturbance waveform, FFT spectrum center C FFT is the spectrum center of the absolute value of the discrete Fourier transform of the disturbance data, and the formula of the n-point discrete Fourier transform is:

[0031]

[0032] Where, e is the natural base;

[0033] Power spectral density S xx (m) measures the power spectrum distribution of the disturbance waveform, power spectrum density S xx The formula for (m) is:

[0034]

[0035] Where, * indicates conjugation;

[0036] The standard deviation σ measures the statistical distribution of the disturbance waveform. The formula for the standard deviation σ is:

[0037]

[0038] Furthermore, the deep neural network includes deep neural network 1, deep neural network 2 and deep neural network 0, deep neural network 1 outputs the positive sequence data classification result, deep neural network 2 outputs the negative sequence data classification result, deep neural network 0 outputs the zero sequence data classification result, and the positive sequence data classification result, negative sequence data classification result and zero sequence data classification result are respectively input into the heuristic judgment layer.

[0039] Furthermore, the positive sequence test set data is compared with the positive sequence training set data corresponding to the positive sequence data classification result, the negative sequence test set data is compared with the negative sequence training set data corresponding to the negative sequence data classification result, and the zero sequence test set data is compared with the zero sequence training set data corresponding to the zero sequence data classification result, and the positive sequence similarity measurement value, the negative sequence similarity measurement value and the zero sequence similarity measurement value are obtained respectively:

[0040]

[0041] Where sim1 is the positive sequence similarity measure, sim2 is the negative sequence similarity measure, and sim0 is the zero sequence similarity measure. is the positive sequence test set data, is the negative sequence test set data, is the zero-sequence test set data, x y1 is the positive sequence training set data, x y2 is the negative sequence training set data, x y0 is the zero-sequence training set data.

[0042] Furthermore, any two data in the positive sequence training set data, the negative sequence training set data, and the zero sequence training set data are compared respectively to obtain the positive sequence similarity threshold, the negative sequence similarity threshold, and the zero sequence similarity threshold, respectively:

[0043]

[0044] Where, is the positive sequence similarity threshold, is the negative sequence similarity threshold, is the zero-sequence similarity threshold, x a and x b are any two data in the corresponding training set data.

[0045] Furthermore, by comparing the positive sequence similarity measure and the positive sequence similarity threshold, the positive sequence disturbance recognition result is obtained as follows:

[0046]

[0047] Where, is the positive sequence data classification result, y1 is the positive sequence disturbance recognition result;

[0048] Comparing the negative sequence similarity measure and the negative sequence similarity threshold, the negative sequence disturbance recognition result is obtained as follows:

[0049]

[0050] Where, is the negative sequence data classification result, and y2 is the negative sequence disturbance recognition result;

[0051] Comparing the zero-sequence similarity measure and the zero-sequence similarity threshold, the zero-sequence disturbance recognition result is obtained as follows:

[0052]

[0053] Where, is the zero-sequence data classification result, and y0 is the zero-sequence disturbance identification result.

[0054] Furthermore, by comparing the positive sequence disturbance identification results, the negative sequence disturbance identification results, and the zero sequence disturbance identification results, the output of the heuristic judgment layer is obtained as follows:

[0055]

[0056] Where y out is the output of the heuristic judgment layer.

[0057] Compared with the existing technology, the present invention does not require an accurate mathematical model of the power grid and a detailed analysis of the power system disturbance process. Instead, it is driven by simulation data and historical operation data, and uses artificial intelligence technology to achieve multi-classification recognition of disturbances. It can perform offline learning and online learning at the same time, providing assistance for the safety and decision-making of the power system; it takes into account the judgment of unknown disturbances, thereby overcoming the deficiency of traditional disturbance recognition methods that can only identify disturbances as a certain known disturbance, and avoiding the misclassification of unknown disturbance types. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a schematic diagram of a power system disturbance identification method based on deep learning and considering unknown disturbance types according to the present invention.

[0059] Figure 2 Schematic diagram of the structure of the composite model for power system disturbance identification based on deep neural network.

[0060] Figure 3 Schematic diagram of the heuristic judgment layer. DETAILED DESCRIPTION

[0061] The power system disturbance identification method considering unknown disturbance types based on deep learning of the present invention is further described below with reference to the accompanying drawings and specific embodiments.

[0062] See also Figure 1 The present invention discloses a method for identifying power system disturbances based on deep learning and considering unknown disturbance types, comprising the following steps:

[0063] S1. Based on the power system parameters, using the grid transient simulation data and the grid historical operation data, obtain the time series data of the positive sequence voltage, negative sequence voltage, and zero sequence voltage when various disturbances occur, and generate the time series disturbance data set of the power system.

[0064] S2. Use filtering technology to denoise the time series disturbance dataset of the power system.

[0065] S3. Extract features from the time series disturbance dataset according to the feature index, obtain disturbance feature data, and use the disturbance type corresponding to the disturbance feature data as a label to generate a disturbance identification feature sample set, which is divided into a training set and a test set.

[0066] S4. Construct a composite model for power system disturbance identification based on a deep neural network, which includes a deep neural network and a heuristic judgment layer, and train the deep neural network using the training set data and its corresponding disturbance type labels.

[0067] S5. Input the test set data into the power system disturbance identification composite model based on deep neural network. The deep neural network outputs the pre-classification result. Based on the pre-classification result and the similarity between the test set data and the training set data, the heuristic judgment layer outputs the final disturbance identification result taking into account the unknown disturbance type.

[0068] Specifically, power system parameters include the grid structure of the power system, the physical parameters of the electrical equipment contained in the power system, and the output and load demand data of the generator sets under different operating modes. The power grid transient simulation data is the data generated by the power grid through transient calculation simulation through various analysis software (such as PSASP, PSD-BPA, MATLAB, Simulink) that can perform transient simulation calculations on the power system. The power grid historical operation data is the historical operation data of the bus voltage recorded in the time series when disturbances occur during the actual operation of the power grid. The time series data of the positive sequence voltage, negative sequence voltage, and zero sequence voltage when various disturbances occur are the time series data of the positive sequence voltage amplitude, negative sequence voltage amplitude, zero sequence voltage amplitude and phase of the bus when the bus has single-phase short circuit to ground, two-phase short circuit, two-phase short circuit to ground, three-phase short circuit, induction motor self-start or other different types of disturbances.

[0069] Input the parameters of the IEEE standard 3-machine 9-bus test system, including the grid structure of the power system, the physical parameters of the electrical equipment contained in the power system, and the generator output and load demand data under different operating modes.

[0070] For various disturbance types such as single-phase short circuit to grounding, two-phase short circuit, two-phase short circuit to grounding, three-phase short circuit, and induction motor self-starting, the fault bus, fault start and clearing time, and fault resistance are randomly set within a certain range to perform electromagnetic transient simulation. The simulation result data is saved, including the time series data of the positive-sequence, negative-sequence, and zero-sequence voltage amplitudes and phases of the fault bus within a certain time window.

[0071] The method of the present invention does not require an accurate mathematical model of the power grid and a detailed analysis of the power system disturbance process. Therefore, when the actual power system historical data is insufficient, any power system transient simulation platform can be used to generate simulation data and expand the data set.

[0072] Considering the noise problem in the transient simulation data and historical operation data of the power grid, a sliding average filter is performed on each time series disturbance data set. The filtering formula is:

[0073]

[0074] Where k is the filter window length, is the data before filtering, {x i ,i=1,2,...,N-k+1} is the filtered data.

[0075] When performing feature extraction, the selected feature indicators include kurtosis factor, total time series change, C3 coefficient, FFT spectrum center, power spectrum density and standard deviation.

[0076] For each time series disturbance data set after denoising, calculate its kurtosis factor K, total time series change C abs , C3 coefficient, FFT spectrum center C FFT , power spectral density S xx (m) and standard deviation σ, a total of 6 characteristic indicators.

[0077] For a time series perturbation dataset {x i ,i=1,2,...,n}, the kurtosis factor K measures the smoothness of the disturbance waveform. The formula of the kurtosis factor K is:

[0078]

[0079] Where x i is the i-th data in the time series perturbation data set, n is the length of a time series perturbation data set, is the mean value of the time series perturbation dataset.

[0080] Total time series change C abs Measure the change amplitude of the disturbance waveform, the total change in time series C abs The formula is:

[0081]

[0082] Where x i+1 is the i+1th data in the time series perturbation dataset.

[0083] The C3 coefficient measures the nonlinearity of the disturbance waveform. The formula for the C3 coefficient is:

[0084]

[0085] Where x i+2lag is the i+2lag data in the time series perturbation data set, x i+lag is the i+lag th data in the time series disturbance data set, where lag is an integer set artificially to represent the phase lag.

[0086] FFT spectrum center C FFT Measure the frequency domain distribution of the disturbance waveform, FFT spectrum center C FFT is the spectrum center of the absolute value of the discrete Fourier transform of the disturbance data, and the formula of the n-point discrete Fourier transform is:

[0087]

[0088] Where e is the natural base.

[0089] Power spectral density S xx (m) measures the power spectrum distribution of the disturbance waveform, power spectrum density S xx The formula for (m) is:

[0090]

[0091] Where, * Indicates conjugation.

[0092] The standard deviation σ measures the statistical distribution of the disturbance waveform. The formula for the standard deviation σ is:

[0093]

[0094] Thus, each high-dimensional time series perturbation dataset {x i ,i=1,2,...,n} are all reduced to 6-dimensional perturbation feature data {K,C abs ,C3,C FFT ,Sxx Then, the disturbance type corresponding to each disturbance feature data is used as the label y, and the disturbance feature data is aggregated to generate a power system disturbance identification feature sample set, which is divided into a training set and a test set.

[0095] See also Figure 2 In order to comprehensively utilize the disturbance information contained in the positive sequence voltage, negative sequence voltage and zero sequence voltage and improve the accuracy and robustness of the classification model, the present invention constructs a power system disturbance identification composite model based on deep neural network, which includes a deep neural network and a heuristic judgment layer. Figure 2 As shown, the deep neural network includes deep neural network 1, deep neural network 2 and deep neural network 0.

[0096] The deep neural network is trained using the training set data and its corresponding perturbation type labels. The training set data includes positive sequence training set data, negative sequence training set data and zero sequence training set data. The positive sequence training set data and its corresponding perturbation type labels are used to train deep neural network 1, the negative sequence training set data and its corresponding perturbation type labels are used to train deep neural network 2, and the zero sequence training set data and its corresponding perturbation type labels are used to train deep neural network 0.

[0097] The test set data is input into the power system disturbance identification composite model based on deep neural network. The test set data includes positive sequence test set data, negative sequence test set data and zero sequence test set data. The positive sequence test set data is input into deep neural network 1 to obtain the positive sequence data classification result, the negative sequence test set data is input into deep neural network 2 to obtain the negative sequence data classification result, and the zero sequence test set data is input into deep neural network 0 to obtain the zero sequence data classification result. The positive sequence data classification result, negative sequence data classification result and zero sequence data classification result are respectively input into the heuristic judgment layer. The process of the heuristic judgment layer is as follows: Figure 3 shown.

[0098] See also Figure 3 In the heuristic judgment layer, the positive sequence test set data is compared with the positive sequence training set data corresponding to the positive sequence data classification result, the negative sequence test set data is compared with the negative sequence training set data corresponding to the negative sequence data classification result, and the zero sequence test set data is compared with the zero sequence training set data corresponding to the zero sequence data classification result. The positive sequence similarity measurement value, negative sequence similarity measurement value and zero sequence similarity measurement value are obtained respectively:

[0099]

[0100] Where sim1 is the positive sequence similarity measure, sim2 is the negative sequence similarity measure, and sim0 is the zero sequence similarity measure. is the positive sequence test set data, is the negative sequence test set data, is the zero-sequence test set data, x y1 is the positive sequence training set data, x y2 is the negative sequence training set data, x y0 is the zero-sequence training set data.

[0101] In the heuristic judgment layer, any two data in the positive sequence training set data, negative sequence training set data, and zero sequence training set data are compared respectively, and the positive sequence similarity threshold, negative sequence similarity threshold, and zero sequence similarity threshold are obtained respectively:

[0102]

[0103] Where, is the positive sequence similarity threshold, is the negative sequence similarity threshold, is the zero-sequence similarity threshold, x a and x b are any two data in the corresponding training set data.

[0104] In the heuristic judgment layer, the positive sequence similarity measure sim1 is compared with the positive sequence similarity threshold By comparing and judging the similarity between the positive sequence test set data and the positive sequence training set data, the positive sequence disturbance recognition result is obtained as follows:

[0105]

[0106] Where, is the positive sequence data classification result, and y1 is the positive sequence disturbance recognition result.

[0107] In the heuristic judgment layer, the negative sequence similarity measure sim2 is compared with the negative sequence similarity threshold By comparing and judging the similarity between the negative sequence test data and the negative sequence training set data, the negative sequence disturbance recognition result is obtained as follows:

[0108]

[0109] Where, is the negative sequence data classification result, and y2 is the negative sequence disturbance recognition result.

[0110] In the heuristic judgment layer, the zero-sequence similarity measure sim0 is compared with the zero-sequence similarity threshold By comparing and judging the similarity between the zero-sequence test data and the zero-sequence training set data, the zero-sequence disturbance recognition result is obtained as follows:

[0111]

[0112] Where, is the zero-sequence data classification result, and y0 is the zero-sequence disturbance identification result.

[0113] In the heuristic judgment layer, the positive sequence disturbance identification results, negative sequence disturbance identification results and zero sequence disturbance identification results are compared, and the output of the heuristic judgment layer is obtained as follows:

[0114]

[0115] Where y out is the output of the heuristic judgment layer.

[0116] The output y of the heuristic judgment layer out This is the final disturbance identification result output by the power system disturbance identification composite model based on deep neural network.

[0117] A comparative experiment was conducted on the method of the present invention and two other methods, wherein comparative method 1 performed feature extraction but only used an ordinary fully-linked neural network as a classifier, and comparative method 2 neither performed feature extraction nor used only a fully-linked neural network as a classifier. The comparative experimental results are shown in Table 1. From the comparison results of the model training time, it can be seen that the dimension reduction of the input data by feature extraction can greatly reduce the length of model training and a single run. From the comparison results of the classification accuracy, it can be seen that the use of robust feature extraction and a composite classification model can improve the accuracy of the model. This is because the method proposed in the present invention is more capable of capturing the core features of the data, and the composite model has higher stability. For unknown disturbances, traditional neural network classifiers can only output known categories. Therefore, comparative methods 1 and comparative methods 2 are completely unable to distinguish unknown disturbances, while the heuristic judgment layer in the method proposed in the present invention can well identify unknown disturbances.

[0118] Table 1

[0119]

[0120] In summary, the present invention does not require an accurate mathematical model of the power grid and a detailed analysis of the disturbance process of the power system. Instead, it is driven by simulation data and historical operation data, and uses artificial intelligence technology to achieve multi-classification recognition of disturbances. It can perform offline learning and online learning at the same time, providing assistance for the safety and decision-making of the power system; it takes into account the judgment of unknown disturbances, thereby overcoming the deficiency of traditional disturbance identification methods that can only identify disturbances as a certain known disturbance, and avoiding the misclassification of unknown disturbance types.

[0121] The above description is a detailed description of the preferred embodiments of the present invention, but the embodiments are not intended to limit the scope of the patent application of the present invention. Any equivalent changes or modifications made under the technical spirit disclosed by the present invention should fall within the patent scope covered by the present invention.

Claims

1. A method for identifying power system disturbances based on deep learning and considering unknown disturbance types, characterized in that: The following steps are involved: S1. Based on the power system parameters, using the power grid transient simulation data and the power grid historical operation data, obtain the time series data of the positive sequence voltage, negative sequence voltage, and zero sequence voltage when various disturbances occur, and generate the time series disturbance data set of the power system; S2, using filtering technology to denoise the time series disturbance dataset of the power system; S3. Extract features from the time series disturbance dataset based on the feature index to obtain disturbance feature data and use the disturbance type corresponding to the disturbance feature data as a label to generate a disturbance identification feature sample set, which is divided into a training set and a test set. S4. Construct a composite model for power system disturbance identification based on a deep neural network, which includes a deep neural network and a heuristic judgment layer, and train the deep neural network using the training set data and its corresponding disturbance type labels; S5. Input the test set data into the power system disturbance identification composite model based on the deep neural network. The deep neural network outputs the pre-classification result. Based on the pre-classification result and the similarity between the test set data and the training set data, the heuristic judgment layer outputs the final disturbance identification result taking into account the unknown disturbance type. The deep neural network includes deep neural network 1, deep neural network 2 and deep neural network 0. Deep neural network 1 outputs the classification result of positive sequence data, deep neural network 2 outputs the classification result of negative sequence data, and deep neural network 0 outputs the classification result of zero sequence data. The classification results of positive sequence data, negative sequence data and zero sequence data are input into the heuristic judgment layer respectively; Compare the positive sequence test set data with the positive sequence training set data corresponding to the positive sequence data classification results, the negative sequence test set data with the negative sequence data classification results, and the zero sequence test set data with the zero sequence data classification results. The positive sequence similarity measure value, negative sequence similarity measure value, and zero sequence similarity measure value are obtained respectively: Where sim1 is the positive sequence similarity measure, sim2 is the negative sequence similarity measure, and sim0 is the zero sequence similarity measure. is the positive sequence test set data, is the negative sequence test set data, is the zero-sequence test set data, x y1 is the positive sequence training set data, x y2 is the negative sequence training set data, x y0 is the zero-sequence training set data; Compare any two data in the positive sequence training set data, negative sequence training set data, and zero sequence training set data respectively, and obtain the positive sequence similarity threshold, negative sequence similarity threshold, and zero sequence similarity threshold respectively: Where, is the positive sequence similarity threshold, is the negative sequence similarity threshold, is the zero-sequence similarity threshold, x a and x b are any two data in the corresponding training set data; Comparing the positive sequence similarity measure and the positive sequence similarity threshold, the positive sequence disturbance recognition result is obtained as follows: Where, is the positive sequence data classification result, y1 is the positive sequence disturbance recognition result; Comparing the negative sequence similarity measure and the negative sequence similarity threshold, the negative sequence disturbance recognition result is obtained as follows: Where, is the negative sequence data classification result, and y2 is the negative sequence disturbance recognition result; Comparing the zero-sequence similarity measure and the zero-sequence similarity threshold, the zero-sequence disturbance recognition result is obtained as follows: Where, is the zero-sequence data classification result, y0 is the zero-sequence disturbance identification result; Comparing the positive sequence disturbance identification results, negative sequence disturbance identification results and zero sequence disturbance identification results, the output of the heuristic judgment layer is obtained as follows: Where y out is the output of the heuristic judgment layer.

2. The method for identifying power system disturbances based on deep learning and considering unknown disturbance types according to claim 1, characterized in that: Power system parameters include the grid structure of the power system, the physical parameters of the electrical equipment contained in the power system, and the output and load demand data of the generator sets under different operating modes; The power grid transient simulation data is generated by the power grid through transient calculation simulation using various analysis software that can perform power system transient simulation calculations; The historical operation data of the power grid is the historical operation data of the bus voltage in the time series recorded when disturbances occur during the actual operation of the power grid; The time series data of the positive-sequence voltage, negative-sequence voltage, and zero-sequence voltage when various disturbances occur are the time series data of the positive-sequence voltage amplitude, negative-sequence voltage amplitude, zero-sequence voltage amplitude, and phase of the busbar when a single-phase short circuit to grounding, two-phase short circuit, two-phase short circuit to grounding, three-phase short circuit, induction motor self-starting, or other different types of disturbances occur.

3. The method for identifying power system disturbances based on deep learning and considering unknown disturbance types according to claim 1, characterized in that: The filtering technology is sliding average filtering, and the filtering formula is: Where k is the filter window length, is the data before filtering, {x i ,i=1,2,...,N-k+1} is the filtered data.

4. The method for identifying power system disturbances based on deep learning and considering unknown disturbance types according to claim 1, characterized in that: The characteristic indicators include kurtosis factor, total time series variation, C3 coefficient, FFT spectrum center, power spectrum density and standard deviation.

5. The method for identifying power system disturbances based on deep learning and considering unknown disturbance types according to claim 4, characterized in that: For a time series perturbation dataset {x i ,i=1,2,...,n}, the kurtosis factor K measures the smoothness of the disturbance waveform. The formula of the kurtosis factor K is: Where x i is the i-th data in the time series perturbation data set, n is the length of a time series perturbation data set, is the average value of the time series perturbation data set; Total time series change C abs Measure the change amplitude of the disturbance waveform, the total change in time series C abs The formula is: Where x i+1 is the i+1th data in the time series perturbation dataset; The C3 coefficient measures the nonlinearity of the disturbance waveform. The formula for the C3 coefficient is: Where x i+2lag is the i+2lag data in the time series perturbation data set, x i+lag is the i+lag data in the time series disturbance data set, where lag is an integer set artificially to represent the phase lag; FFT spectrum center C FFT Measure the frequency domain distribution of the disturbance waveform, FFT spectrum center C FFT is the spectrum center of the absolute value of the discrete Fourier transform of the disturbance data, and the formula of the n-point discrete Fourier transform is: Where, e is the natural base; Power spectral density S xx (m) measures the power spectrum distribution of the disturbance waveform, power spectrum density S xx The formula for (m) is: Where, * indicates conjugation; The standard deviation σ measures the statistical distribution of the disturbance waveform. The formula for the standard deviation σ is:

Images