A power distribution network harmonic detection method and system based on smart meter data

Abstract

The present application relates to the field of power distribution network harmonic detection, and more particularly to a power distribution network harmonic detection method and system based on smart meter data, wherein the method comprises: obtaining multi-phase voltage, current waveform and active power, reactive power, power factor and other original power consumption data uploaded by smart meters at multiple monitoring points of the power distribution network; using an improved spectral clustering algorithm optimized according to harmonic propagation and attenuation characteristics to measure similarity, to identify potential harmonic source access points and pollution time periods; performing high-precision spectral analysis on waveform data in the pollution time periods, extracting harmonic characteristic parameters, inputting the harmonic state evaluation network to generate a harmonic state atlas, outputting responsibility quantization indicators and governance priority suggestions through a correlation model, and converting the indicators and suggestions into a structured report to be pushed to a management system. The method can accurately identify harmonic sources, comprehensively grasp harmonic distribution, quantify pollution responsibility, and improve detection and governance efficiency without the need for additional hardware investment.

Description

Technical Field

[0001] This invention relates to the field of harmonic detection in power distribution networks. In particular, it relates to a method and system for harmonic detection in power distribution networks based on smart meter data. Background Technology

[0002] Harmonic pollution in distribution networks can disrupt the operational stability of power equipment, reduce power quality, and interfere with the stable operation of the power system. Therefore, harmonic detection is a necessary part of distribution network operation and maintenance. Currently, harmonic detection in distribution networks mainly relies on dedicated harmonic monitoring equipment. This equipment is costly to deploy, and the collected data is mostly limited to basic data such as multiphase voltage and current waveforms, lacking the integrated utilization of multi-dimensional power consumption data such as active power, reactive power, and power factor. Existing detection methods generally use traditional spectral clustering algorithms to process the collected waveform data, extract harmonic features through conventional spectrum analysis, determine the harmonic pollution status using simple analytical models, and finally push the detection results to the distribution network management system in the form of a simplified report.

[0003] Traditional spectral clustering algorithms fail to consider the topological characteristics of distribution networks in their similarity metrics, making it impossible to match the attenuation patterns of harmonic propagation within the network. This leads to inaccurate identification of harmonic source access points and pollution periods, resulting in false positives and false negatives. Furthermore, existing detection methods rely solely on harmonic characteristic parameters for analysis, neglecting to integrate multi-dimensional electricity consumption data. This makes it difficult to comprehensively understand the harmonic distribution within the distribution network, quantify the pollution responsibility of each potential harmonic source, and provide reasonable prioritization for remediation. Therefore, a distribution network harmonic detection solution is needed that can accurately identify harmonic source access points and pollution periods, comprehensively integrate multi-dimensional electricity consumption data, quantify harmonic source responsibility, and provide remediation priority recommendations. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a method and system for detecting harmonics in power distribution networks based on smart meter data.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for detecting harmonics in a distribution network based on smart meter data, comprising: The raw electricity consumption data set uploaded by smart meters deployed at multiple monitoring points in the power distribution network is obtained. The raw electricity consumption data set includes multiphase voltage waveform data, multiphase current waveform data, active power data, reactive power data, and power factor data. An improved spectral clustering algorithm is applied to the multiphase voltage waveform data and multiphase current waveform data to identify potential harmonic source access points and harmonic pollution periods. The improved spectral clustering algorithm optimizes the similarity metric based on the propagation attenuation characteristics of harmonics in the distribution network topology. High-precision spectrum analysis is performed on the voltage and current waveform data during the identified harmonic pollution period to extract the amplitude, phase, harmonic distortion rate and harmonic power characteristics of each harmonic, forming a set of harmonic characteristic parameters; The set of harmonic characteristic parameters, the active power data, the reactive power data, and the power factor data are input into the harmonic state assessment network to generate a harmonic state map describing the overall harmonic distribution state and pollution severity of the distribution network. Based on the aforementioned harmonic state spectrum, the responsibility quantification index and governance priority suggestions for each potential harmonic source are calculated and output through the harmonic source location and responsibility allocation model. The responsibility quantification indicators and governance priority recommendations for the harmonic sources are converted into a structured harmonic detection report and pushed to the distribution network management system.

[0006] Furthermore, an improved spectral clustering algorithm is applied to the multiphase voltage waveform data and multiphase current waveform data to identify potential harmonic source access points and harmonic pollution periods, including: The multiphase voltage waveform data and multiphase current waveform data of each monitoring point are preprocessed to extract time-domain and frequency-domain features that characterize waveform distortion, forming a waveform feature vector for each monitoring point. Based on the topological connections and electrical distances of the power distribution network, a connection graph is constructed between monitoring points. In the graph, nodes represent monitoring points, and the weights of the edges are determined by the electrical distances and connections. The improved spectral clustering algorithm is used to cluster the waveform feature vectors of all monitoring points. When calculating the similarity between feature vectors, the improved spectral clustering algorithm not only considers the Euclidean distance of the feature vectors themselves, but also corrects the similarity according to the weight of the edges in the connection graph, so that monitoring points that are electrically adjacent and have similar waveform features are clustered into one class. After clustering is completed, each cluster is considered to correspond to the influence range of one or more harmonic sources. The change of the cluster center in the time dimension is used to identify the period of harmonic activity, thereby determining the potential harmonic source access point and the period of harmonic pollution.

[0007] Furthermore, the high-precision spectrum analysis of the voltage and current waveform data during the identified harmonic pollution period, extracting the amplitude, phase, harmonic distortion rate, and harmonic power characteristics of each harmonic, includes: The voltage and current waveform data during the harmonic pollution period are resampled and windowed to reduce spectral leakage. High-resolution fast Fourier transforms are performed on the processed voltage waveform data and current waveform data respectively to obtain the voltage spectrum and current spectrum; In the voltage and current spectra, the position of the fundamental frequency is identified, and based on this, the integer harmonic frequency components are located. Extract the voltage amplitude, current amplitude, voltage phase, and current phase corresponding to each integer harmonic frequency component; Based on the extracted amplitude and phase, calculate the harmonic content, total harmonic distortion, power of each harmonic, and harmonic directional power. The characteristic calculation results of all harmonics are collected to form the set of harmonic characteristic parameters.

[0008] Furthermore, the set of harmonic characteristic parameters, the active power data, the reactive power data, and the power factor data are input into the harmonic state assessment network to generate a harmonic state map describing the overall harmonic distribution and pollution severity of the distribution network, including: The harmonic state assessment network includes a feature encoding layer, a spatial correlation layer, and a spectrum generation layer. The feature coding layer receives the set of harmonic feature parameters, active power data, reactive power data and power factor data of each monitoring point, and encodes them into the node state vector of each monitoring point. The spatial association layer receives the topology data of the distribution network and the node state vectors of all monitoring points. Using a graph neural network model, it simulates the propagation and superposition process of harmonic features on the distribution network topology and updates the node state vector of each monitoring point to include neighborhood information. The map generation layer will spatially arrange the updated node state vectors of all monitoring points according to the actual geographical or electrical location of the monitoring points, and use interpolation technology to generate a continuous and visualized harmonic state map covering the entire power distribution network area. In the harmonic state map, different colors or contour lines represent the severity of harmonic pollution at different locations.

[0009] Furthermore, based on the aforementioned harmonic state spectrum, a harmonic source location and responsibility allocation model is used to calculate and output quantitative responsibility indicators and governance priority recommendations for each potential harmonic source, including: In the harmonic state spectrum, the harmonic source region is identified based on the gradient change of the severity of harmonic pollution; Extract the harmonic phase and harmonic power flow information of the monitoring points located in the harmonic source region from the set of harmonic characteristic parameters; Using a harmonic power flow calculation model, combined with the distribution network topology and impedance parameters, reverse power flow tracing is performed to estimate the magnitude of the injected harmonic current from each potential harmonic source. Based on the estimated magnitude of the injected harmonic current of each potential harmonic source and its contribution ratio to the total harmonic pollution, the responsibility quantification index of each potential harmonic source is calculated. By combining the pollution range shown in the harmonic state spectrum, the magnitude of the responsibility quantification index, and the user type of each potential harmonic source, the governance priority recommendation is generated through preset rules or scoring models.

[0010] Furthermore, when calculating the similarity between feature vectors, the improved spectral clustering algorithm not only considers the Euclidean distance between the feature vectors themselves, but also corrects the similarity based on the weights of the edges in the connectivity graph, including: For any two monitoring points, calculate their Euclidean distance using waveform feature vectors and convert it into an initial similarity based on the Gaussian kernel function; Obtain the weight of the edge connecting the two monitoring points in the connection graph, where the weight is inversely proportional to the electrical distance; Design a similarity enhancement function based on edge weights. The larger the edge weight between two monitoring points, the closer the electrical connection or the closer the distance. The greater the enhancement of the initial similarity by the similarity enhancement function. The value corrected by the similarity enhancement function is used as the similarity between the waveform feature vectors of the two monitoring points in the final improved spectral clustering algorithm.

[0011] Furthermore, the design of a similarity enhancement function based on edge weights includes: Define an edge weight threshold. When the edge weight connecting two monitoring points is greater than the threshold, the two monitoring points are considered to be electrically closely connected. For electricalally closely related monitoring point pairs, the similarity correction formula is that the enhanced similarity is equal to the initial similarity multiplied by an enhancement coefficient greater than one, where the enhancement coefficient is a monotonically increasing function of the edge weights. For monitoring point pairs that are not closely related by electricity, the enhanced similarity is equal to the initial similarity.

[0012] Furthermore, the spatial association layer receives the topology data of the distribution network and the node state vectors of all monitoring points, and uses a graph neural network model to simulate the propagation and superposition process of harmonic characteristics on the distribution network topology, including: The distribution network topology is abstracted as a graph, where nodes are monitoring points or power grid buses, and edges are feeders or transformer branches. The node state vector of each monitoring point is used as the initial feature of the corresponding graph node; The graph neural network model consists of multiple graph convolutional layers. Each graph convolutional layer performs the following operations: aggregates the features of all neighboring nodes of each node, and updates the node's features after linear transformation and nonlinear activation. Through multi-layer graph convolution operations, the feature information of each node is propagated and interacted along the topological edges in multiple hops.

[0013] Furthermore, the method of using a harmonic power flow calculation model, combined with distribution network topology and impedance parameters, to perform reverse power flow tracing and estimate the magnitude of injected harmonic current from each potential harmonic source includes: Construct a harmonic impedance network model of the target distribution network at a specific harmonic frequency; The harmonic voltages of each monitoring point identified in the harmonic state spectrum are taken as known node voltages; The location of potential harmonic sources is considered as the source node into which harmonic currents are injected; Based on the harmonic impedance network model, the harmonic voltage equations of the nodes are established, where the harmonic voltage at each monitoring point is a known quantity, and the injected harmonic current at each potential harmonic source is a quantity to be determined.

[0014] Furthermore, the present invention also includes a distribution network harmonic detection system based on smart meter data. The system includes a processor and a memory, the memory and the processor being connected. The memory is used to store programs, instructions or code, and the processor is used to run the programs, instructions or code in the memory to implement the distribution network harmonic detection method based on smart meter data as described above.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: Based on the propagation and attenuation characteristics of harmonics in distribution network topology, a similarity metric is optimized for processing multiphase voltage and current waveform data. This algorithm accurately captures the attenuation patterns of harmonic propagation in the distribution network, making the similarity metric more closely aligned with actual distribution network operation scenarios. It effectively avoids the identification bias caused by traditional spectral clustering algorithms that do not consider this characteristic, and can more accurately pinpoint potential harmonic source access points and harmonic pollution periods, reducing false positives and false negatives, thus making harmonic detection more targeted. This optimized algorithm is fully adaptable to the topological characteristics of distribution networks, improving the accuracy of waveform data processing without requiring additional hardware. This makes subsequent harmonic feature extraction more targeted, providing more reliable basic data support for subsequent harmonic state analysis.

[0016] By simultaneously inputting multi-dimensional data, including harmonic characteristic parameters, active power data, reactive power data, and power factor data, into the harmonic state assessment network, the network can integrate the correlation information of multi-source power consumption data to generate a harmonic state map that comprehensively reflects the overall harmonic distribution and pollution severity of the distribution network. This overcomes the limitations of conventional detection methods that rely solely on a single harmonic characteristic parameter. Based on a harmonic source location and responsibility allocation model, the responsibility of each potential harmonic source can be quantitatively calculated, outputting clear governance priority recommendations. This allows harmonic governance to be carried out systematically according to the severity of responsibility, avoiding blind governance and improving the rationality and efficiency of harmonic governance. The integrated use of multi-dimensional data can compensate for the limitations of single-data analysis, comprehensively presenting the overall situation of harmonic pollution in the distribution network. Quantitative responsibility and priority allocation make governance work more directional, improving governance effectiveness without incurring additional governance costs. Attached Figure Description

[0017] The following sections will describe some specific embodiments of the invention in a detailed manner by way of example and not limitation, with reference to the accompanying drawings. The same reference numerals in the drawings denote the same or similar parts or portions. Those skilled in the art should understand that these drawings are not necessarily drawn to scale. In the drawings: Figure 1 This is a flowchart of a distribution network harmonic detection method based on smart meter data, as described in this invention. Figure 2 This is a flowchart of an improved spectral clustering algorithm for identifying harmonic source access points and harmonic pollution periods; Figure 3 This is a flowchart for high-precision spectrum analysis to extract harmonic features. Detailed Implementation

[0018] The following reference Figures 1 to 3 This invention describes a method and system for detecting harmonics in a distribution network based on smart meter data, according to an embodiment of the present invention. In this description, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature, that is, include one or more of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. When a feature "includes or contains" one or more of the features it encompasses, unless otherwise specifically described, this indicates that other features are not excluded and may be further included.

[0019] In the description of this embodiment, the terms "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0020] See Figure 1 A general implementation scheme for a distribution network harmonic detection method based on smart meter data involves acquiring a set of raw electricity consumption data uploaded by smart meters deployed at multiple monitoring points in the distribution network. This raw electricity consumption data set includes multiphase voltage waveform data, multiphase current waveform data, active power data, reactive power data, and power factor data. An improved spectral clustering algorithm is then applied to the multiphase voltage and current waveform data to identify potential harmonic source access points and harmonic pollution periods. This improved spectral clustering algorithm optimizes the similarity metric based on the propagation and attenuation characteristics of harmonics in the distribution network topology. Finally, the voltage and current waveform data within the identified harmonic pollution periods are processed... High-precision spectrum analysis is performed on waveform data to extract the amplitude, phase, harmonic distortion rate, and harmonic power characteristics of each harmonic, forming a set of harmonic characteristic parameters. This set of harmonic characteristic parameters, the active power data, the reactive power data, and the power factor data are input into a harmonic state assessment network to generate a harmonic state map describing the overall harmonic distribution and pollution severity of the distribution network. Based on this harmonic state map, a harmonic source location and responsibility allocation model is used to calculate and output quantitative indicators of responsibility and priority recommendations for each potential harmonic source. These quantitative indicators of responsibility and priority recommendations are then converted into a structured harmonic detection report and pushed to the distribution network management system.

[0021] In one embodiment of the present invention, see [reference] Figure 2The multiphase voltage and current waveform data of each monitoring point are preprocessed to extract time-domain and frequency-domain features characterizing waveform distortion, forming waveform feature vectors for each monitoring point. Based on the topological connections and electrical distances of the distribution network, a connection graph between monitoring points is constructed, where nodes represent monitoring points and edge weights are determined by electrical distances and connections. The improved spectral clustering algorithm is used to cluster the waveform feature vectors of all monitoring points. When calculating the similarity between feature vectors, the improved spectral clustering algorithm considers not only the Euclidean distance of the feature vectors themselves but also corrects the similarity based on the edge weights in the connection graph, so that monitoring points that are electrically adjacent and have similar waveform characteristics are clustered together. After clustering, each cluster is considered to correspond to the influence range of one or more harmonic sources. The change of the cluster center in the time dimension is used to identify the periods of harmonic activity, thereby determining the potential harmonic source access point and harmonic pollution period.

[0022] In the specific implementation, a 10kV distribution network feeder area containing 15 monitoring points was selected as the implementation scenario. Smart meters deployed at each monitoring point continuously collected three-phase voltage and current waveform data at a sampling rate of 256 points per power frequency cycle. The multi-phase voltage and current waveform data from each monitoring point were preprocessed. The preprocessing steps included using a median filter to remove impulse noise, using linear interpolation to complete missing sampling points, and extracting three-level detail coefficients from the waveform distortion features as time-domain features using discrete wavelet transform. Simultaneously, a 512-point Fast Fourier Transform was performed on the complete waveform. The content of the 3rd, 5th, 7th, 9th, 11th, and 13th harmonics was extracted as frequency domain features. These six specific odd harmonic orders were selected as frequency domain features because, in the actual operation of the distribution network, the harmonic currents generated by typical nonlinear loads such as three-phase bridge rectifiers, frequency converters, and electric arc furnaces are mainly odd-order. Among them, the amplitudes of the 3rd, 5th, and 7th harmonics are usually high and are the main components causing voltage distortion in the distribution network. Although the amplitudes of the 9th, 11th, and 13th harmonics are relatively low, they may still be significantly amplified under some resonance conditions. The changes in their content can reflect whether there is a high-frequency resonance risk in the distribution network. The discrete wavelet transform is used to extract three levels of detail coefficients as time-domain features because waveform distortion caused by harmonic sources often manifests as non-stationary transient disturbances or abrupt changes, and general time-domain statistics are difficult to effectively capture such local transient characteristics. The discrete wavelet transform can decompose the waveform signal into detail coefficients in different frequency bands through multi-resolution analysis. The three levels of detail coefficients correspond to the transient components in the higher, middle and lower frequency bands, respectively, which can completely characterize the instantaneous shape and attenuation process of waveform distortion caused by harmonic source switching and load abrupt changes from different time scales.

[0023] These six harmonics constitute an odd-order harmonic sequence covering low to high frequencies. This sequence can characterize the typical spectral distribution of conventional nonlinear loads while also meeting the monitoring needs of high-frequency abnormal states. From the perspective of feature selection, redundancy can be effectively avoided, and it has good distinguishability of harmonic source types. The time-domain features and frequency-domain features are concatenated to form the waveform feature vector of each monitoring point. The waveform feature vector has a dimension of 42.

[0024] In some embodiments, a connection graph between monitoring points is constructed based on the topological connections and electrical distances of the distribution network. Each monitoring point is a node in the connection graph. If there is a direct feeder connection or an electrical path through a transformer branch between any two monitoring points, an edge is established between the two nodes. The weight of the edge is calculated using the following formula: in: This represents the weight of the edge connecting monitoring point i and monitoring point j. This weight is a dimensionless value. This represents the per-unit value of the electrical distance between monitoring point i and monitoring point j. The electrical distance is defined as the sum of the magnitudes of all feeder impedances along the path from monitoring point i to monitoring point j, plus the sum of the impedances converted from the number of transformers passed through. This sum is then divided by the reference impedance value. Obtain the dimensionless per-unit value of electrical distance The reference impedance value is taken as the modulus of the short-circuit impedance at the beginning of the distribution network feeder. It takes the value 0.001 and is a dimensionless constant.

[0025] Optionally, an improved spectral clustering algorithm is used to cluster the waveform feature vectors of all monitoring points. Specifically, the clustering process first calculates the degree matrix based on the constructed modified similarity matrix, that is, summing the elements of each row of the similarity matrix to obtain the degree value of each node, and arranging the degree values ​​in sequence to form a diagonal matrix; then, the denormalized Laplacian matrix is ​​calculated, that is, the degree matrix is ​​subtracted from the similarity matrix; the Laplacian matrix is ​​subjected to eigenvalue decomposition to obtain all eigenvalues ​​and corresponding eigenvectors; then, the eigenvalues ​​are sorted from smallest to largest, the zero eigenvalues ​​are discarded, and the eigenvectors corresponding to the k smallest eigenvalues ​​are arranged in columns to form a feature matrix. Each row of this matrix corresponds to a low-dimensional embedding representation of an original monitoring point in the new feature space; in this low-dimensional embedding space, the feature vectors of monitoring points within the same harmonic source influence range will show a more compact clustered distribution, and subsequent clustering can be completed by simply partitioning the row vectors of the feature matrix. The improved spectral clustering algorithm, when calculating the similarity between feature vectors, not only considers the Euclidean distance between the feature vectors themselves, but also corrects the similarity based on the weights of the edges in the connectivity graph. The specific implementation steps include: constructing an initial similarity matrix between the waveform feature vectors of all monitoring points; and for any two monitoring points i and j, calculating their waveform feature vectors. and Euclidean distance And based on the Gaussian kernel function, it is converted into an initial similarity. ,in The value is taken as 1.5 times the standard deviation of the Euclidean distance between all monitoring points; the weight of the edge connecting monitoring point i and monitoring point j in the connectivity graph is obtained. If there is no direct edge between two monitoring points, then For each monitoring point, calculate the corrected similarity. ,in The value is 2.0.

[0026] In some embodiments, a Laplacian matrix is ​​constructed using the modified similarity matrix. Eigenvalue decomposition is performed on the Laplacian matrix, and the eigenvectors corresponding to the first k smallest eigenvalues ​​are used to form the eigenma matrix. Specifically, k is determined by analyzing the distribution of the eigenvalues ​​of the Laplacian matrix. After arranging the first 20 smallest eigenvalues ​​in ascending order, the differences between adjacent eigenvalues ​​are calculated, and the trend of the difference sequence is observed. When the difference between the i-th eigenvalue and the (i+1)-th eigenvalue is significantly greater than the previous difference variation, it indicates the existence of a feature gap at the i-th eigenvalue. In the 15 monitoring points of the implementation scenario, the difference between the 4th and 5th eigenvalues ​​shows a significant jump, indicating the largest feature gap. This suggests that the intrinsic dimension of the data in the low-dimensional embedding space should be 4. Setting k to 4 most naturally reflects the clustering structure formed by the waveform feature vectors of the monitoring points after topological enhancement, effectively separating the influence range of different harmonic sources. The k-means algorithm is used to cluster the row vectors of the feature matrix. Specifically, each row of the feature matrix is ​​treated as a data point in a 15-dimensional space, and four data points are randomly selected as initial cluster centers. For each data point, the Euclidean distance to the four cluster centers is calculated, and the data point is assigned to the cluster to which the nearest cluster center belongs. After all data points are assigned, for each cluster, the mean of the coordinates of all data points in the cluster is taken as the new cluster center. The assignment and update steps are repeated until the sum of the squared distances of all cluster centers between two consecutive iterations is less than a preset convergence threshold. The preset convergence threshold is selected based on factors such as the scale of the power distribution network monitoring points, the dimension of the waveform feature vector, and the real-time requirements of harmonic detection. In this embodiment, the number of monitoring points is 15, and the waveform feature vector dimension is 42. To ensure a balance between clustering accuracy and computational efficiency, the preset convergence threshold is set to 0.01. The preset convergence threshold directly affects the stability of the clustering results and the computation time: if the threshold is set too small (e.g., less than 0.001), more iterations are needed to stabilize the cluster centers, increasing computational resource consumption; if the threshold is set too large (e.g., greater than 0.1), the iteration may terminate prematurely, resulting in insufficient convergence of cluster centers and affecting the accuracy of identifying harmonic source access points and pollution periods. Finally, each monitoring point obtains a cluster label based on the cluster to which its corresponding data point ultimately belongs, thus completing the clustering of the 15 monitoring points.After clustering, each cluster is considered to correspond to the influence range of one or more harmonic sources. In the implementation scenario, 15 monitoring points are divided into 3 clusters: cluster A contains monitoring points 2, 3, 4, and 5; cluster B contains monitoring points 7, 8, 9, 10, and 11; and cluster C contains monitoring points 1, 6, 12, 13, 14, and 15. For each cluster, the waveform feature vectors of all monitoring points within the cluster are used to calculate the cluster center according to a time window (each time window is 10 minutes). The change of the cluster center in the time dimension is used to identify the period when harmonics are active. The specific method is to calculate the Euclidean distance between the cluster center vector of each time window and the baseline cluster center vector (take the average value of the non-pollution period from 0:00 to 6:00 on the same day). The selection of the preset threshold of 0.35 is based on the statistical characteristics of the Euclidean distance during the non-pollution period and the harmonic detection sensitivity requirements. In the implementation scenario, by analyzing the non-pollution period data from 0:00 to 6:00 for 7 consecutive days from 15 monitoring points, the mean Euclidean distance between the cluster centers of each time window and the baseline was calculated to be 0.18, with a standard deviation of 0.07. To avoid misjudging normal load fluctuations as pollution, the threshold needs to be higher than the mean plus 2 times the standard deviation (0.18 + 2 × 0.07 = 0.32). In conjunction with the actual observation results that this distance generally exceeds 0.3 in historical pollution events, 0.35 was finally selected as the threshold, balancing anti-interference ability and detection sensitivity. When the distance exceeds a preset threshold of 0.35, the time window is marked as a harmonic pollution period. The monitoring point with the highest frequency among all marked harmonic pollution periods is identified as a potential harmonic source access point. The size of this preset threshold directly affects the accuracy of pollution period identification: if the threshold is set too small (e.g., <0.3), normal load fluctuations during non-pollution periods (e.g., motor starting, air conditioner start-stop) may be misjudged as harmonic pollution, increasing the false alarm rate; if the threshold is set too large (e.g., >0.4), slight harmonic pollution periods (e.g., small-capacity nonlinear load switching) may be missed, reducing detection sensitivity. In the implementation scenario, monitoring points 3, 9, and 12 were identified as potential harmonic source access points, with harmonic pollution periods concentrated between 9:00 and 11:00 and between 14:00 and 16:00 daily.

[0027] It is understandable that all preprocessing results of waveform feature vectors, edge weight calculation results of connection graphs, and cluster partitioning results obtained from cluster analysis are stored in the data server on the distribution network side for subsequent harmonic detection processes. Optionally, for each cluster obtained after clustering, the cosine similarity between the waveform feature vector of the monitoring points within the cluster and the cluster center is further calculated. The cosine similarity is calculated as follows: for a cluster, the mean vector of the waveform feature vectors of all monitoring points within the cluster in each dimension is taken as the cluster center vector; for the waveform feature vector of each monitoring point within the cluster, the dot product between it and the cluster center vector is calculated, and the magnitudes of these two vectors are calculated separately. The dot product is divided by the product of the magnitudes of the two vectors, and the quotient is the cosine similarity between the waveform feature vector of the monitoring point and the cluster center. Its value ranges from negative one to positive one. The closer the value is to positive one, the more consistent the waveform distortion characteristics of the monitoring point are with the typical pattern within the cluster. When the cosine similarity of a monitoring point is less than 0.6, the monitoring point is marked as a boundary point, indicating that the monitoring point is located in the edge region of the harmonic source's influence range. The boundary information is used for accuracy correction during subsequent harmonic source localization. It can be understood that the changes in cluster centers over time are smoothed using a sliding window averaging method, with a smoothing window width of three time windows, to reduce the impact of instantaneous fluctuations on the identification of harmonic pollution periods.

[0028] In one embodiment of the present invention, see [reference] Figure 3 The voltage and current waveform data during the harmonic pollution period were resampled and windowed to reduce spectral leakage. High-resolution fast Fourier transforms were then performed on the processed voltage and current waveform data to obtain the voltage and current spectra. The fundamental frequency position was identified within these spectra, and the integer harmonic frequency components were located based on this. The voltage amplitude, current amplitude, voltage phase, and current phase corresponding to each integer harmonic frequency component were extracted. Based on the extracted amplitude and phase, the harmonic content, total harmonic distortion rate, harmonic power, and harmonic directional power of each harmonic were calculated. The characteristic calculation results of all harmonics were then compiled to form a set of characteristic parameters for the harmonic.

[0029] In the specific implementation, voltage and current waveform data identified at monitoring point M3 during the harmonic pollution period between 9:40 AM and 9:45 AM were selected as the analysis objects. The original sampling frequency of the voltage and current waveform data was 12800 Hz, with 256 points sampled per power frequency cycle. Resampling processing was performed on the voltage and current waveform data during the harmonic pollution period. The resampling process uniformly increased the sampling frequency to 25600 Hz using cubic spline interpolation. The resampled voltage and current waveform data were then multiplied by a Hanning window function for windowing processing. The length of the Hanning window function was 1024 sampling points. The purpose of windowing processing was to reduce spectral leakage.

[0030] In some embodiments, high-resolution fast Fourier transforms (HFTs) are performed on the windowed voltage and current waveform data, respectively. The HFTs use 8192 points to obtain voltage and current spectra. The voltage spectrum contains the voltage amplitude and phase information for each frequency component, and the current spectrum contains the current amplitude and phase information for each frequency component. The fundamental frequency (50 Hz) is identified in both the voltage and current spectra. The frequency point with the largest amplitude in the vicinity of 50 Hz is determined as the precise location of the fundamental frequency. Based on this precise location, 25 integer harmonic frequency components are located, with each harmonic frequency component's position being an integer multiple of the fundamental frequency. Optionally, the voltage amplitude, current amplitude, voltage phase, and current phase corresponding to each integer harmonic frequency component are extracted. For the h-th harmonic, the spectral amplitude at frequency h × 50 Hz is read from the voltage spectrum as the voltage amplitude. The amplitude of the spectral line at the corresponding frequency is read from the current spectrum as the current amplitude. Read the voltage phase from the phase spectrum of the voltage spectrum Read the current phase from the phase spectrum of the current spectrum. Calculate the phase difference between the h-th harmonic voltage and current. Based on the extracted voltage amplitude, current amplitude, and phase difference, the harmonic content, total harmonic distortion (THD), power of each harmonic, and harmonic directional power are calculated. The formula for calculating the total harmonic distortion (THD) is as follows: in: This represents the voltage amplitude of the h-th harmonic. This represents the amplitude of the fundamental voltage; both are measured in volts, and the ratio is... The value is dimensionless. Taking the square root and multiplying it by 100% gives a dimensionless percentage value. The total harmonic distortion (THD) is expressed as a percentage.

[0031] In practical implementation, for the third harmonic, the voltage amplitude It is 12.6 volts, and the current amplitude is... It is 5.8 amperes, with a phase difference of 1. The calculated power of the third harmonic is 66.7 watts, and the directional harmonic power is 29.8 var. For the fifth harmonic, the voltage amplitude is... It is 8.3 volts, and the current amplitude is... The phase difference is 3.2 amperes. With a radius of 0.78 radians, the power of the 5th harmonic was calculated to be 18.9 watts, and the directional harmonic power was 18.7 vars. The characteristic values ​​of the 2nd to 25th harmonics were calculated in sequence.

[0032] In some embodiments, the characteristic calculation results of all harmonics are aggregated to form a harmonic characteristic parameter set. This set includes the harmonic order identifier, harmonic voltage amplitude sequence, harmonic current amplitude sequence, harmonic voltage phase sequence, harmonic current phase sequence, harmonic content rate sequence, total harmonic distortion rate, harmonic active power sequence, harmonic reactive power sequence, and harmonic directional power sequence for each monitoring point during each harmonic pollution period. The harmonic characteristic parameter set is stored in a structured data format on the distribution network side data server. Each data record includes a timestamp, monitoring point identifier, harmonic order, and all corresponding characteristic values. It can be understood that when performing windowing processing after resampling, a Hanning window is selected, and the main lobe width of the Hanning window is... (Where N is the window function length), the main lobe width corresponds to a frequency resolution of 25600 / 8192 = 3.125 Hz under the 8192-point Fast Fourier Transform, effectively distinguishing the frequency interval between the 50 Hz fundamental wave and the 100 Hz second harmonic. Optionally, for harmonic frequency components with voltage amplitudes below 0.2 volts, the contribution of this harmonic is ignored when calculating the total harmonic distortion rate to avoid the influence of noise interference on the calculation results. It can be understood that the positive or negative sign of the harmonic directional power is used to determine the flow direction of harmonic power. When the harmonic directional power is positive, it indicates that the harmonic reactive power flows from the monitoring point to the distribution network bus; when the harmonic directional power is negative, it indicates that the harmonic reactive power flows from the distribution network bus to the monitoring point. This directional information is used for flow direction analysis in the subsequent harmonic source location process.

[0033] In one embodiment of the present invention, the harmonic state assessment network includes a feature encoding layer, a spatial correlation layer, and a graph generation layer. The feature encoding layer receives the set of harmonic feature parameters, active power data, reactive power data, and power factor data of each monitoring point, and encodes them into a node state vector for each monitoring point. The spatial correlation layer receives the topology data of the distribution network and the node state vectors of all monitoring points, and uses a graph neural network model to simulate the propagation and superposition process of harmonic features on the distribution network topology, updating the node state vector of each monitoring point to include neighborhood information. The graph generation layer spatially arranges the updated node state vectors of all monitoring points according to their actual geographical or electrical locations, and uses interpolation techniques to generate a continuous and visualized harmonic state graph covering the entire distribution network area. In the harmonic state graph, different colors or contour lines represent harmonics at different locations. The severity of pollution is assessed. The spatial association layer receives topology data of the distribution network and node state vectors of all monitoring points. Using a graph neural network model, the specific implementation of simulating the propagation and superposition of harmonic features on the distribution network topology is as follows: the distribution network topology is abstracted as a graph, where nodes are monitoring points or grid buses, and edges are feeders or transformer branches. The node state vector of each monitoring point is used as the initial feature of the corresponding graph node. This graph neural network model consists of multiple graph convolutional layers. Each graph convolutional layer performs the following operations: aggregating the features of all neighboring nodes of each node; after linear transformation and nonlinear activation, updating the node's features; through multi-layer graph convolution operations, the feature information of each node propagates and interacts along the topology edges in multiple hops. Ultimately, the updated features of each node not only contain its own harmonic state information but also integrate the harmonic state information of its upstream and downstream neighboring areas.

[0034] In the specific implementation, a distribution network feeder area containing 10 monitoring points was selected as the implementation scenario. The monitoring points were numbered from M01 to M10. At each monitoring point, the harmonic characteristic parameter set, active power data, reactive power data, and power factor data were obtained through the aforementioned steps. The harmonic characteristic parameter set includes the total harmonic distortion rate, the voltage amplitude of the 3rd harmonic, the voltage amplitude of the 5th harmonic, the voltage amplitude of the 7th harmonic, and the active power sequence of each harmonic. The harmonic state assessment network includes a feature encoding layer, a spatial correlation layer, and a spectrum generation layer. The feature encoding layer receives the harmonic characteristic parameter set, active power data, reactive power data, and power factor data of each monitoring point and encodes them into a node state vector for each monitoring point. The node state vector is set to 64 dimensions, of which the first 32 dimensions are obtained by linear transformation of the harmonic characteristic parameter set, the middle 16 dimensions are obtained by nonlinear mapping of active power data and reactive power data, and the last 16 dimensions are obtained by normalization of power factor data.

[0035] In some embodiments, the spatial association layer receives the topology data of the distribution network and the node state vectors of all monitoring points. The topology data of the distribution network includes feeder connection relationships, transformer branch connection relationships, and electrical distance information between each monitoring point. The spatial association layer uses a graph neural network model to simulate the propagation and superposition process of harmonic features on the distribution network topology, and updates the node state vector of each monitoring point to include neighborhood information. Specifically, the distribution network topology is abstracted as a graph, where the nodes are monitoring points (M01 to M10), and the edges are feeders or transformer branches. An undirected edge is established between monitoring points with direct electrical connections. The node state vector of each monitoring point is used as the initial feature vector of the corresponding graph node. Here, v represents the node number. The graph neural network model consists of two graph convolutional layers. The first graph convolutional layer transforms the feature vector of each node from 64 dimensions to 128 dimensions, and the second graph convolutional layer transforms the 128-dimensional feature vector back to 64 dimensions. The operation performed by each graph convolutional layer is to aggregate the features of all neighboring nodes of each node, and update the node's features after linear transformation and nonlinear activation. Optionally, the mean aggregation function is used to aggregate the features of neighboring nodes in the graph convolutional layer. For a node v, its set of neighboring nodes is denoted as . The update formula for the l-th graph convolutional layer is: in: This represents the feature vector of node v at layer l. This represents the weight matrix of the l-th layer. This represents the bias vector of the l-th layer. Represents the ReLU nonlinear activation function. This represents the element-wise average of all vectors in the set. This represents the feature vector of node v at layer l+1, and the weight matrix is... The size is 128×64, and the weight matrix is... The size is 64×128, and the bias vector is... and The dimensions are 128 and 64 respectively.

[0036] In the specific implementation, through two layers of graph convolution, the feature information of each node is propagated and interacted along the topological edges in two hops. For node M03, after the first layer of graph convolution, the features of its direct neighboring nodes M02 and M04 are aggregated. After the second layer of graph convolution, the features of M01, M05 and more distant nodes are further aggregated. The updated node state vector of each node not only contains its own harmonic state information, but also integrates the harmonic state information of the upstream and downstream neighboring regions. The updated state vectors of the 10 nodes are all 64-dimensional. In some embodiments, the map generation layer spatially arranges the updated node state vectors of all monitoring points according to the actual geographical location of the monitoring points. The latitude and longitude coordinates of monitoring points M01 to M10 have been pre-entered into the power distribution network geographic information system. The map generation layer uses the first component (i.e., the total harmonic distortion rate encoding value) in the node state vector of each monitoring point as the value of the location of that monitoring point. It uses inverse distance weighted interpolation technology to generate a continuous and visualized harmonic state map covering the entire power distribution network area. The interpolation grid resolution is 200×200, and the interpolation weight is the reciprocal of the square of the distance. Different colors in the harmonic state map represent the severity of harmonic pollution at different locations. The color transitions from green (low pollution) through yellow to red (high pollution). The resolution of the harmonic state map is 1920×1080 pixels.

[0037] It is understandable that the parameters of the inverse distance weighted interpolation technique are set to a power exponent of 2 and a search radius of 500 meters. The value of each grid point is obtained by weighting the values ​​of surrounding monitoring points by the inverse square of the distance. For areas not covered by monitoring points, the extrapolation distance of the interpolation result does not exceed 100 meters. Optionally, for monitoring points on both sides of the transformer branch, when establishing edges in the graph neural network model, the edge weight is assigned as the inverse of the transformer impedance to reflect the attenuation effect of the transformer on harmonics. This edge weight is used as a normalization coefficient when aggregating neighbor features. It is understandable that the harmonic state map output by the graph generation layer is stored in PNG format on the distribution network side server. The file name includes the generation timestamp. At the same time, the pollution level label at each monitoring point location is output. This label is obtained by thresholding the total harmonic distortion (THD) component in the node state vector: THD less than 3% is marked as light pollution, 3% to 5% is marked as moderate pollution, and greater than 5% is marked as heavy pollution (see Table 1).

[0038] Table 1: Initial Feature Table of Node State Vectors at Each Monitoring Point Optionally, the total harmonic distortion rate values ​​of each monitoring point in the above table can be used as one of the input features when encoding the node state vector. The active power, reactive power, and power factor data are normalized by the maximum and minimum values, respectively, and then concatenated with other features in the harmonic feature parameter set to form a complete initial node state vector.

[0039] In one embodiment of the present invention, the harmonic source region is identified in the harmonic state spectrum based on the gradient change of the severity of harmonic pollution; the harmonic phase and harmonic power flow direction information of the monitoring points located in the harmonic source region are extracted from the harmonic characteristic parameter set; using a harmonic power flow calculation model, combined with the distribution network topology and impedance parameters, reverse power flow tracing is performed to estimate the magnitude of the injected harmonic current of each potential harmonic source; based on the estimated magnitude of the injected harmonic current of each potential harmonic source and its contribution ratio in the total harmonic pollution, the responsibility quantification index of each potential harmonic source is calculated; combining the pollution range shown in the harmonic state spectrum, the magnitude of the responsibility quantification index, and the user type of each potential harmonic source, the governance priority recommendation is generated through preset rules or a scoring model; wherein, using the harmonic power flow calculation model, the harmonic power flow... The flow calculation model is a power grid harmonic analysis model established for each harmonic frequency. Its inputs include the node admittance matrix or impedance matrix of the distribution network at a specific harmonic frequency, the amplitude and phase of the harmonic voltage measured at each monitoring point, and the location of potential harmonic source access nodes preliminarily determined based on the harmonic state spectrum and harmonic power flow information. The model is based on Kirchhoff's current law and describes the relationship between node injected harmonic current and node harmonic voltage in the harmonic frequency domain. The node injected harmonic current vector has a non-zero component to be determined only at the node corresponding to the potential harmonic source. For the overdetermined case where the number of monitoring points is greater than the number of unknown harmonic sources, the model aims to minimize the weighted sum of squares of the difference between the measured harmonic voltage and the model-calculated harmonic voltage. It uses the weighted least squares method to iteratively solve for the estimated value of injected harmonic current at each potential harmonic source node. The specific implementation method for calculating the injected harmonic current magnitude of each potential harmonic source by combining the distribution network topology and impedance parameters and performing reverse power flow tracing is as follows: A harmonic impedance network model of the target distribution network at a specific harmonic frequency is constructed. This model is based on the single-line diagram of the distribution network, abstracting the connection points between each bus and feeder as nodes, and abstracting feeder segments, transformer branches, and reactive power compensation capacitor branches as branches. For each feeder branch, its harmonic impedance parameter is obtained by multiplying the resistance and reactance per unit length by the feeder length to obtain the fundamental impedance, and then multiplying by the harmonic order. The impedance values ​​are converted to the corresponding harmonic frequencies. For transformer branches, the harmonic impedance is obtained by multiplying the per-unit value of the transformer short-circuit impedance to the actual value by the harmonic order, while also considering the influence of the transformer winding connection group on the transmission of harmonic sequence components. For parallel capacitor or cable charging capacitor branches, the harmonic impedance decreases inversely with the increase of the harmonic order. The impedance parameters of all nodes and branches are summarized to form the node admittance matrix or impedance matrix of the distribution network at a specific harmonic frequency, which constitutes the harmonic impedance network model used for subsequent reverse power flow tracing calculations.The harmonic voltages at each monitoring point identified in the harmonic state spectrum are taken as known node voltages. The locations of potential harmonic sources are considered as source nodes for injected harmonic currents. Based on the harmonic impedance network model, node harmonic voltage equations are established, where the harmonic voltages at each monitoring point are known quantities, and the injected harmonic currents of each potential harmonic source are unknown quantities. Since the number of monitoring points is usually less than the number of unknown quantities, a least squares estimation method based on the harmonic power direction is introduced to solve the node harmonic voltage equations, obtaining estimated values ​​of injected harmonic currents of each potential harmonic source that minimize the difference between the calculated voltage and the measured voltage.

[0040] In the specific implementation, a 10kV distribution network feeder area containing 10 monitoring points (numbered M01 to M10) was selected as the implementation scenario. The aforementioned steps have generated a harmonic state map describing the overall harmonic distribution status and pollution severity of the distribution network. The total harmonic distortion rate (THD) value at each location is represented by a color gradient in the harmonic state map. The harmonic source region is identified in the harmonic state map based on the gradient change of the severity of harmonic pollution. The specific identification method is as follows: calculate the gradient magnitude of the THD value of each pixel in the harmonic state map, mark the region with the gradient magnitude greater than the threshold 0.5 (the rate of change of THD per pixel per unit distance) as the edge region, trace along the gradient descent direction to the local maximum point, and determine the harmonic source region as the circular region with a radius of 50 meters centered on the maximum point. In this implementation scenario, three harmonic source regions were identified, located near monitoring point M03, near monitoring point M07, and near monitoring point M09, respectively.

[0041] In some embodiments, harmonic phase and harmonic power flow information of monitoring points located in the harmonic source region are extracted from the harmonic characteristic parameter set. For the harmonic source region located near M03, the phase of the 3rd harmonic voltage at monitoring point M03 is extracted as 1.23 radians, the phase of the 3rd harmonic current is extracted as 0.81 radians, the phase of the 5th harmonic voltage is extracted as 2.56 radians, the phase of the 5th harmonic current is extracted as 1.94 radians, and the active power of the 3rd harmonic is extracted as 66.7 watts and the reactive power of the 3rd harmonic is extracted as 29.8 vars. The harmonic power flow information is determined by the sign of the harmonic reactive power. The 3rd harmonic reactive power at monitoring point M03 is positive, indicating that the harmonic power flows from monitoring point M03 to the distribution network bus. For the harmonic source region located near M07, the phase and harmonic power flow information of the monitoring point are extracted as follows: At point M07, the phase of the 3rd harmonic voltage is 2.87 radians, the phase of the 3rd harmonic current is 3.42 radians, the phase of the 5th harmonic voltage is 4.12 radians, and the phase of the 5th harmonic current is 4.89 radians. The active power of the 3rd harmonic is 112.3 watts, and the reactive power of the 3rd harmonic is -45.6 var (the negative value indicates that the harmonic power flows from the distribution network bus to monitoring point M07). For the harmonic source area located near M09, the phase of the 3rd harmonic voltage at monitoring point M09 is extracted as 0.78 radians, the phase of the 3rd harmonic current is 0.35 radians, the phase of the 5th harmonic voltage is 1.94 radians, the phase of the 5th harmonic current is 1.52 radians, the active power of the 3rd harmonic is 89.4 watts, and the reactive power of the 3rd harmonic is 12.7 var.

[0042] In practical implementation, a harmonic power flow calculation model is used in conjunction with the distribution network topology and impedance parameters to perform reverse power flow tracing and estimate the magnitude of the injected harmonic current from each potential harmonic source. Specifically, a harmonic impedance network model of the target distribution network at the third harmonic frequency is constructed. This model includes 10 nodes (corresponding to the busbars where monitoring points M01 to M10 are located) and the impedance of the feeder branches between the nodes. The impedance value of each feeder branch at the third harmonic frequency is the fundamental impedance multiplied by the harmonic order 3. The harmonic impedance of the transformer branch is the transformer short-circuit impedance multiplied by the harmonic order 3. The harmonic state spectrum is then used to calculate the injected harmonic current magnitude from the harmonic state spectrum. The third harmonic voltage at each identified monitoring point is taken as the known node voltage. The measured amplitudes of the third harmonic voltages at monitoring points M01 to M10 are 2.3 volts, 4.7 volts, 12.6 volts, 6.5 volts, 3.1 volts, 1.8 volts, 8.9 volts, 4.3 volts, 10.2 volts, and 2.9 volts, respectively. The location of the potential harmonic source is considered as the source node for injecting harmonic current. The monitoring points M03, M07, and M09 corresponding to the three harmonic source regions identified in the previous steps are the locations of the potential harmonic sources. The node harmonic voltage equations are established based on the harmonic impedance network model. in: The node harmonic voltage vector is 10×1 (a known quantity). The harmonic impedance matrix is ​​10×10 (a known quantity, calculated from the harmonic impedance network model). A 10×1 node harmonic current vector is injected, where only the injected harmonic current at the nodes containing potential harmonic sources (M03, M07, M09) is a non-zero unknown, while the injected harmonic current at the other nodes is zero. Since there are 10 monitoring points and 3 unknowns, the number of equations exceeds the number of unknowns. Therefore, a least-squares estimation method based on the harmonic power direction is introduced to solve the harmonic voltage equation for this node. The objective function is... The estimated values ​​of the injected harmonic currents of each potential harmonic source that minimize the difference between the calculated voltage and the measured voltage are obtained by solving the problem. The results are: the third harmonic injection current at monitoring point M03 is 4.2 amps, the third harmonic injection current at monitoring point M07 is 2.8 amps, and the third harmonic injection current at monitoring point M09 is 1.5 amps.

[0043] Optionally, when solving the nodal harmonic voltage equations, the least squares estimation method uses a weighted least squares method. The weight coefficients are determined by the harmonic power direction information. Potential harmonic sources with harmonic power flowing from the monitoring point to the bus (positive harmonic reactive power) are assigned a higher weight (1.0), while potential harmonic sources with harmonic power flowing from the bus to the monitoring point (negative harmonic reactive power) are assigned a lower weight (0.5). The weighted objective function is: ,in It is a diagonal weight matrix.

[0044] In some embodiments, the responsibility quantification index of each potential harmonic source is calculated based on the estimated magnitude of the injected harmonic current and its contribution ratio in the total harmonic pollution. The total harmonic pollution is taken as the sum of the squares of the third harmonic voltage amplitudes at monitoring points M03, M07, and M09 (12.6² + 8.9² + 10.2² = 324.8). The contribution ratio of monitoring point M03 is 4.2 amps of injected harmonic current, which accounts for 49.4% of the sum of the three injected currents (4.2 + 2.8 + 1.5 = 8.5 amps). Combining this ratio with the harmonic voltage amplitude weight at monitoring point M03, the responsibility quantification index of monitoring point M03 is calculated to be 52.8%. The injected harmonic current of monitoring point M07 is 2.8 amps, with a contribution ratio of 32.9% and a responsibility quantification index of 31.5%. The injected harmonic current of monitoring point M09 is 1.5 amps, with a contribution ratio of 17.6% and a responsibility quantification index of 15.7%.

[0045] In practical implementation, combining the pollution range displayed by the harmonic state spectrum, the magnitude of the responsibility quantification index, and the user type of each potential harmonic source, a governance priority suggestion is generated through a preset rule model. The pollution range displayed by the harmonic state spectrum is measured by the number of affected monitoring points. The pollution range of monitoring point M03 covers four monitoring points: M02, M03, M04, and M05; the pollution range of monitoring point M07 covers three monitoring points: M06, M07, and M08; and the pollution range of monitoring point M09 covers three monitoring points: M01, M09, and M10. The responsibility quantification indices are 52.8%, 31.5%, and 15.7%, respectively. The user type is obtained from the smart meter user files. The user type corresponding to monitoring point M03 is an industrial user (electric arc furnace load), the user type corresponding to monitoring point M07 is a commercial user (variable frequency air conditioning group), and the user type corresponding to monitoring point M09 is a residential user (switching power supply load). The preset rule model uses a scoring formula. Where R is the normalized value of the responsibility quantification index (divided by 100), C is the normalized value of the pollution range (the number of affected monitoring points divided by the total number of monitoring points 10), and U is the user type coefficient (1.0 for industrial users, 0.6 for commercial users, and 0.3 for residential users). The calculated score for monitoring point M03 is 0.584, the score for monitoring point M07 is 0.3675, and the score for monitoring point M09 is 0.2285. Based on the scores from high to low, the recommended priority for treatment is: treat the harmonic source corresponding to monitoring point M03, then treat the harmonic source corresponding to monitoring point M07, and finally treat the harmonic source corresponding to monitoring point M09.

[0046] It is understandable that, for the 5th harmonic and other integer harmonics, the above harmonic power flow calculation and reverse power flow tracing process is repeated to obtain the estimated injected harmonic current of each potential harmonic source under each harmonic and the corresponding responsibility quantification index. Finally, the overall responsibility quantification index integrating the weights of each harmonic is output, where the weight is the normalized value of the product of the voltage amplitude of each harmonic and the total harmonic distortion rate. Optionally, when the estimated injected harmonic current is negative, it indicates that the location absorbs harmonic current rather than injects harmonic current. This location should not be considered as a harmonic source for mitigation, and its responsibility quantification index is set to zero in the responsibility allocation model, as shown in Table 2.

[0047] Table 2: Quantitative Indicators of Responsibility and Priority Scoring Table for Each Potential Harmonic Source It is understandable that the above-mentioned governance priority scoring results are consistent with the trend of pollution severity gradient changes shown in the harmonic state map. The area near monitoring point M03 is a red high-pollution area in the harmonic state map, which is consistent with the scoring results.

[0048] In one embodiment of the present invention, for any two monitoring points' waveform feature vectors, their Euclidean distance is calculated, and converted into an initial similarity based on a Gaussian kernel function; the weights of the edges connecting these two monitoring points in the connection graph are obtained, and these weights are inversely proportional to the electrical distance; a similarity enhancement function based on edge weights is designed, where a larger edge weight between two monitoring points indicates a closer electrical connection or a shorter distance, and the greater the enhancement of the initial similarity by the similarity enhancement function; the value corrected by the similarity enhancement function is used as the final waveform feature vector of the two monitoring points in the improved spectral clustering algorithm. The similarity between feature vectors is designed. A specific implementation of a similarity enhancement function based on edge weights is as follows: An edge weight threshold is defined. When the edge weight connecting two monitoring points is greater than the threshold, the two monitoring points are considered to be electrically closely related. For electrically closely related monitoring point pairs, the similarity correction formula is that the enhanced similarity equals the initial similarity multiplied by an enhancement coefficient greater than one. This enhancement coefficient is a monotonically increasing function of the edge weights. For electrically loosely related monitoring point pairs, the similarity correction formula is that the enhanced similarity equals the initial similarity multiplied by a coefficient equal to one, i.e., no enhancement is performed.

[0049] In the specific implementation, a 10kV distribution network feeder area containing 15 monitoring points was selected as the implementation scenario. The waveform feature vector of each monitoring point has been extracted according to the aforementioned steps. The waveform feature vector has a dimension of 42. For any two monitoring points i and j, the waveform feature vectors are... and Calculate vector with vector Euclidean distance between And based on the Gaussian kernel function, the Euclidean distance is converted into an initial similarity. Among them, Gaussian kernel parameters The value is taken as 1.2 times the standard deviation of the Euclidean distance between all monitoring points. This represents the dimensionless initial similarity value.

[0050] In some embodiments, the weights of the edges connecting monitoring point i and monitoring point j in the connection graph are obtained. The connection graph is constructed as follows: each monitoring point is treated as a node in the graph. For two monitoring points that are directly connected by a feeder or pass through a transformer branch, an edge is established between the two nodes. The weight of the edge is calculated using the following formula: ,in This represents the electrical distance (in kilometers) between monitoring point i and monitoring point j. The electrical distance is defined as the sum of the feeder path lengths between the two monitoring points. When passing through a transformer branch, the equivalent electrical distance is calculated based on the transformer capacity. The unit is the reciprocal of kilometers, which is a dimensional value and needs to be adjusted in subsequent similarity corrections. Dimensionless processing is performed, specifically by dividing each... Divide by the maximum value of all non-zero edge weights in the entire connectivity graph. The dimensionless normalized edge weights are obtained. Normalized edge weights The value ranges from 0 to 1. Optionally, a similarity enhancement function based on edge weights can be designed, where the similarity enhancement function normalizes the edge weights. As input, the output enhancement coefficient is used. The enhancement function takes the form of: in: This represents the enhancement coefficient, which is a dimensionless value. The preset enhancement amplitude factor is set to 1.8. For indicator functions, when It takes the value 1 if the condition is met, and 0 otherwise. The edge weight threshold is set to 0.3. Normalized edge weights (dimensionless). With a power exponent of 0.5, the indicator function ensures that similarity enhancement is only performed on electrically closely related monitoring point pairs (i.e., normalized edge weights are greater than the threshold of 0.3), while for electrically loosely related monitoring point pairs (normalized edge weights are less than or equal to 0.3), the enhancement coefficient is equal to 1, meaning no enhancement is performed. The selection of the edge weight threshold τ=0.3 is based on the following: Through statistical analysis of the edge weights of the 10kV distribution network where 15 monitoring points are located, the mean normalized edge weight between monitoring points directly connected by feeders in this distribution network is 0.65 (standard deviation 0.15), the mean normalized edge weight between monitoring points connected by a single transformer is 0.42 (standard deviation 0.08), and the mean normalized edge weight between monitoring points spanning two or more transformers is 0.18 (standard deviation 0.05). To avoid misjudging loosely related points spanning multiple transformers as tightly related points, and to cover tightly related scenarios of direct feeder and single transformer connections, 0.3 is selected as the threshold, so that the edge weight of approximately 85% of direct / single transformer connection point pairs is greater than this value, while the edge weight of more than 95% of point pairs spanning multiple transformers is less than this value. The magnitude of the edge weight threshold directly affects the accuracy of similarity enhancement: if the threshold is set too small (e.g., <0.25), only a very small number of high-weight pairs (e.g., directly adjacent feeders) will be enhanced, potentially missing some closely related point pairs connected to a single transformer, leading to the inability to correctly merge monitoring points within the same harmonic source's influence range during clustering; if the threshold is set too large (e.g., >0.35), many loosely related point pairs spanning multiple transformer levels will be included in the enhancement range, causing monitoring points within different harmonic source influence ranges to be incorrectly merged, reducing the accuracy of harmonic source access point identification. In specific implementation, when the normalized edge weight between two monitoring points... When the value is greater than the threshold of 0.3, the two monitoring points are considered to be electrically closely related, and the enhancement coefficient is [value missing]. The enhancement coefficient is a monotonically increasing function of the normalized edge weights, for example, when The time enhancement coefficient is ,when The time enhancement coefficient is The enhanced similarity calculation formula is as follows: ,in The similarity between the waveform feature vectors of two monitoring points is used in the final improved spectral clustering algorithm. For the initial similarity, To enhance the coefficients, both are dimensionless values, and the product is also a dimensionless value.

[0051] It is understandable that for normalized edge weights For monitoring point pairs with a threshold of 0.3 or less, the indicator function output is 0, and the enhancement coefficient is... The enhanced similarity is equal to the initial similarity, meaning no enhancement is performed.

[0052] In some embodiments, for monitoring point pairs that do not have direct edges, the weights of edges in the connection graph are... After normalization The threshold judgment result is If the similarity enhancement coefficient is no greater than 0.3, the enhancement coefficient is 1, and no similarity enhancement is performed. Optional parameters in the similarity enhancement function. , , The configuration can be tailored to the density of the distribution network topology. For radial distribution network topologies, the parameters... Setting the threshold to 0.4 is too low; for a ring network distribution network, the parameter... The threshold is set to 0.2, which is considered high. It can be understood that after processing with the aforementioned similarity enhancement function, the similarity of monitoring point pairs that are electrically adjacent (normalized edge weights are greater than the threshold) and have certain similarities in waveform features (high initial similarity) is significantly amplified after correction. Therefore, they are more likely to be classified into the same category during spectral clustering. This correction method conforms to the physical law that the closer the electrical distance, the more similar the waveform distortion characteristics when harmonics propagate in the distribution network.

[0053] In practical implementation, the corrected similarity matrix will be... (Size 15×15) As input to the improved spectral clustering algorithm, a Laplacian matrix is ​​constructed and eigenvalue decomposition is performed. After the clustering is completed, the cluster division results are output. In the cluster division results, the monitoring points in each cluster show continuous distribution characteristics in electrical location, and the cluster boundary is basically consistent with the feeder segment position in the distribution network topology.

[0054] The flowchart provided in this embodiment is not intended to indicate that the operations of the method will be performed in any particular order, or that all operations of the method are included in every case. Furthermore, the method may include additional operations. Within the scope of the technical concept provided by the method in this embodiment, additional variations can be made to the above method.

[0055] It should be understood that in some embodiments, the components may be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods may be implemented using software or firmware stored in memory and executed by a suitable instruction execution system.

[0056] Therefore, those skilled in the art should recognize that although numerous exemplary embodiments of the present invention have been shown and described in detail herein, many other variations or modifications conforming to the principles of the present invention can be directly determined or derived from the disclosure of the present invention without departing from the spirit and scope of the invention. Thus, the scope of the present invention should be understood and construed as covering all such other variations or modifications.

Claims

1. A method for detecting harmonics in a distribution network based on smart meter data, characterized in that, The method includes: The raw electricity consumption data set uploaded by smart meters deployed at multiple monitoring points in the power distribution network is obtained. The raw electricity consumption data set includes multiphase voltage waveform data, multiphase current waveform data, active power data, reactive power data, and power factor data. An improved spectral clustering algorithm is applied to the multiphase voltage and current waveform data to identify potential harmonic source access points and harmonic pollution periods. The improved spectral clustering algorithm optimizes the similarity metric based on the propagation and attenuation characteristics of harmonics in the distribution network topology. This includes: preprocessing the multiphase voltage and current waveform data for each monitoring point to extract time-domain and frequency-domain features characterizing waveform distortion, forming a waveform feature vector for each monitoring point; and constructing a connection graph between monitoring points based on the distribution network topology and electrical distances, where nodes represent monitoring points and edge weights... The similarity between monitoring points is determined by electrical distance and connectivity. The improved spectral clustering algorithm is used to cluster the waveform feature vectors of all monitoring points. When calculating the similarity between feature vectors, the algorithm considers not only the Euclidean distance of the feature vectors themselves but also the weights of the edges in the connectivity graph, thus grouping electrically adjacent monitoring points with similar waveform features into one cluster. After clustering, each cluster is considered to correspond to the influence range of one or more harmonic sources. The change in the cluster center over time is used to identify periods of harmonic activity, thereby determining the potential harmonic source access points and harmonic pollution periods. High-precision spectrum analysis is performed on the voltage and current waveform data during the identified harmonic pollution period to extract the amplitude, phase, harmonic distortion rate and harmonic power characteristics of each harmonic, forming a set of harmonic characteristic parameters; The set of harmonic characteristic parameters, the active power data, the reactive power data, and the power factor data are input into the harmonic state assessment network to generate a harmonic state map describing the overall harmonic distribution state and pollution severity of the distribution network. Based on the aforementioned harmonic state spectrum, the responsibility quantification index and governance priority suggestions for each potential harmonic source are calculated and output through the harmonic source location and responsibility allocation model. The responsibility quantification indicators and governance priority suggestions for the harmonic sources are converted into a structured harmonic detection report and pushed to the distribution network management system. The improved spectral clustering algorithm, when calculating the similarity between feature vectors, not only considers the Euclidean distance between the feature vectors themselves, but also corrects the similarity based on the edge weights in the connectivity graph, including: For any two monitoring points, calculate their Euclidean distance using waveform feature vectors and convert it into an initial similarity based on the Gaussian kernel function; Obtain the weight of the edge connecting the two monitoring points in the connection graph, where the weight is inversely proportional to the electrical distance; Design a similarity enhancement function based on edge weights. The larger the edge weight between two monitoring points, the closer the electrical connection or the closer the distance. The greater the enhancement of the initial similarity by the similarity enhancement function. The value corrected by the similarity enhancement function is used as the similarity between the waveform feature vectors of the two monitoring points in the final improved spectral clustering algorithm. The design of a similarity enhancement function based on edge weights includes: Define an edge weight threshold. When the edge weight connecting two monitoring points is greater than the threshold, the two monitoring points are considered to be electrically closely connected. For electricalally closely related monitoring point pairs, the similarity correction formula is that the enhanced similarity is equal to the initial similarity multiplied by an enhancement coefficient greater than one, where the enhancement coefficient is a monotonically increasing function of the edge weights. For monitoring point pairs that are not closely related by electricity, the enhanced similarity is equal to the initial similarity.

2. The method for detecting harmonics in a distribution network based on smart meter data according to claim 1, characterized in that, The process involves performing high-precision spectrum analysis on the voltage and current waveform data within the identified harmonic pollution period to extract the amplitude, phase, harmonic distortion rate, and harmonic power characteristics of each harmonic, including: The voltage and current waveform data during the harmonic pollution period are resampled and windowed to reduce spectral leakage. High-resolution fast Fourier transforms are performed on the processed voltage waveform data and current waveform data respectively to obtain the voltage spectrum and current spectrum; In the voltage and current spectra, the position of the fundamental frequency is identified, and based on this, the integer harmonic frequency components are located. Extract the voltage amplitude, current amplitude, voltage phase, and current phase corresponding to each integer harmonic frequency component; Based on the extracted amplitude and phase, calculate the harmonic content, total harmonic distortion, power of each harmonic, and harmonic directional power. The characteristic calculation results of all harmonics are collected to form the set of harmonic characteristic parameters.

3. The method for detecting harmonics in a distribution network based on smart meter data according to claim 1, characterized in that, The set of harmonic characteristic parameters, the active power data, the reactive power data, and the power factor data are input into the harmonic state assessment network to generate a harmonic state map describing the overall harmonic distribution and pollution severity of the distribution network, including: The harmonic state assessment network includes a feature encoding layer, a spatial correlation layer, and a spectrum generation layer. The feature coding layer receives the set of harmonic feature parameters, active power data, reactive power data and power factor data of each monitoring point, and encodes them into the node state vector of each monitoring point. The spatial association layer receives the topology data of the distribution network and the node state vectors of all monitoring points. Using a graph neural network model, it simulates the propagation and superposition process of harmonic features on the distribution network topology and updates the node state vector of each monitoring point to include neighborhood information. The map generation layer will spatially arrange the updated node state vectors of all monitoring points according to the actual geographical or electrical location of the monitoring points, and use interpolation technology to generate a continuous and visualized harmonic state map covering the entire power distribution network area. In the harmonic state map, different colors or contour lines represent the severity of harmonic pollution at different locations.

4. The method for detecting harmonics in a distribution network based on smart meter data according to claim 1, characterized in that, Based on the aforementioned harmonic state spectrum, the harmonic source location and responsibility allocation model is used to calculate and output quantitative indicators of responsibility for each potential harmonic source and recommendations for governance priorities, including: In the harmonic state spectrum, the harmonic source region is identified based on the gradient change of the severity of harmonic pollution; Extract the harmonic phase and harmonic power flow information of the monitoring points located in the harmonic source region from the set of harmonic characteristic parameters; Using a harmonic power flow calculation model, combined with the distribution network topology and impedance parameters, reverse power flow tracing is performed to estimate the magnitude of the injected harmonic current from each potential harmonic source. Based on the estimated magnitude of the injected harmonic current of each potential harmonic source and its contribution ratio to the total harmonic pollution, the responsibility quantification index of each potential harmonic source is calculated. By combining the pollution range shown in the harmonic state spectrum, the magnitude of the responsibility quantification index, and the user type of each potential harmonic source, the governance priority recommendation is generated through preset rules or scoring models.

5. The method for detecting harmonics in a distribution network based on smart meter data according to claim 3, characterized in that, The spatial association layer receives the topology data of the distribution network and the node state vectors of all monitoring points. Using a graph neural network model, it simulates the propagation and superposition of harmonic characteristics on the distribution network topology, including: The distribution network topology is abstracted as a graph, where nodes are monitoring points or power grid buses, and edges are feeders or transformer branches. The node state vector of each monitoring point is used as the initial feature of the corresponding graph node; The graph neural network model consists of multiple graph convolutional layers. Each graph convolutional layer performs the following operations: aggregates the features of all neighboring nodes of each node, and updates the node's features after linear transformation and nonlinear activation. Through multi-layer graph convolution operations, the feature information of each node is propagated and interacted along the topological edges in multiple hops.

6. The method for detecting harmonics in a distribution network based on smart meter data according to claim 4, characterized in that, The method utilizes a harmonic power flow calculation model, combined with distribution network topology and impedance parameters, to perform reverse power flow tracing and estimate the magnitude of injected harmonic currents from each potential harmonic source, including: Construct a harmonic impedance network model of the target distribution network at a specific harmonic frequency; The harmonic voltages of each monitoring point identified in the harmonic state spectrum are taken as known node voltages; The location of potential harmonic sources is considered as the source node into which harmonic currents are injected; Based on the harmonic impedance network model, the harmonic voltage equations of the nodes are established, where the harmonic voltage at each monitoring point is a known quantity, and the injected harmonic current at each potential harmonic source is a quantity to be determined.

7. A distribution network harmonic detection system based on smart meter data, characterized in that, The system includes a processor and a memory, the memory and the processor being connected. The memory is used to store programs, instructions or code, and the processor is used to run the programs, instructions or code in the memory to implement the distribution network harmonic detection method based on smart meter data as described in any one of claims 1-6.

Images