A method and system for customer line fault section localization

Abstract

The application provides a customer line fault section positioning method and system, comprising obtaining voltage and current measurement data of multiple monitoring points on a customer line within a preset time window, and performing clock synchronization and normalization processing; performing Hilbert-Huang transform on the transient component of each monitoring point to obtain the instantaneous frequency and instantaneous envelope of the transient traveling wave, and determining the main oscillation period according to the energy spectrum of the instantaneous frequency; constructing a combined feature vector; constructing a gradient boosting decision tree model that fuses feature attention and tree attention, weighting and summing the prediction values of all base learners according to the corresponding tree attention weights to obtain the model output; and using the model to infer the combined feature vector extracted from real-time measurement data, and outputting the fault section with the highest probability as the positioning result. The application uses a combined feature vector and feature attention and tree attention to improve the accuracy of customer line fault section positioning.

Description

Technical Field

[0001] This application pertains to the field of power, and in particular relates to a method and system for locating fault sections in customer power lines. Background Technology

[0002] The structure of customer-side distribution networks is becoming increasingly complex, and the types and causes of line faults are becoming more diverse. Fault location technologies are mainly divided into two categories: model-driven impedance methods and traveling wave methods. Impedance methods determine the fault distance by measuring the voltage and current at the line terminals and calculating the impedance from the fault point to the measurement point. Traveling wave methods utilize the characteristic that transient traveling waves generated at the moment of a fault propagate at near the speed of light on the transmission line. They locate the fault by detecting the time difference in arrival time of the traveling wave at different monitoring points, theoretically offering high accuracy, but they have stringent requirements on the sampling rate and clock synchronization accuracy of the data acquisition equipment. In addition, data-driven methods extract patterns from historical operation and fault data to build intelligent models that can automatically identify fault sections. For example, support vector machines, artificial neural networks, and decision trees are used to classify the electrical features extracted from fault waveform data. However, feature engineering is relatively rudimentary, mostly focusing on steady-state or quasi-steady-state features such as the effective values, phase, and harmonics of voltage and current. It is difficult to detect other information during the transient process of the fault, especially the correlation characteristics exhibited by the traveling wave as it propagates at different spatial monitoring points. In terms of model building, although ensemble learning models, represented by gradient boosting decision trees, have superior performance, they treat all features and samples equally during the iteration process. They cannot focus on the key features that contribute the most to distinguishing faulty sections, nor can they focus on hard samples that the learning model cannot correctly judge, which to some extent limits the model's localization ability. Summary of the Invention

[0003] To address the aforementioned problems, in a first aspect, the present invention proposes a method for locating faulty sections of customer lines, comprising the following steps:

[0004] The system acquires voltage and current measurement data from multiple monitoring points on the customer's line within a preset time window, and performs clock synchronization and normalization processing. It then performs Hilbert-Huang transform on the transient components of each monitoring point to obtain the instantaneous frequency and instantaneous envelope of the transient traveling wave, and determines the main oscillation period based on the energy spectrum of the instantaneous frequency.

[0005] Construct a combined feature vector, including the correlation entropy feature calculated based on the synchronous differential sequence of any two monitoring points within a time sub-window with a length that is a preset multiple of the main oscillation period, the phase amplitude asymmetry feature calculated based on the product of the instantaneous phase difference and the instantaneous envelope of the transient traveling wave of the two monitoring points, and the basic electrical quantity feature;

[0006] Construct a gradient boosting decision tree model that integrates feature attention and tree attention, and sum the predictions of all base learners according to their corresponding tree attention weights to obtain the model output;

[0007] The model is used to infer the combined feature vectors extracted from real-time measurement data, and the fault section with the highest probability is output as the location result.

[0008] In a second aspect, the present invention proposes a customer line fault section location system, comprising the following modules:

[0009] The acquisition module is used to acquire voltage and current measurement data from multiple monitoring points on the customer line within a preset time window, and perform clock synchronization and normalization processing; it performs Hilbert-Huang transform on the transient components of each monitoring point to obtain the instantaneous frequency and instantaneous envelope of the transient traveling wave, and determines the main oscillation period based on the energy spectrum of the instantaneous frequency.

[0010] The feature extraction module is used to construct a combined feature vector, including the correlation entropy feature calculated based on the synchronous differential sequence of any two monitoring points within a time sub-window with a length that is a preset multiple of the main oscillation period, the phase amplitude asymmetry feature calculated based on the product of the instantaneous phase difference and the instantaneous envelope of the transient traveling wave of the two monitoring points, and the basic electrical quantity feature.

[0011] The construction and training module is used to build a gradient boosting decision tree model that integrates feature attention and tree attention. It calculates the weighted sum of the predictions of all base learners according to their corresponding tree attention weights to obtain the model output.

[0012] The inference module is used to infer the combined feature vectors extracted from real-time measurement data using the model, and output the fault section with the highest probability as the location result.

[0013] The fault location method for customer lines proposed in this invention, through Hilbert-Huang transform of transient components, can identify the characteristics of transient traveling waves during propagation from multiple dimensions and in depth. The constructed gradient-boosting decision tree attention model, during training, emphasizes the contribution of key features in node splitting evaluation, enhancing the model's ability to detect core identification information; it focuses on learning fault samples that are difficult to distinguish, improving the model's accuracy by assigning higher decision weights to base learners that perform well on these samples. The reliability of fault location is improved through a combination of feature engineering and optimized model structure. Attached Figure Description

[0014] Figure 1 A flowchart of the first embodiment;

[0015] Figure 2 This is a schematic diagram of a transient component signal;

[0016] Figure 3 A schematic diagram illustrating the construction of combined feature vectors;

[0017] Figure 4 This outputs a schematic diagram of the fault location results. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use, and processing of related data must comply with relevant laws, regulations, and standards, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0019] A specific embodiment of a method for locating faulty sections of a customer's line, such as... Figure 1 This includes the following steps:

[0020] Step 1: Obtain voltage and current measurement data from multiple monitoring points on the customer line within a preset time window, and perform clock synchronization and normalization processing; perform Hilbert-Huang transform on the transient components of each monitoring point to obtain the instantaneous frequency and instantaneous envelope of the transient traveling wave, and determine the main oscillation period based on the energy spectrum of the instantaneous frequency.

[0021] Specifically, synchronous phasor measurement units or intelligent terminals are deployed at key locations such as substations, feeder branch points, and large customer access points to collect three-phase voltage and current waveforms at a sampling rate of no less than 10kHz; the preset time window is set from 100ms before the fault occurrence to 200ms after the fault occurrence; the data of all monitoring points are synchronized using BeiDou or GPS satellite timing signals to ensure that the timestamp error is within 1μs; the minimum-maximum normalization method is used to linearly map the voltage and current measurements of each monitoring point to the interval between 0 and 1, eliminating dimensional differences. The power frequency and low-order harmonic components in the normalized data are filtered out by fast Fourier transform to obtain pure transient components; empirical mode decomposition is performed on this transient component signal, such as... Figure 2The component signal is decomposed into a set of intrinsic mode functions (EMFs) from high frequency to low frequency. A Hilbert transform is performed on each EMF to calculate the corresponding instantaneous frequency and instantaneous envelope. The Hilbert energy spectrum of all EMFs is calculated, and the frequency corresponding to the peak energy in the energy spectrum is identified as the dominant frequency; its reciprocal is the dominant oscillation period. For example, instantaneous frequency data of a monitoring point within a preset time window is acquired to form a time series, such as a series containing 1000 data points with values ​​fluctuating around 50Hz. A Fast Fourier Transform (FFT) algorithm is performed on the instantaneous frequency series composed of 1000 data points, converting the time-domain signal into a frequency-domain signal, thereby obtaining the energy spectral density function (ESD), which represents the distribution of signal energy at different frequencies. The ESD spectrum is then searched for the peak with the highest energy. For example, if a peak is found at 2Hz after analysis, 2Hz is identified as the dominant frequency of the system, indicating the existence of a low-frequency oscillation with a period of 2 times per second. Calculate the reciprocal of the dominant frequency, which is 0.5s, and the 0.5s is determined to be the dominant oscillation period.

[0022] Step 2: Construct a combined feature vector, including the correlation entropy feature calculated based on the synchronous differential sequence of any two monitoring points within a time sub-window with a length that is a preset multiple of the main oscillation period, the phase amplitude asymmetry feature calculated based on the product of the instantaneous phase difference and the instantaneous envelope of the transient traveling wave of the two monitoring points, and the basic electrical quantity feature;

[0023] For the correlation entropy feature, monitoring points A and B are selected, and the time sub-window length is set to twice the main oscillation period, i.e., 0.4 ms. Within this sub-window, the synchronous transient voltage sequence of monitoring point A is subtracted point by point from the synchronous transient voltage sequence of monitoring point B to obtain a synchronous differential sequence. Then, the sample entropy is calculated on this differential sequence to obtain a correlation entropy feature value. More specifically, the correlation entropy feature calculated based on the synchronous differential sequence of any two monitoring points within a time sub-window with a length that is a preset multiple of the main oscillation period includes:

[0024] The sequence of synchronous measurement data from two monitoring points within a time sub-window with a length of 3 times the main oscillation period is extracted; the difference between the two sequences at each sampling time is calculated to obtain the synchronous differential sequence; the sample entropy algorithm is used, with a similarity tolerance of 0.2 times the standard deviation of the synchronous differential sequence, and the entropy value is calculated as the correlation entropy feature.

[0025] Assuming the determined main oscillation period is 0.5s and the preset multiplier is 3, the length of the time sub-window is 1.5s. Voltage measurement data of length 1.5s are simultaneously extracted from monitoring points A and B. If the sampling frequency is 1000Hz, each will yield a sequence containing 1500 data points. The values ​​of the two sequences at each sampling time are subtracted to generate a new synchronous differential sequence of length 1500, reflecting the voltage difference between the two points. The sample entropy algorithm is applied to the synchronous differential sequence to represent its complexity. The standard deviation of the differential sequence is calculated, assumed to be 0.05. The similarity tolerance r is set to 0.01. The embedding dimension m is 2. The logarithms of the two-dimensional and three-dimensional vectors with a distance less than 0.01 in the sequence are counted to calculate an entropy value, for example, 1.8, which represents the correlation entropy characteristic between the two monitoring points. A higher value indicates a more complex and irregular relationship between the two points.

[0026] For the phase-amplitude asymmetry characteristic, the instantaneous phase difference is obtained by subtracting the instantaneous phases of the corresponding intrinsic mode functions at monitoring points A and B. Simultaneously, the instantaneous envelopes are multiplied to obtain the instantaneous envelope product. The instantaneous phase difference sequence and the instantaneous envelope product sequence are then multiplied point-by-point, and the mean is calculated to obtain the phase-amplitude asymmetry characteristic value. For example, applying the Hilbert transform to the transient current measurement data of monitoring points A and B yields their respective instantaneous phase and instantaneous envelope sequences, each with the same length as the original data. For instance, at a certain time t, the instantaneous phase at monitoring point A is 1.5 radians, and the instantaneous envelope is 1.2 units; the instantaneous phase at monitoring point B is 1.1 radians, and the instantaneous envelope is 1.4 units. The instantaneous phase difference at this time is calculated as 0.4 radians. Simultaneously, the average value of the instantaneous envelope is calculated as 1.3 units. Multiplying the instantaneous phase difference of 0.4 by the instantaneous envelope average of 1.3 yields a phase-amplitude product value of 0.52 at that moment. This calculation process is repeated for each sampling point within the entire time window, generating a completely new phase-amplitude product sequence. The arithmetic mean of all values ​​in this phase-amplitude product sequence is calculated; assuming the result is 0.25, this value 0.25 represents the phase-amplitude asymmetry characteristic between the two monitoring points.

[0027] The basic electrical quantity characteristics include the effective values ​​of voltage and current, active power, reactive power, and three-phase imbalance calculated within a preset time window. In another embodiment, the basic electrical quantity characteristics include: calculating the effective value, peak value, waveform factor, and kurtosis factor of voltage and current at each monitoring point within the preset time window for each monitoring point's voltage and current measurement data, and using the calculated values ​​as the basic electrical quantity characteristics. Electrical data from a monitoring point is selected, for example, voltage and current waveform data collected within a preset time window of 1 second. The voltage data is processed, and the effective value is calculated to be 220V, i.e., the root mean square value. The maximum absolute value of the voltage waveform is found as the peak value, for example, 310V. The waveform factor is calculated, i.e., the effective value of 220V divided by the average absolute value of the waveform, resulting in a ratio, for example, 1.12. The kurtosis factor is calculated, which represents the sharpness of the waveform distribution, and may yield a value such as 3.5. After completing the voltage data calculation, the exact same calculation process is repeated for the current measurement data within the same time window. The effective value of the current, for example, 5A; the peak value of the current, for example, 8A; the waveform factor of the current, for example, 1.20; and the kurtosis factor of the current, for example, 4.2, are obtained. For this monitoring point, a total of eight numerical characteristics are obtained, namely the effective value, peak value, waveform factor, and kurtosis factor of the voltage and current respectively. These values ​​together constitute the basic electrical quantity characteristic set of this monitoring point.

[0028] By concatenating the three types of features above in sequence, a one-dimensional combined feature vector is formed, such as... Figure 3 .

[0029] Step 3: Construct a gradient boosting decision tree model that integrates feature attention and tree attention. Sum the predictions of all base learners according to their corresponding tree attention weights to obtain the model output.

[0030] More specifically, the training process of the gradient boosting decision tree model that integrates feature attention and tree attention includes the following steps during the iterative construction of the base learner:

[0031] a) Feature Attention: Calculate the cumulative split gain of each feature in all base learners built up to the current iteration, and calculate the weight of each feature based on the cumulative split gain of each feature; when building a new base learner to split nodes, use the weight of the feature to weight the gradient and second gradient used to calculate the split gain.

[0032] In the t-th iteration, i.e., when constructing the t-th base learner (i.e., decision tree), we obtain the t-1 trees that have already been constructed. Assume that multiple features such as association entropy, phase-amplitude asymmetry, and voltage RMS value are used, and calculate the total gain contributed by each feature to all split nodes in the t-1 trees. For example, suppose the association entropy feature, due to its good discriminative effect, has been frequently used as a split node in past trees, with a cumulative split gain of 150; while the voltage RMS feature has a mediocre effect, with a cumulative gain of only 20. Based on the cumulative gain, a weight is calculated for each feature; for example, after normalization, the weight of association entropy is 0.7, and the weight of voltage RMS is 0.1. When constructing the t-th tree to find the optimal split point, if we want to evaluate the effect of splitting using association entropy, the model will multiply its corresponding weight of 0.7 by the gradient and second gradient of each sample, amplifying the calculated split gain; conversely, when evaluating voltage RMS, it will multiply by 0.1, thus reducing its split gain.

[0033] b) Tree attention: Based on the prediction results of the previous iteration, training samples with prediction confidence lower than the preset quantile threshold are selected as hard samples; based on the loss function values ​​of all the built base learners on the hard samples, tree attention weights are calculated for each base learner.

[0034] After the (t-1)th iteration, the model predicts all training samples, obtaining the predicted probability (confidence score) of each sample belonging to the true class. All confidence scores are sorted from low to high, and a threshold of, for example, the 20th percentile (0.5) is used. Samples with confidence scores below 0.5 are identified as hard samples. For each tree in the constructed (t-1) base learners, its average cross-entropy loss on the hard samples is calculated individually. For example, the loss of the first tree on hard samples is 0.3, and that of the second tree is 0.1. Based on the reciprocal of the loss and normalization, a tree attention weight is assigned to each tree; the smaller the loss, the higher the weight. For example, the weight of the first tree is 0.25, and that of the second tree is 0.75. When a test sample is input into the model, the (t-1) base learners output their predictions, resulting in a probability vector. The probability vector output by the i-th tree is multiplied by its corresponding tree attention weight, and all weighted probability vectors are summed element-wise to obtain the combined prediction probability vector.

[0035] In an optional embodiment, the step of weighting the gradient and second-order gradient used to calculate the split gain using the weights of the features when constructing a new base learner for node splitting includes:

[0036] For each sample i in a node to be split, the original gradient is: The second gradient is If the weights of candidate feature j used to evaluate splitting are... Then the gradient of sample i will be updated. The second gradient is updated to Use the updated gradient and Calculate the splitting gain of candidate feature j.

[0037] In constructing a decision tree, it is assumed that a node containing multiple samples needs to be split. For sample i in the node, the original gradient is calculated based on the current model error. -0.5, second gradient The value is 0.25. Consider using candidate feature j, i.e., the effective voltage value, for splitting. This feature j is pre-assigned a weight. Let's assume the value is 1.2.

[0038] To evaluate the merits of feature j as a split point, the gradient of sample i is temporarily adjusted. The updated gradient... =-0.6. Updated second gradient. =0.3. The weighting process is applied to all samples within a node, but only when evaluating feature j. The algorithm uses the updated gradient. and second gradient This is used to calculate the gain gained from splitting using feature j. In this way, the gradient value of a feature with a higher weight is amplified when calculating the split gain, thus making important features more likely to be selected for building the decision tree.

[0039] In an optional embodiment, the step of selecting training samples with prediction confidence levels lower than a preset quantile threshold as hard samples based on the prediction results of the previous iteration includes:

[0040] For all training samples in the previous iteration, the probability of predicting the true class is obtained as the prediction confidence score. All prediction confidence scores are sorted in ascending order. The confidence score value at the 20th quantile position after sorting is selected as the confidence quantile threshold. Samples with prediction confidence scores less than the threshold are considered hard samples. Assuming there are 1000 samples in the training set, after the model completes one training iteration, the current model is used to predict each of the 1000 samples. For each sample, the probability of the model predicting the true label is examined. For example, for a sample with a true label of "fault," the model predicts the probability of the sample being faulty is 0.35; for another sample with a true label of "normal," the model predicts the probability of the sample being normal is 0.98. These 1000 probability values ​​are collected to form a confidence score list. The list containing the 1000 confidence scores is sorted from low to high. After sorting, the value at the 20th quantile position is found. For 1000 samples, this position is the 200th value. Assuming the 200th confidence level after sorting is 0.62, then 0.62 is set as the confidence quantile threshold. Any sample with a prediction confidence level below 0.62 is defined as a hard sample. For example, the sample with a confidence level of 0.35 mentioned earlier would be identified as a hard sample, while the sample with a confidence level of 0.98 would not.

[0041] In an optional embodiment, calculating tree attention weights for each base learner based on the loss function values ​​of all constructed base learners on the hard samples includes:

[0042] For the k-th base learner, calculate the average of the cross-entropy loss function of the base learner on all hard samples. The negative value of the average loss over all base learners - The Softmax function is applied for normalization to obtain the normalized weights. The weight This refers to the tree attention weights of the k-th base learner.

[0043] Assume the model consists of 3 base learners, i.e., 3 decision trees, and 50 hard samples have already been identified. Evaluate the performance of each tree on these hard samples sequentially. For the first tree, make predictions on the 50 hard samples and calculate the cross-entropy loss for each prediction. Sum these 50 loss values ​​and average them to obtain the average loss. The average loss is 0.9. The average losses of the second and third trees on the hard sample were calculated to be... =1.5 and =0.7. After obtaining the average loss for each tree, negative values ​​are taken: -0.9, -1.5, and -0.7. These three values ​​are then input into the Softmax function. After calculation, the result is... =0.35, =0.19, =0.46. The three weight values ​​are summed to 1, with the third tree having the highest attention weight due to its lowest loss on hard samples. These weights will guide the model in subsequent updates to focus more on base learners that perform better on hard samples.

[0044] Step 4: Use the model to infer the combined feature vectors extracted from the real-time measurement data, and output the fault section with the highest probability as the location result.

[0045] When an actual fault occurs on the line, the real-time monitoring system detects the fault data and immediately executes the entire process described above: synchronization, normalization, transient component extraction, Hilbert-Huang transform, and combined feature vector construction. This generates a feature vector with the same format as during training. This vector is then input into a pre-trained gradient boosting decision tree attention model. The model outputs a probability list covering all possible fault segments; for example, segment 1 has a probability of 0.05, segment 2 has a probability of 0.92, and segment 3 has a probability of 0.03. Segment 2, with the highest probability value, is selected as the fault location result. Figure 4 And report it to the dispatch center.

[0046] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0047] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for locating faulty sections of customer lines, characterized in that, Includes the following steps: The system acquires voltage and current measurement data from multiple monitoring points on the customer's line within a preset time window, and performs clock synchronization and normalization processing. It then performs Hilbert-Huang transform on the transient components of each monitoring point to obtain the instantaneous frequency and instantaneous envelope of the transient traveling wave, and determines the main oscillation period based on the energy spectrum of the instantaneous frequency. Construct a combined feature vector, including the correlation entropy feature calculated based on the synchronous differential sequence of any two monitoring points within a time sub-window with a length that is a preset multiple of the main oscillation period, the phase amplitude asymmetry feature calculated based on the product of the instantaneous phase difference and the instantaneous envelope of the transient traveling wave of the two monitoring points, and the basic electrical quantity feature; Construct a gradient boosting decision tree model that integrates feature attention and tree attention, and sum the predictions of all base learners according to their corresponding tree attention weights to obtain the model output; The model is used to infer the combined feature vectors extracted from real-time measurement data, and the fault section with the highest probability is output as the location result. The training process of the gradient boosting decision tree model that integrates feature attention and tree attention includes the following steps during the iterative construction of the base learner: a) Feature Attention: Calculate the cumulative split gain of each feature in all base learners built up to the current iteration, and calculate the weight of each feature based on the cumulative split gain of each feature; when building a new base learner to split nodes, use the weight of the feature to weight the gradient and second gradient used to calculate the split gain. b) Tree attention: Based on the prediction results of the previous iteration, select training samples with prediction confidence lower than the preset quantile threshold as hard samples; calculate tree attention weights for each base learner based on the loss function values ​​of all constructed base learners on the hard samples. The correlation entropy feature calculated based on the synchronous differential sequence of any two monitoring points within a time sub-window whose length is a preset multiple of the main oscillation period includes: The sequence of synchronous measurement data from two monitoring points within a time sub-window with a length of 3 times the main oscillation period is extracted; the difference between the two sequences at each sampling time is calculated to obtain the synchronous differential sequence; the sample entropy algorithm is used, with a similarity tolerance of 0.2 times the standard deviation of the synchronous differential sequence, and the entropy value is calculated as the correlation entropy feature.

2. The method according to claim 1, characterized in that, The determination of the main oscillation period based on the energy spectrum of the instantaneous frequency includes: Calculate the fast Fourier transform of the instantaneous frequency sequence to obtain the energy spectral density function; search for the peak point with the largest energy value in the energy spectral density function, and take the frequency corresponding to the peak point as the dominant frequency; take the reciprocal of the dominant frequency as the main oscillation period.

3. The method according to claim 1, characterized in that, The phase-amplitude asymmetry characteristic calculated based on the product of the instantaneous phase difference and the instantaneous envelope of the transient traveling wave at the two monitoring points includes: Acquire the instantaneous phase sequence and instantaneous envelope sequence of two monitoring points; calculate the difference between the two instantaneous phase sequences to obtain the instantaneous phase difference sequence; multiply the instantaneous phase difference sequence with the average sequence of the instantaneous envelope sequences of the two monitoring points point by point to obtain the phase-amplitude product sequence; calculate the arithmetic mean of the phase-amplitude product sequence over the entire time window as the phase-amplitude asymmetry feature.

4. The method according to claim 1, characterized in that, The basic electrical quantity characteristics include: For the voltage and current measurement data of each monitoring point, calculate the effective value, peak value, waveform factor and kurtosis factor of the voltage and current of each monitoring point within the preset time window, and use the calculated values ​​as the basic electrical quantity characteristics.

5. The method according to claim 1, characterized in that, When constructing a new base learner for node splitting, the gradient and second-order gradient used to calculate the splitting gain are weighted using the weights of the features, including: For each sample i in a node to be split, the original gradient is: The second gradient is If the weights of candidate feature j used to evaluate splitting are... Then the gradient of sample i will be updated. The second gradient is updated to Use the updated gradient and Calculate the splitting gain of candidate feature j.

6. The method according to claim 1, characterized in that, The step of selecting training samples with prediction confidence levels below a preset quantile threshold as hard samples based on the prediction results of the previous iteration includes: For all training samples in the previous iteration, obtain the probability of being predicted as the true class as the prediction confidence; sort all prediction confidence in ascending order; select the confidence value at the 20th percentile position after sorting as the confidence quantile threshold; and treat samples with prediction confidence less than the threshold as hard samples.

7. The method according to claim 1, characterized in that, The tree attention weights are calculated for each base learner based on the loss function values ​​of all constructed base learners on the hard samples, including: For the k-th base learner, calculate the average of the cross-entropy loss function of the base learner on all hard samples. The negative value of the average loss over all base learners - The Softmax function is applied for normalization to obtain the normalized weights. The weight This refers to the tree attention weights of the k-th base learner.

8. A customer line fault section location system, characterized in that, Includes the following modules: The acquisition module is used to acquire voltage and current measurement data from multiple monitoring points on the customer line within a preset time window, and perform clock synchronization and normalization processing; it performs Hilbert-Huang transform on the transient components of each monitoring point to obtain the instantaneous frequency and instantaneous envelope of the transient traveling wave, and determines the main oscillation period based on the energy spectrum of the instantaneous frequency. The feature extraction module is used to construct a combined feature vector, including the correlation entropy feature calculated based on the synchronous differential sequence of any two monitoring points within a time sub-window with a length that is a preset multiple of the main oscillation period, the phase amplitude asymmetry feature calculated based on the product of the instantaneous phase difference and the instantaneous envelope of the transient traveling wave of the two monitoring points, and the basic electrical quantity feature. The construction and training module is used to build a gradient boosting decision tree model that integrates feature attention and tree attention. It calculates the weighted sum of the predictions of all base learners according to their corresponding tree attention weights to obtain the model output. The inference module is used to infer the combined feature vectors extracted from real-time measurement data using the model, and output the fault section with the highest probability as the location result. The training process of the gradient boosting decision tree model that integrates feature attention and tree attention includes the following steps during the iterative construction of the base learner: a) Feature Attention: Calculate the cumulative split gain of each feature in all base learners built up to the current iteration, and calculate the weight of each feature based on the cumulative split gain of each feature; when building a new base learner to split nodes, use the weight of the feature to weight the gradient and second gradient used to calculate the split gain. b) Tree attention: Based on the prediction results of the previous iteration, select training samples with prediction confidence lower than the preset quantile threshold as hard samples; calculate tree attention weights for each base learner based on the loss function values ​​of all constructed base learners on the hard samples. The correlation entropy feature calculated based on the synchronous differential sequence of any two monitoring points within a time sub-window whose length is a preset multiple of the main oscillation period includes: The sequence of synchronous measurement data from two monitoring points within a time sub-window with a length of 3 times the main oscillation period is extracted; the difference between the two sequences at each sampling time is calculated to obtain the synchronous differential sequence; the sample entropy algorithm is used, with a similarity tolerance of 0.2 times the standard deviation of the synchronous differential sequence, and the entropy value is calculated as the correlation entropy feature.

Images