A partial discharge detection method for distribution network with branched overhead line

Abstract

The application relates to the technical field of line discharge monitoring, in particular to a distribution network partial abnormal discharge detection method for overhead line with branches, which comprises the following steps: installing detection units at overhead line main line section nodes and each branch line head-end tower; calibrating the time axis of all detection units; performing frequency band decomposition and adaptive noise reduction on waveform data collected by each detection unit; performing dynamic time window sliding scanning on the noise-reduced waveform data, calculating the response value of the pulse feature based on the deviation between the waveform feature and the historical reference, dynamically adjusting the time window length according to the response value, and intercepting the partial discharge pulse; determining the specific line category to which the partial discharge pulse belongs; performing reverse sequence simulation on the classified partial discharge pulse to lock the discharge source interval, identifying the defect type, and calculating the discharge source position along the line. The application can improve the distribution network partial discharge fault positioning and type identification capability.

Description

Technical Field

[0001] This application relates to the field of line discharge monitoring technology, specifically to a method for detecting partial abnormal discharge in a distribution network with a branch support. Background Technology

[0002] Overhead power distribution lines typically employ a radial topology with a main trunk and multiple branches, covering complex areas such as farmland, forests, and villages. Due to environmental erosion, external impacts, and insulation aging, these lines are prone to partial discharge defects.

[0003] However, in multi-branch support aerial line scenarios, traditional partial discharge detection methods have limited coverage and are difficult to capture partial discharges inside equipment such as transformers and circuit breakers. Furthermore, the multi-branch structure causes reflections and waveform distortions in the traveling wave signals of partial discharges, making it difficult for traditional partial discharge detection methods to determine whether the partial discharge signal originates from the main line or a branch line. This leads to problems such as missed detections, misjudgments, and large positioning errors in partial discharge locations. In addition, traditional partial discharge detection methods are unable to separate multi-source discharge signals and determine the defect type, failing to meet the needs for efficient and accurate detection of multi-branch support aerial lines. Summary of the Invention

[0004] To address the aforementioned issues, a method for detecting partial abnormal discharge in distribution networks with branch supports and open lines is provided, which can improve the ability to locate and identify partial discharge faults in distribution networks.

[0005] The technical solution adopted in this application is as follows:

[0006] One embodiment of this application provides a method for detecting partial abnormal discharge in a distribution network with a branch support and an empty line. The method includes the following steps:

[0007] Detection units are installed at the segment nodes of the main overhead line and at the beginning of each branch line tower; standard electromagnetic synchronization pulses are sent to the entire line through the detection unit at the beginning of the main line, the traveling wave transmission delay is calculated based on the pulse reception time recorded by each detection unit, and the time axis of all detection units is calibrated based on the transmission time.

[0008] Frequency band decomposition and adaptive noise reduction are performed on the waveform data collected by each detection unit; dynamic time window sliding scanning is used on the noise-reduced waveform data, and the response value of the pulse feature is calculated based on the deviation between the waveform features and the historical benchmark. The time window length is dynamically adjusted according to the response value to extract the partial discharge pulse.

[0009] The time-frequency multidimensional features of partial discharge pulses are extracted, and the feature difference degree is calculated by comparing them with the preset line standard features to determine the specific line category to which the partial discharge pulse belongs. Electromagnetic time inversion technology is used to perform reverse simulation on the classified partial discharge pulses to lock the discharge power source range, and the defect type is identified by a deep learning classification model. Based on the aligned time reference, the position of the discharge power source along the line is calculated using the double-ended traveling wave time difference.

[0010] Preferably, calibrating the time axis of all detection units based on the transmission time includes:

[0011] Record the global zero point time when the detection unit at the beginning of the main line emits the standard electromagnetic synchronization pulse, and the time when the other detection units receive the standard electromagnetic synchronization pulse. Use the time interval between the receiving time and the global zero point time as the traveling wave transmission delay.

[0012] The local time axis of each of the other detection units is shifted forward by using the traveling wave propagation delay as compensation.

[0013] Preferably, the response value for calculating the pulse characteristics includes:

[0014] Calculate various characteristic parameters of waveform data within the current time window;

[0015] Calculate the dispersion of a single characteristic parameter within all time windows under historical partial discharge-free operating conditions, and use the normalized result of the reciprocal of the dispersion as the stability judgment response weight of the single characteristic parameter within the current time window.

[0016] Extract the maximum value of a single feature parameter within all time windows under historical partial discharge-free operating conditions;

[0017] Compare the various characteristic parameters of the pulse data within the current time window with their corresponding maximum values; based on the comparison results and the stability judgment response weights of various characteristic parameters within the current time window, obtain the response value of the pulse characteristics of the pulse data within the current time window.

[0018] Preferably, the method for calculating the response value is as follows:

[0019] Calculate the ratio of various feature parameters of the pulse data within the current time window to their corresponding maximum values, and use the weighted sum of the ratios of all feature parameters of the pulse data within the current time window as the response value of the pulse feature of the pulse data within the current time window, wherein the weight of the ratio of a single feature parameter of the pulse data within the current time window is the stability judgment response weight.

[0020] Preferably, the step of dynamically adjusting the time window length based on the response value includes:

[0021] If the response values ​​calculated within a consecutive preset number of time windows do not exceed the preset limit, the length of the dynamic time window and the sliding step size will be increased synchronously.

[0022] Preferably, the step of dynamically adjusting the time window length based on the response value further includes:

[0023] If the response value calculated within the current time window exceeds the preset limit, the length of the current time window is recalculated based on the normalized value of the response value calculated within the current time window.

[0024] Preferably, the process of recalculating the length of the current time window is as follows:

[0025] Calculate the difference between the length of the current time window and the preset minimum time window length;

[0026] The product of the normalized value of the response value calculated within the current time window and the difference is used as the length adjustment amount of the current time window.

[0027] The length of the current time window is calculated by subtracting the length adjustment amount from the current time window length, and this result is used as the recalculated length of the current time window.

[0028] Preferably, the extraction process of the time-frequency multidimensional features is as follows:

[0029] The time spectrum of each group of partial discharge pulses intercepted by each detection unit is obtained respectively;

[0030] The maximum frequency domain peak value, spectral bandwidth, frequency centroid, and time domain duration are extracted from the time spectrum and used to construct the pulse feature vector of each set of partial discharge pulses intercepted by each detection unit.

[0031] Preferably, the calculation process for the feature difference degree is as follows:

[0032] Standard partial discharge pulses were artificially injected into the main line and each branch line without partial discharge defects in advance to establish the standard pulse characteristic vectors of the standard partial discharge pulses of the main line and each branch line respectively.

[0033] The Euclidean distance between the extracted pulse feature vector and the standard pulse feature vector is used as the feature difference.

[0034] Preferably, the method for determining the specific circuit category to which the partial discharge pulse belongs is as follows:

[0035] The line category corresponding to the minimum characteristic difference value is taken as the specific line category to which the partial discharge pulse belongs.

[0036] This application has at least the following beneficial effects:

[0037] By installing detection units at the segment nodes of the main line and the towers at the beginning of each branch line, the main line and all branch lines can be covered synchronously, expanding the monitoring range of partial discharge and solving the problem of limited coverage of traditional methods. By sending standard electromagnetic synchronization pulses from the detection unit at the beginning of the main line to the entire line, and then calibrating the time axis of all detection units based on the transmission time, local time deviations of each detection unit can be eliminated, providing an aligned time reference for the accurate calculation of the subsequent double-ended traveling wave time difference.

[0038] By performing frequency band decomposition and adaptive noise reduction on the waveform data collected by each detection unit, the mixed signal can be separated according to its frequency characteristics, effectively filtering out noise interference and preserving the weak partial discharge waveform characteristics. By employing dynamic time window sliding scanning on the noise-reduced waveform data, the response value of the pulse characteristics is calculated based on the deviation between the waveform characteristics and historical benchmarks, and the time window length is dynamically adjusted according to the response value. This avoids pulse truncation or feature loss caused by a fixed window, and accurately captures the partial discharge pulse.

[0039] By extracting the time-frequency multidimensional features of partial discharge pulses and calculating the feature difference with preset line standard features, the specific line category to which the partial discharge pulse belongs can be determined. This effectively distinguishes whether the partial discharge pulse originates from the main line or a branch line, solving the problem of line attribution determination caused by downlink reflections and waveform distortions in multi-branch structures. By using electromagnetic time inversion technology to perform reverse simulation on the classified partial discharge pulses to pinpoint the discharge source range, the problem of positioning ambiguity caused by downlink wave diffusion in multi-branch structures can be overcome. Furthermore, the identification of defect types through a deep learning classification model helps in identifying partial discharge defects inside equipment. Based on an aligned time reference, the location of the discharge source along the line is calculated using the time difference of the traveling wave at both ends. This enables precise positioning of the discharge source within the determined specific line category, significantly reducing positioning errors and improving the ability to locate and identify partial discharge faults in the distribution network. Attached Figure Description

[0040] Figure 1 A flowchart illustrating the steps of a method for detecting partial abnormal discharge in a distribution network with a branch support for an open line, as provided in this application;

[0041] Figure 2 A schematic diagram of the distributed detection unit deployment;

[0042] Figure 3 A schematic diagram illustrating the precise location of a localized abnormal discharge. Detailed Implementation

[0043] The following description, in conjunction with the accompanying drawings, details the specific scheme of the method for detecting partial abnormal discharge in a distribution network with branch supports provided in this application.

[0044] Figure 1 The flowchart of a method for detecting partial abnormal discharge in a distribution network with a branch support provided in this application includes the following steps:

[0045] Step 1: Install detection units at the segment nodes of the main overhead line and at the beginning of each branch line tower.

[0046] The topology of the actual overhead lines under inspection was investigated, and the start and end points of the main line, segment nodes, and the locations of the first-level and second-level branch line towers were marked. Inspection units were deployed along the main line at 1.2km intervals; similar inspection units were also installed at the beginning towers of all first-level and second-level branch lines. Each inspection unit included an open-type high-frequency current transformer, a filter amplifier circuit, a data acquisition card, a pulse synchronization device, and a GPS (Global Positioning System) timing module. The high-frequency current transformer was fixed to the outside of the overhead phase line using a snap-fit ​​method, with a bandwidth of 0.5MHz to 10MHz; the high-frequency current transformer was connected to a filter amplifier circuit with a bandwidth of 0.5MHz to 5MHz and a gain of 40dB; the filter amplifier circuit was connected to the data acquisition card at the rear end, with a sampling frequency of 500MHz, two channels, and a storage depth of 500k; the data acquisition card was communicatively connected to the pulse synchronization device. All inspection units were connected to the same wireless local area network to form a traveling wave signal inspection network. Figure 2 A schematic diagram of the deployment of distributed detection units.

[0047] Step 2: The detection unit at the beginning of the main line sends standard electromagnetic synchronization pulses to the entire line. The traveling wave transmission delay is calculated based on the pulse reception time recorded by each detection unit, and the time axis of all detection units is calibrated based on the transmission time.

[0048] Each detection unit receives GPS timing signals to achieve initial synchronization of the local time across the entire line. Commands are issued from the central control unit to the first detection unit at the beginning of the main line to generate a standard electromagnetic synchronization pulse, which is then propagated synchronously to all detection units along the line in the form of a traveling wave via overhead conductors. The pulse transmission time of the first detection unit at the beginning of the main line is recorded. And obtain the pulse reception time of each of the remaining detection units; and set the pulse reception time of the i-th detection unit among the remaining detection units. With pulse transmission time The time interval is used as the propagation delay of the traveling wave from the first detection unit at the beginning to the i-th detection unit among the remaining detection units. If the line distance L between the first detection unit at the beginning of the main line and the i-th detection unit among the remaining detection units is known, it can be calculated using the formula... Calculate the traveling wave propagation velocity v for subsequent fault location calculation and analysis.

[0049] Based on the calculated traveling wave propagation delay, clock synchronization calibration is performed on each detection unit. The specific process is as follows: using the pulse transmission time of the first detection unit at the head end... As the global zero point, the local time axis of the remaining detection units is shifted forward by the corresponding traveling wave transmission delay, that is, the pulse reception time recorded by the remaining detection units is subtracted from the traveling wave transmission delay. Through the translation of the local time axis, the time reference of all remaining detection units is completely aligned with the time reference of the first detection unit at the beginning of the main line, thus completing the clock calibration of all main and branch line detection units.

[0050] Step 3: Perform frequency band decomposition and adaptive noise reduction on the waveform data collected by each detection unit.

[0051] After all detection units complete clock calibration, all detection units are controlled to synchronously acquire raw time-domain waveform data. Since multi-channel power lines are often distributed in complex environments such as suburbs, farmland, forests, and areas surrounding villages and towns, the acquired waveform data may be mixed with periodic harmonic noise, spatial radio frequency white noise, and transient lightning random pulse interference. This can mask the effective partial discharge signal during the partial discharge anomaly detection process of multi-channel power lines. Therefore, this application uses the SGWT (Spectral Graph Wavelet Transform) algorithm to perform frequency band decomposition on the waveform data acquired by each detection unit, dividing it into multiple independent frequency band intervals. This separates the mixed signal according to frequency characteristics, achieving preliminary separation of noise signals and effective partial discharge signals in the frequency domain, laying the foundation for subsequent targeted filtering of various interferences.

[0052] Because the signals from partial discharge anomalies in power distribution lines are typically weak in amplitude and concentrated in frequency band, after frequency band decomposition, the adaptive threshold noise reduction module of the SGWT algorithm is used to perform noise reduction processing on the waveform data of each independent frequency band, as follows:

[0053] (1) Select the db5 wavelet basis function that is adapted to the transient characteristics of partial discharge pulse;

[0054] (2) The adaptive threshold of each frequency band is dynamically calculated by the adaptive threshold noise reduction module; the wavelet coefficients of each frequency band are judged, the noise coefficients below the threshold are set to zero, and the effective discharge signal coefficients above the threshold are retained.

[0055] The SGWT algorithm is used to denoise the waveform data collected by each detection unit, enabling layered filtering of periodic harmonic noise, spatial radio frequency white noise, and transient lightning random pulse interference. This fully preserves the weak partial discharge waveform and its time-domain and frequency-domain characteristics, thereby improving the accuracy of partial discharge signal extraction in complex electromagnetic environments and ensuring the overall accuracy of subsequent pulse identification, signal separation, fault location, and defect type identification. The process of wavelet filtering using the SGWT algorithm is well-known to those skilled in the art, and its specific details will not be described in this application.

[0056] Step 4: Apply dynamic time window sliding scan to the denoised waveform data. Calculate the response value of the pulse characteristics based on the deviation between the waveform characteristics and the historical benchmark. Dynamically adjust the time window length according to the response value to extract the partial discharge pulse.

[0057] After step 3, the denoised waveform data of each detection unit is obtained. Due to line impedance loss, branch node reflections, residual random interference, and signal waveform distortion during actual transmission, there are significant deviations in the pulse characteristic judgment of the waveform data acquired by each detection unit. This affects the accuracy of judging the pulse arrival time, duration, and transmission polarity, leading to errors in identifying the travel time difference of partial discharge at both ends, incorrect fault line division, and decreased defect location accuracy. Therefore, it is necessary to perform segmented analysis on the denoised waveform data. By analyzing the characteristic differences between the signals in each time period and historical normal operating condition benchmarks, the time period of the partial discharge pulse can be determined and located to accurately pinpoint the effective pulse location, reduce residual pseudo-pulse interference, and accurately extract core characteristic parameters such as pulse arrival time, duration, and transmission polarity. This provides a reliable data foundation for subsequent multi-source signal separation and fault location.

[0058] Furthermore, since partial discharge pulses may exhibit varying widths, amplitude fluctuations, and random distributions in actual waveforms, segmenting them using a fixed-length time window can easily lead to pulse truncation or feature loss. Therefore, this application employs a dynamic time window to analyze the interval variation characteristics of the waveform data and adjusts the window length in real time based on the analysis results. The specific process is as follows:

[0059] (1) The basic length of the time window is set according to the typical time-domain parameters of the partial discharge pulse in the distribution network. Considering that the rise time of the partial discharge pulse in the distribution network is about several nanoseconds to tens of nanoseconds, and the duration of the complete pulse is concentrated in the range of 10μs to 50μs, this application sets the basic length of the time window to 20μs and the moving step size to 5μs. During the sliding division process, the length of the time window is dynamically adjusted according to the amplitude and slope change characteristics of the waveform within the time window to avoid the partial discharge pulse being segmented and the loss of features.

[0060] Based on the above analysis, the waveform data is initially divided into time windows of a basic length, and the characteristic parameters of the waveform data within each time window are calculated. These characteristic parameters include the maximum amplitude, average amplitude, waveform rising edge slope, waveform falling edge slope, and signal energy. The specific calculation method for these characteristic parameters is as follows:

[0061] 1) Extract the waveform amplitude of all sampling points within each time window, and calculate the maximum amplitude and average amplitude within each time window;

[0062] 2) Calculate the amplitude difference between adjacent sampling points within each time window, and use the ratio of the amplitude difference to the sampling interval as the instantaneous slope of the waveform. Calculate the maximum slope of the rising phase and the minimum slope of the falling phase within each time window, and record them as the rising edge slope and falling edge slope, respectively. Specifically, if the waveform data within a single time window is monotonically increasing, only the maximum slope of the rising phase is calculated; if the waveform data within a single time window is monotonically decreasing, only the minimum slope of the falling phase is calculated. For slope features not calculated due to the monotonicity of the waveform data, their values ​​are assigned to 0 to complete the feature parameter dimension. Simultaneously, after extracting the maximum slope of the rising phase and the minimum slope of the falling phase, their absolute values ​​are taken, and these absolute values ​​are used as feature parameters in subsequent response value calculations.

[0063] 3) Calculate the signal energy of waveform data within each time window using the signal energy calculation formula.

[0064] (2) Obtain the characteristic parameters of waveform data within each time window under historical partial discharge-free operating conditions, divided by the same base length, and record them as basic characteristic parameters. These parameters are used for comparative analysis with the characteristic parameters of waveform data within the current time window. Specifically, obtain 500 time windows divided by base length under operating conditions without partial discharge and only with conventional environmental noise from the locally stored historical database to calculate the basic characteristic parameters. Here, 500 is merely one embodiment of this application, and the implementer can flexibly set its specific value according to the actual situation. This application does not impose any special limitations on this.

[0065] To avoid the impact of the sensitivity of waveform data to interference under historical partial discharge-free operating conditions on the accuracy of pulse feature comparison, the dispersion of individual feature parameters within all time windows under historical partial discharge-free operating conditions is calculated. This dispersion is used to measure the stability of individual feature parameters. A larger dispersion indicates lower stability of the individual feature parameter's response to changes within the base time window under historical partial discharge-free operating conditions, leading to greater deviation in the analysis and judgment of sudden pulse features. Therefore, in the comprehensive judgment process within the current time window, to avoid judgment errors caused by interference fluctuations during normal operation, the judgment weight assigned to each individual feature parameter should be as small as possible.

[0066] In this embodiment, the dispersion of a single feature parameter is the variance. As for other implementation methods, based on the ability to measure the unevenness of the distribution of a single feature parameter, the implementer may choose other existing feasible techniques, such as standard deviation, coefficient of variation, etc. This application does not impose any special limitations on this.

[0067] Based on the above analysis, the normalized result of the reciprocal of the discreteness of the calculated single feature parameter is used as the stability judgment response weight of the single feature parameter within the current time window.

[0068] Specifically, if the dispersion of a single feature parameter is 0, the sum of the dispersion of the single feature parameter and a preset value greater than 0 is calculated first, and then subsequent calculations are performed. In this embodiment, the preset value greater than 0 is 0.01, and the preset value greater than 0 and the dispersion of the single feature parameter have the same unit. The implementer can flexibly set the value of the preset value greater than 0 according to the actual situation, and this application does not impose any special limitations on this.

[0069] In this embodiment, the reciprocals of all the calculated feature parameters are used as the input of the Softmax function, and the normalized results of the reciprocals of the various feature parameters are output. The Softmax function is a well-known technology, and the specific calculation process will not be described in this application.

[0070] Furthermore, the maximum value of each single feature parameter within all time windows under historical partial discharge-free operating conditions is extracted; the various feature parameters of the pulse data in the current time window are compared with their corresponding maximum values; based on the comparison results and combined with the stability judgment response weights of various feature parameters in the current time window, the response value of the pulse feature of the pulse data in the current time window is obtained, and the calculation formula is as follows:

[0071] In the formula, P represents the response value of the pulse characteristics of the pulse data within the current time window; This represents the i-th characteristic parameter of the pulse data within the current time window; This represents the maximum value of the i-th characteristic parameter within all time windows under historical non-local discharge conditions; The larger the calculation result, the greater the possibility that the waveform data in the current time window deviates from the historical partial discharge-free operating condition. This represents the stable judgment response weight of the i-th feature parameter within the current time window.

[0072] It is understandable that the larger the calculated response value, the greater the probability of a partial discharge pulse occurring within or after the current time window.

[0073] (3) Based on the calculated response value, the presence of partial discharge pulses in the waveform data is determined in real time during the sliding process of the time window, and the length of the time window and the sliding step are dynamically adjusted to prevent the partial discharge pulses from being truncated and to improve the accuracy of pulse feature judgment in the waveform data. The specific process is as follows:

[0074] 1) If the response value calculated within the current time window does not exceed the preset limit, it means that the waveform data within the current time window has not deviated significantly from the historical partial discharge-free operating condition. Therefore, it is determined that the current time window contains only conventional noise, and the length of the current time window is maintained and the sliding division continues.

[0075] In this embodiment, the preset limit is set to 1.

[0076] 2) If the response values ​​calculated within a consecutive preset number of time windows do not exceed the preset limit, it indicates that the waveform data within the current time window has stable response characteristics compared to historical partial discharge-free operating conditions. Considering that the duration of a complete partial discharge pulse is typically concentrated in the range of 10μs to 50μs, this application simultaneously enlarges the length of the time window and the sliding step size to improve the efficiency of comparison and judgment and to preserve the complete pulse to the greatest extent possible during sudden partial discharge pulses. Then, the sliding division continues according to the enlarged time window length and sliding step size, and the response value of the pulse characteristics of the pulse data within the updated time window is calculated in real time. If the response value still does not exceed the preset limit, the enlarged time window length and sliding step size are maintained, thereby accelerating the data scanning and comparison judgment efficiency of the partial discharge-free safe area.

[0077] In this embodiment, the length of the time window is increased to 50 μs, and the sliding step size is increased to 25 μs.

[0078] In this embodiment, the preset quantity is 10. The preset quantity can be flexibly set by the implementer according to the actual situation, and this application does not impose any special limitations on it.

[0079] 3) If the response value calculated within the current time window exceeds the preset limit, it indicates that the waveform data within the current time window deviates significantly from the historical partial discharge-free operating condition. This means that there may be a sudden increase in amplitude and slope, and the signal energy may be significantly higher than the reference value. Therefore, it is determined that a partial discharge pulse is suspected to exist within the current time window. At this time, the larger the calculated response value, the more significant the partial discharge pulse signal characteristics, the more drastic the waveform change, and the smaller the pulse concentration area. Therefore, the time window is dynamically reduced. The specific reduction process is as follows:

[0080] Waveform data within 500 historical time windows suspected of containing partial discharge pulses are acquired, and the response values ​​of the pulse characteristics of the waveform data within these 500 time windows are combined into a response data set. Based on the maximum and minimum values ​​in the response data set, the response values ​​calculated within the current time window are normalized using the Min-Max normalization method.

[0081] Based on the normalized value of the response value calculated within the current time window, the length adjustment of the current time window is calculated using the following formula:

[0082] In the formula, This indicates the adjustment amount for the length of the current time window; p represents the normalized value of the response calculated within the current time window. Indicates the length of the current time window; This indicates the minimum length of the preset time window. Since the duration of a complete pulse of partial discharge usually falls within the range of 10μs to 50μs, the minimum length of the preset time window is set to 10μs.

[0083] Reduce the length of the current time window to The current time window is re-divided according to the shortened time window length, and the response value of the pulse characteristics of the pulse data in the updated time window is calculated in real time. The time window length is re-adjusted based on the calculation results. If the updated response value falls to or below the preset limit during the re-adjustment process, or if the time window length to be adjusted has reached the lower limit standard value of 10μs, the shortening determination is stopped directly.

[0084] The Min-Max normalization method is a well-known technique, and the detailed calculation process will not be described in this application. The value 500 is merely one embodiment of this application; implementers can flexibly set its specific value according to actual circumstances, and this application does not impose any special limitations on it.

[0085] Specifically, if the maximum and minimum values ​​in the response dataset are equal, a preset minimum positive number is used as the denominator in the Min-Max normalization method to avoid the denominator being 0. In this embodiment, the preset minimum positive number is 0.01. Implementers can flexibly set the value of the preset minimum positive number according to the actual situation, and this application does not impose any special restrictions on it.

[0086] Through the aforementioned dynamic adjustment mechanism, the length of the time window is updated in real time during the sliding window scanning of the full-time-domain waveform data. When encountering a pulse-free region, the time window is enlarged to improve scanning efficiency; when encountering a suspected partial discharge pulse region, the time window is adaptively shrunk for refined positioning. When the calculated response value after readjusting the time window length drops to less than or equal to the preset limit, or when the current time window length is reduced to the preset minimum time window length, the dynamic shrinkage adjustment of the time window for the current suspected region is stopped, and a valid partial discharge pulse is truncated according to the currently determined final time window boundary. The above process is repeated until the full-time-domain waveform data scanning and division are completed, thereby avoiding unreasonable truncation or feature loss of the partial discharge pulse waveform.

[0087] (4) The time window corresponding to the response value being greater than the preset limit and the dynamic window adjustment convergence being completed is determined as the time window containing a suspected partial discharge pulse. The waveform data within the time window is then extracted. Specifically, the polarity of the waveform data within each extracted time window is determined, and the pulse polarity is distinguished according to the traveling wave transmission direction. Waveform data whose polarity changes multiple times within the same time window is determined to be a pseudo-pulse caused by random environmental interference and is then removed. The remaining real partial discharge pulse signals after removing pseudo-pulses are marked, and the precise time, peak amplitude, waveform duration, and forward / reverse transmission polarity of the pulse arriving at the corresponding detection unit within each time window are recorded. The pulses are then classified and organized according to the detection unit number to construct a pulse time series dataset containing waveform feature information of all detection units in the entire line.

[0088] Step 5: Extract the time-frequency multidimensional features of the partial discharge pulse, calculate the feature difference degree with the preset line standard features, and determine the specific line category to which the partial discharge pulse belongs.

[0089] In multi-branch support systems with empty lines, the same detection unit may simultaneously acquire multiple partial discharge pulses from the main trunk line, first-level branches, and second-level branches. These multiple partial discharge pulses overlap in the time domain, making it difficult to directly distinguish their line sources. To solve the above-mentioned problem of multi-source signal overlap, the following processing is performed:

[0090] The time spectrum of each group of partial discharge pulses captured by each detection unit in the pulse time series dataset is obtained respectively; various feature values ​​are extracted from the time spectrum, including: maximum frequency domain peak value, spectral bandwidth, frequency centroid, and time domain duration.

[0091] In this embodiment, a short-time Fourier transform is performed on each group of partial discharge pulses captured by each detection unit in the pulse timing dataset to generate a time spectrum. The transform window length is set to 20 μs. The short-time Fourier transform is a well-known technique, and the detailed process will not be described in this application.

[0092] The various feature values ​​of each group of partial discharge pulses captured by each detection unit are normalized. Then, the normalized values ​​of all the feature values ​​of each group of partial discharge pulses captured by each detection unit are used to form the pulse feature vector of each group of partial discharge pulses captured by each detection unit.

[0093] Standard partial discharge pulses are artificially injected into the main line, first-level branch line, and second-level branch line without partial discharge defects in advance, and the waveform data collected by the detection unit closest to the pulse injection point is extracted to establish the standard pulse feature vectors of the standard partial discharge pulses of the main line, first-level branch line, and second-level branch line respectively.

[0094] In this embodiment, the Min-Max normalization method is used to normalize the feature representation values. Taking the normalization process of the maximum frequency domain peak as an example, when normalizing the maximum frequency domain peak, the maximum and minimum values ​​are directly called from the historical global maximum and minimum frequency domain peaks recorded when establishing the historical standard pulse feature vector, so as to ensure that the measured features and the standard features use the same normalization scaling reference scale.

[0095] The Euclidean distance between the pulse feature vector and the standard pulse feature vector is used as the feature difference degree. The line category corresponding to the minimum feature difference degree is identified as the specific line category to which the partial discharge pulse belongs. Specifically: if the feature difference degree between the pulse and the standard pulse feature vector of the main line is the smallest, it is identified as a main line discharge pulse; if the feature difference degree between the pulse and the standard pulse feature vector of the first-level branch line is the smallest, it is identified as a first-level branch discharge pulse; if the feature difference degree between the pulse and the standard pulse feature vector of the second-level branch line is the smallest, it is identified as a second-level branch discharge pulse. In this way, the line attribution of multi-source partial discharge pulses is determined group by group, providing a basis for subsequent group-by-group location of the specific line to which the discharge source belongs.

[0096] Step 6: Use electromagnetic time inversion technology to perform reverse simulation on the classified partial discharge pulses to lock the discharge power range.

[0097] After separation in step 5, although the specific circuit category to which each group of partial discharge pulses belongs has been identified, the following processing is performed to determine the specific location of the power source corresponding to each group of partial discharge pulses:

[0098] All main line segments, first-level branches, and second-level branches are numbered according to the detection unit number. Then, the partial discharge pulses, determined in step 5, are used as inputs. Electromagnetic time inversion technology is used to perform time-domain reverse simulation of the discharge traveling waves of each group of partial discharge pulses, outputting the approximate location area of ​​the discharge source corresponding to each group of partial discharge pulses on the corresponding line, thus clarifying the specific line and section to which each discharge source belongs. For example, after inversion using electromagnetic time inversion technology, the discharge source of a certain group of main line discharge pulses is output as located in the middle section of the main line; after inversion, the discharge source of a certain group of first-level branch 1 discharge pulses is output as located at the beginning of that branch. The electromagnetic time inversion technology is a known technique, and its detailed process will not be described further in this application.

[0099] Step 7: Identify the defect type using a deep learning classification model, and calculate the position of the discharge source along the line based on the aligned time reference using the time difference of the double-ended traveling wave.

[0100] Having determined the specific lines and sections to which each discharge source belongs in step 6, it is still necessary to further clarify the physical defect type corresponding to each discharge source and calculate the precise location of the discharge source. The specific process is as follows:

[0101] (1) Obtain the partial discharge pulse waveforms and time-frequency feature data of four typical defects in historical operation, and construct a defect feature sample library. A deep autoencoder is used to extract the feature depth of each group of partial discharge pulses separated in step 5, and the feature depth extraction results of each group of partial discharge pulses are obtained. The deep autoencoder adopts a three-layer symmetrical network structure, the input layer dimension is consistent with the time-frequency feature data dimension, the intermediate hidden layer is set as a low-dimensional bottleneck structure, the activation function is the ReLU function, and the output layer dimension matches the input layer dimension. The training process adopts the adaptive moment estimation (Adam) optimizer for iterative training, the batch size is set to 32, the number of iterations is set to 200 rounds, the mean squared error (MSE) is used as the loss function, and a random deactivation (Dropout) layer is added with a coefficient of 0.2 to prevent overfitting. The training process of the deep autoencoder is well known to those skilled in the art, and the detailed training process will not be described in detail in this application.

[0102] In this embodiment, the four typical defects specifically include: needle plate defects, bubble defects, suspension defects, and surface defects.

[0103] In this embodiment, the time-frequency feature data includes statistical feature values ​​in the time domain and frequency domain, specifically including variance, root mean square value, peak value, peak-to-peak value, zero-crossing rate, maximum frequency domain peak value, spectral bandwidth, and frequency centroid.

[0104] (2) Based on the feature depth extraction results of each type of partial discharge pulse, the measured features are matched and compared with the defect feature samples by the support vector machine to obtain the defect type corresponding to each discharge source. For example: a discharge source located in the middle section of the main line is identified as insulator surface discharge; a discharge source located in the front section of the No. 1 primary branch is identified as intermittent contact discharge between tree branches and conductors; a discharge source located in the No. 3 primary branch is identified as internal floating discharge of pole-mounted circuit breaker; a discharge source located in the rear section of the No. 2 secondary branch is identified as moisture bubble discharge of surge arrester.

[0105] It should be noted that the kernel function of the above support vector machine is the radial basis function (RBF), and the penalty coefficient C and the kernel function parameter γ are set to C=10 and γ=0.1, respectively. A one-to-one multi-classification strategy is adopted to construct a multi-classification model for four types of defects: needle plate, bubble, suspension, and surface. The classification process of the support vector machine is well known to those skilled in the art, and the detailed process will not be described in this application.

[0106] (3) Using the two-end traveling wave time difference calculation formula, based on the arrival time difference of the partial discharge pulse recorded by the detection units at both ends of the line where the discharge source is located and the line wave velocity, the distance along the line between each discharge source and the nearest detection unit is calculated to complete the fine positioning. It should be noted that: the aligned arrival time of the partial discharge pulse recorded by the detection units at both ends of the line where the discharge source is located is extracted, and the traveling wave transmission delay obtained by the corresponding detection unit in step 2 is added to obtain the real physical time of the partial discharge pulse arriving at the two detection units. The difference between the above real physical times at both ends is taken as the real two-end traveling wave propagation time difference and substituted into the two-end traveling wave time difference calculation formula for calculation. Figure 3 This is a detailed schematic diagram illustrating the precise location of a localized abnormal discharge. On-site maintenance personnel conduct verification and repair based on the located discharge source and defect type, restoring power after repair. The line wave velocity is the traveling wave propagation speed calculated in step 2.

[0107] In summary, by installing detection units at the segment nodes of the main line and the towers at the beginning of each branch line, it is possible to simultaneously cover the main line and all branch lines, expanding the monitoring range of partial discharge and solving the problem of limited coverage of traditional methods. By sending standard electromagnetic synchronization pulses from the detection unit at the beginning of the main line to the entire line, and then calibrating the time axis of all detection units based on the transmission time, the local time deviation of each detection unit can be eliminated, providing an aligned time reference for the accurate calculation of the subsequent double-end traveling wave time difference.

[0108] By performing frequency band decomposition and adaptive noise reduction on the waveform data collected by each detection unit, the mixed signal can be separated according to its frequency characteristics, effectively filtering out noise interference and preserving the weak partial discharge waveform characteristics. By employing dynamic time window sliding scanning on the noise-reduced waveform data, the response value of the pulse characteristics is calculated based on the deviation between the waveform characteristics and historical benchmarks, and the time window length is dynamically adjusted according to the response value. This avoids pulse truncation or feature loss caused by a fixed window, and accurately captures the partial discharge pulse.

[0109] By extracting the time-frequency multidimensional features of partial discharge pulses and calculating the feature difference with preset line standard features, the specific line category to which the partial discharge pulse belongs can be determined. This effectively distinguishes whether the partial discharge pulse originates from the main line or a branch line, solving the problem of line attribution determination caused by downlink reflections and waveform distortions in multi-branch structures. By using electromagnetic time inversion technology to perform reverse simulation on the classified partial discharge pulses to pinpoint the discharge source range, the problem of positioning ambiguity caused by downlink wave diffusion in multi-branch structures can be overcome. Furthermore, the identification of defect types through a deep learning classification model helps in identifying partial discharge defects inside equipment. Based on an aligned time reference, the location of the discharge source along the line is calculated using the time difference of the traveling wave at both ends. This enables precise positioning of the discharge source within the determined specific line category, significantly reducing positioning errors and improving the ability to locate and identify partial discharge faults in the distribution network.

Claims

1. A method for detecting partial abnormal discharge in a distribution network with branch supports, characterized in that, The method includes the following steps: Detection units are installed at the segment nodes of the main overhead line and at the beginning of each branch line tower; standard electromagnetic synchronization pulses are sent to the entire line through the detection unit at the beginning of the main line, the traveling wave transmission delay is calculated based on the pulse reception time recorded by each detection unit, and the time axis of all detection units is calibrated based on the transmission time. Frequency band decomposition and adaptive noise reduction are performed on the waveform data collected by each detection unit; dynamic time window sliding scanning is used on the noise-reduced waveform data, and the response value of the pulse feature is calculated based on the deviation between the waveform features and the historical benchmark. The time window length is dynamically adjusted according to the response value to extract the partial discharge pulse. The time-frequency multidimensional features of partial discharge pulses are extracted, and the feature difference degree is calculated by comparing them with the preset line standard features to determine the specific line category to which the partial discharge pulse belongs. Electromagnetic time inversion technology is used to perform reverse simulation on the classified partial discharge pulses to lock the discharge power source range, and the defect type is identified by a deep learning classification model. Based on the aligned time reference, the position of the discharge power source along the line is calculated using the double-ended traveling wave time difference.

2. The method for detecting partial abnormal discharge in a distribution network with a branch support overhead line as described in claim 1, characterized in that, The calibration of the time axis of all detection units based on the transmission time includes: Record the global zero point time when the detection unit at the beginning of the main line emits the standard electromagnetic synchronization pulse, and the time when the other detection units receive the standard electromagnetic synchronization pulse. Use the time interval between the receiving time and the global zero point time as the traveling wave transmission delay. The local time axis of each of the other detection units is shifted forward by using the traveling wave propagation delay as compensation.

3. The method for detecting partial abnormal discharge in a distribution network with a branch support overhead line as described in claim 1, characterized in that, The response value for calculating the pulse characteristics includes: Calculate various characteristic parameters of waveform data within the current time window; Calculate the dispersion of a single characteristic parameter within all time windows under historical partial discharge-free operating conditions, and use the normalized result of the reciprocal of the dispersion as the stability judgment response weight of the single characteristic parameter within the current time window. Extract the maximum value of a single feature parameter within all time windows under historical partial discharge-free operating conditions; Compare the various characteristic parameters of the pulse data within the current time window with their corresponding maximum values; based on the comparison results and the stability judgment response weights of various characteristic parameters within the current time window, obtain the response value of the pulse characteristics of the pulse data within the current time window.

4. The method for detecting partial abnormal discharge in a distribution network with a branch support overhead line as described in claim 3, characterized in that, The method for calculating the response value is as follows: Calculate the ratio of various feature parameters of the pulse data within the current time window to their corresponding maximum values, and use the weighted sum of the ratios of all feature parameters of the pulse data within the current time window as the response value of the pulse feature of the pulse data within the current time window, wherein the weight of the ratio of a single feature parameter of the pulse data within the current time window is the stability judgment response weight.

5. The method for detecting partial abnormal discharge in a distribution network with a branch support overhead line as described in claim 4, characterized in that, The step of dynamically adjusting the time window length based on the response value includes: If the response values ​​calculated within a consecutive preset number of time windows do not exceed the preset limit, the length of the dynamic time window and the sliding step size will be increased synchronously.

6. The method for detecting partial abnormal discharge in a distribution network with a branch support overhead line as described in claim 5, characterized in that, The method of dynamically adjusting the time window length based on the response value also includes: If the response value calculated within the current time window exceeds the preset limit, the length of the current time window is recalculated based on the normalized value of the response value calculated within the current time window.

7. The method for detecting partial abnormal discharge in a distribution network with a branch support overhead line as described in claim 6, characterized in that, The process of recalculating the length of the current time window is as follows: Calculate the difference between the length of the current time window and the preset minimum time window length; The product of the normalized value of the response value calculated within the current time window and the difference is used as the length adjustment amount of the current time window. The length of the current time window is calculated by subtracting the length adjustment amount from the current time window length, and this result is used as the recalculated length of the current time window.

8. The method for detecting partial abnormal discharge in a distribution network with a branch support overhead line as described in claim 1, characterized in that, The extraction process of the time-frequency multidimensional features is as follows: The time spectrum of each group of partial discharge pulses intercepted by each detection unit is obtained respectively; The maximum frequency domain peak value, spectral bandwidth, frequency centroid, and time domain duration are extracted from the time spectrum and used to construct the pulse feature vector of each set of partial discharge pulses intercepted by each detection unit.

9. The method for detecting partial abnormal discharge in a distribution network with a branch support overhead line as described in claim 8, characterized in that, The calculation process for the feature difference degree is as follows: Standard partial discharge pulses were artificially injected into the main line and each branch line without partial discharge defects in advance to establish the standard pulse characteristic vectors of the standard partial discharge pulses of the main line and each branch line respectively. The Euclidean distance between the extracted pulse feature vector and the standard pulse feature vector is used as the feature difference.

10. The method for detecting partial abnormal discharge in a distribution network with a branch support overhead line as described in claim 1, characterized in that, The method for determining the specific circuit category to which a partial discharge pulse belongs is as follows: The line category corresponding to the minimum characteristic difference value is taken as the specific line category to which the partial discharge pulse belongs.

Images