Fault identification method, device and equipment of power distribution network and medium

Abstract

Embodiments of the present application provide a power distribution network fault identification method, device, equipment and medium. The method first acquires to-be-processed electrical data, wherein the to-be-processed electrical data is the electrical data of a distributed power source grid-connected point in a to-be-processed power distribution network, then performs multi-scale feature extraction on the to-be-processed electrical data to obtain electrical feature data, wherein the electrical feature data includes transient feature data, time-domain physical quantity feature data, dynamic feature data, and frequency-domain distribution feature data, and finally inputs the electrical feature data into a fault identification model to obtain a target fault type of the to-be-processed power distribution network, wherein the fault identification model is obtained by training a plurality of simulation feature data and a labeled fault type corresponding to each simulation feature data. By applying the technical solution of the present application, the to-be-processed electrical data is subjected to feature extraction, the low-voltage ride-through masking effect is decoupled from the physical mechanism, and the fault identification model is used to achieve accurate fault identification.

Description

Technical Field

[0001] This application relates to the field of power system protection, and in particular to a method, apparatus, equipment and medium for fault identification in power distribution networks. Background Technology

[0002] With the increasing penetration of distributed generation (DG) power sources such as photovoltaics and wind power in distribution networks, traditional passive distribution networks with unidirectional power flow are rapidly evolving into active distribution networks with bidirectional power exchange capabilities. When a high proportion of DG is integrated, its grid-connected operation must meet grid-connection requirements such as Low Voltage Ride Through (LVRT). This means that when the distribution network voltage drops, the inverter must maintain grid-connected operation and inject reactive current to support the distribution network. At this point, unplanned islanding protection becomes a core issue in ensuring the safe operation of the distribution network.

[0003] Existing anti-islanding protection methods are mainly divided into two categories: active and passive methods. Active methods inject harmonic or frequency disturbance signals into the distribution network to actively disrupt the power balance and trigger protection actions; passive methods mainly identify faults based on the deviation of steady-state electrical quantities of the power grid such as voltage and frequency.

[0004] However, in distribution network scenarios with a high proportion of inverter-type distributed power sources connected to the grid, especially in complex operating environments where low voltage ride-through control is required, existing fault identification methods are prone to misjudging external faults as islanded faults, resulting in poor fault identification accuracy. Summary of the Invention

[0005] This application provides a method, apparatus, equipment, and medium for fault identification in power distribution networks, in order to improve the accuracy of fault identification.

[0006] In a first aspect, embodiments of this application provide a method for fault identification in a power distribution network, including:

[0007] Acquire electrical data to be processed, wherein the electrical data to be processed is the electrical data of the distributed power generation grid connection point in the distribution network to be processed;

[0008] Multi-scale feature extraction is performed on the electrical data to be processed to obtain electrical feature data. The multi-scale feature extraction includes extracting transient energy, time-domain physical quantities, dynamic electrical quantities, and frequency domain distribution features for a first preset duration. The second preset duration is longer than the first preset duration. The electrical feature data includes transient feature data, time-domain physical quantity feature data, dynamic feature data, and frequency domain distribution feature data.

[0009] The electrical feature data is input into the fault identification model to obtain the target fault type of the distribution network to be processed. The fault identification model is trained based on multiple simulated feature data and the labeled fault type corresponding to each simulated feature data.

[0010] In one or more embodiments, the step of performing multi-scale feature extraction on the electrical data to be processed to obtain electrical feature data includes:

[0011] The Teager-Kaiser energy operator and Hinkley cumulative test are used to extract the features of the transient energy of the electrical data to be processed in the first preset time period to obtain the transient feature data, which includes: voltage transient energy, frequency transient energy, active power transient energy, and reactive power transient energy.

[0012] The first-order difference is used to extract the features of the electrical data to be processed in the time-domain physical quantities to obtain the time-domain physical quantity feature data, which includes: voltage change rate, frequency change rate, active power change rate, and reactive power change rate.

[0013] The features of the electrical data to be processed in the dynamic electrical quantities of the second preset time period are extracted to obtain the dynamic feature data, which includes: apparent power, power factor, equivalent dynamic resistance, and equivalent dynamic reactance;

[0014] Wavelet packet transform is used to extract the features of the electrical data to be processed in the frequency domain distribution to obtain the frequency domain distribution feature data, which includes time-frequency energy entropy.

[0015] In one or more embodiments, before inputting the electrical characteristic data into the fault identification model to obtain the target fault type of the distribution network to be processed, the method further includes:

[0016] Based on the simulation model of the distribution network to be processed, the multiple simulation feature data under multiple operating conditions corresponding to different labeled fault types are determined;

[0017] The backpropagation neural network model is trained based on the multiple simulated feature data and the labeled fault type corresponding to each simulated feature data to obtain the fault identification model.

[0018] In one or more embodiments, training the backpropagation neural network model based on the plurality of simulated feature data and the labeled fault type corresponding to each simulated feature data to obtain the fault identification model includes:

[0019] The slime mold algorithm is used to optimize the hyperparameters of the backpropagation neural network model and determine the target hyperparameters, which include: the number of hidden layer nodes and the L2 regularization coefficient.

[0020] Based on the target hyperparameter values, the backpropagation neural network model is hyperparameter adjusted to obtain the adjusted backpropagation neural network model.

[0021] The adjusted backpropagation neural network model is trained based on the multiple simulated feature data and the labeled fault type corresponding to each simulated feature data to obtain the fault identification model.

[0022] In one or more embodiments, the adjusted backpropagation neural network model includes: an input layer, a hidden layer, and an output layer;

[0023] Accordingly, the step of training the adjusted backpropagation neural network model based on the plurality of simulated feature data and the labeled fault type corresponding to each simulated feature data to obtain the fault identification model includes:

[0024] S1, for each simulated feature data, the input layer receives the simulated feature data and transmits it to the hidden layer;

[0025] S2, the hidden layer performs nonlinear feature extraction on the simulated feature data to obtain nonlinear feature data;

[0026] S3, input the nonlinear characteristic data to the output layer to obtain the simulated fault type;

[0027] S4. Based on all simulated fault types and the labeled fault type corresponding to each simulated feature data, update the parameters of the adjusted backpropagation neural network model. Repeat steps S1-S4 until the preset training stopping condition is met to obtain the fault identification model.

[0028] In one or more embodiments, before inputting the electrical characteristic data into the fault identification model to obtain the target fault type of the distribution network to be processed, the method further includes:

[0029] The ReliefF algorithm is used to perform feature selection on the electrical feature data to obtain multiple candidate feature data.

[0030] Determine multiple Pearson correlation coefficient values ​​between every two candidate feature data in the plurality of candidate feature data;

[0031] The electrical characteristic data is updated based on the Pearson correlation coefficient value that is greater than a preset threshold.

[0032] In one or more embodiments, after acquiring the electrical data to be processed, the method further includes:

[0033] Perform signal mutation analysis on the electrical data to be processed to determine the magnitude of the mutation.

[0034] Based on the magnitude of the mutation, the preset sliding time window is adjusted to obtain the adjusted sliding time window;

[0035] The electrical data to be processed is truncated according to the adjusted sliding time window to obtain the truncated electrical data to be processed.

[0036] Secondly, embodiments of this application provide a fault identification device for a power distribution network, comprising:

[0037] The acquisition module is used to acquire electrical data to be processed, which is the electrical data of the distributed power source grid connection point in the distribution network to be processed;

[0038] An extraction module is used to perform multi-scale feature extraction on the electrical data to be processed to obtain electrical feature data. The multi-scale feature extraction includes extracting transient energy, time-domain physical quantities, dynamic electrical quantities, and frequency domain distribution features for a first preset duration. The second preset duration is longer than the first preset duration. The electrical feature data includes: transient feature data, time-domain physical quantity feature data, dynamic feature data, and frequency domain distribution feature data.

[0039] The processing module is used to input the electrical feature data into the fault identification model to obtain the target fault type of the distribution network to be processed. The fault identification model is trained based on multiple simulated feature data and the labeled fault type corresponding to each simulated feature data.

[0040] In one or more embodiments, the extraction module is specifically used for:

[0041] The Teager-Kaiser energy operator and Hinkley cumulative test are used to extract the features of the transient energy of the electrical data to be processed in the first preset time period to obtain the transient feature data, which includes: voltage transient energy, frequency transient energy, active power transient energy, and reactive power transient energy.

[0042] The first-order difference is used to extract the features of the electrical data to be processed in the time-domain physical quantities to obtain the time-domain physical quantity feature data, which includes: voltage change rate, frequency change rate, active power change rate, and reactive power change rate.

[0043] The features of the electrical data to be processed in the dynamic electrical quantities of the second preset time period are extracted to obtain the dynamic feature data, which includes: apparent power, power factor, equivalent dynamic resistance, and equivalent dynamic reactance;

[0044] Wavelet packet transform is used to extract the features of the electrical data to be processed in the frequency domain distribution to obtain the frequency domain distribution feature data, which includes time-frequency energy entropy.

[0045] In one or more embodiments, before inputting the electrical characteristic data into the fault identification model to obtain the target fault type of the distribution network to be processed, the processing module is further configured to:

[0046] Based on the simulation model of the distribution network to be processed, the multiple simulation feature data under multiple operating conditions corresponding to different labeled fault types are determined;

[0047] The backpropagation neural network model is trained based on the multiple simulated feature data and the labeled fault type corresponding to each simulated feature data to obtain the fault identification model.

[0048] In one or more embodiments, the processing module trains a backpropagation neural network model based on the plurality of simulated feature data and the labeled fault type corresponding to each simulated feature data to obtain the fault identification model, specifically for:

[0049] The slime mold algorithm is used to optimize the hyperparameters of the backpropagation neural network model and determine the target hyperparameters, which include: the number of hidden layer nodes and the L2 regularization coefficient.

[0050] Based on the target hyperparameter values, the backpropagation neural network model is hyperparameter adjusted to obtain the adjusted backpropagation neural network model.

[0051] The adjusted backpropagation neural network model is trained based on the multiple simulated feature data and the labeled fault type corresponding to each simulated feature data to obtain the fault identification model.

[0052] In one or more embodiments, the adjusted backpropagation neural network model includes: an input layer, a hidden layer, and an output layer;

[0053] Accordingly, the processing module trains the adjusted backpropagation neural network model based on the multiple simulated feature data and the labeled fault type corresponding to each simulated feature data to obtain the fault identification model, specifically for:

[0054] S1, for each simulated feature data, the input layer receives the simulated feature data and transmits it to the hidden layer;

[0055] S2, the hidden layer performs nonlinear feature extraction on the simulated feature data to obtain nonlinear feature data;

[0056] S3, input the nonlinear characteristic data to the output layer to obtain the simulated fault type;

[0057] S4. Based on all simulated fault types and the labeled fault type corresponding to each simulated feature data, update the parameters of the adjusted backpropagation neural network model. Repeat steps S1-S4 until the preset training stopping condition is met to obtain the fault identification model.

[0058] In one or more embodiments, before inputting the electrical characteristic data into the fault identification model to obtain the target fault type of the distribution network to be processed, the processing module is further configured to:

[0059] The ReliefF algorithm is used to perform feature selection on the electrical feature data to obtain multiple candidate feature data.

[0060] Determine multiple Pearson correlation coefficient values ​​between every two candidate feature data in the plurality of candidate feature data;

[0061] The electrical characteristic data is updated based on the Pearson correlation coefficient value that is greater than a preset threshold.

[0062] In one or more embodiments, after acquiring the electrical data to be processed, the acquisition module is further configured to:

[0063] Perform signal mutation analysis on the electrical data to be processed to determine the magnitude of the mutation.

[0064] Based on the magnitude of the mutation, the preset sliding time window is adjusted to obtain the adjusted sliding time window;

[0065] The electrical data to be processed is truncated according to the adjusted sliding time window to obtain the truncated electrical data to be processed.

[0066] Thirdly, embodiments of this application provide an electronic device, including: a memory and a processor;

[0067] The memory stores computer-executed instructions;

[0068] The processor executes computer execution instructions stored in the memory, such that the processor, when executed, is used to implement the method described in the first aspect and any of the embodiments described above.

[0069] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the methods described in the first aspect and any of the embodiments above.

[0070] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, is used to implement a fault identification method for a power distribution network as described in the first aspect and various possible implementations of the first aspect.

[0071] This application provides a method, apparatus, device, and medium for fault identification in a distribution network. The method first acquires electrical data to be processed, which is the electrical data of distributed power source grid connection points in the distribution network. Next, it extracts features from the electrical data to obtain electrical feature data. The multi-scale feature extraction includes extracting transient energy, time-domain physical quantities, dynamic electrical quantities, and frequency-domain distribution features for a first preset duration, where the second preset duration is longer than the first preset duration. The electrical feature data includes transient feature data, time-domain physical quantity feature data, dynamic feature data, and frequency-domain distribution feature data. Finally, the electrical feature data is input into a fault identification model to obtain the target fault type of the distribution network. The fault identification model is trained based on multiple simulated feature data and the labeled fault types corresponding to each simulated feature data. In the above method, electrical data of distributed power generation grid connection points in the distribution network to be processed is acquired, and multi-scale feature extraction is performed to obtain electrical feature data including transient feature data, time-domain physical quantity feature data, dynamic feature data, and frequency-domain distribution feature data. Then, the electrical feature data is input into the fault identification model trained based on multiple simulated feature data and corresponding labeled fault types to obtain the target fault type of the distribution network to be processed. By extracting features from different time scales or different representation dimensions through multi-scale feature extraction, the low voltage ride-through masking effect is decoupled from the physical mechanism, improving the differentiation accuracy between islanded faults and external faults in the complex operating environment of low voltage ride-through control, thereby effectively improving the accuracy of distribution network fault identification. Attached Figure Description

[0072] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0073] Figure 1 A flowchart illustrating the fault identification method for a power distribution network provided in this application embodiment. Figure 1 ;

[0074] Figure 2 A flowchart illustrating the fault identification method for a power distribution network provided in this application embodiment. Figure 2 ;

[0075] Figure 3 A flowchart illustrating the fault identification method for a power distribution network provided in this application embodiment. Figure 3 ;

[0076] Figure 4 A schematic diagram of the structure of the adjusted backpropagation neural network model provided in the embodiments of this application;

[0077] Figure 5 A flowchart illustrating the fault identification method for a power distribution network provided in this application embodiment. Figure 4 ;

[0078] Figure 6 A schematic diagram of the structure of the fault identification device for the power distribution network provided in the embodiments of this application;

[0079] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0080] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0081] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0082] Before introducing the embodiments of this application, the application background of the embodiments of this application will be explained first:

[0083] With the increasing penetration of distributed generation (DG) power sources such as photovoltaics and wind power in distribution networks, traditional passive distribution networks with unidirectional power flow are rapidly evolving into active distribution networks with bidirectional power exchange capabilities. When a high proportion of DG is integrated, its grid-connected operation must meet grid-connection requirements such as Low Voltage Ride Through (LVRT). This means that when the distribution network voltage drops, the inverter must maintain grid-connected operation and inject reactive current to support the distribution network. At this point, unplanned islanding protection becomes a core issue in ensuring the safe operation of the distribution network.

[0084] Existing anti-islanding protection methods are mainly divided into two categories: active and passive methods. Active methods inject harmonic or frequency disturbance signals into the distribution network to actively disrupt the power balance and trigger protection actions, thus identifying the fault based on the difference between the local network and the islanded state, where the network is more sensitive to disturbances. Passive methods mainly identify faults based on the deviation of steady-state electrical quantities of the power grid, such as voltage and frequency.

[0085] However, active methods suffer from drawbacks such as power quality degradation due to injected disturbances and easy cancellation of disturbances in multi-machine parallel operation. Although passive methods do not have harmonic pollution, they have serious theoretical defects: 1) Detection blind zone under power balance conditions: When the output power of the DG is highly matched with the local load, the steady-state deviation of electrical quantities (voltage, frequency) before and after the islanding outage is extremely small, and the traditional criteria based on over / under voltage protection (OVP / UVP) and over / under frequency protection (OFP / UFP) fail; 2) Coupling maloperation during low voltage ride-through: When an external short-circuit fault causes a voltage dip, the grid-connected inductor forces the inverter to start low voltage ride-through control, injecting reactive current into the distribution network. This process causes the short-circuit transient trajectory and the power matching islanding trajectory to highly overlap in the low-dimensional observation space.

[0086] In summary, in distribution network scenarios with a high proportion of inverter-type distributed power sources connected to the grid, especially in complex operating environments where low-voltage ride-through control is required, existing fault identification methods are prone to misidentifying external faults as islanded faults, resulting in poor fault identification accuracy.

[0087] The fault identification method for distribution networks provided in this application aims to solve the aforementioned technical problems of existing technologies. The technical concept of this application is as follows: Existing methods are prone to misclassifying external faults as islanding faults in complex operating environments requiring low-voltage ride-through control, resulting in poor fault identification accuracy. Fault identification in distribution networks mainly relies on electrical data from distributed generation points in the distribution network to be processed, analyzing the differences in physical characteristics between external faults and islanding faults during low-voltage ride-through control to determine the fault type of the distribution network. If the low-voltage ride-through masking effect can be decoupled from the physical mechanism, external faults and islanding faults can be effectively distinguished, improving the accuracy of fault identification. Therefore, this application embodiment considers acquiring electrical data of distributed power generation grid connection points in the distribution network to be processed, so that fault identification is based on continuously collected electrical information during actual operation; then, multi-scale feature extraction is performed on the electrical data to be processed to obtain electrical feature data for characterizing the fault state. The electrical feature data includes transient feature data, time-domain physical quantity feature data, dynamic feature data, and frequency-domain distribution feature data, which can comprehensively reflect the fault process of the distribution network from different time scales or different characterization dimensions, decouple the low voltage ride-through masking effect from the physical mechanism, and improve the accuracy of distinguishing between islanded faults and external faults; then, the electrical feature data is input into the fault identification model trained based on multiple simulated feature data and corresponding labeled fault types to accurately identify the target fault type of the distribution network to be processed.

[0088] The execution subject of this application embodiment is an electronic device, which can be a terminal device, such as a laptop, desktop computer, or tablet computer, or a server. In practical applications, whether the electronic device is a terminal device or a server can be determined according to the actual situation, and no specific limitation is imposed on it.

[0089] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0090] Figure 1 A flowchart illustrating the fault identification method for a power distribution network provided in this application embodiment. Figure 1 .like Figure 1 As shown, the fault identification method for this power distribution network includes the following steps:

[0091] S101. Obtain the electrical data to be processed;

[0092] Among them, the electrical data to be processed is the electrical data of the distributed power generation grid connection point in the distribution network.

[0093] In this step, in order to identify faults in the distribution network, the electrical data of the distributed power generation grid connection points in the distribution network to be processed can be obtained first as the electrical data to be processed, so as to provide a reliable data basis for subsequent fault identification of the distribution network.

[0094] For example, the grid connection point of a distributed power source is the electrical connection node where the distributed power source is connected to the distribution network, and it is also a key observation location for fault identification.

[0095] In one possible implementation, the electrical data to be processed mainly includes discrete sampling sequences corresponding to the three-phase voltage data and three-phase current data at the grid connection point of the distributed power source. Using the three-phase voltage data and three-phase current data as the basic sampling quantities, electrical quantities such as voltage RMS value, frequency, phase angle, active power, reactive power, and apparent power can be obtained through mathematical calculations, thereby ensuring that the electrical data to be processed has a unified time reference.

[0096] Assuming the sampling frequency is Discrete sampling period Then the current sampling time The sampled electrical data to be processed includes: a vector of instantaneous three-phase voltage values. and the instantaneous value vector of three-phase current ,in These represent phase A, phase B, and phase C, respectively.

[0097] To ensure the reliability of subsequent fault identification, basic preprocessing operations can be performed on the electrical data to be processed, which may include: filling in missing values, removing outliers, aligning timestamps, and unifying units.

[0098] In one possible implementation, after step S101 described above, the fault identification method for the power distribution network may further include the following steps:

[0099] Step 1: Perform signal mutation analysis on the electrical data to be processed to determine the amplitude of the mutation.

[0100] For example, signal mutation analysis is used to identify sudden increases, decreases, or inflection point changes in electrical data to be processed within a short period of time, in order to characterize the transient response of distributed power grid connection points during fault disturbances; the mutation magnitude is used to quantify the degree of mutation, and the larger the value, the more drastic the change in electrical quantities such as voltage, current, power, or frequency.

[0101] In one possible implementation, signal mutation analysis can be performed based on the difference values ​​between adjacent sampling points, where the difference values ​​can reflect the instantaneous changes between consecutive sampling points, making it easy to quickly locate the moment of mutation.

[0102] In the At each analysis time point, the magnitude of the mutation It can be defined as: the current sampling time The instantaneous value vector of the three-phase voltage obtained by sampling and the instantaneous value vector of three-phase current The maximum term of the absolute value of the first-order forward difference is calculated using the following formula:

[0103]

[0104] in, , These represent phase A, phase B, and phase C, respectively.

[0105] Step 2: Adjust the preset sliding time window according to the magnitude of the mutation to obtain the adjusted sliding time window;

[0106] For example, a preset sliding time window is used to limit the time range of data interception. The initial window length is usually preset based on the sampling frequency or sampling period. The length of the preset sliding time window can be adaptively adjusted according to the amplitude of the mutation to adapt to transient processes of different intensities.

[0107] In one possible implementation, in the first... At each analysis time point, the window length, adjusted based on the magnitude of the mutation, is the length of the preset sliding time window. (Unit is the number of sampling points), the specific calculation formula is as follows:

[0108]

[0109] in, This indicates the floor function. and These represent the upper and lower limits of the initial window length for the preset sliding time window, respectively. This represents the sensitivity adjustment attenuation coefficient.

[0110] Based on the above calculation formula, through the negative exponential function Achieve continuous and smooth distinction between high-frequency transient and steady-state conditions without setting a hard threshold: When a high-frequency transient change occurs in the system, the magnitude of the change... The window length increases sharply, the negative exponent term approaches 0, and the adjusted window length... The lower limit of the adaptive shrinking to the initial window length Capture sudden jump characteristics; under steady-state conditions, the perturbation noise floor reduces the amplitude of sudden changes. Approaching 0, the negative exponent term approaches 1, and the adjusted window length... Adaptive setting of upper limit for initial window length Extend it to enhance noise resistance.

[0111] Step 3: Based on the adjusted sliding time window, perform data truncation processing on the electrical data to be processed to obtain the truncated electrical data to be processed.

[0112] For example, data capture processing involves extracting data segments corresponding to the time period from continuously collected electrical data to be processed according to the adjusted sliding time window.

[0113] In one possible implementation, the extracted electrical data to be processed retains local transient data that is highly correlated with the mutation event, which can provide a more centralized data foundation for subsequent processing.

[0114] In the At each analysis point, a local transient data matrix is ​​constructed based on the adjusted sliding time window. Specifically, it is expressed as follows:

[0115]

[0116] Among them, matrix The number of rows is set to 6, and each row corresponds to a physical quantity (including the three-phase voltage and three-phase current of A / B / C); matrix The number of columns is set to the adjusted sliding time window. The number of sampling points within each column corresponds to the instantaneous values ​​of three-phase voltage and current at the same sampling time.

[0117] In this embodiment, the preset sliding time window is adaptively adjusted based on the magnitude of the mutation, so as to avoid the fixed time window from intercepting redundant steady-state data under strong disturbances or missing key transient information under weak disturbances, thereby improving the effectiveness and representativeness of the electrical data to be processed.

[0118] S102. Perform multi-scale feature extraction on the electrical data to be processed to obtain electrical feature data;

[0119] The multi-scale feature extraction includes extracting transient energy, time-domain physical quantities, dynamic electrical quantities, and frequency-domain distribution features for a first preset duration, and the second preset duration is longer than the first preset duration.

[0120] Electrical characteristic data includes: transient characteristic data, time-domain physical quantity characteristic data, dynamic characteristic data, and frequency domain distribution characteristic data.

[0121] In this step, multi-scale feature extraction is performed on the electrical data to be processed, including the extraction or transformation of features of transient energy, time-domain physical quantities, dynamic electrical quantities, and frequency-domain distribution for a first preset duration, to obtain electrical feature data including transient feature data, time-domain physical quantity feature data, dynamic feature data, and frequency-domain distribution feature data.

[0122] For example, transient feature data is obtained by extracting features of transient energy for a first preset duration, wherein the first preset duration is a short or instantaneous duration, and the feature extraction of transient energy for the first preset duration is short-time transient feature extraction. Transient feature data is mainly used to capture abrupt changes, impacts and instantaneous energy changes in the electrical data to be processed.

[0123] Time-domain physical quantity characteristic data are used to characterize the variation of three-phase voltage, three-phase current, frequency, and power physical quantities on the time axis;

[0124] The dynamic feature data is obtained by extracting features of dynamic electrical quantities for a second preset duration. The second preset duration is a longer duration than the first preset duration. The feature extraction of dynamic electrical quantities for the second preset duration is called long-term dynamic feature extraction. The dynamic feature data is mainly used to reflect the evolution trend of impedance characteristics over a longer time scale.

[0125] Frequency domain distribution characteristic data is used to characterize the differences in energy distribution of electrical data to be processed across multiple frequency bands.

[0126] In one possible implementation, different categories of feature data correspond to the external manifestations of different fault types at the physical mechanism level. When combined, they can reflect the differences between islanded faults, external faults, and other disturbance states from multiple time scales and multiple representation dimensions.

[0127] Transient characteristic data emphasizes the abrupt changes in the initial stage of a fault; time-domain physical quantity characteristic data emphasizes the evolution of physical quantities during the fault process; dynamic characteristic data emphasizes the operating trend of the distribution network under continuous disturbance; and frequency domain distribution characteristic data emphasizes the redistribution of energy in each frequency band. These four types of characteristic data together constitute complete electrical characteristic data.

[0128] Furthermore, post-processing can be performed on feature data of different categories, including feature normalization, outlier smoothing, and feature concatenation. For example, maximum-minimum normalization or mean-variance standardization can be used to eliminate dimensional differences and normalize features of different categories. For feature data that is significantly affected by noise, stability can be improved by using moving average or median filtering methods. For multiple sets of feature data generated by different sliding time windows, they can be concatenated in chronological order to form feature vectors, or fixed-dimensional inputs can be formed through statistical aggregation.

[0129] S103. Input the electrical characteristic data into the fault identification model to obtain the target fault type of the distribution network to be processed;

[0130] The fault identification model is trained based on multiple simulated feature data and the labeled fault type corresponding to each simulated feature data.

[0131] In this step, a fault identification model is trained based on multiple simulated feature data and the labeled fault type corresponding to each simulated feature data. The electrical feature data is then input into the fault identification model to identify the fault and obtain the target fault type of the distribution network to be processed.

[0132] For example, the target fault type is used to represent the actual fault state of the distribution network to be processed, which may include real islanded faults and non-islanded faults, and non-islanded faults include: external faults and other non-islanded disturbances.

[0133] In one possible implementation, multiple simulated feature data can originate from an electromagnetic transient simulation platform or a digital simulation device. Simulations are performed for the operating conditions corresponding to each fault type. By changing conditions such as the distribution network operating mode, line impedance, distributed generation penetration rate, inverter control parameters, local load matching degree, fault location, and fault duration, simulated electrical data covering different operating conditions is generated. Then, feature extraction is performed on the simulated electrical data to obtain multiple simulated feature data, and each simulated feature data is labeled with its corresponding fault type.

[0134] The fault identification model can adopt a backpropagation neural network model. A multi-layer neural network is used to receive multiple normalized simulated feature data. The underlying parameters of the backpropagation neural network model are continuously adjusted through error backpropagation, so that the loss function between the model output fault type and the labeled fault type corresponding to each simulated feature data gradually decreases until the preset convergence condition is met, the training is completed, and the fault identification model is obtained.

[0135] Fault identification models can also employ support vector machine models, ensemble tree models, or other model structures capable of supervised classification.

[0136] In one possible implementation, during the online identification phase of the fault identification model, electrical characteristic data is input into the fault identification model, and the confidence level of the islanding fault type is output. Combined with the preset sliding time window in the above embodiment, a continuous confidence sequence of the continuous sliding time window is determined. When the confidence level of a preset number of sliding time windows is detected to be greater than a preset threshold, the target fault type is determined to be a real islanding fault, and a trip command is generated; otherwise, the target fault type is determined to be an external fault or other non-islanding disturbance, and low voltage ride-through support is maintained.

[0137] For example, when the target fault type is determined to be a real islanding fault, the main control unit of the microprocessor protection device in the distribution network to be processed can drive the high-power output relay to close and output the anti-islanding trip hard contact command; when the target fault type is determined to be an external fault or other non-islanding disturbance, the microprocessor protection device refuses to issue a trip pulse, thereby being compatible with and deeply supporting the safe operation requirements of the low voltage ride-through guideline.

[0138] The fault identification method for distribution networks provided in this application first acquires electrical data to be processed, wherein the electrical data to be processed is the electrical data of the distributed power generation grid connection point in the distribution network to be processed. Then, feature extraction is performed on the electrical data to be processed to obtain electrical feature data, wherein the electrical feature data includes: transient feature data, time-domain physical quantity feature data, dynamic feature data, and frequency-domain distribution feature data. Finally, the electrical feature data is input into the fault identification model to obtain the target fault type of the distribution network to be processed. The fault identification model is trained based on multiple simulated feature data and the labeled fault type corresponding to each simulated feature data. In this embodiment, electrical data of distributed power generation grid connection points in the distribution network to be processed is acquired, and the electrical data to be processed is jointly characterized from multiple scales including short-term transient, time-domain physical quantities, long-term dynamics, and frequency domain distribution. This decouples the low-voltage ride-through masking effect from a physical mechanism perspective, enhancing the distinguishability of fault differences under complex transient disturbances. The electrical feature data is then input into a fault identification model for fault identification, accurately outputting the target fault type of the distribution network to be processed, thereby improving the accuracy of fault identification in the complex operating environment of low-voltage ride-through control.

[0139] Based on the above embodiments, Figure 2 A flowchart illustrating the fault identification method for a power distribution network provided in this application embodiment. Figure 2 .like Figure 2 As shown, one possible implementation of step S102 above includes the following steps:

[0140] S201. The Teager-Kaiser energy operator and Hinkley cumulative test are used to extract the features of the transient energy of the electrical data to be processed in the first preset time period to obtain transient feature data.

[0141] The transient characteristic data includes: voltage transient energy, frequency transient energy, active power transient energy, and reactive power transient energy.

[0142] In this step, in order to capture extremely weak high-frequency phase jumps, the Teager-Kaiser energy operator is used, combined with the Hinkley cumulative test to extract features from the transient energy of the electrical data to be processed in the first preset duration, to obtain transient feature data including voltage transient energy, frequency transient energy, active power transient energy, and reactive power transient energy.

[0143] For example, the Teager-Kaiser energy operator is used to characterize transient energy changes in electrical data on short time scales, and it is sensitive to local abrupt changes in voltage, current, frequency, and power.

[0144] Hinkley cumulative test is used to assist in the identification of abrupt changes. By tracking cumulative deviations, it enhances the ability to identify the moment of failure and the starting point of disturbance.

[0145] In one possible implementation, the features of the transient energy of the electrical data to be processed are extracted in the first preset duration, which is called short-time transient feature extraction of the electrical data to be processed. If the electrical data to be processed includes: voltage RMS value, frequency, active power, and reactive power, when performing short-time transient feature extraction of the electrical data to be processed, the Teager-Kaiser energy operator can be applied to the voltage RMS value, frequency, active power, and reactive power respectively to obtain the corresponding transient energy sequence. Then, the Hinkley cumulative test is combined to locate the energy change interval, and then the voltage transient energy, frequency transient energy, active power transient energy, and reactive power transient energy are extracted.

[0146] For example, for the extremely weak high-frequency phase jumps generated by physical behaviors such as the moment the circuit breaker disconnects, the extinction of the arc of the mechanical contacts, and the re-tracking of the inverter's phase-locked loop (PLL), the Teager-Kaiser energy operator is applied to the effective voltage value, frequency, active power, and reactive power respectively.

[0147] With a single variable Taking the general formula derivation as an example, its discrete Teager-Kaiser instantaneous energy The calculation formula is as follows:

[0148]

[0149] Among them, univariate That is, the effective value of voltage ,frequency Active power and reactive power one of the, This represents the local time cursor index within the current sliding time window.

[0150] Then, univariate Discrete Teager-Kaiser instantaneous energy The nonlinear mapping is applied to the Hinkley cumulative test model, which tracks and calculates the local statistical mean of the preceding background noise within the current sliding time window in real time. with standard deviation Dynamically calculate the Hinkley cumulative sum sequence With test statistic :

[0151]

[0152]

[0153] in, This indicates the preset sensitivity adjustment coefficient. This indicates an adaptively updated drift parameter. This adaptive drift mechanism can forcibly filter out "pseudo-energy jumps" caused by grid-connected inverter harmonics or low-voltage ride-through forced current injection.

[0154] Within the current sliding time window, for the effective voltage value ,frequency Active power and reactive power Calculate the corresponding test statistics respectively. The effective voltage value is obtained by taking the maximum value and standard deviation. The maximum value of the corresponding test statistic and standard deviation That is, voltage transient energy; frequency The maximum value of the corresponding test statistic and standard deviation That is, frequency transient energy; active power. The maximum value of the corresponding test statistic and standard deviation Active power is the transient energy; reactive power is the transient energy. The maximum value of the corresponding test statistic and standard deviation That is, reactive power transient energy.

[0155] S202. First-order difference is used to extract the features of the electrical data to be processed in the time domain physical quantity to obtain the time domain physical quantity feature data.

[0156] Among them, the time-domain physical quantity characteristic data include: voltage change rate, frequency change rate, active power change rate, and reactive power change rate.

[0157] In this step, in order to perceive the impact caused by the initial occurrence of islanding, the first-order difference is used to extract the features of the electrical data to be processed in the time domain physical quantities, and the time domain physical quantity feature data including voltage change rate, frequency change rate, active power change rate, and reactive power change rate are obtained.

[0158] In one possible implementation, the electrical data to be processed includes: effective voltage value, frequency, active power, and reactive power. When extracting the features of the electrical data to be processed in the time domain physical quantities, the first-order difference is performed on the adjacent sampling points within the current sliding time window and the maximum absolute value is taken to obtain the voltage change rate, frequency change rate, active power change rate, and reactive power change rate.

[0159] The formulas for calculating the rate of change of frequency (ROCOF) and the rate of change of voltage (ROCOV) within the current sliding time window are as follows:

[0160]

[0161]

[0162] in, This represents the step difference interval parameter (a known set constant) used to smooth unavoidable high-frequency white noise in measurements. Representing the discrete sampling period, at the instant an island occurs, ROCOF will surge from a near-zero steady-state noise floor to a step signal with a steep, monotonic slope.

[0163] Rate of change of active power within the current sliding time window and reactive power change rate This is used to characterize the severity of the power supply and demand imbalance during a network outage in an isolated area. The specific calculation formula is shown below:

[0164]

[0165]

[0166] Furthermore, to amplify hidden disturbances within the power balance dead zone and break the low-value stability suppression of voltage / reactive power and frequency / active power under grid-connected conditions, a power cross-coupling transient sensitivity is constructed, including: voltage-reactive power transient sensitivity. Frequency-active transient sensitivity The specific calculation formula is as follows:

[0167]

[0168]

[0169] in, The operator representing the mathematical expectation within the current sliding time window. , , ,as well as These are the effective voltage values ​​within the current sliding time window. ,frequency Active power and reactive power The forward difference sequence, This represents a very small constant (with a known set value) to prevent the denominator from being zero.

[0170] S203. Extract the features of the dynamic electrical quantities of the electrical data to be processed in the second preset time period to obtain dynamic feature data;

[0171] The dynamic characteristic data includes: apparent power, power factor, equivalent dynamic resistance, and equivalent dynamic reactance.

[0172] In this step, in order to delineate the macroscopic physical boundaries between islanded faults and external faults, features of the electrical data to be processed in the dynamic electrical quantities of the second preset time period are extracted to obtain dynamic feature data including apparent power, power factor, equivalent dynamic resistance, and equivalent dynamic reactance.

[0173] In one possible implementation, under the low-voltage ride-through condition of LVRT triggered by an external short circuit, the low-impedance loop of the main distribution network remains connected, and the reactive power regulation of the inverter only causes a slight drop in the equivalent impedance of the distributed power source grid connection point and the voltage change rate converges rapidly. Under the islanding condition, the physical branch of the main distribution network is disconnected, the equivalent impedance of the distributed power source grid connection point increases sharply by a step and a quadrant shift occurs in the RX complex plane, and the voltage change rate fluctuates significantly over a long period of time. Therefore, the feature extraction of dynamic electrical quantities for the second preset time period of the electrical data to be processed can be based on the essential physical difference between the equivalent impedance and the voltage change rate, i.e., long-term dynamic feature extraction.

[0174] The electrical data to be processed may include: the fundamental voltage RMS value, apparent power, and power factor, and then the equivalent dynamic resistance can be calculated based on the fundamental voltage RMS value and apparent power. and equivalent dynamic reactance The specific calculation formula is as follows:

[0175]

[0176]

[0177] in, Indicates the effective value of the fundamental voltage. This represents the mean of apparent power. It represents the real-time dynamic power factor angle during transient distortion.

[0178] S204. Wavelet packet transform is used to extract the features of the electrical data to be processed in the frequency domain to obtain frequency domain distribution feature data.

[0179] Among them, the frequency domain distribution characteristic data includes time-frequency energy entropy.

[0180] In this step, in order to characterize the high-frequency energy scattering phenomenon caused by island topological fracture from the frequency domain mechanism, wavelet packet transform is used to extract the frequency domain distribution features of the electrical data to be processed, and obtain frequency domain distribution feature data including time-frequency energy entropy.

[0181] For example, wavelet packet transform is used to decompose electrical data to be processed into multiple frequency bands and calculate time-frequency energy entropy based on the energy distribution of each frequency band to characterize the discreteness and complexity of signal energy in the time-frequency domain.

[0182] In one possible implementation, given the strong asymmetry of transient jumps caused by network outages in isolated areas and the specific physical requirement to extract weak high-frequency components from the fundamental energy forced by low-voltage crossing, a fifth-order Daubechies5 (db5) wavelet with high-order vanishing moments and asymmetric compact support characteristics is selected as the mother wavelet to perform multi-scale lossless orthogonal decomposition of the non-stationary voltage transient waveforms in the electrical data to be processed.

[0183] For example, the aforementioned local transient data matrix The single-phase voltage discrete sampling sequence (such as the A-phase voltage sequence, i.e., the extraction matrix) in the data. The first row element corresponding to the phase A voltage is used as the input data for wavelet packet decomposition, and is denoted as... ,in, This represents the discrete cursor index within the adjusted sliding time window.

[0184] Using db5 mother wavelet For layer orthogonal decomposition, given the low-pass filter coefficients h[k] and the high-pass filter coefficients g[k] (both are known constant sequences determined by the db5 wavelet basis), the recursive equation for the decomposition of discrete coefficients is as follows:

[0185]

[0186]

[0187] in, This indicates that the transient voltage signal is mapped to the first... Layer, First A sequence of discrete wavelet coefficients in an orthogonal time-frequency subspace.

[0188] By using the two recursive formulas for decomposing discrete coefficients, the wideband voltage sequence that was originally aliased in the time domain can be refined and reconstructed into multiple independent sub-band signals covering low-frequency fundamental waves to high-frequency harmonics.

[0189] When the wavelet decomposition reaches a preset depth level, the reconstructed discrete wavelet coefficient sequence within each independent sub-band is denoted as: ,in, This represents the discrete cursor index within the current sliding time window. Represents the sub-band index, for the first... Within each sub-band The transient energy of each sampling point is quantized, and its normalized relative probability measure relative to the total energy of that frequency band is calculated. :

[0190]

[0191] in, This represents the energy of the wavelet coefficients at a single sampling point. Indicates the first time within the adjusted sliding time window The total global energy contained in each sub-band.

[0192] Furthermore, based on Shannon's information entropy principle, the first... Layer, First Normalized relative logarithm time-frequency energy entropy of each sub-band :

[0193]

[0194] When an external fault occurs, the physical structure of the distribution network remains unchanged, and transient energy is still highly concentrated and clamped in the power frequency fundamental band, with the time-frequency energy entropy maintaining a low baseline level. However, when a real topological break occurs, a large number of broadband harmonics surge in and are rapidly scattered to the high-frequency subband, causing a sharp increase in the uncertainty and disorder of energy distribution, and the time-frequency energy entropy will undergo a pulse-like and violent jump.

[0195] In one possible implementation, Table 1 is an example table of electrical characteristic data, which includes: transient characteristic data (including voltage transient energy, frequency transient energy, active power transient energy, and reactive power transient energy), time-domain physical quantity characteristic data (including voltage change rate, frequency change rate, active power change rate, and reactive power change rate), dynamic characteristic data (including apparent power, power factor, equivalent dynamic resistance, and equivalent dynamic reactance), and frequency domain distribution characteristic data (including time-frequency energy entropy).

[0196] Table 1

[0197]

[0198] The fault identification method for distribution networks provided in this application first uses the Teager-Kaiser energy operator and Hinkley cumulative test to extract features of the transient energy of the electrical data to be processed in the first preset duration, obtaining transient feature data, which includes: voltage transient energy, frequency transient energy, active power transient energy, and reactive power transient energy. Next, the first-order difference is used to extract features of the electrical data to be processed in the time-domain physical quantities, obtaining time-domain physical quantity feature data, which includes: voltage change rate, frequency change rate, active power change rate, and reactive power change rate. Then, the features of the electrical data to be processed in the dynamic electrical quantities in the second preset duration are extracted to obtain dynamic feature data, which includes: apparent power, power factor, equivalent dynamic resistance, and equivalent dynamic reactance. Finally, wavelet packet transform is used to extract features of the electrical data to be processed in the frequency domain distribution, obtaining frequency domain distribution feature data, which includes time-frequency energy entropy. In this embodiment, the operating status of the distribution network is characterized simultaneously at four scales: short-term transient, time-domain physical quantity, long-term dynamic, and frequency-domain energy distribution. This allows the feature data to reflect both the initial disturbance of the fault and the subsequent dynamic recovery and energy distribution differences. From a physical mechanism perspective, the low-voltage ride-through masking effect is decoupled, reducing the instability of fault identification in low-voltage ride-through, power balance, and noise interference scenarios caused by a single threshold criterion. This improves the ability to distinguish between real island faults and faults outside the area, thereby enhancing the accuracy and anti-interference capability of distribution network fault identification.

[0199] Based on the above embodiments, Figure 3 A flowchart illustrating the fault identification method for a power distribution network provided in this application embodiment. Figure 3 .like Figure 3 As shown, prior to step S103 above, the fault identification method for this power distribution network further includes the following steps:

[0200] S301. Based on the simulation model of the distribution network to be processed, determine multiple simulated feature data under multiple operating conditions corresponding to different labeled fault types.

[0201] In this step, based on the simulation model, a simulation mapping relationship consistent with the distribution network to be processed is pre-constructed. This can simulate the electrical response process of different labeled fault types under different operating conditions, collect electrical quantities during the electrical response process, and obtain multiple simulation feature data.

[0202] For example, the simulation model is a power grid calculation model built based on the topology, line parameters, distributed generation grid connection parameters, and load parameters of the distribution network to be processed; the operating conditions are used to characterize the state differences of the distribution network to be processed during the simulation. The state differences of the distribution network to be processed may include load fluctuations, changes in distributed generation output, changes in voltage levels at the distributed generation grid connection points, and changes in fault access locations, etc.

[0203] In one possible implementation, a dynamic perturbation engine based on Monte Carlo random tracing is embedded in the simulation model. Within a limited physical range, the underlying control parameters of the simulation model, including transition resistance, load equivalent quality factor, and background nonlinear harmonics, are subjected to high-frequency random perturbation to generate multiple operating conditions corresponding to the labeled fault types.

[0204] Under each operating condition, the simulation model outputs corresponding three-phase voltage, three-phase current, power, frequency, or combinations thereof as electrical data. After feature extraction from these electrical data as described in the previous embodiments, multiple simulated feature data are obtained. These multiple simulated feature data and the electrical feature data share the same feature dimensions, facilitating subsequent model learning.

[0205] The following are examples illustrating several operating conditions:

[0206] The first type of operating condition is constrained by the non-detection zone (NDZ):

[0207] The true islanding evolution boundary condition belongs to the classic NDZ extreme islanding fault type of operation condition, which is difficult to detect. Under the true islanding evolution boundary condition, the source and load power in the distribution network is almost balanced, and the voltage / frequency fluctuations are extremely weak. LVRT control can easily completely mask the islanding characteristics. The NDZ constraint formula is as follows:

[0208]

[0209] in, and These represent the lower limit operating thresholds for the protection settings: voltage and frequency, respectively. Indicates the load quality factor. and These represent the active power difference and reactive power difference between the distributed power source and the local load after islanding is formed, respectively. Indicates the rated active power of the local load. and These represent the rated voltage and rated frequency, respectively.

[0210] If the power difference between the source and load is within the upper limit of the NDZ constraint formula, the islanding voltage and frequency will never reach the protection setting value, the traditional passive islanding protection will fail, and it will fall into the NDZ blind zone, which is the operating condition with the most severe LVRT masking effect.

[0211] Further subdividing the real-world island evolution boundary conditions, we can categorize them into the following types of island conditions:

[0212] (1) Power balance islanding mode: Set the source and load power to be extremely matched, apply a very small load fluctuation, the source and load imbalance is very small, the voltage / frequency drift is very smooth, and the LVRT reactive power compensation almost completely smooths out the electrical changes.

[0213] (2) High Q value islanding condition: set the load quality factor Randomly distributed within a preset range, such as 2.0~3.0, to excite strong LC parallel resonance to mask island frequency drift, LVRT further compensates for voltage, and the waveform highly overlaps with normal grid connection;

[0214] (3) Islanding operation in weak network environment: Dynamically change the equivalent impedance of the power grid side to limit the short circuit ratio (SCR) of the power grid to a preset range, such as 10~20. When there is a fault outside the area, the power grid clamping capability is poor. After the LVRT voltage regulation, the fault waveform is infinitely close to the island waveform, which is a high-incidence scenario of LVRT masking.

[0215] (4) Cascaded fault islanding: simulates the extreme evolution scenario of cascaded tripping caused by a short circuit outside the zone.

[0216] The above values ​​are for illustrative purposes only and do not limit the specific implementation methods.

[0217] The second type of operating condition is not subject to NDZ constraints:

[0218] Operating conditions with non-islanding faults are not subject to NDZ constraints and mainly include operating conditions with external fault types and other non-islanding disturbance types.

[0219] Operating conditions with faults outside the zone include the following conditions:

[0220] (1) External short circuit fault condition: The duration of the three-phase short circuit is set to meet the low voltage ride-through test time. The transition resistance varies randomly within the first preset resistance range, such as 0.01~0.5Ω.

[0221] (2) Power grid fluctuation conditions: Set the preset phase drop range, such as -20°~-10°, the preset voltage drop percentage, such as 60%~80%, and the preset frequency drop value, such as 0.5Hz~1.5Hz;

[0222] (3) Power grid harmonic injection condition: The injected content fluctuates within the preset first fluctuation percentage of the second harmonic and the preset second fluctuation percentage of the third harmonic, forming a nonlinear distorted pollution waveform, wherein the preset first fluctuation percentage is, for example, 1%~5%, and the preset second fluctuation percentage is, for example, 1%~4%;

[0223] (4) Long line long-term fault condition: Set the length of the long overhead line, such as 3~10km, and the short-circuit transition resistance at the far end varies randomly within the second preset resistance range, such as 5~20Ω.

[0224] (5) Circuit breaker phase loss condition: Set the fluctuation range of the main load level, such as -5% to +5%, disconnect the phase line to generate random unbalanced current impact;

[0225] (6) High-resistance arc grounding condition: Introduce nonlinear edge samples, set the breakdown voltage range, such as 1500V~3700V, and the grounding impedance distribution range, such as 1500~2500Ω.

[0226] The above values ​​are for illustrative purposes only and do not limit the specific implementation methods.

[0227] Other non-islanding disturbance types of operating conditions include the following:

[0228] (1) Continuous disturbance islanding condition: Load is introduced in advance at random times within a preset time range before the islanding action is triggered to create a transient operation. The preset time range is, for example, 0.4s to 0.6s.

[0229] (2) Asynchronous motor starting condition: Set the rated power of the asynchronous motor to a preset power percentage of the total capacity, such as 10%~20%, and full-voltage start-up brings shock transient;

[0230] (3) Load step condition: a resistive load is suddenly switched on, causing a jump in active power that varies randomly within a preset amplitude range, such as 14kW~26kW;

[0231] (4) Capacitor switching operation: When the power grid is at any phase angle, the reactive power compensation capacitor is switched on and off at random times;

[0232] (5) DG internal switching condition: The number of parallel groups inside the inverter is set to undergo a slight topology change at random times.

[0233] The above values ​​are for illustrative purposes only and do not limit the specific implementation methods.

[0234] Furthermore, a global transient perturbation eigenvector is defined to characterize the source-load coupling depth and fault intensity. This includes the operating condition control inputs of the Monte Carlo simulation (i.e., the underlying control parameters of the simulation), used to randomly generate the different operating conditions mentioned above, controlling the source-load imbalance degree, fault strength, and grid distortion level of each simulation group. The global transient perturbation feature vector can be represented as follows:

[0235]

[0236] in, and These represent the active power mismatch and reactive power mismatch, respectively. Indicates the load quality factor. Indicates the fault transition resistance. This represents a background distortion operator that includes phase jumps and frequency drops.

[0237] S302. Train the backpropagation neural network model based on multiple simulated feature data and the labeled fault type corresponding to each simulated feature data to obtain the fault identification model.

[0238] In this step, multiple simulated feature data are input into the backpropagation neural network model for fault identification, and the fault type output by the backpropagation neural network model is obtained. The labeled fault type corresponding to each simulated feature data is used as the supervision target to train the backpropagation neural network model, and the trained backpropagation neural network model is used as the fault identification model.

[0239] In one possible implementation,

[0240] Each simulated feature data can be input into the backpropagation neural network model, and its corresponding labeled fault type can be used as the expected output. By calculating the error between the fault type output by the backpropagation neural network model and the labeled fault type, the network weights and thresholds of the backpropagation neural network model are updated layer by layer using the backpropagation algorithm until the training error meets the preset convergence condition, thereby obtaining the fault identification model.

[0241] In one possible implementation, step S302 described above may include the following steps:

[0242] Step 1: Use the slime mold algorithm to optimize the hyperparameters of the backpropagation neural network model and determine the target hyperparameters;

[0243] The hyperparameters include: the number of hidden layer nodes and the L2 regularization coefficient;

[0244] For example, the number of hidden layer nodes can affect the ability of the backpropagation neural network model to express the nonlinear mapping of fault features, and the L2 regularization coefficient is used to suppress excessive weight growth and alleviate overfitting.

[0245] In one possible implementation, the slime mold algorithm is used to jointly search and optimize the number of hidden layer nodes and L2 regularization coefficient of the backpropagation neural network model, so as to reduce the risk of local optima during training and improve the generalization ability of the model.

[0246] Preset number of hidden layer nodes The search scope, such as Preset L2 regularization coefficient The search scope, such as The above values ​​are only for the purpose of understanding the scheme and do not limit the specific implementation method.

[0247] Define a composite fitness functional to characterize the current combination of hyperparameters. The overall trial-and-error cost of the backpropagation neural network model, specifically expressed by the composite fitness functional, is as follows:

[0248]

[0249] in, This represents the current hyperparameter position vector of an individual in the Slime Mould Algorithm (SMA) population. This represents the cross-entropy classification error of the backpropagation neural network model under the current hyperparameter combination. This represents the percentage of hidden layer nodes and is used to penalize overly complex network topologies. Indicates the L2 regularization penalty term ( This represents the underlying weight matrix of the backpropagation neural network model, used to suppress overfitting caused by excessively large weights. and These represent the penalty weighting factors.

[0250] The composite fitness functional calculated The smaller the scalar value, the better the combination of hyperparameters (smaller error and simpler structure). The specific value calculated by the composite fitness functional is used to evaluate the "fitness value" of the location of each slime mold population individual.

[0251] SMA simulates the venous contraction and oscillation mechanism of *Vorticella velutipes* during foraging in nature. Through adaptive weighted feedback oscillation, it achieves dynamic decoupling between wide-area exploration of the solution space and local depth exploitation. This drives the population position vector. The discrete dynamics piecewise equations of evolution can be characterized as:

[0252]

[0253] in, and They represent the first generation and first The hyperparameter location vector of the generation (including the number of hidden layer nodes to be optimized and the L2 regularization coefficient). This indicates the position of the individual with the best fitness (calculated from the composite fitness functional above) in the current population. Represents a uniformly distributed random number within the interval [0, 1]. This represents a preset minimum probability constant, used to control the probability that the algorithm will force it to exit a local optimum and be randomly reinitialized. and Let represent the position vectors of two individuals randomly selected from the current population. and Let represent the upper and lower bounds of the hyperparameter search space, respectively. The adaptive weighting operator representing the thickness of simulated slime mold veins is not preset, but rather dynamically generated logarithmic mapping values ​​based on the fitness ranking of all individuals in the current iteration. This represents the oscillation control vector, used to drive the population to conduct large-scale wide-area exploration (escaping local optima), prompting the population to... Perform localized fine-grained convergence in the vicinity. This represents the feedback decay vector, whose value converges linearly to 0 with the number of iterations. The dynamic probability threshold representing the control over whether a slime mold individual switches between exploration (wide-area search) and exploitation (local convergence) is calculated using the following formula: The closer the current individual fitness is to the optimal solution, The closer it gets to 0.

[0254] With the number of iterations When the increase of , SMA stops iterating and is determined to be "converged" when any of the following preset conditions are met, and the target hyperparameter value is obtained. The preset conditions include:

[0255] (1) Reach the preset maximum total number of iterations;

[0256] (2) Preset the fitness function value of the globally optimal individual for a consecutive number of iterations. The variation is less than the preset tolerance.

[0257] Step 2: Adjust the hyperparameters of the backpropagation neural network model based on the target hyperparameter values ​​to obtain the adjusted backpropagation neural network model;

[0258] For example, by completing the structural reconstruction and training constraint configuration based on the target hyperparameters, an adjusted backpropagation neural network model is obtained, which facilitates stable convergence of subsequent supervised learning.

[0259] In one possible implementation, the target number of hidden layer nodes and the target L2 regularization coefficient are determined based on the target hyperparameter values. After configuring the connection relationships and regularization terms of the input layer, hidden layer and output layer of the backpropagation neural network model, the adjusted backpropagation neural network model is obtained.

[0260] Step 3: Train the adjusted backpropagation neural network model based on multiple simulated feature data and the labeled fault type corresponding to each simulated feature data to obtain the fault identification model.

[0261] For example, multiple simulated feature data and the labeled fault type corresponding to each simulated feature data are input into the adjusted backpropagation neural network model for supervised training. The network weights and biases are updated using error backpropagation until the preset convergence condition is met, thus obtaining the fault identification model.

[0262] In this embodiment, the hyperparameters of the backpropagation neural network model are optimized using the slime mold algorithm, so that the model has a more stable classification boundary when facing simulated feature data under different operating conditions and improves the ability to represent complex fault modes. Furthermore, after the hyperparameters are adjusted and then supervised training is performed, the fault identification model can effectively improve its ability to distinguish between real islanded faults and non-islanded faults (including external faults and other disturbance states).

[0263] In one possible implementation, Figure 4 This is a schematic diagram of the structure of the adjusted backpropagation neural network model provided in an embodiment of this application. Figure 4 As shown, the layers are the input layer, hidden layer, and output layer.

[0264] For example, a fault identification model is obtained by training the adjusted backpropagation neural network model, and the application process of the fault identification model includes:

[0265] First, the input layer receives power grid feature data (the power grid feature data includes feature data with a preset number of dimensions). The input layer contains a preset number of neuron nodes and transmits them to the hidden layer according to the preset dimensions.

[0266] Next, the hidden layer performs nonlinear mapping on the power grid characteristic data to extract nonlinear feature data that characterizes the differences in faults. The hidden layer includes multiple independent neuron computing nodes to complete the nonlinear spatial mapping of high-dimensional transient features.

[0267] Subsequently, the output layer outputs the scalar confidence level of the fault type corresponding to the nonlinear feature data of the power grid, and then determines the target fault type. For example, the scalar confidence level is compared with a preset threshold: if the threshold is exceeded, it indicates that the fault has entered the real island feature space and the corresponding target fault type is a real island fault; otherwise, it is characterized as a grid-connected disturbance state such as low voltage ride-through and the corresponding target fault type is a non-island fault.

[0268] The above values ​​are for illustrative purposes only and do not limit the specific implementation methods.

[0269] Accordingly, one possible implementation of step 3 above includes the following steps:

[0270] S1, for each simulated feature data, the input layer receives the simulated feature data and transmits it to the hidden layer;

[0271] For example, the input layer acts as a data receiving layer, receiving simulated feature data and transmitting it to the hidden layer.

[0272] S2, the hidden layer performs nonlinear feature extraction on the simulated feature data to obtain nonlinear feature data;

[0273] For example, nonlinear feature data is obtained by weighted summation and nonlinear activation of the simulated feature data by the hidden layer.

[0274] In one possible implementation, the nonlinear activation function could be a Sigmoid function, a tanh function, or a ReLU (Rectified Linear Unit) function, to improve the fitting ability to complex transient features.

[0275] S3 inputs the nonlinear characteristic data to the output layer to obtain the simulated fault type;

[0276] For example, nonlinear feature data is input to the output layer, which outputs a scalar confidence functional characterizing islanded faults. Then, the simulated fault type is determined by comparing the scalar confidence functional value with a preset benchmark threshold.

[0277] In one possible implementation, the discrete Boolean decision function for a single sliding time window It can be defined as:

[0278]

[0279] in, This represents the preset baseline threshold for binary classification, which distinguishes between real islanded faults and non-islanded faults (including out-of-area faults and other disturbance states). When, it indicates that the transient trajectory within the current single sliding time window has entered the high-dimensional feature space of the unplanned island; when At this time, it represents the grid-connected state where the distribution network is disturbed but the physical topology is not broken.

[0280] To avoid transient oscillations at the decision boundary caused by high-frequency glitches, strong white noise superposition, or specific-depth grid faults (such as the low-voltage ride-through period) in single-sliding-time-window measurements, a continuous sliding-window cumulative tripping function based on a finite-state machine is established. :

[0281]

[0282] in, This indicates the number of consecutive sliding time windows; only when consecutive... If the discrete Boolean decision results of adjacent dynamic sliding time windows are true when strictly performing the logical AND operation, then the fault type is determined to be a true islanding fault, that is:

[0283]

[0284] S4. Based on all simulated fault types and the labeled fault type corresponding to each simulated feature data, update the parameters of the adjusted backpropagation neural network model. Repeat steps S1-S4 until the preset training stopping condition is met to obtain the fault identification model.

[0285] For example, preset training stopping conditions may include: reaching the maximum number of iterations, the overall error being below a threshold, or the error decreasing by less than a set value over multiple consecutive iterations.

[0286] In one possible implementation, the simulated fault type is compared with the labeled fault type corresponding to each simulated feature data, the classification error is calculated, and the network weights and biases are updated by backpropagation based on the classification error, so that the underlying weight parameters of the adjusted backpropagation neural network model are iteratively adjusted in the direction of reducing the classification error.

[0287] In this embodiment, the adjusted backpropagation neural network model can learn the differences in fault modes in multiple simulated feature data during the training process, enabling the input layer, hidden layer and output layer to work together to complete the mapping from features to fault types. The fault identification model obtained by training can effectively distinguish between real island faults and similar states such as faults outside the area, improve the identification accuracy and anti-interference ability under complex transient conditions, and reduce the probability of misjudgment and missed judgment.

[0288] The fault identification method for distribution networks provided in this application first determines multiple simulated feature data under multiple operating conditions corresponding to different labeled fault types based on a simulation model of the distribution network to be processed. Then, a backpropagation neural network model is trained based on the multiple simulated feature data and the labeled fault type corresponding to each simulated feature data to obtain a fault identification model. In this embodiment, training the backpropagation neural network model based on multiple simulated feature data under multiple operating conditions corresponding to different labeled fault types can enable the model to cover fault evolution forms and operating boundary conditions, effectively improving the model's ability to distinguish between real islanded faults and non-islanded faults, and enabling it to output a more stable target fault type when subsequent input of actual electrical feature data.

[0289] Based on the above embodiments, Figure 5 A flowchart illustrating the fault identification method for a power distribution network provided in this application embodiment. Figure 4 .like Figure 5 As shown, prior to step S103 above, the fault identification method for this power distribution network further includes the following steps:

[0290] S501. The ReliefF algorithm is used to select features from electrical feature data to obtain multiple candidate feature data.

[0291] In this step, the ReliefF algorithm is used to evaluate the importance of electrical feature data, including transient feature data, time-domain physical quantity feature data, dynamic feature data, and frequency-domain distribution feature data. Different weights are assigned to different feature data according to their importance, and multiple candidate feature data are obtained by filtering based on the weights.

[0292] In one possible implementation, the electrical feature data and multiple simulated feature data have the same feature dimensions. When using the ReliefF algorithm to select features from the electrical feature data, a target simulated feature data can be randomly extracted from the multiple simulated feature data. Then, based on the multiple simulated feature data and the labeled fault type corresponding to each simulated feature data, the similar nearest neighbor feature data and dissimilar nearest neighbor feature data of the target simulated feature data are determined. The fault types include two types: real islanded faults and non-islanded faults. If the labeled fault type of the target simulated feature data is a real islanded fault, then its similar nearest neighbor feature data is the simulated feature data corresponding to the other real islanded fault that has the closest Euclidean distance to the target simulated feature data among the multiple simulated feature data. The dissimilar nearest neighbor feature data is the simulated feature data corresponding to the non-islanded fault that has the closest Euclidean distance to the target simulated feature data among the multiple simulated feature data.

[0293] Then, the target simulated feature data is extracted cyclically from multiple simulated feature data. After each extraction of target simulated feature data, the global weights of different feature dimensions are updated and iterated. When the number of cyclic extractions reaches the preset number, the update and iteration of the global weights of different feature dimensions ends, and the global weights of all feature dimensions that have finally converged are obtained. Then, they are sorted in descending order according to the global weights, and the top K feature dimensions are selected to obtain multiple candidate feature data.

[0294] When updating and iterating the global weights for different feature dimensions, the global weights for all feature dimensions are first initialized to zero, that is, the weights for the first feature dimension are set to zero. Global weights of dimensional features After that, for the first Global weights of dimensional features The evolution equations are updated as follows:

[0295]

[0296] in, Indicates the fault type Prior probability ( This indicates a real-world isolated failure. (Indicates a non-islanded fault). It is calculated a priori by statistically analyzing the proportion of each type of fault-related simulated feature data in multiple simulated feature data sets to all simulated feature data sets; This represents the prior probability of the fault type to which the target simulated feature data belongs; Indicates the preset number of times to extract in a loop; , ,as well as They represent the first The target simulation feature data, the same type of nearest neighbor feature data, and the different type of nearest neighbor feature data are in the first... The specific numerical value of the feature.

[0297] S502. Determine multiple Pearson correlation coefficient values ​​between every two candidate feature data in multiple candidate feature data.

[0298] In this step, in order to further filter multiple candidate feature data, multiple Pearson correlation coefficient values ​​between every two candidate feature data can be calculated to characterize the degree of linear correlation between any two candidate feature data.

[0299] In one possible implementation, candidate feature data and candidate feature data Pearson correlation coefficient between The calculation formula is as follows:

[0300]

[0301] in, and These represent features of the same dimension in the candidate feature data. and candidate feature data The specific values ​​on, and These represent the characteristic mean.

[0302] S503. Update the electrical characteristic data based on the Pearson correlation coefficient value that is greater than the preset threshold.

[0303] In this step, multiple Pearson correlation coefficient values ​​are compared with a preset threshold. Based on the Pearson correlation coefficient values ​​that are greater than the preset threshold, the two candidate feature data corresponding to the Pearson correlation coefficient values ​​are retained, and then the electrical feature data is updated.

[0304] In one possible implementation, for candidate feature data pairs with a Pearson correlation coefficient greater than a preset threshold, the electrical feature data can be updated by combining feature weights, stability, or missing data. This can typically involve removing feature data with weak discriminative contributions or dynamically weighting and fusing highly correlated feature data.

[0305] The fault identification method for power distribution networks provided in this application first uses the ReliefF algorithm to select features from electrical feature data, obtaining multiple candidate feature data. Then, it determines multiple Pearson correlation coefficient values ​​between every two candidate feature data. Finally, it updates the electrical feature data based on Pearson correlation coefficient values ​​greater than a preset threshold. In this embodiment, the ReliefF algorithm first filters out feature data with low discriminative power, obtaining multiple candidate feature data. Then, Pearson correlation analysis further eliminates candidate feature data pairs corresponding to Pearson correlation coefficient values ​​not greater than a preset threshold, improving the discriminative power of the electrical feature data and reducing feature data redundancy, thereby improving the accuracy of fault identification.

[0306] Based on the above embodiments, the following are embodiments of the apparatus involved in this application:

[0307] Figure 6 This is a schematic diagram of the structure of a fault identification device for a power distribution network provided in an embodiment of this application. Figure 6 As shown, the fault identification device 600 of the power distribution network includes:

[0308] The acquisition module 601 is used to acquire electrical data to be processed, wherein the electrical data to be processed is the electrical data of the distributed power generation grid connection point in the distribution network to be processed;

[0309] The extraction module 602 is used to perform multi-scale feature extraction on the electrical data to be processed to obtain electrical feature data. The multi-scale feature extraction includes extracting transient energy, time-domain physical quantities, dynamic electrical quantities, and frequency-domain distribution features for a first preset duration, and the second preset duration is longer than the first preset duration. The electrical feature data includes: transient feature data, time-domain physical quantity feature data, dynamic feature data, and frequency-domain distribution feature data.

[0310] The processing module 603 is used to input electrical feature data into the fault identification model to obtain the target fault type of the distribution network to be processed. The fault identification model is trained based on multiple simulated feature data and the labeled fault type corresponding to each simulated feature data.

[0311] In one or more embodiments, the extraction module 602 is specifically used for:

[0312] The Teager-Kaiser energy operator and Hinkley cumulative test are used to extract features from the transient energy of the electrical data to be processed in the first preset time period to obtain transient feature data. The transient feature data includes: voltage transient energy, frequency transient energy, active power transient energy, and reactive power transient energy.

[0313] The first-order difference method is used to extract the features of the electrical data to be processed in the time domain physical quantities to obtain the time domain physical quantity feature data, which includes: voltage change rate, frequency change rate, active power change rate, and reactive power change rate.

[0314] Features of the electrical data to be processed in the dynamic electrical quantities of the second preset time period are extracted to obtain dynamic feature data, which includes: apparent power, power factor, equivalent dynamic resistance, and equivalent dynamic reactance.

[0315] Wavelet packet transform is used to extract the features of the electrical data to be processed in the frequency domain to obtain frequency domain distribution feature data, which includes time-frequency energy entropy.

[0316] In one or more embodiments, before inputting electrical characteristic data into the fault identification model to obtain the target fault type of the distribution network to be processed, the processing module 603 is further configured to:

[0317] Based on the simulation model of the distribution network to be processed, multiple simulated feature data under multiple operating conditions corresponding to different labeled fault types are determined.

[0318] The backpropagation neural network model is trained based on multiple simulated feature data and the labeled fault type corresponding to each simulated feature data to obtain a fault identification model.

[0319] In one or more embodiments, the processing module 603 trains a backpropagation neural network model based on multiple simulated feature data and the labeled fault type corresponding to each simulated feature data to obtain a fault identification model, specifically for:

[0320] The slime mold algorithm is used to optimize the hyperparameters of the backpropagation neural network model and determine the target hyperparameters, including the number of hidden layer nodes and the L2 regularization coefficient.

[0321] Based on the target hyperparameter values, the backpropagation neural network model is adjusted to obtain the adjusted backpropagation neural network model.

[0322] The adjusted backpropagation neural network model is trained based on multiple simulated feature data and the labeled fault type corresponding to each simulated feature data to obtain the fault identification model.

[0323] In one or more embodiments, the adjusted backpropagation neural network model includes: an input layer, a hidden layer, and an output layer;

[0324] Correspondingly, processing module 603 trains the adjusted backpropagation neural network model based on multiple simulated feature data and the labeled fault type corresponding to each simulated feature data to obtain a fault identification model, specifically used for:

[0325] S1, for each simulated feature data, the input layer receives the simulated feature data and transmits it to the hidden layer;

[0326] S2, the hidden layer performs nonlinear feature extraction on the simulated feature data to obtain nonlinear feature data;

[0327] S3 inputs the nonlinear characteristic data to the output layer to obtain the simulated fault type;

[0328] S4. Based on all simulated fault types and the labeled fault type corresponding to each simulated feature data, update the parameters of the adjusted backpropagation neural network model. Repeat steps S1-S4 until the preset training stopping condition is met to obtain the fault identification model.

[0329] In one or more embodiments, before inputting electrical characteristic data into the fault identification model to obtain the target fault type of the distribution network to be processed, the processing module 603 is further configured to:

[0330] The ReliefF algorithm is used to select features from electrical feature data to obtain multiple candidate feature data.

[0331] Determine multiple Pearson correlation coefficient values ​​between every two candidate feature data in a multi-candidate feature data set;

[0332] The electrical characteristic data is updated based on the Pearson correlation coefficient value that is greater than a preset threshold.

[0333] In one or more embodiments, after acquiring the electrical data to be processed, the acquisition module 601 is further configured to:

[0334] Perform signal mutation analysis on the electrical data to be processed to determine the magnitude of the mutation;

[0335] The preset sliding time window is adjusted based on the magnitude of the mutation to obtain the adjusted sliding time window;

[0336] The electrical data to be processed is truncated according to the adjusted sliding time window to obtain the truncated electrical data to be processed.

[0337] The fault identification device for the power distribution network provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0338] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 7 As shown, the electronic device 700 includes: a processor 701, a memory 702, and a bus 703;

[0339] The memory 702 is used to store the computer-executed instructions of the processor 701;

[0340] The processor 701 is configured to execute the technical solutions of any of the foregoing method embodiments by executing computer execution instructions.

[0341] The specific implementation process of processor 701 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0342] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0343] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0344] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0345] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0346] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0347] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random-Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0348] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an application-specific integrated circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0349] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0350] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0351] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0352] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory (RAM), magnetic disks, or optical disks.

[0353] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0354] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A fault identification method for a power distribution network, characterized in that, include: Acquire electrical data to be processed, wherein the electrical data to be processed is the electrical data of the distributed power generation grid connection point in the distribution network to be processed; Multi-scale feature extraction is performed on the electrical data to be processed to obtain electrical feature data. The multi-scale feature extraction includes extracting transient energy, time-domain physical quantities, dynamic electrical quantities, and frequency domain distribution features for a first preset duration. The second preset duration is longer than the first preset duration. The electrical feature data includes transient feature data, time-domain physical quantity feature data, dynamic feature data, and frequency domain distribution feature data. The electrical feature data is input into the fault identification model to obtain the target fault type of the distribution network to be processed. The fault identification model is trained based on multiple simulated feature data and the labeled fault type corresponding to each simulated feature data.

2. The method according to claim 1, characterized in that, The step of performing multi-scale feature extraction on the electrical data to be processed to obtain electrical feature data includes: The Teager-Kaiser energy operator and Hinkley cumulative test are used to extract the features of the transient energy of the electrical data to be processed in the first preset time period to obtain the transient feature data, which includes: voltage transient energy, frequency transient energy, active power transient energy, and reactive power transient energy. The first-order difference is used to extract the features of the electrical data to be processed in the time-domain physical quantities to obtain the time-domain physical quantity feature data, which includes: voltage change rate, frequency change rate, active power change rate, and reactive power change rate. The features of the electrical data to be processed in the dynamic electrical quantities of the second preset time period are extracted to obtain the dynamic feature data, which includes: apparent power, power factor, equivalent dynamic resistance, and equivalent dynamic reactance; Wavelet packet transform is used to extract the features of the electrical data to be processed in the frequency domain distribution to obtain the frequency domain distribution feature data, which includes time-frequency energy entropy.

3. The method according to claim 2, characterized in that, Before inputting the electrical characteristic data into the fault identification model to obtain the target fault type of the distribution network to be processed, the method further includes: Based on the simulation model of the distribution network to be processed, the multiple simulation feature data under multiple operating conditions corresponding to different labeled fault types are determined; The backpropagation neural network model is trained based on the multiple simulated feature data and the labeled fault type corresponding to each simulated feature data to obtain the fault identification model.

4. The method according to claim 3, characterized in that, The step of training the backpropagation neural network model based on the plurality of simulated feature data and the labeled fault type corresponding to each simulated feature data to obtain the fault identification model includes: The slime mold algorithm is used to optimize the hyperparameters of the backpropagation neural network model and determine the target hyperparameters, which include: the number of hidden layer nodes and the L2 regularization coefficient. Based on the target hyperparameter values, the backpropagation neural network model is hyperparameter adjusted to obtain the adjusted backpropagation neural network model. The adjusted backpropagation neural network model is trained based on the multiple simulated feature data and the labeled fault type corresponding to each simulated feature data to obtain the fault identification model.

5. The method according to claim 4, characterized in that, The adjusted backpropagation neural network model includes: an input layer, a hidden layer, and an output layer; Accordingly, the step of training the adjusted backpropagation neural network model based on the plurality of simulated feature data and the labeled fault type corresponding to each simulated feature data to obtain the fault identification model includes: S1, for each simulated feature data, the input layer receives the simulated feature data and transmits it to the hidden layer; S2, the hidden layer performs nonlinear feature extraction on the simulated feature data to obtain nonlinear feature data; S3, input the nonlinear characteristic data to the output layer to obtain the simulated fault type; S4. Based on all simulated fault types and the labeled fault type corresponding to each simulated feature data, update the parameters of the adjusted backpropagation neural network model. Repeat steps S1-S4 until the preset training stopping condition is met to obtain the fault identification model.

6. The method according to any one of claims 1-5, characterized in that, Before inputting the electrical characteristic data into the fault identification model to obtain the target fault type of the distribution network to be processed, the method further includes: The ReliefF algorithm is used to perform feature selection on the electrical feature data to obtain multiple candidate feature data. Determine multiple Pearson correlation coefficient values ​​between every two candidate feature data in the plurality of candidate feature data; The electrical characteristic data is updated based on the Pearson correlation coefficient value that is greater than a preset threshold.

7. The method according to any one of claims 1-5, characterized in that, After acquiring the electrical data to be processed, the method further includes: Perform signal mutation analysis on the electrical data to be processed to determine the magnitude of the mutation. Based on the magnitude of the mutation, the preset sliding time window is adjusted to obtain the adjusted sliding time window; The electrical data to be processed is truncated according to the adjusted sliding time window to obtain the truncated electrical data to be processed.

8. A fault identification device for a power distribution network, characterized in that, include: The acquisition module is used to acquire electrical data to be processed, which is the electrical data of the distributed power source grid connection point in the distribution network to be processed; An extraction module is used to perform multi-scale feature extraction on the electrical data to be processed to obtain electrical feature data. The multi-scale feature extraction includes extracting transient energy, time-domain physical quantities, dynamic electrical quantities, and frequency domain distribution features for a first preset duration. The second preset duration is longer than the first preset duration. The electrical feature data includes: transient feature data, time-domain physical quantity feature data, dynamic feature data, and frequency domain distribution feature data. The processing module is used to input the electrical feature data into the fault identification model to obtain the target fault type of the distribution network to be processed. The fault identification model is trained based on multiple simulated feature data and the labeled fault type corresponding to each simulated feature data.

9. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-7.

Images