A power distribution network three-phase imbalance traceability analysis and treatment method based on user behavior patterns

Abstract

This application relates to the technical field of power system operation and control, and in particular to a method for the tracing and mitigation of three-phase imbalance in distribution networks based on user behavior patterns. The method includes: deploying a pre-trained diagnostic model to the edge gateway of the distribution area or the cloud dispatch master station; real-time monitoring of electrical data, generating tracing results when a three-phase imbalance occurs, and performing three-phase imbalance mitigation; the diagnostic model, based on a sliding window segmentation mechanism, divides real-time electrical data into several time slices and maps them to time-series vectors; employs a weighted granular ball mechanism, using the time-series vectors as granules for adaptive splitting and evolution, modeling a graph structure and extracting typical behavioral model clusters; performing tracing analysis to generate interpretable diagnostic results; this application addresses the problems of insufficient existing three-phase imbalance tracing methods and the lack of interpretability in mitigation decisions.

Description

Technical Field

[0001] This application relates to the technical field of power system operation control, specifically to a method for the analysis and management of three-phase imbalance in distribution networks based on user behavior patterns. Background Technology

[0002] Three-phase load imbalance in a distribution network refers to a situation where the three-phase currents (or voltages) in a distribution network system are unbalanced, and the difference in amplitude between the three phases exceeds a certain threshold. In my country, there are numerous single-phase power supply users, and the randomness and significant regional differences in individual power loads, along with the connection of a large number of high-power loads in multiple phases, often lead to varying degrees of three-phase load imbalance in distribution transformer areas. Three-phase load imbalance increases the energy loss of lines and transformers, reduces transformer output capacity and motor operating efficiency, and generates zero-sequence current, accelerating the aging of equipment insulation, ultimately affecting the reliability and quality of power supply.

[0003] In the existing technology, there are a variety of three-phase imbalance management solutions. For example, load testing and transformer area optimization and redistribution can alleviate the imbalance to a certain extent; increasing inter-phase reactive power compensation can reduce losses and voltage deviations caused by imbalance; load compensation can be performed by introducing load compensation devices according to control strategies; and load balance can be quickly achieved by controlling commutation methods.

[0004] However, methods such as load testing, transformer area optimization and redistribution, and increasing phase-to-phase reactive power compensation have limited effectiveness. Introducing load compensation devices is costly and difficult to implement on a large scale. While controlled commutation can effectively improve balance, it is mostly guided by engineering experience and lacks data-driven quantitative analysis and automation solutions. Furthermore, these methods mainly focus on early warning and prediction of imbalances, while rarely addressing the root cause analysis of imbalances that have already occurred. Summary of the Invention

[0005] In view of this, this application discloses a method for the analysis and management of three-phase imbalance in distribution networks based on user behavior patterns, in order to solve the problems of the lack of existing three-phase imbalance tracing methods, the low efficiency and poor noise resistance of traditional clustering calculations, and the lack of interpretability of traditional management decisions. The method includes: deploying a pre-trained diagnostic model to the distribution area edge gateway or cloud dispatch master station; the distribution area edge gateway or cloud dispatch master station monitors electrical data in real time, and the diagnostic model processes the real-time monitoring data; when a three-phase imbalance occurs, the diagnostic model generates interpretable tracing results, and three-phase imbalance management is carried out based on the tracing results.

[0006] The diagnostic model, based on a sliding window segmentation mechanism, divides real-time electrical data into several time slices and maps the time slices to time-series vectors. It adopts a weighted granular ball mechanism, using the time-series vectors as granular balls for adaptive splitting and evolution, modeling a graph structure and extracting typical behavioral model clusters. It then performs causal analysis on the behavioral model clusters to generate interpretable diagnostic results.

[0007] The beneficial effects of this application include:

[0008] This application extracts historical load time-series data of users in distribution substations through time-series feature extraction, segments the data using a sliding window, and vectorizes the time series into low-dimensional feature vectors. It then introduces graph structure learning into a granular clustering framework, clustering electrical data based on multi-granularity granularity computation theory. A large-range-priority granularity computation is introduced to construct a boundary topology graph structure for high-dimensional feature vectors. Graph convolutional neural networks are used to extract topological features and perform joint clustering, generating typical behavior model clusters. These clusters provide a reference for demand response and power grid planning based on typical electricity consumption patterns. The multi-granularity granularity mechanism is efficient and robust, exhibiting significant advantages in clustering massive amounts of electrical data. Based on this, a three-phase imbalance management method for distribution networks is constructed, achieving interpretable correlation and closed-loop management from macroscopic electrical imbalance to microscopic user behavior. This reduces the blind spot of maintenance personnel in on-site phase adjustment and addresses the lack of existing three-phase imbalance tracing methods and the lack of interpretability in management decisions.

[0009] This provides a new design approach for those skilled in the art. Attached Figure Description

[0010] Figure 1 This is a schematic diagram of the three-phase imbalance cause analysis and management method for power distribution networks based on user behavior patterns in the embodiments of this application;

[0011] Figure 2 This is a schematic diagram of stage two in an embodiment of this application;

[0012] Figure 3 This is a schematic diagram of the optimization process in the embodiments of this application. Detailed Implementation

[0013] To make the objectives, technical solutions, features, and advantages of this application clearer and to facilitate a better understanding of the technical solutions of this application by those skilled in the art, the following detailed description of this application is provided in conjunction with the accompanying drawings and embodiments.

[0014] Example 1:

[0015] This embodiment includes a method for analyzing and managing three-phase imbalance in a distribution network based on user behavior patterns. The method includes: deploying a pre-trained diagnostic model to a distribution area edge gateway or a cloud dispatch master station; the distribution area edge gateway or cloud dispatch master station monitors electrical data in real time, and the diagnostic model processes the real-time monitoring data; when a three-phase imbalance occurs, the diagnostic model generates interpretable tracing results, and the three-phase imbalance is managed based on the tracing results.

[0016] The diagnostic model, based on a sliding window segmentation mechanism, divides real-time electrical data into several time slices and maps the time slices to time-series vectors. It adopts a weighted granular ball mechanism, using the time-series vectors as granular balls for adaptive splitting and evolution, modeling a graph structure and extracting typical behavioral model clusters. It then performs causal analysis on the behavioral model clusters to generate interpretable diagnostic results.

[0017] Specifically, the data processing procedure of the diagnostic model is divided into three progressive levels and explained below, including:

[0018] In the first stage, based on the sliding window sequence segmentation and time-series vectorization representation, the original meter data is transformed into a high-dimensional time-series vector.

[0019] Considering that three-phase imbalance is often caused by brief, localized load surges, such as concentrated electric vehicle charging during specific evening hours or large-scale start-ups and shutdowns of industrial motors, this application employs a fixed-step local sliding window segmentation mechanism to accurately capture transient behavior characteristics. A fixed-span sliding window is set on the historical timeline, with a sliding step size set within the window. The sliding window is shifted along the historical timeline, strictly truncating the continuous time series into several short time series segments of fixed length. In this embodiment, a sliding window with a time span of 2 hours is set, with a corresponding sliding step size of 15 minutes.

[0020] After segmentation, the original sequence fragments still have extremely high dimensionality redundancy and inevitably contain measurement errors and communication noise. Therefore, feature engineering and time series vectorization models are used to process the short time series fragments.

[0021] The feature engineering process extracts statistical features, frequency domain distribution features, and imbalance characterization features reflecting the physical properties of three-phase imbalance within short time series segments to construct a high-dimensional time series feature sequence. Specifically, statistical features and frequency domain distribution features constitute the feature sequence. , representing the statistical information corresponding to the i-th behavioral segment, to characterize the waveform distortion and fluctuation pattern generated by the nonlinear load; the imbalance characterization features include at least the phase load mean offset, phase active power difference, phase current difference, load surge direction, and contribution ratio to the unbalanced phase within a certain time window, used to characterize the potential influence direction and intensity of the behavioral segment on the three-phase imbalance, and used for subsequent particle splitting discrimination, graph edge weight construction, and construction of imbalance contribution constraints.

[0022] The time series vectorization model compresses high-dimensional time series feature sequences into a low-dimensional continuous latent representation space through nonlinear mapping, nonlinearly mapping the original high-dimensional feature sequences into behavioral representation vectors. , This represents the low-dimensional feature vector corresponding to the i-th behavioral segment. Further, the imbalance representation vector corresponding to the i-th behavioral segment is constructed from the imbalance representation features. This is used to characterize the direction and relative intensity of the effect of the behavioral segment on the three-phase imbalance. In some embodiments, the feature sequence is... With behavioral representation vector The vectors are spliced ​​or merged to form the main input vector of the behavior. The imbalance representation vector serves as the main input for particle generation, representative node selection, and basic behavior graph construction. It can also be used for particle splitting regulation, graph edge weight correction, and physical consistency constraint construction.

[0023] To filter out irrelevant interference dimensions in power data features, this embodiment pre-evaluates the high-dimensional feature sequence using LW-K-means before mapping, and assigns a weight vector w to each feature in the sequence, satisfying... LW-K-means is a weighted K-means method with feature weight learning, used to evaluate the contribution of each feature to the clustering results. The assigned weight vector is used to suppress the interference of noise and weakly correlated dimensions on subsequent vectorization and particle construction. The time series vectorization model can use a Long Short-Term Memory (LSTM) autoencoder, and in some embodiments, it can also use encoder architectures such as Temporal Convolutional Network (TCN) or Transformer.

[0024] Through sliding window sequence segmentation and time-series vectorization representation, the first stage transforms the infinite, continuous raw meter data into finite, discrete high-dimensional time-series vectors.

[0025] Phase Two: Based on the Weighted Particle-Sphere (WGB) mechanism, massive, disordered meter feature vectors are modeled into a reliable graph structure, and then clustered using GCN to extract clusters of typical behavioral models; for example... Figure 2 As shown, it includes:

[0026] Step 1: Set all behavior fragment vectors in the latent space. As the initial coarse-grained granules.

[0027] Step 2: Adaptively split and evolve the coarse-grained spheres to obtain a set WGBS containing m highly homogeneous spheres (typically...). , where N represents the total number of behavior segment vectors in the input latent space.

[0028] The adaptive splitting and evolution, for any particle in the latent space The purity and variance of the spheres are calculated. If either the purity or the variance is outside the preset threshold range, the sphere is split in two. The splitting and evolution process is recursively iterated until no spheres meet the splitting condition, resulting in a set of highly homogeneous spheres.

[0029] During the iteration process, the formulas for calculating the center and radius of the sphere are as follows:

[0030]

[0031]

[0032] in, Indicates the center of the ball. Indicates radius, Indicates the th in the current ball One element, Indicates the current number of elements in the sphere. This represents the weight vector assigned by LW-K-means.

[0033] In this embodiment, if the purity of the samples within the current sphere is too low or the variance is too large, a splitting mechanism is triggered, employing an unsupervised dual-center splitting strategy based on K-Means++. Particles that are too small can contain a limited number of samples, resulting in insufficient coverage, while particles that are too large suffer from poor specificity and inconsistent characteristics. The aforementioned adaptive splitting and evolution process addresses the balance between sphere size and sample coverage.

[0034] Step 3: Select representative nodes from the set of highly homogeneous spheres. By searching for representative nodes The nearest neighbors are then fully connected to obtain a global graph skeleton that spans different user behavior patterns.

[0035] In this implementation, considering that the theoretical sphere center c obtained through calculation may not be a real feature vector point, in order to maintain data purity, the real feature node closest to the sphere center is selected through an approximate substitution formula. Representing the current particle, the formula is:

[0036]

[0037] in, This represents the approximate center of the sphere. Selected representative nodes are used as graph nodes. First, a global graph skeleton is constructed based on the nearest neighbor relationships between representative nodes: for any representative node... and ,like belong If a node is one of its k nearest neighbors, then it is in the node... With nodes Establish candidate connections between them.

[0038] Step 4: Combine the global graph skeleton with the connection probability distributions within and between spheres to generate a directed graph P; then symmetric P into an undirected graph. ,from Extract the normalized adjacency matrix of the feature map. .

[0039] The connection probabilities are defined as follows: intra-sphere connection probability characterizes the density of clustering of samples within a corresponding sphere around a representative node, while inter-sphere connection probability characterizes the similarity between representative nodes of different spheres. For any candidate connection edge... The edge weights are determined by combining the global graph skeleton constraints, the compactness within spheres, and the similarity between spheres, thus obtaining the connection probabilities in the directed graph P. In this embodiment, the formula for the connection probability between spheres is:

[0040]

[0041] in, This represents the connection indicator in the global graph skeleton, when a node... For nodes When the nearest neighbor node, =1, otherwise =0; Indicates granules With granules The similarity between them; and They represent granules and balls respectively. With granules The degree of compactness within the sphere; This represents the weighting coefficient, with a value range of 0. 1.

[0042] After constructing a directed graph P based on the above connection probabilities, P is then symmetrically transformed into an undirected graph. ,from Extract the normalized adjacency matrix of the feature map. , represents the global graph skeleton / graph connection relationship.

[0043] Step 5: Use an encoder to process the node feature matrix V and the adjacency matrix. The data is processed by extracting the latent low-dimensional embedding Z from the meter data through graph convolution operations. The formula is as follows:

[0044]

[0045] Where V represents the node feature matrix, and the behavior of the node is formed by stacking the main input vector hi in rows. It is formed by stacking all the feature vectors representing the nodes selected in step 3 in rows. That is, V corresponds to the feature representation of each node in the global graph skeleton. This represents the normalized adjacency matrix extracted from the undirected graph W obtained in step 4; and Indicates the activation function; and Z represents the learnable network weight matrix; Z represents the latent low-dimensional embedding representation matrix of the graph convolutional network output.

[0046] Furthermore, each row of Z corresponds to a low-dimensional embedding feature representing a node, used to characterize the comprehensive behavioral pattern features of that node and its corresponding sphere under graph structure constraints. In other words, Z is neither the original meter data nor the final cluster label, but a low-dimensional discriminative feature representation used for subsequent clustering.

[0047] Step 6: Use the K-means algorithm to divide the feature embeddings to obtain K typical behavior model clusters for the current transformer area.

[0048] Specifically, the low-dimensional vectors in each row of Z are used as clustering samples. Euclidean distance is used to measure the distance between the samples and the cluster centers. Sample allocation and center updates are performed iteratively until the cluster centers converge or the maximum number of iterations is reached, thus completing the clustering. After the representative node completes clustering, the original behavioral fragments in its corresponding spheres inherit the cluster labels, thereby forming K typical behavioral model clusters of the station areas.

[0049] In the above steps, the representative node obtained in step 3 serves as both a node object in the global graph skeleton and an input node for the subsequent graph convolutional network; step 4 completes the edge weight construction and adjacency matrix extraction; step 5 completes node embedding learning; and finally, step 6 completes the division of typical behavior model clusters.

[0050] Furthermore, Phase Two is implemented based on an optimized neural network model; the optimization process is as follows: Figure 3As shown, the optimization process employs joint gradient optimization, which minimizes the KL divergence between the original constructed graph distribution and the reconstructed connected distribution, supplemented by a local consistency penalty term and a physical consistency constraint term, to perform joint gradient optimization. The specific formula is as follows:

[0051] ;

[0052]

[0053]

[0054]

[0055] in, Let represent the joint gradient optimization function. The first term is used to constrain the reconstructed connectivity distribution in the low-dimensional embedding space to remain consistent with the original constructed graph distribution, so as to preserve the global structural information in the original graph. The second term is a local consistency penalty term, which is used to constrain locally adjacent or similar nodes in the original graph to remain adjacent in the low-dimensional embedding space, thereby enhancing the embedding result's ability to preserve the local topology. The third term is a physical consistency constraint term, which is used to constrain nodes with similar imbalance contribution relationships in imbalance representation features to maintain similar representations in the low-dimensional embedding space, so as to improve the sensitivity of the clustering result to the identification of the three-phase imbalance cause. This is the original constructed graph distribution, representing the nodes. Pointing to node The connection probability, This represents the nodes in the original graph obtained from step 4. Pointing to node Connection weights. This represents the reconstructed connected distribution, which is constructed based on the Euclidean distance in the decoder; Represents a node The Euclidean distance between them; This represents a local consistency penalty term, used to constrain nodes that are close to each other and locally adjacent in the original graph to remain close in the low-dimensional embedding space. This represents a potential low-dimensional embedding representation; This represents the graph Laplacian matrix constructed based on the local adjacency relationships of nodes; This represents the graph Laplacian matrix corresponding to the physical consistency graph; Indicates the local consistency penalty weight; This represents a physical consistency constraint term, used to constrain nodes with similar imbalance contribution characteristics to maintain similar representations in the low-dimensional embedding space. This represents the graph Laplacian matrix corresponding to the physical consistency graph. This represents the weight of the physical consistency constraint. , Typically, these are positive real numbers. Based on the aforementioned joint gradient optimization function... The network is optimized, and when the network converges, the final hidden layer representation Z is extracted using this network.

[0056] Phase 3: Mapping and interpretable cause diagnosis of abnormal events.

[0057] Finally, after constructing typical behaviors from offline historical data, the diagnostic model is deployed on the edge gateway of the distribution area or the cloud scheduling master station to achieve real-time targeted governance decisions.

[0058] Example 2:

[0059] This embodiment includes a method for analyzing and managing the root causes of three-phase imbalance in a distribution network based on user behavior patterns. The difference from Embodiment 1 is that this embodiment will be explained in conjunction with a pre-deployed diagnostic model at the distribution network edge gateway or cloud-based dispatching master station. For example... Figure 1 As shown, when the edge gateway of the distribution area or the cloud dispatch master station with the diagnostic model deployed performs three-phase imbalance management, the following steps are executed:

[0060] S1. Set a three-phase imbalance threshold according to a preset three-phase imbalance criterion; the three-phase imbalance threshold is set based on at least one of national standards, industry specifications, allowable operating deviations of equipment, and historical operating data of the distribution area. Calculate the three-phase voltage and current imbalance in the distribution area in real time through the distribution transformer terminal or the low-voltage side monitoring device of the transformer; when a three-phase imbalance threshold alarm is triggered, obtain the time window of imbalance and the specific unbalanced phase.

[0061] S2. Retrieve electrical data sequences of all users on the unbalanced phase within the unbalanced time window, such as voltage, current, power and power factor sequences, and extract real-time behavior segments that are consistent with the time span of the diagnostic model; use a pre-trained vectorized encoder to map the real-time behavior segments into low-dimensional feature vectors.

[0062] S3. Perform inverse mapping of clusters, projecting low-dimensional feature vectors into the latent space containing spheres and clusters, and the diagnostic model generates interpretable attribution results.

[0063] In this stage of scoring, imbalance business indicators are used to make the clustering results more sensitive to grid imbalance than to general load shape. Combining relevant imbalance business indicators, the distribution of each behavioral cluster on different phases and its load contribution are analyzed when an imbalance event occurs. This determines whether a certain type of electricity consumption behavior is concentrated on a certain phase and aligns with the direction of imbalance, thereby identifying the main causal behaviors leading to three-phase imbalance and the corresponding user groups. This step is used to achieve a precise causal mapping from macroscopic electrical imbalance to microscopic behavioral patterns, thus making the diagnostic model designed in this application interpretable. Real-time feature vectors are mapped to the latent space where the spheres and typical behavioral model clusters reside, and the causal score corresponding to each behavioral model cluster is calculated to identify the main causal user groups leading to this event.

[0064] S4. Based on the interpretable causal results, identify the causal user groups and their behavioral characteristics, and select appropriate governance strategies according to constraints such as the upper limit of phase switching frequency, phase switchability, user priority, user scale, and compensation capacity.

[0065] The governance strategies, for example, include implementing phase commutation adjustments for user groups with significantly unbalanced phase loads, generating phase commutation or compensation control commands that meet the action constraints, implementing peak-shifting or demand response for user groups with obvious time-based concentrated load characteristics, and adjusting user groups with continuous high-power load characteristics in combination with energy storage or load transfer, so as to achieve the redistribution and balance of the three-phase load in the distribution area, and ultimately achieve closed-loop governance decision generation.

[0066] Finally, it should be noted that the above description only depicts some embodiments of this application. For those skilled in the art, various changes, modifications, substitutions, and variations can be conceived of these embodiments without departing from the principles and spirit of this application. The scope of protection of this application is defined by the appended claims and their equivalents, and all the above-mentioned behaviors should be covered within the scope of protection of this application.

Claims

1. A method for analyzing and managing the root causes of three-phase imbalance in a power distribution network based on user behavior patterns, characterized in that, include: The pre-trained diagnostic model is deployed to the distribution area edge gateway or cloud dispatch master station; the distribution area edge gateway or cloud dispatch master station monitors electrical data in real time, and the diagnostic model processes the real-time monitoring data; when a three-phase imbalance occurs, the diagnostic model generates interpretable cause-finding results, and the three-phase imbalance is addressed based on the cause-finding results. The diagnostic model, based on a sliding window segmentation mechanism, divides real-time electrical data into several time slices and maps the time slices to time-series vectors. It adopts a weighted granular ball mechanism, using the time-series vectors as granular balls for adaptive splitting and evolution, modeling a graph structure and extracting typical behavioral model clusters. It then performs causal analysis on the behavioral model clusters to generate interpretable diagnostic results.

2. The method for tracing and managing three-phase imbalance of power distribution network based on user behavior pattern according to claim 1, characterized in that, The process of dividing real-time electrical data into several time slices includes: setting a sliding window with a fixed span and sliding step size on the historical time axis, synchronously extracting the phase voltage, phase current, phase power and power factor sequences of each user, and forming short time series segments consistent with the time scale of three-phase imbalance event analysis.

3. The method for tracing and managing three-phase imbalance of power distribution network based on user behavior pattern according to claim 1, characterized in that, Feature engineering and time series vectorization models are used to process short time series segments; The feature engineering includes: extracting statistical features, frequency domain distribution features, and imbalance characterization features reflecting the physical characteristics of three-phase imbalance within short time series segments, and constructing a high-dimensional time series feature sequence; the imbalance characterization features include: phase current deviation, phase power deviation, phase load bias degree, and phase contribution difference within the event window; The time series vectorization model compresses high-dimensional time series feature sequences into a low-dimensional continuous latent representation space through nonlinear mapping, and nonlinearly maps the original high-dimensional feature sequences into a feature vector.

4. The method for tracing and managing three-phase imbalance of power distribution network based on user behavior pattern according to claim 1, characterized in that, The adaptive splitting and evolution process includes: for any particle in the latent space, calculating the sample purity and variance of the particle; if either the sample purity or the variance is not within a preset threshold range, then splitting the particle in two; the splitting and evolution process is recursively iterated until no particle satisfies the splitting condition, thus obtaining a set of highly homogeneous particles.

5. The method for analyzing and managing three-phase imbalance in a distribution network based on user behavior patterns according to claim 1, characterized in that, The modeling process yields a graph structure and extracts a cluster of typical behavioral models, including: Representative nodes are selected from the set of highly homogeneous spheres. By searching the nearest neighbors of the representative nodes and performing full connections, a global graph skeleton spanning different user behavior patterns is obtained. The set of highly homogeneous spheres is obtained through adaptive splitting and evolution. The global graph skeleton is combined with the connection probability distribution within and between spheres to generate a directed graph; the directed graph is then symmetricized into an undirected graph, and the normalized adjacency matrix of the feature graph is extracted from the undirected graph. An encoder is used to process the node feature matrix and adjacency matrix, and graph convolution operation is used to extract the latent low-dimensional embeddings in the meter data. The K-means algorithm is used to divide the feature embeddings to obtain K typical behavioral model clusters for the current transformer area.

6. The method for tracing and managing three-phase imbalance of power distribution network based on user behavior pattern according to claim 5, characterized in that, The representative node is selected from the set of high homogeneity granules, and the nearest real feature node to the center of the granule is selected by an approximate replacement formula The representative current granule is selected by the formula ； wherein, denotes the approximated granule sphere center, denotes the weight vector assigned by the LW-K-means, denotes the current granule, denotes the set of granules, denotes the original sphere center of the granule.

7. The method for tracing and managing three-phase imbalance of power distribution network based on user behavior pattern according to claim 5, characterized in that, The latent low-dimensional embedding is formulated as follows: ； in, Represents a potential low-dimensional embedding. Represents the adjacency matrix. Represents the node feature matrix, and This represents the activation function. and This represents the learnable network weight matrix.

8. The method for analyzing and managing three-phase imbalance in a distribution network based on user behavior patterns according to claim 1, characterized in that, The process employs a weighted particle-sphere mechanism, using time-series vectors as particles for adaptive splitting and evolution. This modeling yields a graph structure and extracts clusters of typical behavioral models, implemented based on an optimized neural network model. The optimization process utilizes joint gradient optimization, as shown in the formula: ； in, Denotes the joint gradient optimization function. Represents the original construction graph distribution. Indicates the reconstructed connected distribution. Represents a node The Euclidean distance between them This represents the local consistency penalty term. This represents a potential low-dimensional embedding representation. The Laplace matrix of the graph is represented. Indicates the local consistency penalty weight; Represents physical consistency constraints. This represents the graph Laplacian matrix corresponding to the physical consistency graph. This represents the weight of the physical consistency constraint.

9. The method for tracing and managing three-phase imbalance of power distribution network based on user behavior pattern according to claim 1, characterized in that, The generation of interpretable diagnostic results includes: performing abnormal event mapping and interpretable cause-finding diagnosis.

10. The method for tracing and managing three-phase imbalance in power distribution networks based on user behavior patterns as claimed in claim 1, wherein, When performing three-phase imbalance management on the edge gateway of the distribution area or the cloud dispatch master station with the diagnostic model already deployed, the following steps shall be performed: S1. Set the three-phase unbalance threshold; calculate the three-phase voltage and current imbalance in the distribution area in real time through the distribution transformer terminal or the low-voltage side monitoring device of the transformer; when the three-phase unbalance threshold alarm is triggered, obtain the time window of the imbalance and the specific unbalanced phase. S2. Retrieve the electrical data sequence of all users on the unbalanced phase within the unbalanced time window, and extract the electrical data sequence into real-time behavior segments that are consistent with the time span of the diagnostic model; use a pre-trained vectorized encoder to map the real-time behavior segments into low-dimensional feature vectors. S3. Perform inverse mapping of clusters, projecting low-dimensional feature vectors into the latent space containing spheres and clusters, and the diagnostic model generates interpretable attribution results. S4. Based on interpretable attribution results, identify the causal user groups and their behavioral characteristics, and select appropriate governance strategies.

Images