A power distribution overhead network partial discharge diagnosis method based on multi-terminal data fusion

Abstract

The application is suitable for the technical field of partial discharge diagnosis, and provides a power distribution overhead network partial discharge diagnosis method based on multi-end data fusion, which comprises the following steps: synchronously collecting partial discharge signals through at least two monitoring nodes arranged at different positions of a power distribution overhead network; performing noise suppression processing on the partial discharge signals collected by each monitoring node; extracting index parameters for distinguishing discharge types based on the noise-reduced partial discharge signals; performing fusion analysis on the calculated index parameters to obtain source type, severity and position information of the partial discharge; determining a target discharge diagnosis scheme according to the source type, severity and position information of the partial discharge; and ensuring comprehensive and accurate diagnosis results to provide effective support for operation and maintenance of the partial discharge of the power distribution overhead network.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of partial discharge diagnosis, and particularly relates to a partial discharge diagnosis method for power distribution overhead network based on multi-end data fusion. BACKGROUND

[0002] The partial discharge of the power distribution overhead network refers to the partial corona and arc discharge phenomenon caused by the electric field concentration in the insulation structure of the power distribution overhead line and insulator. The discharge will gradually deteriorate the insulation of the equipment, and if not diagnosed and treated in time, may cause insulation breakdown, line trip and other faults, threatening the safe and stable operation of the power distribution network. The purpose of the partial discharge diagnosis is to find the insulation hidden danger of the equipment in advance, assess the severity and development trend of the fault, and provide data basis for operation and maintenance decision, so as to prevent major power failure, prolong the service life of the equipment and improve the power supply reliability of the power distribution network.

[0003] In the prior art, the partial discharge diagnosis research has limited ability to distinguish complex noise in the noise processing link of the discharge signal, and it is difficult to accurately adapt to the characteristics of various noises, resulting in residual noise in the denoised signal, affecting the accuracy of subsequent discharge type identification, severity assessment and other analyses, and leading to inaccurate partial discharge diagnosis results. SUMMARY

[0004] The present application provides a partial discharge diagnosis method for power distribution overhead network based on multi-end data fusion, which is used to solve the problem of inaccurate partial discharge diagnosis results.

[0005] The present application provides a partial discharge diagnosis method for power distribution overhead network based on multi-end data fusion, which comprises:

[0006] Synchronously collecting the partial discharge signals by at least two monitoring nodes deployed at different positions of the power distribution overhead network;

[0007] Performing noise suppression processing on the partial discharge signals collected by each monitoring node; wherein: based on a plurality of partial discharge signal pulses collected by the monitoring nodes within a continuous time window, calculating the feature parameters of each pulse in time domain and frequency domain, and generating a plurality of pulse feature vector sets; performing clustering analysis on the plurality of pulse feature vector sets, and dividing the pulses into different clusters based on a preset clustering division rule; identifying and filtering out the noise pulse clusters according to the statistical characteristics of each cluster, and obtaining effective discharge pulses;

[0008] Extracting index parameters for distinguishing discharge types based on the denoised partial discharge signals;

[0009] Performing fusion analysis on the calculated index parameters to obtain the source type, severity and position information of the partial discharge;

[0010] Determine a target discharge diagnosis scheme according to the source type, severity and location information of the partial discharge.

[0011] Further, the clustering analysis on the plurality of pulse feature vector sets includes dividing pulses into different clusters based on preset clustering division rules, comprising:

[0012] An initial clustering result of the plurality of pulse feature vector sets is generated by using a density-based clustering algorithm, and the initial clustering result includes an internal discharge set, a surface discharge set and a discrete noise point set;

[0013] Based on the power frequency phase distribution characteristics of the pulses, the internal discharge set is divided into a typical internal discharge cluster and a suspended potential discharge cluster; wherein the pulse phase distribution range of the typical internal discharge cluster conforms to a preset first phase threshold range, and the pulse phase distribution range of the suspended potential discharge cluster is wider than the first phase threshold range and the amplitude variation coefficient is higher than a first variation threshold;

[0014] Based on the waveform symmetry and repetition rate of the pulses, the surface discharge set is divided into a typical surface discharge cluster and a surface creeping cluster; wherein the pulse rise time of the typical surface discharge cluster is longer than a first rise time threshold and the pulse repetition rate is higher than a first repetition rate threshold, and the pulse amplitude of the surface creeping cluster is lower than a first amplitude threshold and the number of pulses per unit time is higher than a first density threshold;

[0015] Based on the time interval regularity of the pulses, the discrete noise point set is divided into a random noise point set and a periodic interference point set; wherein the random noise point set is irregularly distributed in the time domain, and the pulse interval of the periodic interference point set has a fixed frequency.

[0016] Further, the internal discharge set is divided into a typical internal discharge cluster and a suspended potential discharge cluster based on the power frequency phase distribution characteristics of the pulses, comprising:

[0017] Pulse waveform correlation analysis is performed on the typical internal discharge cluster and the suspended potential discharge cluster, and a normalized cross-correlation coefficient of each pulse waveform in the cluster and the corresponding cluster center waveform is calculated;

[0018] Pulses with a cross-correlation coefficient lower than a preset correlation threshold are determined as abnormal noise points embedded in the discharge cluster and are filtered out;

[0019] Phase distribution consistency test is performed on the suspended potential discharge cluster, and pulse points deviating from the main phase distribution area in the cluster are determined as phase abnormal noise points and are filtered out.

[0020] Further, the surface discharge set is divided into a typical surface discharge cluster and a surface creeping cluster based on the waveform symmetry and repetition rate of the pulses, comprising:

[0021] establishing an amplitude and phase distribution model of the typical surface discharge cluster and the creepage discharge cluster;

[0022] calculating Mahalanobis distance between each pulse point in the cluster and the amplitude and phase distribution model, and determining a pulse point as a distribution abnormal point if the distance exceeds a preset abnormal threshold;

[0023] performing pulse time interval statistical analysis on the creepage discharge cluster, determining a pulse as a time sequence abnormal noise point if the time interval distribution is different from the main distribution mode in the cluster, and filtering out the pulse.

[0024] Further, based on the time interval regularity of the pulses, the discrete noise point set is divided into a random noise point set and a periodic interference point set, including:

[0025] performing frequency spectrum analysis on the periodic interference point set to identify corresponding fundamental frequency components and harmonic components;

[0026] constructing an adaptive comb filter based on the identified fundamental frequency components and harmonic components to filter out interference components of corresponding frequencies in the original signal;

[0027] performing amplitude statistical distribution analysis on the random noise point set to establish an amplitude probability distribution model, and determining a pulse point as an abnormal amplitude noise point if the amplitude deviates from a preset signal interval of the amplitude probability distribution model and filtering out the pulse point.

[0028] Further, the index parameters for distinguishing the discharge types include at least one of the following features:

[0029] time domain features, including pulse amplitude, pulse rise time, pulse fall time, pulse width, discharge repetition rate, and power frequency phase distribution;

[0030] frequency domain features, including main frequency, frequency band energy distribution, spectral centroid, and spectral asymmetry;

[0031] time-frequency domain features, including wavelet packet energy entropy, wavelet coefficient variance, and time-frequency matrix singular value;

[0032] statistical features, including discharge quantity skewness, discharge quantity kurtosis, pulse interval coefficient of variation, and discharge phase distribution confidence;

[0033] multi-terminal collaborative features, including discharge signal amplitude ratio detected by each monitoring node, pulse arrival time difference, and discharge mode spatial correlation.

[0034] Further, the target discharge diagnosis scheme is determined according to the source type, severity, and location information of the partial discharge, including:

[0035] A multi-objective optimization function is established, aiming at minimizing the system comprehensive cost and maximizing the system power supply availability rate;

[0036] Multiple constraint conditions are set, including maintenance resource constraints, power outage time constraints, risk priority constraints and network operation constraints;

[0037] Based on the multi-objective optimization function and multiple constraint conditions, a multi-objective decision model for strategy optimization is constructed;

[0038] A non-dominated sorting genetic algorithm with an elitist strategy is used to solve the multi-objective decision model, obtaining a set of non-dominated solutions, which constitute a Pareto front;

[0039] Based on the preset operation and maintenance decision preference, an optimal compromise solution is selected from the Pareto front as the final target discharge diagnosis scheme.

[0040] Further, the system comprehensive cost minimization Objective function expression is:

[0041]

[0042] Wherein: is the decision variable, is the total number of discharge sources, is the maintenance strategy of the th discharge source, is the average implementation cost coefficient of unit maintenance strategy, is the local discharge characteristic weight of the th discharge source, is the risk reduction rate of the th discharge source after adopting strategy , is the unit risk cost coefficient, is the power outage time required by the th discharge source to adopt strategy , is the load level affected by the maintenance of the th discharge source, is the power loss cost coefficient of unit power outage time, is the unit diagnosis cost benchmark, is the effective discharge pulse number of the th discharge source screened out after noise suppression and clustering, is the total pulse number collected by the th discharge source, is the multi-end collaborative diagnosis cost coefficient of the th discharge source.

[0043] Further, the system power availability maximization The expression of the objective function is:

[0044]

[0045] Wherein: is the total statistical cycle, is the discharge development coefficient of the th discharge source, , is the amplitude change amount, is the initial amplitude, is the time interval, is the interval time from the diagnosis discovery to the execution of the repair of the th discharge source.

[0046] From the above technical solution, the present application has the following advantages:

[0047] The present application firstly acquires partial discharge signals synchronously by deploying at least two monitoring nodes at different positions of the distribution overhead network, so as to obtain more abundant discharge information; the signals collected by each monitoring node are subjected to noise suppression processing, so as to weaken noise interference and provide more accurate data for subsequent analysis; secondly, based on the signals after noise reduction, index parameters capable of distinguishing discharge types are extracted, so as to describe discharge characteristics from multiple dimensions; then these index parameters are subjected to fusion analysis, so as to comprehensively analyze multi-terminal and multi-feature information, identify the source type of partial discharge, evaluate the severity, and locate the discharge position; finally, according to the obtained source type, severity and position information, a target discharge diagnosis scheme is determined in a targeted manner, so as to guarantee the comprehensive and accurate diagnosis result, and provide effective support for the operation and maintenance of partial discharge of the distribution overhead network. BRIEF DESCRIPTION OF DRAWINGS

[0048] Fig. 1 is an embodiment flowchart of a partial discharge diagnosis method for a distribution overhead network based on multi-terminal data fusion in the present application;

[0049] Fig. 2 is another embodiment flowchart of a partial discharge diagnosis method for a distribution overhead network based on multi-terminal data fusion in the present application;

[0050] Fig. 3 is another embodiment flowchart of a partial discharge diagnosis method for a distribution overhead network based on multi-terminal data fusion in the present application;

[0051] Fig. 4 is another embodiment flowchart of a partial discharge diagnosis method for a distribution overhead network based on multi-terminal data fusion in the present application. DETAILED DESCRIPTION

[0052] The terms "first", "second", "third", "fourth" and the like in the description of this application and in the claims, if any, are used for distinguishing between similar elements and not necessarily for describing a particular sequential or chronological order. It is to be understood that the use of these terms herein is to be construed to cover a generalised use of such terms to refer to similar elements independently of each other occurrence in the description and claims. It is to be understood that the data thus described with the term "first", "second", "third", "fourth" and the like can be interchanged with each other in the description and claims to describe any of the embodiments of the application. Furthermore, the terms "comprising", "including", "containing", and "having" and their conjugates, as used herein, are intended to encompass the presence of one or more elements, integers, steps, processes, actions, features, objects, components, units, or the like, without necessarily excluding the presence of one or more other elements, integers, steps, processes, actions, features, objects, components, units, or the like.

[0053] Embodiment one

[0054] The method implemented in the embodiment can be implemented in a system, and can be implemented in a server or in a terminal, and the specific implementation is not limited. From the perspective of system implementation, the power distribution overhead network partial discharge diagnosis method based on multi-terminal data fusion in the application will be introduced. Please refer to Figs. 1 to 4 The method provided in the embodiment of the application includes the following steps:

[0055] S1. Synchronously collecting partial discharge signals through at least two monitoring nodes deployed at different positions of the power distribution overhead network;

[0056] In the embodiment, the monitoring nodes should be deployed at key electrical nodes of the power distribution overhead network and positions prone to insulation deterioration, for example, insulator suspension points of line towers, wiring terminals of circuit breakers, disconnectors, load switches, connections of cable terminals, proximity of outlet bushings of transformers and mutual inductors, old sections of lines, sections of lines prone to environmental influences such as crossing forest areas, etc. Each monitoring node is composed of a sensor unit and a data acquisition unit. The sensor unit is fixed on the grounding downlead, the armored layer or the body of the device to be measured through a non-intrusive clamp or bracket for sensing signals; the data acquisition unit is installed on the tower or the device box, and is responsible for signal conditioning, digitization and communication.

[0057] The signals synchronously collected by the monitoring nodes are not a single type, but in order to comprehensively capture various physical effects of partial discharge, usually include the following multi-modal signals: (1) The electromagnetic wave signal of 300 MHz-3 GHz frequency band radiated by partial discharge is received through the ultra-high frequency antenna sensor, the signal has strong penetration and anti-low frequency interference ability, is suitable for detecting the discharge inside the equipment, and can effectively sense the starting moment of the discharge, and is the preferred signal for time difference positioning. (2) The transient ground voltage generated by the high-frequency current propagating along the equipment metal shell on the grounding down lead is detected by the capacitive coupling type sensor, which is very sensitive to the surface discharge and internal discharge of the metal packaging equipment such as switch cabinet. (3) The partial discharge pulse current flowing through the grounding wire is measured by clamping the high-frequency current transformer on the grounding wire of the equipment, which can effectively measure the current amplitude and waveform of the discharge pulse, and is used to evaluate the discharge amount. (4) The sound wave or vibration signal generated by the partial discharge in the air or solid insulation is received by the ultrasonic sensor, the propagation speed is much slower than the electromagnetic wave, but it has good directionality, which is suitable for assisting in positioning the physical position of the discharge source in a complex electromagnetic environment.

[0058] S2. The partial discharge signals collected by each monitoring node are subjected to noise suppression processing;

[0059] In this embodiment, the partial discharge signal is denoised based on the noise suppression rule, which is realized by the following steps:

[0060] S21. Based on the multiple partial discharge signal pulses collected by the monitoring node in the continuous time window, the feature parameters of each pulse in the time domain and the frequency domain are calculated, and multiple pulse feature vector sets are generated;

[0061] The essence of partial discharge is a discrete and individual insulation breakdown event, and each event will generate a detectable pulse. Taking the pulse as the basic analysis unit can most directly capture the individual characteristics of each discharge, thereby providing basic information for distinguishing discharge types and identifying noise. In contrast, directly processing the signal of the entire time length will confuse the characteristics of different pulses and cannot utilize the correlation information between pulses. Specifically, the amplitude threshold method is used here to segment a single pulse from the continuous time domain signal. The amplitude threshold method is to set a voltage or current threshold higher than the background noise level. When the signal amplitude exceeds the threshold, it is determined that it is the starting point of a pulse, and the waveform in this time period is cut off as a pulse segment until the signal falls below the threshold.

[0062] For each extracted pulse segment , a set of quantitative feature parameters is calculated from the time domain and the frequency domain, respectively. The time domain feature parameters include pulse amplitude , pulse rise time , pulse fall time Pulse width and pulse energy At this point, the frequency domain characteristic parameters include the dominant frequency. Average frequency Frequency centroid and bandwidth To make a single pulse The feature parameters are combined in a fixed order to generate 3D feature vectors are A monitoring node extracts data within a time window. There are pulses, and the corresponding feature vector set is: .

[0063] S22. Perform cluster analysis on multiple pulse feature vector sets, and divide the pulses into different clusters based on preset clustering rules;

[0064] Cluster analysis is performed through the following steps:

[0065] S221. A density-based clustering algorithm is used to generate initial clustering results for multiple pulse feature vector sets. The initial clustering results include internal discharge sets, surface discharge sets, and discrete noise point sets.

[0066] Each vector in the pulse feature vector set is mapped to a high-dimensional feature space. In this space, a density-based clustering algorithm is used to calculate the density of each data point within a preset neighborhood radius. The algorithm identifies data points with densities higher than a preset threshold as core points and divides interconnected core points and their density-reachable boundary points into different clusters. These form high-density compact clusters in the feature space, corresponding to discharge types with stable and similar physical characteristics. Data points that cannot be classified into any high-density core cluster are marked as discrete noise points, forming a discrete noise point set. For the identified core clusters, based on the position of the statistical average of all pulse feature vectors within the cluster in the feature space, combined with prior knowledge in the field of partial discharge diagnosis, they are initially divided into internal discharge sets and surface discharge sets.

[0067] This step utilizes a density-based clustering algorithm to unsupervisedly separate high-density regions (core clusters) and low-density regions (noise) based on the distribution of data in the feature space. The division of high-density core clusters into internal discharge sets and surface discharge sets is not a function of the algorithm, but rather a result of the principle that the pulse characteristics of different discharge types will cluster in different regions of the feature space.

[0068] S222. Based on the pulse-based power frequency phase distribution characteristics, the internal discharge set is divided into a typical internal discharge cluster and a floating potential discharge cluster; wherein the pulse phase distribution range of the typical internal discharge cluster meets a preset first phase threshold range, and the pulse phase distribution range of the floating potential discharge cluster is wider than the first phase threshold range and the amplitude variation coefficient is higher than a first variation threshold;

[0069] The above internal discharge is a broad category, which contains subtypes of different physical mechanisms, and there are subtle but key differences in power frequency phase distribution and pulse characteristics, which need to be separated by the second step of clustering or threshold judgment. The typical internal discharge cluster refers to the discharge occurring in the closed air gap inside the solid or liquid insulating medium, and the pulse strictly concentrates in a specific phase interval (first and third quadrants) near the peak value of the power frequency voltage. The phase distribution range is very narrow; due to the fixed discharge environment (air gap size, air pressure), the amplitude of the discharge pulse is also relatively stable. The floating potential discharge cluster is caused by the poor grounding of a certain metal part in the device, which obtains electric energy through capacitive coupling and accumulates electric charge, and when its potential to ground rises to the breakdown threshold, a discharge occurs; the discharge pulse has weak correlation with the power frequency phase, and the phase distribution range is significantly wider than that of the typical internal discharge, and may span multiple quadrants; due to the instability of the discharge process, the fluctuation of the pulse amplitude is also more violent.

[0070] The preset first phase threshold range is set to 45° to 90° and 225° to 270° of the power frequency phase, and if more than 80% of the pulses in a cluster fall within this phase range, the phase distribution is considered to meet the threshold range. The first phase threshold range refers to the 45-90° and 225°-270° range mentioned above, and the first variation threshold refers to the threshold of the coefficient of variation (CV), which is set to CV>0.5. Due to its instability, the CV value of the floating potential discharge is greater than 0.5, while the CV value of the typical internal discharge is less than 0.3.

[0071] This step is implemented through the following sub-steps:

[0072] 1. Perform pulse waveform correlation analysis on the typical internal discharge cluster and the floating potential discharge cluster, and calculate the normalized cross-correlation coefficient of each pulse waveform in the cluster and the corresponding cluster center waveform;

[0073] 2. Determine the pulses with a cross-correlation coefficient lower than a preset correlation threshold as abnormal noise points embedded in the discharge cluster and filter them out;

[0074] 3. Perform phase distribution consistency test on the floating potential discharge cluster, and determine the pulse points deviating from the main phase distribution area in the cluster as phase abnormal noise points and filter them out.

[0075] Calculate the power frequency phase of all pulses in the initial internal discharge cluster and count the phase distribution histogram. Calculate the standard deviation and amplitude variation coefficient of the phase distribution; if the phase standard deviation of the subset is < 25° and simultaneously more than 80% of the pulses are located within the first phase threshold range (45°-90° and 225°-270°), and its CV < 0.4, then the subset is divided into a typical internal discharge cluster as a whole; if the phase standard deviation of the subset is > 40° or CV > 0.5, it is divided into a suspended potential discharge cluster. For the two new clusters after division, the cluster centers (i.e. the mean value of the feature vector) are recalculated to more accurately represent the typical characteristics of the discharge.

[0076] S223. Based on the waveform symmetry and repetition rate of the pulses, the surface discharge set is divided into a typical surface discharge cluster and a surface creeping discharge cluster; wherein the pulse rise time of the typical surface discharge cluster is longer than the first rise time threshold and the pulse repetition rate is higher than the first repetition rate threshold, and the pulse amplitude of the surface creeping discharge cluster is lower than the first amplitude threshold and the number of pulses per unit time is higher than the first density threshold;

[0077] This step is a further refinement of the surface discharge phenomenon, although it is also a surface discharge, but the typical surface discharge and the surface creeping discharge have differences in physical process, development speed and hazard degree, and their pulse characteristics have distinguishable patterns in waveform and statistical rules. The typical surface discharge cluster refers to the persistent and relatively stable discharge that occurs on the surface of insulators, cable terminals and other locations due to the formation of conductive paths by contaminants under humid conditions; the discharge development requires time, so the pulse rising edge is relatively slow. The discharge continues to occur, so the pulse repetition frequency is high. Because the discharge channel is relatively stable, the pulse amplitude fluctuates within a certain range, but there will be pulses with larger amplitudes. The surface creeping discharge cluster refers to the discharge phenomenon that does not fix at a point, but rapidly and randomly creeps and extends on the insulating surface in a branch-like form; the existence time of a single discharge channel is extremely short, so the energy of a single pulse is low and the amplitude is generally small. But in the moment of rapid development, a large number of small discharge pulses will be generated in a very short time, forming a high-density pulse group or pulse train. From a macro statistical point of view, its repetition rate may be very high, but microscopically it is a burst mode.

[0078] The first rise time threshold is set to 100 ns. The ion migration process of typical surface discharge leads to a longer pulse rise time, usually greater than 100 ns. The first repetition rate threshold is set to 100 pulses per second. To distinguish between active sustained discharge and sparse accidental discharge, the repetition rate of typical surface discharge is usually higher than this value. The first amplitude threshold is set to 30% of the average amplitude of the effective discharge pulse of the monitoring node. The pulse amplitude of the surface creepage is significantly lower than other types of surface discharge. The first density threshold is set to 50 pulses per millisecond. This threshold is used to identify explosive discharge. If a pulse sequence produces more than 50 pulses in a 1 millisecond time window, it is considered to have explosive discharge, which is a typical characteristic of surface creepage.

[0079] This step is achieved by the following three processes:

[0080] 1. Establish the amplitude and phase distribution model of typical surface discharge clusters and surface creepage clusters;

[0081] 2. Calculate the Mahalanobis distance of each pulse point in the cluster to the amplitude and phase distribution model, and determine the pulse point with a distance exceeding the preset abnormal threshold as a distribution abnormal point;

[0082] 3. Perform pulse time interval statistical analysis on the surface creepage cluster, and determine the pulse with different time interval distribution from the main distribution mode in the cluster as a time sequence abnormal noise point and filter it out.

[0083] For each pulse in the initial surface discharge set, the rise time and amplitude are calculated; the instantaneous pulse repetition rate and burst density of the entire signal segment are calculated using a sliding time window; the rise time and repetition rate are used as the main feature dimension, and the amplitude and burst density are used as the auxiliary dimension to construct a new feature space; in this space, the pulse group that satisfies the rise time > 100 ns and the repetition rate > 100 Pulse / s is divided into a typical surface discharge cluster; the pulse group that satisfies the amplitude < (average amplitude * 0.3) and the burst density > 50 Pulse / ms is divided into a surface creepage cluster. The cluster centers of the two new clusters after division are recalculated.

[0084] S224. Based on the regularity of the time interval of the pulses, the discrete noise point set is divided into a random noise point set and a periodic interference point set; wherein the random noise point set is irregularly distributed in the time domain, and the pulse interval of the periodic interference point set has a fixed frequency.

[0085] This step is to classify the noise which has been preliminarily separated. Different types of noise have different mechanisms, so different strategies must be used to effectively filter them out. Discrete noise point sets are derived from random electromagnetic interference in the environment, such as: surges caused by lightning, overvoltage caused by the operation of other unrelated equipment, automobile ignition systems, random corona discharge, etc. These events occur randomly in time, and periodic interference point sets are derived from periodic pulse interference generated by other power electronic devices or carrier communication devices in the power system. For example: switching operation of rectifiers or inverters, control signals of series compensation devices, power line carrier communication signals, etc. These disturbances have a fixed operating frequency or an integer multiple relationship with the power frequency.

[0086] 1. Perform spectral analysis on the periodic interference point set to identify the corresponding fundamental frequency component and harmonic component;

[0087] 2. Based on the identified fundamental frequency component and harmonic component, construct an adaptive comb filter to filter out the interference components at the corresponding frequencies in the original signal;

[0088] 3. Perform amplitude statistical distribution analysis on the random noise point set to establish an amplitude probability distribution model; and determine abnormal amplitude noise points as pulse points whose amplitude deviates from the pre-set signal interval of the amplitude probability distribution model and filter them out.

[0089] For the pulses in the discrete noise point set, first sort them by their occurrence time stamp, then calculate the time interval between all adjacent pulses to form a time interval sequence; calculate the coefficient of variation of the time interval sequence, the formula is where is the standard deviation of the time interval sequence, is its average value If 0.5, it indicates that the time interval is very unstable, then the noise point set is divided into a random noise point set; if 0.2, it indicates that the time interval is very stable, then the noise point set is divided into a periodic interference point set. For the set divided into periodicity, calculate its average value , the reciprocal of which is the fundamental frequency of the interference. Further check whether the fundamental frequency is associated with the power grid frequency or its harmonic frequency, common power electronic switching frequency, to finally confirm.

[0090] S23. Identify and filter out noise pulse clusters based on the statistical characteristics of each cluster to obtain effective discharge pulses.

[0091] Firstly, the pulse eigenvectors are clustered by multi-level clustering analysis, and the pulses are accurately divided into different clusters with typical internal discharge, suspended potential discharge, typical surface discharge, surface creeping, periodic interference and random noise according to the statistical characteristics of power frequency phase, waveform and time interval. Various algorithms such as waveform correlation analysis, Mahalanobis distance detection and time interval statistical analysis are used to remove embedded abnormal noise points from each discharge cluster. Finally, by filtering all identified noise clusters and abnormal points, a pure and highly reliable effective discharge pulse set is obtained. The set contains all the discharge pulses judged to be from real insulation defects, and its amplitude, phase, waveform and other characteristics are real and reliable, providing a high-quality data basis for subsequent calculation of diagnostic index parameters.

[0092] S3. Extracting index parameters for distinguishing discharge types based on the denoised partial discharge signals;

[0093] S4. Fusing and analyzing the calculated index parameters to obtain the source type, severity and location information of the partial discharge;

[0094] In this embodiment, the index parameters for distinguishing discharge types include at least one of the following features: time domain features including pulse amplitude, pulse rise time, pulse fall time, pulse width, discharge repetition rate and power frequency phase distribution; frequency domain features including main frequency, frequency band energy distribution, spectral centroid and spectral asymmetry; time-frequency domain features including wavelet packet energy entropy, wavelet coefficient variance and time-frequency matrix singular value; statistical features including discharge quantity skewness, discharge quantity kurtosis, pulse interval coefficient of variation and discharge phase distribution confidence; multi-terminal collaborative features including discharge signal amplitude ratio detected by each monitoring node, pulse arrival time difference and discharge mode spatial correlation.

[0095] Specifically, the same type of index parameters from all monitoring nodes are normalized to eliminate the influence of dimension and numerical range caused by the difference in propagation path of different nodes, and are spliced to form a high-dimensional comprehensive feature vector, which represents the overall discharge state of the measured object. The high-dimensional comprehensive feature vector is input into a pre-trained multi-class machine learning model; wherein, the model training process includes: 1. Constructing a sample library; the training samples in the sample library contain known types of partial discharge signals, and have been subjected to the same noise suppression processing as described above, that is, effective discharge pulses are screened through clustering analysis to ensure sample quality; 2. Feature preprocessing; sample features need to be normalized to eliminate dimensional differences, that is, to eliminate the influence of dimension and numerical range caused by the difference in propagation path of different nodes as mentioned above; 3. Training the model and optimizing the parameters; using supervised learning, the classification error is minimized by adjusting the model hyperparameters, and finally the mapping from multi-dimensional features to discharge types is realized. The model is trained on a large number of known types of partial discharge sample library, and can establish a complex nonlinear mapping relationship between multi-dimensional features and discharge types (such as internal discharge, surface discharge, corona discharge, etc.); the output of the model is the most likely type of partial discharge source identified. The pulse arrival time difference in the multi-end collaborative feature is used for positioning, specifically: select a monitoring node as the time reference point, calculate the absolute time difference of the same discharge pulse arriving at other nodes and the reference node; based on the known spatial coordinates of the monitoring nodes and the propagation speed of electromagnetic waves or sound waves in air, the hyperbolic time difference positioning algorithm is used to solve the spatial coordinates of the discharge source, and finally the position information of the partial discharge source is determined.

[0096] According to the identified discharge source type, select the feature index most sensitive to this type of discharge, such as for internal discharge, select discharge quantity and repetition rate; for surface discharge, select pulse amplitude maximum and average current; use a weighted fusion algorithm to calculate a comprehensive severity index; compare the index with the preset severity level threshold, and finally map to a quantitative severity level.

[0097] S5. Determine the target discharge diagnosis scheme according to the source type, severity and position information of the partial discharge.

[0098] This step is the final output and application link of the diagnosis method, and the purpose is to recommend a globally optimal or most satisfactory action plan under the current situation from countless possible coping strategies after comprehensively weighing multiple factors such as economic cost, power supply reliability, maintenance risk and on-site operation restrictions, thereby improving the intelligent level and economic benefit of power grid operation and maintenance. Specifically as follows:

[0099] S51. Establish a multi-objective optimization function with the minimum system comprehensive cost and the maximum system power supply availability rate as the target;

[0100] This step quantifies the operation and maintenance decision-making target into a mathematical expression, and the system minimizes the comprehensive cost The objective function expression is:

[0101]

[0102] Wherein: is the decision variable, is the total number of discharge sources, is the maintenance strategy of the th discharge source, is the average implementation cost coefficient of the unit maintenance strategy, is the local discharge characteristic weight of the th discharge source, is the risk reduction rate of the th discharge source after adopting strategy , is the unit risk cost coefficient, is the outage time required by the th discharge source to adopt strategy , is the load level affected by the maintenance of the th discharge source, is the power loss cost coefficient of the unit outage time, is the unit diagnosis cost benchmark, is the number of effective discharge pulses screened out after noise suppression and clustering of the th discharge source, is the total number of pulses collected by the th discharge source, is the multi-end cooperative diagnosis cost coefficient of the th discharge source.

[0103] System power supply availability maximization The expression of the objective function is:

[0104]

[0105] Wherein: is the total statistical period set, is the discharge development coefficient of the th discharge source, , is the amplitude change amount, is the initial amplitude, is the time interval, is the interval time from diagnosis discovery to maintenance execution of the th discharge source.

[0106] S52. Set multiple constraints including maintenance resource constraint, outage time constraint, risk priority constraint and network operation constraint.

[0107] Maintenance resource constraint: The total demand of resources such as personnel, special operation vehicles, test equipment and the like planned in the scheme must not exceed the total amount of resources actually available in the current time period.

[0108] The expression of the maintenance resource constraint is as follows:

[0109]

[0110] Wherein: is the maintenance strategy of the th discharge source, is the diagnosis strategy of the th discharge source, is the amount of the th maintenance resource consumed by executing the maintenance strategy , is the total amount of the th maintenance resource available, is the amount of the th diagnosis resource consumed by executing the diagnosis strategy , is the total amount of the th diagnosis resource available.

[0111] Outage time constraint: The maintenance operation must be limited to the time window in the off-peak period of the power grid load (such as late at night) to minimize the impact on users. The total outage time must not exceed the window period.

[0112] The expression of the outage time constraint is as follows:

[0113]

[0114] Wherein: , are the estimated start and end times of the strategy , which must fall within the preset off-peak time window , is the partial discharge time-frequency distribution function of the th discharge source, is the maximum discharge amount threshold allowed in the outage window.

[0115] Risk priority constraint: For discharge defects diagnosed as high severity, the maintenance priority must be forcibly raised, and the scheme must include timely handling of them, and the risk reduction rate after handling must reach the safety standard.

[0116] The expression of the risk priority constraint is as follows:

[0117]

[0118] wherein: is the severity level of the th discharge source, is the severity threshold, is the discharge type of the th discharge source, is the risk reduction rate, is the dedication reduction rate minimum requirement, set higher threshold (1.2 times) for internal discharge (higher threat), lower threshold (0.8 times) for surface discharge (relatively lower threat), adapting to the fault evolution characteristics of different discharge types.

[0119] Network operation constraints: the topology of the distribution network must remain radial (typical operation mode of distribution network) and cannot have line overload problems, etc. during and after the execution of any maintenance scheme, and must comply with the safe operation regulations of the power grid.

[0120] The expression of the network operation constraints is:

[0121]

[0122] wherein: The pulse arrival time difference fluctuation constraint must be met: ; is the topology of the distribution network after maintenance, which must belong to the feasible radial topology set and the discharge locatable topology set ; , are the pulse arrival time differences from the th monitoring node to the th discharge source before and after the topology change, respectively; is the maximum allowed time difference fluctuation value to avoid excessive topology changes caused by maintenance, which would destroy the basic conditions for multi-terminal location of partial discharge.

[0123] S53. Based on the multi-objective optimization function and multiple constraints, a multi-objective decision model for diagnostic strategy optimization is constructed.

[0124] The objectives and constraints in S51 and S52 are integrated to generate a complete mathematical optimization model, with the decision variables being various possible maintenance strategies. The optimization process is to find strategies that can simultaneously make as small as possible, as large as possible among the strategies that meet all the constraints.

[0125] S54. A non-dominated sorting genetic algorithm with elitist strategy is used to solve the multi-objective decision model, and a set of non-dominated solutions is obtained, which constitutes the Pareto front;

[0126] Since it is a multi-objective optimization, there is usually no unique solution that is optimal on all objectives, but a set of optimal trade-off solutions. Here, a genetic algorithm is used to simulate the search for better solutions in the solution space. The elitist strategy ensures that good individuals are not lost, and the non-dominated sorting can distinguish the level of solutions. The algorithm eventually outputs a set of Pareto optimal solutions on the surface, on which any improvement in one target will inevitably lead to the deterioration of at least another target.

[0127] S55. Based on the preset operation and maintenance decision preference, an optimal trade-off solution is selected from the Pareto front as the final target discharge diagnosis scheme.

[0128] After the algorithm provides multiple optimal trade-off solutions, the final decision needs to be based on the preferences of the operation and maintenance personnel. These preferences can be preset, such as "priority power supply this month, try to control the outage time" or "there is a surplus of maintenance resources this month, prioritize all high-risk defects". According to these preferences, a utility function or trade-off rule is defined to automatically select the one that best meets the current management goals from the Pareto solution set, for example, select the one with the lowest cost from the solutions with less than 1% loss of availability. The selected solution is the target discharge diagnosis scheme recommended by the system, which may be specific to: "for the serious surface discharge located in the XX tower, it is recommended to send 2 workers and 1 insulation test vehicle to the site at 0:00-4:00 on Friday this week, and use the live washing and reinforcement operation scheme to handle it."

[0129] It can be understood that those skilled in the art can combine various embodiments in the above embodiments according to the teachings of the above embodiments to obtain various technical solutions of the embodiments.

[0130] The above only describes the preferred embodiments of the present application and should not be used to limit the present application. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A partial discharge diagnosis method for power distribution overhead networks based on multi-terminal data fusion, characterized in that, The method comprises the steps of: Synchronously collecting partial discharge signals through at least two monitoring nodes arranged at different positions of an overhead power distribution network; Performing noise suppression processing on the partial discharge signals collected by each monitoring node; wherein, based on a plurality of partial discharge signal pulses collected by the monitoring node within a continuous time window, the characteristic parameters of each pulse in the time domain and the frequency domain are calculated to generate a plurality of pulse feature vector sets; clustering analysis is performed on the plurality of pulse feature vector sets, and the pulses are divided into different clusters based on a preset clustering division rule; noise pulse clusters are identified and filtered out according to the statistical characteristics of each cluster to obtain effective discharge pulses; Extracting index parameters for distinguishing discharge types based on the partial discharge signals after noise suppression processing; Performing fusion analysis on the extracted index parameters to obtain source type, severity and location information of the partial discharge; Determining a target discharge diagnosis scheme according to the source type, severity and location information of the partial discharge.

2. The partial discharge diagnostic method for power distribution overhead network based on multi-terminal data fusion according to claim 1, characterized in that, The clustering analysis of the plurality of pulse feature vector sets, and the division of the pulses into different clusters based on a preset clustering division rule, comprises: An initial clustering result of the plurality of pulse feature vector sets is generated by using a density-based clustering algorithm, and the initial clustering result includes an internal discharge set, a surface discharge set and a discrete noise point set; Based on the power frequency phase distribution characteristics of the pulses, the internal discharge set is divided into a typical internal discharge cluster and a suspended potential discharge cluster; wherein, the pulse phase distribution range of the typical internal discharge cluster conforms to a preset first phase threshold range, and the pulse phase distribution range of the suspended potential discharge cluster is wider than the first phase threshold range and the amplitude variation coefficient is higher than a first variation threshold; Based on the waveform symmetry and repetition rate of the pulses, the surface discharge set is divided into a typical surface discharge cluster and a surface creeping cluster; wherein, the pulse rise time of the typical surface discharge cluster is longer than a first rise time threshold and the pulse repetition rate is higher than a first repetition rate threshold, and the pulse amplitude of the surface creeping cluster is lower than a first amplitude threshold and the number of pulses per unit time is higher than a first density threshold; Based on the time interval regularity of the pulses, the discrete noise point set is divided into a random noise point set and a periodic interference point set; wherein, the random noise point set is irregularly distributed in the time domain, and the pulse interval of the periodic interference point set has a fixed frequency.

3. The partial discharge diagnostic method for power distribution overhead network based on multi-terminal data fusion according to claim 2, characterized in that, The division of the internal discharge set into a typical internal discharge cluster and a suspended potential discharge cluster based on the power frequency phase distribution characteristics of the pulses comprises: Pulse waveform correlation analysis is performed on the typical internal discharge cluster and the suspended potential discharge cluster, and the normalized cross-correlation coefficient of each pulse waveform in the cluster and the corresponding cluster center waveform is calculated; Pulses with a cross-correlation coefficient lower than a preset correlation threshold are determined as abnormal noise points embedded in the discharge cluster and are filtered out; Phase distribution consistency test is performed on the suspended potential discharge cluster, and pulse points deviating from the main phase distribution area in the cluster are determined as phase abnormal noise points and are filtered out.

4. The partial discharge diagnostic method for power distribution overhead network based on multi-terminal data fusion according to claim 2, characterized in that, The division of the surface discharge set into a typical surface discharge cluster and a surface creeping cluster based on the waveform symmetry and repetition rate of the pulses comprises: Establishing amplitude and phase distribution model of the typical surface discharge cluster and the creeping discharge cluster; Calculating Mahalanobis distance between each pulse point in the cluster and the amplitude and phase distribution model, and determining the pulse point as a distribution abnormal point if the distance exceeds a preset abnormal threshold; Performing pulse time interval statistical analysis on the creeping discharge cluster, determining the pulse with different time interval distribution from the main distribution mode in the cluster as a time sequence abnormal noise point and filtering it out.

5. The partial discharge diagnostic method for power distribution overhead network based on multi-terminal data fusion according to claim 2, characterized in that, Based on the time interval regularity of the pulse, the discrete noise point set is divided into a random noise point set and a periodic interference point set, including: Performing frequency spectrum analysis on the periodic interference point set to identify corresponding fundamental frequency components and harmonic components; Based on the identified fundamental frequency components and harmonic components, constructing an adaptive comb filter to filter out the interference components of the corresponding frequency in the original signal; Performing amplitude statistical distribution analysis on the random noise point set to establish an amplitude probability distribution model, and determining the pulse point with amplitude deviating from the preset signal interval of the amplitude probability distribution model as an abnormal amplitude noise point and filtering it out.

6. The multi-end data fusion based power distribution overhead network partial discharge diagnostic method according to claim 1, wherein, The index parameters for distinguishing the discharge types include at least one of the following features: Time domain features, including pulse amplitude, pulse rise time, pulse fall time, pulse width, discharge repetition rate, and power frequency phase distribution; Frequency domain features, including main frequency, frequency band energy distribution, spectral centroid, and spectral asymmetry; Time-frequency domain features, including wavelet packet energy entropy, wavelet coefficient variance, and time-frequency matrix singular value; Statistical features, including discharge quantity skewness, discharge quantity kurtosis, pulse interval coefficient of variation, and discharge phase distribution confidence; Multi-terminal collaborative features, including discharge signal amplitude ratio detected by each monitoring node, pulse arrival time difference, and discharge mode spatial correlation.

7. The partial discharge diagnostic method for power distribution overhead network based on multi-terminal data fusion according to claim 1, characterized in that, The target discharge diagnosis scheme is determined according to the source type, severity, and location information of the partial discharge, including: Establishing a multi-objective optimization function with the minimum system comprehensive cost and the maximum system power supply availability as the target; Setting multiple constraint conditions including maintenance resource constraints, power outage time constraints, risk priority constraints, and network operation constraints; Based on the multi-objective optimization function and the multiple constraint conditions, constructing a multi-objective decision model for diagnosis strategy optimization; Solving the multi-objective decision model by using a non-dominated sorting genetic algorithm with an elitist strategy to obtain a set of non-dominated solutions, which constitute a Pareto front; Based on the preset operation and maintenance decision preference, selecting an optimal compromise solution from the Pareto front as the final target discharge diagnosis scheme.

8. The partial discharge diagnostic method for power distribution overhead network based on multi-terminal data fusion according to claim 7, characterized in that, The system synthesizes cost minimization The objective function expression is: wherein: is a decision variable, is the total number of discharge sources, is the maintenance strategy of the th discharge source, is the average implementation cost coefficient of the unit maintenance strategy, is the partial discharge characteristic weight of the th discharge source, is the risk reduction rate of the th discharge source after taking the strategy , is the unit risk cost coefficient, is the outage time required for the th discharge source to take the strategy , is the load level affected by the maintenance of the th discharge source, is the power loss cost coefficient of the unit outage time, is the unit diagnosis cost benchmark, is the number of effective discharge pulses screened out after noise suppression and clustering of the th discharge source, is the total number of pulses collected by the th discharge source, is the multi-terminal collaborative diagnosis cost coefficient of the th discharge source.

9. The partial discharge diagnostic method for power distribution overhead network based on multi-terminal data fusion according to claim 8, characterized in that, The system power availability maximization The expression of the objective function is: wherein: is a set total statistics period, is a discharge development coefficient of the discharge source, , is a magnitude change amount, is an initial magnitude, is a time interval, is an interval time from a diagnostic finding to performing a repair of the discharge source.

Images