Power distribution network insulation defect early warning method and device based on instantaneous ground event sequence

Abstract

The application provides a power distribution network insulation defect early warning method and device based on a transient grounding event sequence, comprising: obtaining transient grounding data based on distributed synchronous recording, using a transient method to judge and output the grounding event of the power distribution network, and taking it as a current event; judging whether the transient grounding times of the related grounding section and grounding phase within a preset time limit exceed a threshold value, and outputting an alarm if yes; taking the grounding section and grounding phase of the current event as a keyword to construct a transient grounding event sequence in combination with the current event and four previous transient grounding history events; extracting multi-dimensional features to construct a feature vector of each event for the waveform data of each event in the sequence; inputting the feature vector sequence formed by the transient grounding event sequence into a trained hidden Markov model group, and outputting the insulation defect state of each feature dimension of the current event; and using a decision tree to infer, determine the insulation defect state corresponding to the current event, and output an alarm according to the state.

Description

Technical Field

[0001] This invention relates to the field of early warning technology for insulation defects in distribution networks, and more specifically, to a method and device for early warning of insulation defects in distribution networks based on a sequence of instantaneous grounding events. Background Technology

[0002] Due to their long lines, numerous branches, and complex network structure, power distribution networks are the most prone to failure. Insulation defects are a major cause of power grid failures. Generally, insulation defect deterioration has a development process, which, from a monitoring perspective, is usually manifested in the form of instantaneous grounding. Insulation defect deterioration has distinct stage evolution characteristics, and instantaneous grounding exhibits different energy and time-frequency characteristics. Therefore, studying the relationship between instantaneous grounding and the development of insulation defect deterioration is of great significance for proactive repair of power distribution networks, improving the ability to handle faults in advance, and enhancing power supply reliability.

[0003] In existing technologies, a method has been proposed to monitor electrical quantities through station terminals (or integrated primary and secondary terminals) to provide early warning of cable insulation anomalies and latent faults. However, this method is essentially based on static modeling of waveform data from a single instantaneous grounding event, which makes it difficult to fully utilize the temporal correlation information between events, thus affecting the accuracy and reliability of subsequent identification.

[0004] With the development of technology, the transient judgment method based on distributed synchronous recording has achieved excellent application results in the location of low-current grounding. The accuracy of the location of low-current grounding with grounding resistance of 5kΩ and below can reach 100%, and it can identify whether the grounding type is transient or permanent. This lays a solid foundation for further research on the relationship between transient grounding and the development of insulation defect deterioration. Summary of the Invention

[0005] In view of the above, the present invention provides a method for early warning of insulation defects in distribution networks based on instantaneous grounding event sequences, comprising the following steps:

[0006] S1: Acquire the instantaneous grounding data output by analyzing the grounding event of the distribution network based on distributed synchronous waveform recording and using the transient method, record the data and use it as the current event; the instantaneous grounding data includes: grounding time, grounding section, grounding phase, grounding duration, transient zero-sequence voltage waveform data, and transient zero-sequence current waveform data of the head node of the grounding section;

[0007] S2: Determine whether the number of instantaneous groundings of the above-mentioned grounding section and grounding phase within the preset time limit exceeds the threshold. If so, determine that there is a serious insulation defect in the section and phase, issue an alarm signal and jump to step S7.

[0008] S3: Construct an instantaneous grounding event sequence by combining the current event with the previous four instantaneous grounding historical events, using the current event's grounding section and grounding phase as keywords;

[0009] S4: For the waveform data of each instantaneous grounding event in the sequence, calculate, record, and extract multidimensional energy and time-frequency features to construct the feature vector of each instantaneous grounding event; the multidimensional energy and time-frequency features include: grounding duration, zero-sequence voltage peak value, zero-sequence voltage mean value, zero-sequence current peak value, zero-sequence current mean value, zero-sequence current crest coefficient, zero-sequence current over-limit half-wave number, and zero-sequence electrical energy;

[0010] S5: Input the feature vector sequence formed by the above event sequence into a pre-trained Hidden Markov Model group, decode it using the Viterbi algorithm, and output the insulation defect state corresponding to each feature dimension of the current event; the insulation defect state includes: severe and general.

[0011] S6: Input the feature dimension label vector of the current event and the corresponding insulation defect state vector into the pre-trained decision tree for inference, and determine whether there is a serious insulation defect in the grounding section and grounding phase corresponding to the current instantaneous grounding event. If so, issue an alarm signal.

[0012] S7: Report alarm signals and related data to the host system according to communication conditions.

[0013] This invention utilizes instantaneous grounding data obtained by analyzing grounding events in distribution networks using a transient method based on distributed synchronous waveform recording to provide early warning of insulation defects in distribution networks. The technical details of the transient analysis method based on distributed synchronous waveform recording will be described in detail in the embodiments.

[0014] The above instantaneous grounding data includes: grounding time, grounding section, grounding phase, grounding duration, transient zero-sequence voltage waveform data, and transient zero-sequence current waveform data at the head node of the grounding section;

[0015] Record the above instantaneous grounding data and treat it as the current event.

[0016] First, a preliminary judgment is made to determine whether the number of instantaneous groundings of the above-mentioned grounding section and grounding phase within the preset time limit exceeds the threshold. If so, it is determined that there is a serious insulation defect in the section and phase, an alarm signal is issued and an alarm is output, and then the process ends. Here, the preset time limit is set to 10 minutes, and the threshold for the number of instantaneous groundings is set to 3 times.

[0017] Then, using the current event's grounding segment and grounding phase as keywords, and combining the current event with the previous four instantaneous grounding historical events, a sequence of instantaneous grounding events is constructed.

[0018] To avoid interference from distant transient grounding events or transient grounding events caused by previous power grid faults on current events, a transient grounding historical data cleanup strategy is adopted to optimize this problem. The data cleanup strategy includes: first, regularly deleting all transient grounding data history records older than three months; second, once a power grid fault occurs, deleting all transient grounding data history records under the corresponding fault section.

[0019] Based on the above, it can be considered that the instantaneous grounding event sequence constructed using the grounding section and grounding phase as keywords represents the manifestation of the same insulation defect at different times under that section and that phase.

[0020] Next, data processing is performed. For each instantaneous grounding event in the sequence, based on its transient zero-sequence voltage waveform data and the transient zero-sequence current waveform data of the grounding section head node, multi-dimensional energy and time-frequency features are calculated, recorded, and extracted to construct a feature vector for each instantaneous grounding event, including:

[0021] Grounding duration The duration of the instantaneous grounding has been determined using the transient method.

[0022] Zero-sequence voltage peak Take the maximum value of the zero-sequence voltage peak.

[0023] Zero-sequence voltage mean Take the root mean square value of the zero-sequence voltage.

[0024]

[0025] The first one representing the zero-sequence voltage waveform k One value, n This represents the number of waveform sampling points.

[0026] Zero-sequence current peak Take the maximum value of the zero-sequence current peak.

[0027] Zero-sequence current mean Take the root mean square value of the zero-sequence current.

[0028]

[0029] The first one representing the zero-sequence current waveform k One value, n This represents the number of waveform sampling points.

[0030] Zero-sequence current crest coefficient Take the maximum value of the ratio of the half-wave peak value to the mean value of the zero-sequence current;

[0031] Zero-sequence current exceeds half-wavelength : The number of half-waves whose zero-sequence current half-wave peak value exceeds a preset value;

[0032] Zero-sequence electrical energy Take the product of the instantaneous zero-power accumulation value and the sampling interval.

[0033]

[0034] The first one representing the zero-sequence voltage waveform k The absolute value of each value, The first one representing the zero-sequence current waveform k The absolute value of each value, Indicates the sampling interval. n This represents the number of waveform sampling points.

[0035] To avoid redundant calculations, transient grounding events for which feature calculations have already been completed are labeled. For labeled transient grounding events, feature calculations are not performed again, but the features are directly extracted.

[0036] After data processing, the aforementioned feature vectors are extracted from each instantaneous grounding event in the above instantaneous grounding event sequence, forming a feature vector sequence, which in turn forms an observation sequence matrix, represented as follows:

[0037]

[0038] The sequence length is 5. t Indicates the position in the sequence. Indicates the first k One feature in t The value of the position, .

[0039] The above observation sequence matrix is ​​input into a pre-trained Hidden Markov Model group and decoded using the Viterbi algorithm to output the insulation defect state corresponding to each feature dimension of the current event. The insulation defect state includes: severe and general.

[0040] The Hidden Markov Model (HMM) group comprises eight independent Gaussian Hidden Markov (GMM) sub-models, each corresponding to a feature dimension. A separate GMM is trained for each feature dimension, and the model representation is as follows:

[0041]

[0042] in, For the first k The parameter set of the Gaussian Hidden Markov Model corresponding to each feature dimension includes the state transition probability matrix. Observation probability model and initial state probability vector ;

[0043] 1) The set of hidden states in the model is defined as:

[0044]

[0045] These correspond to two states of insulation defects: severe and moderate.

[0046] 2) The state transition probability matrix is ​​expressed as:

[0047]

[0048] in,

[0049]

[0050] Representing the hidden state from each feature dimension q i Transferred to q j The conditional probability;

[0051] 3) A Gaussian distribution is used as the observation probability model, and the observed values... o In the hidden state q j The probability density function under the given condition is expressed as:

[0052]

[0053]

[0054] in, , They are respectively states q j Lower features The mean and standard deviation are used to describe the statistical distribution characteristics of each feature dimension under different insulation defect states when a transient grounding occurs;

[0055] 4) The initial state probability is expressed as:

[0056]

[0057] This represents the prior distribution of the hidden states during model initialization. This indicates that the initial state is in each feature dimension. q i The probability of;

[0058] 5) The observation sequence matrix composed of 8 independent eigenvector sequences is represented as:

[0059]

[0060] The sequence length is 5.t Indicates the position in the sequence. Indicates the first k One feature in t The value of the position, ;

[0061] 6) The corresponding state sequence matrix is ​​represented as:

[0062]

[0063] The sequence length is 5. t Indicates the position in the sequence. Indicates the first k One feature in t The state corresponding to the position, , These correspond to two states of insulation defects: severe and moderate.

[0064] 7) Decode using the Viterbi algorithm to obtain the optimal state sequence inference. We only care about the insulation defect state corresponding to each feature dimension of the current event, represented as:

[0065]

[0066] in, 5 represents the state vector of the current event, and 5 represents the sequence length. For the current event k The state corresponding to each feature , These correspond to two states of insulation defects: severe and moderate.

[0067] The feature dimension label vector of the current event and the corresponding insulation defect state vector are input into a pre-trained decision tree for inference to determine whether there is a serious insulation defect in the corresponding grounding section and grounding phase of the current instantaneous grounding event. If so, an alarm signal is issued.

[0068] The decision tree described above is generated by training using the principle of minimizing the Gini coefficient;

[0069] 1) Assuming the training dataset D have There are categories, and the probability of each category appearing is... The Gini coefficient is then defined as:

[0070]

[0071] 2) For binary classification problems, assume the training dataset is... D There are two categories, and the probability of the first category appearing is... Then the Gini coefficient is:

[0072]

[0073] 3) If features A Each possible value v , training dataset D Divided into and Two parts, then in terms of features Under these conditions, the training dataset D The Gini coefficient is:

[0074]

[0075] 4) Based on the training dataset, starting from the root node, recursively perform the following operations on each node to construct a binary decision tree:

[0076] Step 1, let the node training dataset be... D Calculate each feature for the dataset D The Gini coefficient;

[0077] Step 2: Select the feature with the smallest Gini coefficient and its corresponding split point as the optimal feature and optimal split point, generate two child nodes from the current node, and distribute the training dataset to the two child nodes according to the feature.

[0078] Step 3: Recursively call Step 1 and Step 2 on the two child nodes until the condition is met.

[0079] Step 4: Generate a decision tree.

[0080] Finally, based on the communication conditions, the alarm signal and related data are reported to the upper-level system.

[0081] This invention also provides a distribution network insulation defect early warning device based on a transient grounding event sequence. The device is used to execute the distribution network insulation defect early warning method based on a transient grounding event sequence as described in any of the foregoing embodiments. The device includes:

[0082] The data acquisition unit is used to acquire instantaneous grounding data output by analyzing grounding events in the distribution network based on distributed synchronous waveform recording and using the transient method, record the data and use it as the current event; the instantaneous grounding data includes: grounding time, grounding section, grounding phase, grounding duration, transient zero-sequence voltage waveform data, and transient zero-sequence current waveform data of the head node of the grounding section;

[0083] The insulation defect preliminary judgment unit is used to determine whether the number of instantaneous groundings of the above-mentioned grounding section and grounding phase within a preset time limit exceeds the threshold. If so, it is determined that there is a serious insulation defect in the section and phase, an alarm signal is issued and the alarm is output through the output unit and then the process ends.

[0084] The event sequence construction unit is used to construct an instantaneous grounding event sequence by combining the current event with the previous four instantaneous grounding historical events, using the current event's grounding segment and grounding phase as keywords.

[0085] The feature extraction unit is used to calculate, record, and extract multi-dimensional energy and time-frequency features to construct a feature vector for each instantaneous grounding event in the sequence; the multi-dimensional energy and time-frequency features include: grounding duration, zero-sequence voltage peak value, zero-sequence voltage mean value, zero-sequence current peak value, zero-sequence current mean value, zero-sequence current crest coefficient, zero-sequence current over-limit half-wave number, and zero-sequence electrical energy;

[0086] The Hidden Markov Model Group Analysis Unit is used to input the feature vector sequence formed by the above event sequence into a pre-trained Hidden Markov Model Group, decode it using the Viterbi algorithm, and output the insulation defect state corresponding to each feature dimension of the current event. The insulation defect state includes: severe and general.

[0087] The decision tree determination unit is used to input the feature dimension label vector of the current event and the corresponding insulation defect state vector into a pre-trained decision tree for inference, and to determine whether there is a serious insulation defect in the grounding section and grounding phase corresponding to the current instantaneous grounding event. If so, an alarm signal is issued.

[0088] The output unit is used to report alarm signals and related data to the host system according to communication conditions.

[0089] Compared with existing technologies, this invention proposes a method and device for early warning of insulation defects in distribution networks based on transient grounding event sequences. By making full use of the temporal correlation information between transient grounding events to solve the problem of insulation defect identification and early warning in distribution networks, it ensures more stable insulation defect identification results, locates the early warning of serious insulation defects to specific sections and phases, and is easy to implement in engineering. It helps to proactively repair distribution networks, improve the ability to handle faults in advance, and enhance power supply reliability. Attached Figure Description

[0090] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0091] Figure 1 The flowchart of the method proposed in this invention is shown;

[0092] Figure 2 A schematic diagram of a low-current grounding project monitoring device configuration is shown in an embodiment of the present invention;

[0093] Figure 3 The selection principles and status labeling methods for strongly correlated and weakly correlated event sequences are shown.

[0094] Figure 4 The method of splitting strongly correlated event sequences and weakly correlated event sequences into subsets is shown;

[0095] Figure 5(a) shows the transient waveform of the first instantaneous grounding event in a real-world case.

[0096] Figure 5(b) shows the transient waveform of a second instantaneous grounding event in a real-world case.

[0097] Figure 5(c) shows the transient waveform of the third instantaneous grounding event in a real-world case.

[0098] Figure 5(d) shows the transient waveform of the fourth instantaneous grounding event in a real-world case.

[0099] Figure 5(e) shows the transient waveform of the fifth instantaneous grounding event in a real-world case.

[0100] Figure 5(f) shows the transient waveform of the sixth instantaneous grounding event in a real-world case.

[0101] Figure 5(g) shows the transient waveform of the seventh instantaneous grounding event in a real-world case. Detailed Implementation

[0102] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0103] like Figure 1 As shown, the present invention provides a method for early warning of insulation defects in distribution networks based on instantaneous grounding event sequences, comprising the following steps:

[0104] S1: Acquire the instantaneous grounding data output by analyzing the grounding event of the distribution network based on distributed synchronous waveform recording and using the transient method, record the data and use it as the current event; the instantaneous grounding data includes: grounding time, grounding section, grounding phase, grounding duration, transient zero-sequence voltage waveform data, and transient zero-sequence current waveform data of the head node of the grounding section;

[0105] S2: Determine whether the number of instantaneous groundings of the above-mentioned grounding section and grounding phase within the preset time limit exceeds the threshold. If so, determine that there is a serious insulation defect in the section and phase, issue an alarm signal and jump to step S7.

[0106] S3: Construct an instantaneous grounding event sequence by combining the current event with the previous four instantaneous grounding historical events, using the current event's grounding section and grounding phase as keywords;

[0107] S4: For the waveform data of each instantaneous grounding event in the sequence, calculate, record, and extract multidimensional energy and time-frequency features to construct the feature vector of each instantaneous grounding event; the multidimensional energy and time-frequency features include: grounding duration, zero-sequence voltage peak value, zero-sequence voltage mean value, zero-sequence current peak value, zero-sequence current mean value, zero-sequence current crest coefficient, zero-sequence current over-limit half-wave number, and zero-sequence electrical energy;

[0108] S5: Input the feature vector sequence formed by the above event sequence into a pre-trained Hidden Markov Model group, decode it using the Viterbi algorithm, and output the insulation defect state corresponding to each feature dimension of the current event; the insulation defect state includes: severe and general.

[0109] S6: Input the feature dimension label vector of the current event and the corresponding insulation defect state vector into the pre-trained decision tree for inference, and determine whether there is a serious insulation defect in the grounding section and grounding phase corresponding to the current instantaneous grounding event. If so, issue an alarm signal.

[0110] S7: Report alarm signals and related data to the host system according to communication conditions.

[0111] The following points will be emphasized.

[0112] I. Acquisition of instantaneous grounding data:

[0113] This invention achieves the acquisition of instantaneous grounding data by carrying out a small current grounding line location project (hereinafter referred to as "small current grounding project") based on a PMU-type high-precision line parameter tester at a 110kV substation in Tianfu New Area, Sichuan Province (hereinafter referred to as "Substation A"). The neutral point of the 10kV system of Substation A is grounded through an arc suppression coil, and the 10kV outgoing lines are mainly overhead lines.

[0114] This low-current grounding project conducts real-time monitoring of three busbar sections and 13 subordinate 10kV lines of substation A. A total of 83 monitoring nodes are set up. Three side nodes synchronously sample and record the zero-sequence voltage of the three busbar sections, and the other 80 end nodes are deployed in sections on the 13 10kV lines to synchronously sample and record the three-phase current of each node. The side nodes and end nodes are established in a hierarchical relationship according to the busbar-line.

[0115] Side node equipment (hereinafter referred to as "side equipment") can be station type, installed in ring main unit, or pole type, installed near primary and secondary integrated circuit breaker, sharing working power supply and zero-sequence voltage signal with FTU; end node equipment (hereinafter referred to as "end equipment") can be cable type or overhead type, corresponding to cable line and overhead line application scenarios respectively.

[0116] Both the edge devices and the terminal devices achieve synchronous sampling via BeiDou / GPS timing, with a sampling pulse time error of less than 1 microsecond from UTC. The edge devices are used for synchronous recording of zero-sequence voltage, with a sampling rate of 12.8kHz and a measurement accuracy of 1.0 class. The terminal devices are used for synchronous recording of phase current, with a sampling rate of 12.8kHz and a measurement accuracy of 0.5%. The edge devices and terminal devices establish a connection via 4G / 5G wireless communication. The edge devices also establish a connection with the upper-level system (distribution automation master station) via 4G / 5G wireless communication. The configuration diagram is shown below. Figure 2 As shown. It should be noted that... Figure 2 The schematic diagram of the monitoring device configuration shown is a result cited in the low-current grounding project.

[0117] The edge device uses whether the collected zero-sequence voltage amplitude exceeds a preset threshold as the start condition for the distribution network grounding event. When the start condition is met, the edge device generates a zero-sequence voltage waveform file according to the start time time stamp, and recalls the phase current waveform data of the corresponding end devices of each line node according to the start time time stamp to synthesize the zero-sequence current waveform file of each line node. The zero-sequence voltage waveform file and the zero-sequence current waveform file contain data 100ms before and 180ms after the start time.

[0118] The device comprehensively uses methods such as transient power direction, zero-sequence current waveform similarity, and phase current asymmetry to analyze grounding events. It determines whether the grounding event is permanent or transient based on the grounding duration, i.e., the zero-sequence voltage duration, and outputs relevant data. The output transient grounding data includes at least: grounding time, grounding section, grounding phase, grounding duration, transient zero-sequence voltage waveform, and transient zero-sequence current waveform at the head node of the grounding section.

[0119] This invention is actually the result of a deeper application of this low-current grounding project.

[0120] II. Selection, generation, and labeling of training samples:

[0121] From a monitoring perspective, instantaneous grounding is very likely to occur in distribution networks, but most instantaneous grounding events, especially isolated instantaneous grounding events, can be considered as system disturbances and have no practical research significance.

[0122] Those instantaneous grounding events that occur repeatedly in the same section and phase often have a clear correlation with actual power grid faults. The sequence of instantaneous grounding events in the same section and phase that have a clear correlation with actual power grid faults (sequence length >= 5) is called the "fault-correlated instantaneous grounding event sequence", or simply "strongly correlated event sequence".

[0123] Of course, the occurrence of consecutive instantaneous grounding events in the same section and phase does not necessarily develop into a power grid fault. Those sequences of instantaneous grounding events in the same section and phase (sequence length >= 5) that have no clear correlation with actual power grid faults or have no clear correlation at the moment are referred to here as "fault weakly correlated instantaneous grounding event sequences", or simply "weakly correlated event sequences".

[0124] It can be considered that a strongly correlated event sequence is the manifestation of the same insulation defect at different times under the corresponding segment and phase, while the power grid fault is the final manifestation of the insulation defect, while the insulation defect corresponding to a weakly correlated event sequence has not or has not yet manifested the form of a power grid fault.

[0125] Based on the actual monitoring data of the above-mentioned low-current grounding project, on the one hand, "strongly correlated event sequences" were selected and added to the initial dataset of the training samples and their states were labeled. On the other hand, correspondingly, "weakly correlated event sequences" were also selected and added to the initial dataset of the training samples and their states were labeled.

[0126] The set of labeled states for each instantaneous grounding event in each strongly / weakly correlated event sequence is defined as follows:

[0127]

[0128] These correspond to two states of insulation defects: severe and moderate.

[0129] Figure 3 The selection principles and status labeling methods for strongly correlated and weakly correlated event sequences are shown.

[0130] Table 1 shows an example of a strongly correlated event sequence for a completed status label (partial data).

[0131] Table 2 shows an example of a weakly correlated event sequence for a completed state label (partial data).

[0132] Due to the small scope and short operating time of this low-current grounding project, from October 2024 to October 2025, the project monitored more than 300 instantaneous grounding events. During the same period, the project monitored 20 actual power grid faults, including 17 short-circuit faults, 4 permanent grounding faults (2 of which developed into short-circuit faults through ground), and 1 phase loss fault. Five strongly correlated event sequences and three weakly correlated event sequences were obtained. This sample size was clearly insufficient. Referring to the operating data of 15 other low-current grounding projects with similar neutral grounding methods, outgoing line properties (overhead / cable), and outgoing line scales in the past two years, 136 strongly correlated event sequences and 67 weakly correlated event sequences were collected. Finally, an initial training sample dataset consisting of 141 strongly correlated event sequences and 70 weakly correlated event sequences was formed.

[0133] For each instantaneous grounding event in these 141 strongly correlated event sequences and 70 weakly correlated event sequences, state labeling should be performed according to the actual situation, and 8-dimensional feature vectors should be extracted and recorded.

[0134] Note: (1) Because the data was collected from similar low-current grounding projects, the consistency of waveform data of instantaneous grounding events was ensured, including the number of sampling points, sampling frequency, measurement accuracy, and data length before and after startup; (2) For substations with different neutral grounding methods (ungrounded / grounded through arc suppression coil / grounded with small resistance) or different outgoing line properties (overhead / cable), theoretically, training samples need to be reselected for model training.

[0135] The 141 strongly correlated event sequences were randomly divided into three groups in a ratio of approximately 10:1:1, and the 70 weakly correlated event sequences were randomly divided into three groups in a ratio of approximately 10:1:1. The first group of strongly / weakly correlated event sequences was used to train the Hidden Markov Model, the second group of strongly / weakly correlated event sequences was used to generate the decision tree, and the third group of strongly / weakly correlated event sequences was used for overall testing.

[0136] Table 1. Example of a strongly correlated event sequence for a completion status label (partial data)

[0137]

[0138] Table 2. Example of a weakly correlated event sequence for a completion status label (partial data)

[0139]

[0140] III. Training of Hidden Markov Model Groups:

[0141] Supervised learning is employed, using the first set of strongly / weakly correlated event sequences as training samples. For each feature dimension, the parameters of the corresponding Gaussian Hidden Markov Model are estimated using the maximum likelihood method.

[0142] 1) Transition probability Estimate:

[0143] After repeated testing, the transition probability was determined. The estimate is:

[0144]

[0145]

[0146] .

[0147] 2) Estimation of observation probability model parameters:

[0148] The observation probability model is expressed as:

[0149]

[0150]

[0151] for That is, features labeled as "severe" The sample mean, ;

[0152] for That is, features labeled as "severe" The sample standard deviation ;

[0153] for That is, features labeled as "normal" The sample mean, ;

[0154] for That is, features labeled as "normal" The sample standard deviation .

[0155] 3) Initial state probability Estimate:

[0156] After repeated testing, the prior distribution of the hidden states is estimated as follows:

[0157]

[0158] .

[0159] IV. Decision Tree Training:

[0160] The second set of strongly / weakly correlated event sequences is used to generate the decision tree.

[0161] 1) First, split each strongly / weakly correlated event sequence into subsets. Each subset contains 5 consecutive instantaneous grounding events, and each subset serves as a training sample for a decision tree.

[0162] Figure 4 The method of splitting strongly correlated event sequences and weakly correlated event sequences into subsets is shown.

[0163] Taking the strongly correlated event sequences described in Table 1 as an example, after subset splitting, two training samples are obtained:

[0164] The labeled state sequence of sample 1 is as follows: ， Status Mark These correspond to two actual states of insulation defects: severe and moderate.

[0165] The observation sequence matrix for sample 1 is as follows:

[0166]

[0167] The sequence length is 5. t Indicates the position in the sequence. Indicates sample 1 k One feature in t The value of the position, .

[0168] The labeled state sequence of sample 2 is as follows: ， Status Mark These correspond to two actual states of insulation defects: severe and moderate.

[0169] The observation sequence matrix for sample 2 is as follows:

[0170]

[0171] The sequence length is 5. t Indicates the position in the sequence. Indicates sample 2 k One feature in t The value of the position, .

[0172] 2) Input the observation sequence matrix of each decision tree training sample into the Gaussian Hidden Markov Model group whose parameters have been estimated using the maximum likelihood method, and perform state decoding using the Viterbi algorithm to obtain the optimal state sequence inference. The insulation defect state corresponding to each feature dimension of each training sample is combined with the actual labeled state to obtain the mode output as shown in Table 3:

[0173] Table 3. Decision tree training samples after decoding by the Gaussian Hidden Markov Model group Table comparing the insulation defect state of each training sample with the actual labeled state for each feature dimension

[0174]

[0175] For each training sample Time k The state corresponding to each feature , These correspond to two states of insulation defects: severe and moderate. The actual marked states are manually labeled based on the actual situation. See the method for marking states. Figure 3 As shown.

[0176] 3) The decision tree is generated as follows:

[0177] Input: Training dataset D =Table 3 Dataset, Feature Dimension Label Set The condition for stopping the calculation;

[0178] Based on the training dataset, starting from the root node, recursively perform the following operations on each node to construct a binary decision tree:

[0179] Step 1, let the node training dataset be... D Calculate the relationship between each feature dimension in the feature dimension label set and the dataset. D Gini coefficient:

[0180]

[0181]

[0182]

[0183] This indicates a sample with a status label of 1, meaning the insulation defect status is marked as severe. This indicates that the status is marked as 2, meaning the insulation defect status is marked as general.

[0184] Step 2: Select the feature with the smallest Gini coefficient and its corresponding split point as the optimal feature and optimal split point. Generate two child nodes from the current node, distribute the training dataset to the two child nodes according to the feature, and remove the optimal feature label from the feature dimension label set.

[0185] Step 3: Recursively call Step 1 and Step 2 on the two child nodes until the condition is met.

[0186] Step 4: Generate a decision tree.

[0187] The algorithm stops computation when the number of samples in a node is less than a predetermined threshold, or the Gini coefficient of the sample set is less than a predetermined threshold, or there are no more features.

[0188] V. Overall Model Testing:

[0189] The third set of strongly / weakly correlated event sequences was used for overall model testing.

[0190] First, each strongly / weakly correlated event sequence is divided into subsets, with each subset containing 5 consecutive instantaneous grounding events, and each subset serves as a test sample.

[0191] Figure 4 The method of splitting strongly correlated event sequences and weakly correlated event sequences into subsets is shown.

[0192] The observation sequence matrix of each test sample is input into a Gaussian Hidden Markov Model group whose parameters have been estimated using the maximum likelihood method. The Viterbi algorithm is then used for state decoding to obtain the optimal state sequence. Then, each test sample... Each feature dimension label and its corresponding insulation defect state are input into a pre-generated decision tree for inference, and the inference results are compared with the actual situation.

[0193] The training and testing processes described above can be largely completed by computer. The number of groups can also be adjusted appropriately. After repeated training, testing and parameter tuning, the prediction success rate for test samples is 91% to 93%, which is limited by the initial dataset size of the training samples.

[0194] VI. Practical Application Cases:

[0195] The trained Gaussian Hidden Markov Model and decision tree were applied to the aforementioned low-current grounding project, and a serious insulation defect was successfully predicted in December 2025.

[0196] Between November 14 and December 24, 2025, this low-current grounding project continuously reported instantaneous grounding of phase A after the Wangchang branch line #1 of the Songjie line from busbar IV section of substation A, as shown in the table below:

[0197] Table 4. Sequence of instantaneous grounding events in actual cases

[0198]

[0199] The transient waveforms of the seven instantaneous grounding events are shown in Figures 5(a) to 5(g).

[0200] The model works as follows:

[0201] At 11:49 on December 16, 2025, after the instantaneous grounding event of phase A after No. 1 of the Wangchang branch of the Songjie line was obtained, the observation sequence length was 5. It was determined whether the number of instantaneous grounding events of this section and phase within the preset time limit exceeded the threshold. After obtaining a negative conclusion, the instantaneous grounding event sequence was formed by combining the previous 4 instantaneous grounding events of phase A after No. 1 of the Wangchang branch of the Songjie line. An observation sequence matrix was generated, and model analysis was started. After decoding by the Gaussian Hidden Markov Model Group and inference by the decision tree, the insulation defect state was obtained as 2.

[0202] At 21:10 on December 20, 2025, after receiving the instantaneous grounding event of phase A after No. 1 of the Wangchang branch of the Songjie line, it was determined whether the number of instantaneous grounding events of the grounding section and the grounding phase within the preset time limit exceeded the threshold. After obtaining a negative conclusion, the instantaneous grounding event sequence was formed by combining the previous four instantaneous grounding events of phase A after No. 1 of the Wangchang branch of the Songjie line, and an observation sequence matrix was generated. The model analysis was started. After decoding by the Gaussian Hidden Markov Model Group and inference by the decision tree, the insulation defect state was obtained as 2.

[0203] At 11:46 on December 24, 2025, after receiving the instantaneous grounding event of phase A after No. 1 of the Wangchang branch of the Songjie Line, it was determined whether the number of instantaneous grounding events of the grounding section and the grounding phase within the preset time limit exceeded the threshold. After obtaining a negative conclusion, the instantaneous grounding event sequence was formed by combining the previous four instantaneous grounding events of phase A after No. 1 of the Wangchang branch of the Songjie Line, and an observation sequence matrix was generated. The model analysis was started. After decoding by the Gaussian Hidden Markov Model Group and inference by the decision tree, the insulation defect status was obtained as 1. It was determined that there was a serious insulation defect in phase A after No. 1 of the Wangchang branch of the Songjie Line and an alarm was output.

[0204] Upon inspection, the drop-out fuse of phase A at the rear end of line 1 of Wangchang branch line was found to have blown, confirming the accuracy of the model prediction.

[0205] Based on the above-described methodological concept, this invention also provides a distribution network insulation defect early warning device based on a transient grounding event sequence, used to execute the distribution network insulation defect early warning method based on a transient grounding event sequence as described in any of the foregoing embodiments. The device includes:

[0206] The data acquisition unit is used to acquire instantaneous grounding data output by analyzing grounding events in the distribution network based on distributed synchronous waveform recording and using the transient method, record the data and use it as the current event; the instantaneous grounding data includes: grounding time, grounding section, grounding phase, grounding duration, transient zero-sequence voltage waveform data, and transient zero-sequence current waveform data of the head node of the grounding section;

[0207] The insulation defect preliminary judgment unit is used to determine whether the number of instantaneous groundings of the above-mentioned grounding section and grounding phase within a preset time limit exceeds the threshold. If so, it is determined that there is a serious insulation defect in the section and phase, an alarm signal is issued and the alarm is output through the output unit and then the process ends.

[0208] The event sequence construction unit is used to construct an instantaneous grounding event sequence by combining the current event with the previous four instantaneous grounding historical events, using the current event's grounding segment and grounding phase as keywords.

[0209] The feature extraction unit is used to calculate, record, and extract multi-dimensional energy and time-frequency features to construct a feature vector for each instantaneous grounding event in the sequence; the multi-dimensional energy and time-frequency features include: grounding duration, zero-sequence voltage peak value, zero-sequence voltage mean value, zero-sequence current peak value, zero-sequence current mean value, zero-sequence current crest coefficient, zero-sequence current over-limit half-wave number, and zero-sequence electrical energy;

[0210] The Hidden Markov Model Group Analysis Unit is used to input the feature vector sequence formed by the above event sequence into a pre-trained Hidden Markov Model Group, decode it using the Viterbi algorithm, and output the insulation defect state corresponding to each feature dimension of the current event. The insulation defect state includes: severe and general.

[0211] The decision tree determination unit is used to input the feature dimension label vector of the current event and the corresponding insulation defect state vector into a pre-trained decision tree for inference, and to determine whether there is a serious insulation defect in the grounding section and grounding phase corresponding to the current instantaneous grounding event. If so, an alarm signal is issued.

[0212] The output unit is used to report alarm signals and related data to the host system according to communication conditions.

[0213] The above is a description of the invention and should not be considered as a limitation thereof. Although several exemplary embodiments of the invention have been described, those skilled in the art will readily understand that many modifications can be made to the exemplary embodiments without departing from the novel teachings and advantages of the invention. Therefore, all such modifications are intended to be included within the scope of the invention as defined in the claims.

Claims

1. A method for early warning of insulation defects in distribution networks based on instantaneous grounding event sequences, characterized in that, Includes the following steps: S1: Obtain the instantaneous grounding data output by analyzing the grounding event of the distribution network based on distributed synchronous waveform recording and transient method, record the data and use it as the current event; The instantaneous grounding data includes: grounding time, grounding section, grounding phase, grounding duration, transient zero-sequence voltage waveform data, and transient zero-sequence current waveform data at the head node of the grounding section; S2: Determine whether the number of instantaneous groundings of the above-mentioned grounding section and grounding phase within the preset time limit exceeds the threshold. If so, determine that there is a serious insulation defect in the section and phase, issue an alarm signal and jump to step S7. S3: Construct an instantaneous grounding event sequence by combining the current event with the previous four instantaneous grounding historical events, using the current event's grounding section and grounding phase as keywords; S4: For the waveform data of each instantaneous grounding event in the sequence, calculate, record, and extract multidimensional energy and time-frequency features to construct the feature vector of each instantaneous grounding event; the multidimensional energy and time-frequency features include: grounding duration, zero-sequence voltage peak value, zero-sequence voltage mean value, zero-sequence current peak value, zero-sequence current mean value, zero-sequence current crest coefficient, zero-sequence current over-limit half-wave number, and zero-sequence electrical energy; S5: Input the feature vector sequence formed by the above event sequence into a pre-trained Hidden Markov Model group, decode it using the Viterbi algorithm, and output the insulation defect state corresponding to each feature dimension of the current event; the insulation defect state includes: severe and general. S6: Input each feature dimension of the current event and the corresponding insulation defect status into the pre-trained decision tree for inference, and determine whether there is a serious insulation defect in the grounding section and grounding phase corresponding to the current instantaneous grounding event. If so, issue an alarm signal. S7: Report alarm signals and related data to the host system according to communication conditions.

2. The method for early warning of insulation defects in distribution networks based on instantaneous grounding event sequences according to claim 1, characterized in that: In step S1, the distributed synchronous waveform recording refers to the distributed synchronous sampling and recording achieved through BeiDou / GPS timing, with the sampling pulse and UTC time error being less than 1 microsecond and the sampling rate being no less than 12.8kHz. The transient method refers to the use of the transient power direction and zero-sequence current waveform similarity method for zero-sequence voltage start-up. The waveform data includes at least 80ms before and 160ms after the start-up time.

3. The method for early warning of insulation defects in distribution networks based on instantaneous grounding event sequences according to claim 1, characterized in that: In step S2, the preset time limit is set to 10 minutes, and the threshold for the number of instantaneous grounding events is set to 3.

4. The method for early warning of insulation defects in distribution networks based on instantaneous grounding event sequences according to claim 1, characterized in that: Step S3 is optimized by adopting a transient grounding historical data cleanup strategy to avoid interference from distant transient grounding historical events or transient grounding historical events caused by the previous power grid fault on the current event; The data cleanup strategy includes: first, regularly deleting all instantaneous grounding data history records older than three months; and second, deleting all instantaneous grounding data history records in the corresponding fault section once a power grid fault occurs.

5. The method for early warning of insulation defects in distribution networks based on instantaneous grounding event sequences according to claim 1, characterized in that: In step S4, the multidimensional energy and time-frequency features are calculated and extracted using the following method: Grounding duration The duration of the instantaneous grounding has been determined using the transient method. Zero-sequence voltage peak Take the maximum value of the zero-sequence voltage peak. Zero-sequence voltage mean Take the root mean square value of the zero-sequence voltage. The first one representing the zero-sequence voltage waveform k One value, n This represents the number of waveform sampling points. Zero-sequence current peak Take the maximum value of the zero-sequence current peak. Zero-sequence current mean Take the root mean square value of the zero-sequence current. The first one representing the zero-sequence current waveform k One value, n This represents the number of waveform sampling points. Zero-sequence current crest coefficient Take the maximum value of the ratio of the half-wave peak value to the mean value of the zero-sequence current; Zero-sequence current exceeds half-wavelength : The number of half-waves whose zero-sequence current half-wave peak value exceeds a preset value; Zero-sequence electrical energy Take the product of the instantaneous zero-power accumulation value and the sampling interval. The first one representing the zero-sequence voltage waveform k The absolute value of each value, The first one representing the zero-sequence current waveform k The absolute value of each value, Indicates the sampling interval. n This represents the number of waveform sampling points.

6. The method for early warning of insulation defects in distribution networks based on instantaneous grounding event sequences according to claim 1, characterized in that: In step S4, instantaneous grounding events for which feature calculations have been completed are marked to avoid duplicate calculations.

7. The method for early warning of insulation defects in distribution networks based on instantaneous grounding event sequences according to claim 1, characterized in that: In step S5, the Hidden Markov Model group comprises 8 independent Gaussian Hidden Markov sub-models, each corresponding to a feature dimension. A Gaussian Hidden Markov Model is trained for each feature dimension, and the model is represented as follows: in, For the first k The parameter set of the Gaussian Hidden Markov Model corresponding to each feature dimension includes the state transition probability matrix. Observation probability model and initial state probability vector ; 1) The set of hidden states in the model is defined as: These correspond to two states of insulation defects: severe and moderate. 2) The state transition probability matrix is ​​expressed as: in, Representing the hidden state from each feature dimension q i Transferred to q j The conditional probability; 3) A Gaussian distribution is used as the observation probability model, and the observed values... o In the hidden state q j The probability density function under the given condition is expressed as: in, , Hidden state q j Lower features The mean and standard deviation are used to describe the statistical distribution characteristics of each feature dimension under different insulation defect states when a transient grounding occurs; 4) The initial state probability is expressed as: This represents the prior distribution of the hidden states during model initialization. This indicates that the initial state is in each feature dimension. q i The probability of; 5) The observation sequence matrix composed of 8 independent eigenvector sequences is represented as: The sequence length is 5. t Indicates the position in the sequence. Indicates the first k One feature in t The value of the position, ; 6) The corresponding state sequence matrix is ​​represented as: The sequence length is 5. t Indicates the position in the sequence. Indicates the first k One feature in t The state corresponding to the position, , These correspond to two states of insulation defects: severe and moderate. 7) Decode using the Viterbi algorithm to obtain the optimal state sequence inference. We only care about the insulation defect state corresponding to each feature dimension of the current event, represented as: Where 5 is the sequence length. For the current event k The state corresponding to each feature , These correspond to two states of insulation defects: severe and moderate.

8. The method for early warning of insulation defects in distribution networks based on instantaneous grounding event sequences according to claim 1, characterized in that: In step S6, the decision tree is trained and generated using the principle of minimizing the Gini coefficient; 1) Assuming the training dataset D have There are categories, and the probability of each category appearing is... The Gini coefficient is then defined as: 2) For binary classification problems, assume the training dataset is... D There are two categories, and the probability of the first category appearing is... Then the Gini coefficient is: 3) If features A Each possible value v , training dataset D Divided into and Two parts, then in terms of features Under these conditions, the training dataset D The Gini coefficient is: 4) Based on the training dataset, starting from the root node, recursively perform the following operations on each node to construct a binary decision tree: Step 1, let the node training dataset be... D Calculate each feature for the dataset D The Gini coefficient; Step 2: Select the feature with the smallest Gini coefficient and its corresponding split point as the optimal feature and optimal split point, generate two child nodes from the current node, and distribute the training dataset to the two child nodes according to the feature. Step 3: Recursively call Step 1 and Step 2 on the two child nodes until the condition is met. Step 4: Generate a decision tree.

9. The method for early warning of insulation defects in distribution networks based on instantaneous grounding event sequences according to claim 5, characterized in that: The preset value is set to 50A.

10. A distribution network insulation defect early warning device based on instantaneous grounding event sequence, characterized in that, The apparatus is used to execute the distribution network insulation defect early warning method based on instantaneous grounding event sequence as described in any one of claims 1-9, the apparatus comprising: The data acquisition unit is used to acquire instantaneous grounding data output by analyzing grounding events in the distribution network based on distributed synchronous waveform recording and using the transient method, record the data and use it as the current event; the instantaneous grounding data includes: grounding time, grounding section, grounding phase, grounding duration, transient zero-sequence voltage waveform data, and transient zero-sequence current waveform data of the head node of the grounding section; The insulation defect preliminary judgment unit is used to determine whether the number of instantaneous groundings of the above-mentioned grounding section and grounding phase within a preset time limit exceeds the threshold. If so, it is determined that there is a serious insulation defect in the section and phase, an alarm signal is issued and the alarm is output through the output unit and then the process ends. The event sequence construction unit is used to construct an instantaneous grounding event sequence by combining the current event with the previous four instantaneous grounding historical events, using the current event's grounding segment and grounding phase as keywords. The feature extraction unit is used to calculate, record, and extract multi-dimensional energy and time-frequency features to construct a feature vector for each instantaneous grounding event in the sequence; the multi-dimensional energy and time-frequency features include: grounding duration, zero-sequence voltage peak value, zero-sequence voltage mean value, zero-sequence current peak value, zero-sequence current mean value, zero-sequence current crest coefficient, zero-sequence current over-limit half-wave number, and zero-sequence electrical energy; The Hidden Markov Model Group Analysis Unit is used to input the feature vector sequence formed by the above event sequence into a pre-trained Hidden Markov Model Group, decode it using the Viterbi algorithm, and output the insulation defect state corresponding to each feature dimension of the current event. The insulation defect state includes: severe and general. The decision tree determination unit is used to input each feature dimension of the current event and the corresponding insulation defect status into a pre-trained decision tree for inference, and to determine whether there is a serious insulation defect in the grounding section and grounding phase corresponding to the current instantaneous grounding event. If so, an alarm signal is issued. The output unit is used to report alarm signals and related data to the host system according to communication conditions.

Images