A power distribution network fault monitoring method and system based on an electrical quantity atlas

Abstract

The application discloses a power distribution network fault monitoring method and system based on an electrical quantity atlas, and belongs to the technical field of power grid fault detection. The method comprises the following steps: collecting electrical quantity time sequence data with time and space labels and storing the data in layers; constructing a node dependence matrix based on topology, combining dynamic feature extraction and abnormal marking to generate a multi-dimensional atlas; fusing time, space and electrical characteristics, and adaptively weighting to generate a feature vector according to a high-resistance / short-circuit mode; utilizing cloud distributed computing classification storage to form a fault mode library; matching a preset template through flow processing and dimension reduction; deeply analyzing time sequence data to calculate a fault probability; mapping a high-risk probability into a three-dimensional view, combining time and space correlation analysis and historical atlas backtracking to accurately locate a risk section where multiple mutation points are gathered. The system comprises a time and space data acquisition and layered storage module and the like. The application realizes closed-loop identification from fault probability prediction to physical location, and improves the fault detection rate and positioning accuracy of a remote branch line high-resistance fault.

Description

Technical Field

[0001] This invention relates to the field of power grid fault detection technology, and in particular to a method and system for monitoring distribution network faults based on electrical quantity spectra. Background Technology

[0002] Existing power distribution network fault monitoring and defect identification technologies typically rely on distributed terminal acquisition devices to independently collect time-series data of electrical quantities such as voltage, current, and power factor, and then detect anomalies using simple threshold comparisons or statistical analysis methods. While timestamps and geographical coordinates can be recorded during data acquisition, the lack of a unified spatiotemporal hierarchical storage mechanism leads to problems such as time misalignment and coordinate mismatch during the transmission and aggregation of multi-source heterogeneous data. Fault identification often employs single-dimensional threshold judgment or traditional machine learning models to analyze local electrical signals in isolation, failing to effectively integrate the dynamic dependencies between nodes in the electrical network topology and making it difficult to reveal the propagation path and cascading effects of faults in the power grid. Furthermore, due to the realities of weak signals, high noise levels, and severe sampling gaps in power distribution networks, especially in remote branches, traditional preprocessing methods, using only linear interpolation or sliding filtering, cannot achieve high-quality data restoration while preserving temporal continuity and spatial correlation, thus distorting the input data for subsequent analysis.

[0003] Building upon this, while existing technologies attempt to introduce multi-dimensional feature fusion strategies, they often employ fixed weights to simply concatenate temporal, spatial, and electrical features. They fail to adaptively adjust feature contributions based on the dynamic characteristics of typical fault modes such as high-resistance faults (manifested as slow voltage drift and weak current changes) and short-circuit anomalies (manifested as instantaneous current surges and voltage drops), resulting in feature vectors that cannot truly reflect the essence of the fault. Furthermore, fault mode recognition generally relies on offline-trained static models, lacking efficient processing capabilities for real-time streaming data, making it difficult to achieve low-latency online matching and classification. Moreover, due to the explosion of high-dimensional features, effective dimensionality reduction optimization mechanisms are generally not introduced, leading to a surge in computational complexity and failing to meet the real-time requirements of edge and cloud collaboration. Ultimately, fault location remains at the regional level with coarse-grained early warning. There is a lack of analytical methods to deeply correlate probability prediction results with spatiotemporal graph structure and historical anomaly trajectories. This makes it impossible to accurately locate the physical sections where abrupt changes and outliers are clustered. As a result, maintenance personnel have to rely on experience to conduct large-scale manual inspections. The early warning is delayed, the location is inaccurate, and the false alarm rate is high, which seriously restricts the improvement of the intelligent operation and maintenance level of the distribution network. Summary of the Invention

[0004] This invention aims to provide a method and system for monitoring distribution network faults based on electrical quantity maps. By constructing an integrated processing flow of spatiotemporal hierarchical storage, node-dependent maps, adaptive feature weighting, cloud parallel fusion, and three-dimensional spatiotemporal positioning, it achieves accurate identification of high-resistance faults and short-circuit anomalies in distribution networks and precise location of risk sections. It overcomes the problems of high false alarm rate, delayed response, and difficulty in covering remote branches caused by traditional methods due to isolated data, fixed weights, and ambiguous positioning.

[0005] The technical solution adopted in this invention is:

[0006] A method for monitoring distribution network faults based on electrical quantity spectra includes the following steps:

[0007] Step S1: Collect electrical quantity time-series data containing timestamps and geographic coordinates, and store it in layers; wherein, the electrical quantity time-series data includes voltage, current, and power factor;

[0008] Step S2: Construct a node dependency matrix based on the electrical network topology, and combine dynamic feature extraction and anomaly marking to generate a multi-dimensional graph structure containing node associations and anomaly markings;

[0009] Step S3: Extract time, space and electrical features from the multidimensional graph structure, and dynamically adjust the feature weights of each dimension according to the high-resistance fault mode or short-circuit anomaly mode through an adaptive weight allocation mechanism to generate a weighted feature vector set.

[0010] Step S4: Use a cloud-based distributed computing framework to process and summarize the weighted feature vector set in parallel, forming a structured fault feature vector set, and store it according to high-resistance fault and short-circuit anomaly modes to build a fault mode library.

[0011] Step S5: The continuous electrical quantity time series data is segmented into time windows through a real-time streaming pipeline, and the optimized feature vector is obtained after dimensionality reduction by principal component analysis and matched with the preset template in the fault mode library.

[0012] Step S6: Use a long short-term memory network to perform in-depth analysis on the continuous electrical quantity time series data to calculate the probability of high-resistance faults and short-circuit anomalies; when the probability exceeds a preset risk threshold, it is determined to be a high-risk fault type.

[0013] Step S7: Map the risk probability distribution of high-risk fault types into a three-dimensional visualization view of time, space and electrical characteristics. Combine the spatiotemporal correlation analysis method to match its spatial location and time window, and backtrack the historical map data to locate the high-risk segment formed by the aggregation of multiple mutation values ​​and outliers, and output the final risk segment location information.

[0014] A power distribution network fault monitoring system based on electrical quantity graphs includes:

[0015] The spatiotemporal data acquisition and hierarchical storage module is used to acquire electrical quantity time-series data containing timestamps and geographic coordinates, and to store it in hierarchical layers; among which, the electrical quantity time-series data includes voltage, current, and power factor;

[0016] The multidimensional graph construction module is used to generate a dependency matrix between nodes based on the electrical network topology, and combine dynamic feature extraction and outlier marking to generate a multidimensional graph structure containing node associations and outlier markings.

[0017] The cross-dimensional feature fusion and weighting module is used to extract temporal, spatial and electrical features from the multi-dimensional map structure. Through an adaptive weight allocation mechanism, the feature weights of each dimension are dynamically adjusted according to the high-resistance fault mode or short-circuit anomaly mode to generate a set of weighted feature vectors.

[0018] The cloud-based distributed feature processing module is used to perform parallel processing and aggregation of the weighted feature vector set using a cloud-based distributed computing framework, forming a structured fault feature vector set, and storing it according to high-resistance fault and short-circuit anomaly modes to build a fault mode library.

[0019] The real-time streaming feature matching module is used to segment the continuous electrical quantity time series data into time windows through the real-time streaming processing pipeline, obtain optimized feature vectors after dimensionality reduction using principal component analysis, and match them with preset templates in the fault mode library.

[0020] The deep time series probability prediction module is used to perform deep analysis on continuous electrical quantity time series data using a long short time memory network, calculate the probability of occurrence of high resistance faults and short circuit anomalies, and determine the fault type as high risk when the probability exceeds a preset risk threshold.

[0021] The 3D visualization and risk location module is used to map the risk probability distribution of high-risk fault types into a 3D view of time, space and electrical characteristics. It combines spatiotemporal correlation analysis methods to match its spatial location and time window, and backtracks historical map data to locate high-risk segments formed by the aggregation of multiple mutation values ​​and outliers, and outputs the final risk segment location information.

[0022] The beneficial effects of this invention are:

[0023] This invention designs a method and system for monitoring distribution network faults based on electrical quantity maps. By constructing an integrated processing flow of spatiotemporal hierarchical storage, node-dependent maps, adaptive feature weighting, cloud-based parallel fusion, and three-dimensional spatiotemporal positioning, it achieves accurate identification of high-resistance faults and short-circuit anomalies in distribution networks and precise location of risk sections. It overcomes the problems of high false alarm rate, delayed response, and difficulty in covering remote branches caused by traditional methods due to isolated data, fixed weights, and ambiguous positioning. It improves the accuracy, real-time performance, and engineering practicality of fault identification and effectively supports the closed-loop management of intelligent operation and maintenance of distribution networks. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a flowchart illustrating the distribution network fault monitoring method based on electrical quantity graphs in this embodiment. Detailed Implementation

[0026] The embodiments of the invention will now be described in detail with reference to the accompanying drawings.

[0027] One embodiment of this application proposes a method for monitoring distribution network faults based on electrical quantity maps. For example... Figure 1 As shown, the method includes the following steps:

[0028] Step S1: Collect electrical quantity time-series data containing timestamps and geographic coordinates, and store it in layers; wherein, the electrical quantity time-series data includes voltage, current, and power factor.

[0029] In this embodiment, "layered storage" refers to dividing electrical quantity time-series data into continuous time slices based on timestamps. Within each time slice, it is further divided into several geographic grid units based on geographic coordinates (for example, setting the size of the geographic grid unit to 500m×500m, generating the geographic grid code using the H3 geographic grid system, with each grid corresponding to a unique 64-bit integer code, and aligning the grid boundary with the location of the distribution network feeder topology node). This forms a spatiotemporal dual-dimensional data block. Each data block is stored independently and an index mapping relationship is established. The index includes timestamp intervals, geographic grid codes, and data offset fields. By employing the aforementioned hierarchical storage method, when tracing back historical map data, the corresponding time slice can be quickly locked by matching the spatial location and time window in the 3D view through the timestamp interval. The target geographic grid unit can be accurately located through geographic grid coding. Then, the historical electrical quantity time series data within the data block can be directly read with the help of data offset. This allows for the efficient extraction of the evolution information of abrupt changes and outliers in a specific spatiotemporal region without the need for a full scan of massive amounts of data. This improves the speed and accuracy of locating risk sections, effectively solves the problem of historical backtracking delay caused by the fuzzy data storage structure, and enables accurate tracking and rapid response to fault modes in remote branches of the distribution network.

[0030] In this embodiment, this step first acquires time-series data of electrical quantities such as voltage, current, and power factor from key nodes in the distribution network (e.g., the low-voltage outgoing terminals of distribution transformers, feeder sectionalizing switches or tie switches, DTU / FTU installation points, user-specific / public transformer metering points, distributed power grid connection points, cable joints / overhead line splicing points, etc.) through the acquisition terminal. Simultaneously, the timestamp and geographic coordinates corresponding to each set of electrical quantity time-series data are recorded to ensure the spatiotemporal traceability of the data source. The acquired raw data is not merged or integrated across nodes in any way. Instead, it is stored independently according to its timestamp and geographic coordinates, following a hierarchical structure. This ensures that the time and spatial dimensions of the data records remain independent and separable, providing a clear and well-defined raw data foundation for subsequent processing.

[0031] Step S2: Construct a dependency matrix between nodes based on the electrical network topology, perform real-time calculations on voltage and current data using a dynamic feature extraction algorithm, and identify abrupt changes and outliers using an anomaly marking mechanism to generate a multi-dimensional graph structure containing node associations and anomaly markers. The dependency matrix is ​​constructed based on hierarchically stored electrical quantity time-series data, and the input to the anomaly marking mechanism is the real-time calculated voltage and current data.

[0032] The construction of the "dependency matrix" in this embodiment includes: based on the electrical impedance coupling degree between nodes, the weight of energy transmission direction, and the topological adjacency relationship, a weighted graph model is used (for example, the weight coefficients of impedance coupling degree, energy transmission direction weight, and adjacency relationship weight are set to 0.5, 0.3, and 0.2 respectively in the weighted graph model, and the inter-node associations with a power factor lower than 0.5 are eliminated during matrix construction to exclude ineffective coupling paths dominated by reactive power) to quantify the influence strength between nodes. The impedance coupling degree is obtained by normalizing the inverse of the line impedance between nodes, reflecting the physical coupling strength of the electrical distance. The energy transmission direction weight is determined by the active power flow direction: when node i transmits active power to node j (Pij>0), the weight is 1.0; when node j transmits active power to node i (Pji>0), the weight is 0.4 (reverse power flow attenuation factor); when both bidirectional power is less than 5% of the rated power, it is considered that there is no effective energy transmission, and the weight is 0. Topological adjacency: If node i and node j have a direct electrical connection (including feeders, switches, transformers, etc.), the value is 1; otherwise, it is 0. Matrix element Mij = α × impedance coupling degree + β × energy transmission direction weight + γ × topological adjacency; where α = 0.5, β = 0.3, and γ = 0.2 are measured and calibrated weight coefficients. To avoid interference from reactive power-dominated connections, if the power factor between nodes is below 0.5, Mij is forcibly set to 0 to eliminate invalid coupling paths. The above design makes the dependency matrix not only reflect topological connectivity but also realistically depict the asymmetric influence of active power flow direction on electrical coupling strength. It is particularly suitable for modern distribution networks with distributed power sources, providing a physically interpretable graph structure basis for identifying the back propagation path of high-resistance faults and analyzing multi-power source cooperative disturbances. This enhances the accuracy and reliability of anomaly correlation analysis in multidimensional graph structures, ultimately achieving more accurate identification and location of high-resistance faults and short-circuit anomalies in the fault feature extraction and pattern recognition stages.

[0033] In this embodiment, the "dynamic feature extraction algorithm" includes: real-time calculation of the voltage change rate, current harmonic distortion rate, power factor slope, and time-varying characteristics of voltage and current phase difference at each node; wherein, the voltage change rate is obtained by dividing the difference between adjacent sampling points by the sampling period; the current harmonic distortion rate is calculated based on the ratio of the root mean square value of the 2nd to 25th harmonic components to the fundamental amplitude using a fast Fourier transform (which can use a Hamming window function for data weighting, with a sampling window length of 1024 points, a frequency resolution of 5Hz, and only retaining components with amplitudes greater than 1% of the fundamental amplitude for calculation); the power factor slope is obtained by fitting the local linear trend of the power factor time series using the least squares method; and the time-varying characteristics of the phase difference are obtained by using the Histogram method. The Elbert transform extracts the instantaneous phase difference between voltage and current signals and calculates their first-order difference, thereby accurately reflecting the dynamic phase deviation characteristics. These refined electrical time-varying features together constitute a multi-dimensional electrical characterization with strong distinguishing ability for high-resistance faults (manifested as slow slope of power factor change and phase drift) and short-circuit anomalies (manifested as sudden voltage change and surge of harmonic distortion). This enhances the discriminative power of features in the multi-dimensional graph structure, provides high-precision input for the subsequent adaptive weight allocation mechanism, and enables the system to dynamically optimize feature weights according to the fault mode type, thereby improving the accuracy and response speed of fault identification. It effectively solves the problems of mode confusion and early warning lag caused by coarse features, and realizes accurate location and early warning of distribution network faults.

[0034] In this embodiment, the "outlier marking mechanism" includes: using a sliding window (e.g., setting the sliding window length to 20 sampling points, the sampling frequency to 1kHz, the σ threshold dynamically calculated based on the local data standard deviation, with each node independently setting its σ threshold, and adaptively adjusting it based on the dispersion of historical normal operation data) statistical method to identify outliers in voltage or current data that deviate from the 3σ threshold; combining the Cumulative Sum Control Chart (CUSUM) algorithm to detect asymptotic abrupt changes in voltage or current sequences; and the Exponential Weighted Moving Average (EWMA) algorithm to detect small, persistent deviations. When the CUSUM statistic exceeds a preset control limit or the EWMA value... When the value exceeds ±2.5σ for five consecutive sampling periods, it is marked as a transient mutation value, thereby enabling simultaneous identification and accurate classification of three types of electrical anomalies: outliers, gradual mutations, and minor persistent anomalies. This effectively compensates for the problem of missed detection of early slow evolution characteristics of high-resistance faults caused by traditional methods that rely solely on a single threshold or sudden detection. It also improves the completeness and spatiotemporal consistency of anomaly markings in the multidimensional map structure, providing a high-precision, multi-dimensional input basis for subsequent adaptive feature weight allocation, fault mode matching, and risk section location. Ultimately, it enables early warning and accurate identification of highly concealed and slowly evolving electrical defects in the distribution network.

[0035] In this embodiment, the step of constructing a dependency matrix between nodes based on the electrical network topology involves analyzing the physical connections and energy transmission paths between nodes in the electrical network, quantifying the impact of node state changes on adjacent nodes, and thus forming a structured matrix reflecting the strength of electrical associations between nodes. Simultaneously, a dynamic feature extraction algorithm is used to perform real-time calculations on voltage and current data. This involves continuously acquiring the time-series voltage and current data of each node and using the algorithm to deduce trend characteristics reflecting changes in electrical state in real time. Based on this, an anomaly marking mechanism is used to scan the real-time calculated voltage and current data, identifying abrupt changes and outliers that deviate from the normal operating range, and associating these anomalies with the corresponding nodes as marker information. Finally, the dependency matrix between nodes and the anomaly marking information are integrated to form a multi-dimensional graph structure that simultaneously contains the node electrical association structure and anomaly state identifiers. This structure directly reflects the mutual influence relationships between nodes and the spatial distribution of abnormal events in the distribution network.

[0036] Step S3: Extract temporal, spatial, and electrical features, and dynamically adjust the feature weights of each dimension according to the high-resistance fault mode or short-circuit anomaly mode through an adaptive weight allocation mechanism to form a weighted feature vector set; wherein, the temporal, spatial, and electrical features come from a multi-dimensional graph structure, and the adjustment of the adaptive weight allocation mechanism is based on the type of high-resistance fault mode or short-circuit anomaly mode.

[0037] In this embodiment, the "adaptive weight allocation mechanism" includes the following: when a high-impedance fault mode is identified, the weight of time features is increased to 0.5~0.7, the weight of electrical features is decreased to 0.2~0.3, and the weight of spatial features remains constant at 0.1~0.2; when a short-circuit anomaly mode is identified, the weight of electrical features is increased to 0.6~0.8, the weight of time features is decreased to 0.1~0.2, and the weight of spatial features remains constant at 0.1~0.2. The time features include voltage change rate, power factor slope, and phase difference change rate; the electrical features include current amplitude, voltage amplitude, harmonic distortion rate, and absolute power factor value; and the spatial features include node geographic density and topological centrality. By dynamically setting the quantification weight ranges of the three dimensions of time, electrical, and space according to the type of high-resistance fault mode or short-circuit anomaly mode, the essential characteristics of the two types of faults are enhanced and interference is suppressed in a targeted manner. This improves the sensitivity of the weighted feature vector set to the gradual changes in high-resistance faults and the ability to respond to the sudden changes in short-circuit anomalies. It effectively enhances the matching accuracy and classification distinction of subsequent fault templates, and ultimately overcomes the problems of pattern recognition ambiguity and positioning lag caused by the fuzziness of feature contribution, so as to achieve more accurate and reliable distribution network fault identification and risk warning.

[0038] Meanwhile, in this embodiment, the adaptive weight allocation mechanism can classify the initial feature vector through a shallow neural network. The network input consists of the voltage, current, power factor, and their first-order differences from the first 10 sampling points. The output layer is a soft classifier, outputting three probabilities: high-resistance fault, short-circuit anomaly, or normal state. The training samples come from a historical fault feature database. The network structure includes one hidden layer with 64 neurons, using ReLU as the activation function and cross-entropy as the loss function. This allows for the automatic identification of the fault mode type corresponding to the current electrical quantity time series data without the need for manually setting thresholds or rules. The probability output is used as the basis for the dynamic adjustment of the adaptive weight allocation mechanism, enabling the weight allocation of time, space, and electrical features to be optimized in real time according to the probability distribution of the actual fault mode. This effectively overcomes the problems of poor generalization and misjudgment of edge scenarios caused by the reliance on fixed rules in traditional methods, improving the accuracy and adaptability of fault mode recognition. This provides more accurate feature input for subsequent cloud-based distributed parallel processing, ultimately achieving high-precision positioning and low-latency early warning of complex faults in the distribution network.

[0039] In this embodiment, this step extracts temporal, spatial, and electrical features from a multi-dimensional graph structure. Based on the type of high-resistance fault mode or short-circuit anomaly mode, the weights of these three dimensions are dynamically adjusted to form a weighted feature vector set. Temporal features reflect the evolution of electrical quantities over time, spatial features characterize the locational correlation of faults within the geographical distribution of the distribution network, and electrical features reflect the operating status of physical quantities such as voltage and current. By identifying whether the current fault belongs to a high-resistance fault mode or a short-circuit anomaly mode, the system automatically assigns different weight ratios to temporal, spatial, or electrical features according to the mode type. This ensures that modes-sensitive dimensions receive higher representational weights during feature fusion, thereby enhancing the feature vector's ability to represent the essence of the fault.

[0040] Step S4: The weighted feature vector set is processed in parallel using a cloud-based distributed computing framework to generate a fault feature vector set. This set is then classified and stored according to high-resistance faults and short-circuit anomalies to construct a fault mode library. The input to the parallel processing is the weighted feature vector set, and the output of the classified storage is a structured fault feature vector set.

[0041] In this embodiment, the "cloud-based distributed computing framework" adopts the Spark architecture. The weighted feature vector set is divided into multiple sub-task units according to the feeder topology. Each sub-task unit corresponds to all nodes under a key node of the distribution network. Feature vector standardization, missing value imputation, and anomaly label verification are performed in parallel within the sub-task. During missing value imputation, a corresponding hierarchical imputation mechanism is used to complete the missing electrical features. After each sub-task completes its local processing, it is merged and output through a Reduce operation to form a structured and highly consistent set of fault feature vectors. Thus, in scenarios where data collection is incomplete and noise interference is severe in remote branches of large-scale distribution networks, topology-constrained accurate completion of missing electrical features is achieved, improving the reliability and classification accuracy of fault feature vectors. This effectively solves the problems of fault identification misjudgment and early warning lag caused by improper imputation methods, and ultimately achieves high precision, high robustness, and strong scalability in distribution network fault monitoring. In this embodiment, the "layered interpolation mechanism" is as follows: (1) For electrical features with a single point or no more than 3 consecutive missing sampling points, the mean interpolation method is used to calculate the local mean based on the historical time series data of neighboring nodes within the same feeder to ensure that the interpolation process is strictly limited by the topological boundary and avoids feature misleading introduced by cross-feeder data; (2) For long-term missing sequences with more than 3 consecutive missing sampling points, a combination of linear interpolation and local weighted regression (LOESS) is used to fill the missing points. The interpolation window is set to 5 sampling points before and after the missing point, and a Gaussian kernel function truncated by three standard deviations is used to assign weights to neighboring points to fit the local nonlinear trend and effectively restore the dynamic evolution process of electrical quantities such as voltage and current. This layered interpolation mechanism ensures computational efficiency while taking into account the integrity of data and the authenticity of electrical dynamic features, improving the reliability of fault features and effectively avoiding fault misjudgment and early warning lag caused by interpolation distortion.

[0042] In this embodiment, "parallel processing" refers to dividing the weighted feature vector set into multiple sub-task units based on the distribution network feeder topology. Each sub-task unit strictly corresponds to an electrical sub-region defined by key nodes in the distribution network, typically all nodes within the power supply range of a distribution transformer, ensuring that the data processing boundary is completely consistent with the physical structure of the electrical network. In Spark's Map phase, each sub-task unit independently performs feature vector normalization verification, anomaly label consistency check, and high-dimensional feature redundancy removal operations. In the Reduce phase, the intermediate results of each sub-task are merged through a convergence node to generate a fault feature vector set in a unified format. Subsequently, according to preset fault mode classification rules, the merged feature vectors are labeled and structurally archived according to two typical modes: "high resistance fault" and "short circuit anomaly," forming a structured fault feature vector set, which is stored in a distributed database. This parallel processing mechanism aligns distributed computing with the physical structure of the power grid through topology-aware subtask partitioning. This avoids communication overhead and topology distortion caused by cross-regional data interaction, and enables efficient parallel processing of massive feature vectors. Compared with traditional centralized processing, the computational efficiency is improved by ≥4 times, providing low-latency feature input for real-time fault identification.

[0043] In this implementation, "classified storage" refers to storing the fault feature vector set according to fault mode type into a high-resistivity fault feature library and a short-circuit anomaly feature library. Each type of feature vector is labeled with fault duration, mutation amplitude, and spatial cluster density. The fault duration is calculated from the time difference between the first and last anomaly marking. The mutation amplitude is the maximum absolute difference between the anomaly point and the previous steady-state value. Spatial cluster density is the ratio of the number of anomaly-marked nodes within a unit geographic grid to the total number of grids. Specifically, it can be estimated using a Gaussian kernel density method with a kernel bandwidth of 500 meters. A two-dimensional Gaussian kernel function is applied to the geographic coordinates of each anomaly-marked node, and the maximum density value after superposition is taken as the spatial cluster density index of the fault event. This transforms the originally discrete node count ratio into a continuous density expression reflecting the spatial distribution gradient and local peaks, effectively overcoming the defects of boundary ambiguity and misjudgment of sparse areas caused by grid division. It upgrades the spatial concentration judgment of fault events from coarse-grained statistics to sub-meter precision peak location, improving the accuracy of identifying mutation points and outlier clusters in risk sections, and thus enhancing the spatiotemporal coordination accuracy of fault location and early warning response. In this embodiment, the fault feature vector set is classified and stored in the corresponding feature library according to high-resistance fault and short-circuit anomaly modes. Each type of feature vector is attached with three types of structured labels: fault duration, mutation amplitude, and spatial clustering density. This makes each template in the fault mode library not only contain multi-dimensional features of electrical quantities, but also fully carry the dynamic evolution information of the fault from occurrence to dissipation: the fault duration is accurately calculated by the time difference between the first anomaly mark and the last anomaly mark, reflecting the continuity of the fault; the mutation amplitude is the maximum absolute difference between the anomaly point and the previous steady-state value, quantifying the severity of the fault; and the spatial clustering density is the ratio of the number of anomaly mark nodes in a unit geographic grid to the total number of grids, which describes the degree of concentration of the fault in the spatial dimension. The aforementioned tags are automatically generated based on the temporal and spatial data obtained through the anomaly marking mechanism and are deeply bound to the classification and storage structure. This enables the typical evolution patterns of high-resistance faults and short-circuit anomalies to be expressed in a structured manner, thereby improving the distinguishability and accuracy of pattern matching. It supports the historical reproduction and risk evolution trend analysis of similar faults, effectively solving the problems of low matching accuracy and weak pattern library evolution capability caused by the lack of dynamic evolution representation of feature vectors in the original technology. Ultimately, it achieves accurate identification, risk classification, and location early warning of distribution network faults.

[0044] In this embodiment, this step uses a cloud-based distributed computing framework to process the weighted feature vector set in parallel. Large-scale feature data is decomposed into multiple sub-task units and distributed to multiple cloud computing nodes for synchronous execution, thereby accelerating the calculation and integration of each feature vector. After parallel processing, the system summarizes and organizes the output results of each sub-task to form a complete fault feature vector set. Subsequently, the fault feature vector set is classified and stored according to the feature patterns of high-resistance faults and short-circuit anomalies, with different types of fault feature vectors assigned to their corresponding storage areas to form a structured fault feature vector set.

[0045] Step S5: The continuous electrical quantity time series data is segmented and extracted through a real-time streaming data processing pipeline, and the feature dimensionality is reduced using principal component analysis to obtain an optimized feature vector. The optimized feature vector is then matched with the high-resistance fault template and short-circuit anomaly template pre-stored in the fault mode library. If the matching degree is lower than a preset threshold, it is marked as an unknown mode and stored in the dataset to be analyzed. The input for segmented extraction is the continuous electrical quantity time series data, the input for feature dimensionality reduction is the result of segmented extraction, and the input for matching is the optimized feature vector.

[0046] In this embodiment, the time window length of the "real-time streaming data processing pipeline" is 10-60 seconds, dynamically adjusted according to the fluctuation of the power grid load. A short window (10-20 seconds) is used during high-load periods, and a long window (40-60 seconds) is used during low-load periods. The window length adjustment is driven by the load forecasting module. The load forecasting model (e.g., using a Long Short-Term Memory network containing two layers of LSTM units, with a hidden layer dimension of 128 and an output layer of a single-node linear regression layer, with prediction error optimized using the MAE loss function) takes historical load, weather temperature, and daily type label as input and outputs the load change trend for the next 5 minutes. If the predicted change rate exceeds 15%, the window is shortened to 10 seconds. This mechanism ensures that the length of the segmented electrical quantity time-series data accurately matches the evolution rate of the actual fault characteristics. This effectively preserves the weak abrupt change signals of high-resistance faults and the high-frequency energy distribution of short-circuit anomalies in the optimized vector generated after feature dimensionality reduction, avoiding the loss of key information or redundant interference caused by a fixed window. This improves the accuracy of fault mode matching and the timeliness of system response, ultimately achieving high-precision identification and low-latency early warning of complex faults in the distribution network.

[0047] In this embodiment, the "principal component analysis method" includes: First, Z-score standardization is performed on the original feature vectors of all key nodes in the distribution network within the time window, encompassing six dimensions: voltage amplitude, current amplitude, power factor, voltage change rate, current harmonic distortion rate, and phase difference change rate. The mean and standard deviation parameters used for standardization are dynamically obtained based on historical electrical quantity time-series data from the same time period over the past seven days to eliminate the interference of instantaneous outliers on the statistical distribution and ensure the stability of the feature scale and the representativeness of the distribution. Then, based on the standardized feature vector set, a covariance matrix is ​​constructed in real time, and eigenvalue decomposition is performed on this covariance matrix to calculate each principal component and its corresponding variance contribution rate. Finally, the top 3 to 5 principal components with a cumulative variance contribution rate of not less than 95% are selected as optimized feature vectors, achieving effective compression of the original feature space. Through the aforementioned dimensionality reduction process, while retaining over 95% of the information in the original features, the feature dimension is compressed from 6 dimensions to 3-5 dimensions, significantly reducing the computational overhead of subsequent fault mode matching and long short-term memory networks. Simultaneously, because the standardization and variance contribution rate constraint mechanisms effectively suppress noise-dominated pseudo-correlation structures, the extracted principal components can accurately reflect the core electrical mode characteristics of high-resistivity faults in distribution networks, such as the subtle voltage gradient trend, current harmonic mutations, and phase jumps in short-circuit anomalies. This avoids information loss or feature distortion caused by excessive dimensionality reduction, thereby improving the accuracy of fault template matching and the stability of fault probability prediction, ultimately achieving a dual improvement in fault identification accuracy and early warning response timeliness.

[0048] The "high-resistance fault template" in this embodiment includes: a gradual voltage drop curve, a slowly rising current harmonic component, and a persistently low power factor. The gradual voltage drop curve is defined as a voltage amplitude monotonically decreasing by more than 8% within 10-120 seconds at a rate less than 0.1% / s. The slowly rising current harmonic distortion rate refers to the total harmonic distortion rate increasing from 5% to over 12% within 10 minutes. The persistently low power factor refers to a power factor below 0.7 for 30 consecutive minutes without periodic fluctuations. This transforms the identification criteria for high-resistance faults in the fault mode library from fuzzy descriptions to precisely comparable dynamic electrical quantity patterns. When this structured template is matched with the optimized feature vector, it can effectively distinguish high-resistance faults from ordinary voltage fluctuations, equipment start-up / shutdown disturbances, or short-circuit anomalies, avoiding misjudgments and missed judgments due to feature ambiguity. This improves the identification accuracy and response reliability of low-abrupt, gradual high-resistance faults, ultimately achieving accurate location and early warning of hidden high-resistance faults in the distribution network.

[0049] The "short-circuit anomaly template" in this embodiment includes a combination of instantaneous features: voltage drop ≥30%, current spike ≥5 times the rated value, and phase change ≥15°. Voltage drop is determined by the voltage drop amplitude between sampling points; current spike is determined by comparing a single-point sample value with the average of the previous 5 points; and phase change is determined by the maximum absolute value of the first-order difference of the instantaneous phase difference exceeding 15°. All three conditions must be met simultaneously within 200ms, providing quantifiable and automatically executable criteria for the fault mode matching process. This allows for accurate identification of real short-circuit faults when comparing the optimized feature vector with the fault mode library, effectively filtering out similar electrical fluctuations caused by load switching, lightning interference, or equipment start-up and shutdown, avoiding mismatches and missed matches, improving the accuracy and timeliness of short-circuit anomaly mode identification, and ultimately achieving rapid and reliable early warning and location of sudden short-circuit faults in the distribution network.

[0050] In this implementation, the "matching degree" is calculated using the Dynamic Time Warping (DTW) algorithm. The DTW path constraint is the Sanchez constraint, the window width can be set to 20% of the template length, the distance metric is Euclidean distance, the matching degree is the normalized path cost, the high-resistance fault matching threshold is set to 0.85, and the short-circuit anomaly matching threshold is set to 0.90. The Dynamic Time Warping (DTW) algorithm is used to match the optimized feature vectors with high-resistance fault templates and short-circuit anomaly templates in the fault mode library. Sanguine constraints are used to limit the path offset range, and the window width is set to 20% of the template length to accommodate the gradual changes in high-resistance faults. Euclidean distance is used as the metric, and path costs are normalized to unify the dimensions. This allows for robust matching of high-resistance faults due to their slow evolution over time through elastic alignment. Short-circuit anomalies, due to their abrupt changes, are distinguished by a high threshold of 0.90 to enhance consistency at specific time points. The lower threshold of 0.85 for high-resistance faults tolerates their non-abrupt, low-amplitude signal morphology differences. This achieves differentiated and accurate identification of two fundamentally different fault modes without relying on manual parameter tuning. This effectively overcomes the misjudgment and missed judgment problems caused by the original unified matching mechanism's neglect of dynamic fault characteristics, improving the accuracy and timeliness of high-resistance fault and short-circuit anomaly identification in the distribution network.

[0051] In this embodiment, this step segments the continuously input electrical quantity time-series data through a real-time streaming data processing pipeline, dividing the uninterrupted electrical quantity data stream into multiple data blocks with time boundaries to facilitate the extraction of fault feature information segment by segment. Subsequently, using the preliminary fault feature set obtained from the segmented extraction as input, principal component analysis is used to perform mathematical transformation on it, preserving the main variation direction while reducing the feature dimension, thereby obtaining a simplified and optimized feature vector that retains the core fault information.

[0052] Simultaneously, in this step of the embodiment, the optimized feature vector is first used as input and compared with the high-resistance fault templates and short-circuit anomaly templates pre-classified and stored in the fault mode library to identify whether the current feature is consistent with a known fault mode. The fault mode library is constructed from a structured set of fault feature vectors, and its content only includes feature templates for the two known modes of high-resistance faults and short-circuit anomalies, excluding other types or undefined modes. When the matching degree between the optimized feature vector and any template is lower than a preset threshold, the system does not force classification, but explicitly marks it as an unknown mode and stores it in the dataset to be analyzed. This operation is based solely on the direct comparison result of the matching degree value and the threshold, without relying on any additional analysis or manual intervention. This process ensures that only inputs that conform to the features of a known mode are identified as known faults, while all features that fail to meet the matching criteria are isolated and saved, preserving the original data basis for subsequent manual or offline analysis.

[0053] Step S6: The Long Short-Term Memory (LSTM) network algorithm is used to perform deep analysis on the continuous electrical quantity time series data to calculate the occurrence probability of high-resistance faults and short-circuit anomalies; if the probability exceeds a preset threshold, it is marked as a high-risk fault type; wherein, the input of the deep analysis is the continuous electrical quantity time series data, and the calculation result of the occurrence probability is used to trigger the marking of high-risk fault types.

[0054] In this embodiment, "deep analysis" includes: First, continuously acquiring electrical quantity time-series data at a sampling frequency of 100Hz, with each 1.2 seconds (i.e., 120 sampling points) serving as an analysis window, and extracting the latest time-series segment as the LSTM input. This window length can completely cover the typical evolution cycle of a high-resistance fault, from a slight voltage drop to abnormal current accumulation, and finally to a short-circuit abrupt change. Then, the initial hidden state h0h0 and cell state c0c0 of the LSTM are not initialized randomly or with zero values, but instead directly reuse the final state h0 at the end of the previous analysis window. T-1 hT-1 and c T-1The cT-1 mechanism forms a cross-window "state memory chain." This mechanism allows the model to retain the operating context of the previous cycle when processing continuous data streams, thereby perceiving the long-term trend of "slowly accumulating" faults and avoiding state gaps and missed detections caused by window switching. Then, LSTM recursively processes the dimensionality-reduced features of 120 time steps through two layers of hidden units (128 dimensions per layer), extracting the nonlinear dependencies of electrical quantity time-series data in the time dimension layer by layer. Residual connections are introduced between layers to stabilize deep gradient propagation, ensuring the model maintains training stability during long-term dependency modeling. Finally, the network's final output layer uses the Sigmoid activation function to independently generate three risk probability values: high-resistance fault probability, short-circuit anomaly probability, and normal state probability. This design allows all three states to coexist (e.g., 0.85, 0.25, 0.10), realistically reflecting the complex operating conditions of "gradual change + sudden change" in the distribution network, avoiding probability crowding and misjudgments caused by Softmax forced normalization. Finally, the normalized occurrence probabilities of the three types of faults are output as direct inputs for risk assessment and section location. Through the above five-step process, LSTM achieves end-to-end deep analysis from raw electrical quantity time-series data to fault evolution probabilities. Its core lies in constraining the input window with physical duration, maintaining the context with state inheritance, stabilizing the model with residual structure, and decoupling the real scene with Sigmoid, thereby achieving high-precision and low-latency identification of latent faults in noisy and sparsely sampled remote branch environments.

[0055] In this embodiment, a Long Short-Term Memory (LSTM) network algorithm is employed for deep analysis of continuous electrical quantity time-series data. The aim is to capture and model the long-term dependencies of electrical quantities over time, thereby extracting implicit fault evolution characteristics from the time-series sequence. The input to the deep analysis is the continuous electrical quantity time-series data, which is directly used as the processing object of the algorithm without undergoing other feature transformations or dimensionality reduction, ensuring that the analysis is based on the original time-series dynamics. Based on this analysis process, the system calculates the probability of high-resistance faults and short-circuit anomalies. This probability is derived from the output layer of the LSM network through a specific function mapping, reflecting the potential likelihood of these two types of faults in the current time-series state. If the calculated probability exceeds a preset threshold, a high-risk fault type marking action is directly triggered. This marking behavior is determined solely by the comparison between the probability value and the threshold, without relying on other auxiliary judgment conditions, thereby achieving automated identification and response to high-resistance fault and short-circuit anomaly risks.

[0056] Step S7: Map the risk probability distribution of high-risk fault types into a three-dimensional view of time, space, and electrical characteristics. Match spatial location and time window using spatiotemporal correlation analysis method, and combine historical map data backtracking to locate segments where multiple abrupt changes and outliers cluster, thus determining the location of risk segments. The input to the three-dimensional view is the risk probability distribution, and the input to the spatiotemporal correlation analysis method is the three-dimensional view and historical map data. Abrupt changes and outliers come from the multidimensional map structure.

[0057] In this embodiment, the "3D view" maps the time dimension to the X-axis, the spatial dimension to the YZ plane, and the electrical features to a color heatmap. The X-axis timestamp is based on a UTC timestamp, the YZ plane coordinates use the WGS84 geographic coordinate system, and the electrical feature color mapping uses the L channel of the CIELAB color space to represent voltage amplitude, the A channel to represent current amplitude, and the B channel to represent power factor. The heatmap uses bilinear interpolation for spatial smoothing. By uniformly mapping the time dimension to an X-axis based on a UTC timestamp and the spatial dimension to a YZ plane using the WGS84 geographic coordinate system, and encoding the three key electrical features—voltage amplitude, current amplitude, and power factor—into the L, A, and B channels of the CIELAB color space to form a color heatmap, combined with bilinear interpolation for spatial smoothing, the 3D view driven by the risk probability distribution achieves a standardized and reproducible visual expression across the three dimensions of time reference, spatial location, and electrical representation. This combined mapping method ensures that different maintenance terminals and heterogeneous acquisition devices can be used simultaneously. When displaying the same risk section, the cross-regional monitoring system can accurately identify the distribution pattern and spatial clustering characteristics of electrical anomalies based on unified spatiotemporal coordinates and color semantics. This effectively eliminates the risk of positioning deviation and misjudgment caused by inconsistent coordinate systems, chaotic color coding, or missing interpolation methods. Especially in remote branch line environments with limited communication, it improves the spatial alignment accuracy of risk sections and the consistency of electrical status interpretation under multi-terminal collaborative operation, thereby achieving accurate spatial positioning and dynamic trend backtracking of high-resistance faults and short-circuit anomalies. Ultimately, it improves the reliability of distribution network fault early warning and the collaborative efficiency of operation and maintenance response.

[0058] The "spatiotemporal correlation analysis method" in this embodiment includes: performing DBSCAN clustering on abnormal clustered areas in the three-dimensional view, with a neighborhood radius of 500 meters and a minimum number of points of 3 in the clustering parameters. After clustering, the core node group with the highest spatial density is identified and matched with its corresponding time window interval to form a spatiotemporal correlation cluster. Each cluster contains a set of node IDs, time start and end stamps, average voltage drop rate and current change amplitude. By applying the DBSCAN clustering algorithm to the abnormal clustering areas in the 3D view, and limiting the clustering parameters to a neighborhood radius of 500 meters and a minimum number of points of 3, the core node group with the highest spatial density is accurately screened out, thereby effectively eliminating isolated noise points and low-density interference areas and improving the identification accuracy of abnormal clustering areas. Furthermore, by matching the time window interval corresponding to the core node group, a strong correlation is achieved between the spatial cluster and the time-series evolution characteristics of electrical quantities, forming a spatiotemporal correlation cluster containing node ID sets, time start and end stamps, average voltage drop rate, and current mutation amplitude. This enables the originally scattered mutation values ​​and outliers to form a fault cluster representation with statistical significance and physical consistency in the spatial and temporal dimensions, thereby overcoming the problem of inaccurate risk section location caused by the fuzzy spatial distribution and loose temporal correlation in traditional methods. This achieves accurate identification and closed-loop solidification of high-density, multi-dimensional collaborative anomalies, improving the location accuracy and early warning response timeliness of fault risk sections in the distribution network.

[0059] In this embodiment, the "historical map data" backtracking interface calls multi-dimensional map structure data within ±30 minutes before and after the time window, extracts nodes that overlap with the current risk segment in spatial location and their associated anomaly marker set. The anomaly marker set includes mutation value timestamp, anomaly type, marker confidence and source algorithm identifier. The backtracking data is stored in a graph database format (e.g., using Neo4j architecture, node types include electrical nodes, anomaly markers, failure modes and geographic grids, edge types include electrical connections, spatiotemporal proximity, anomaly attribution and pattern matching, and each node stores a unique UUID identifier and timestamp index). Nodes and edges are retrieved using the Apache TinkerPop Gremlin query language. By accessing multidimensional graph structure data within ±30 minutes before and after a time window through the historical graph data backtracking interface, and accurately extracting corresponding nodes and their associated anomaly marker sets based on spatial overlap relationships, this set includes mutation value timestamps, anomaly types, marker confidence levels, and source algorithm identifiers. Combined with node and edge structures stored in graph database format, the Apache TinkerPop Gremlin query language is used to achieve efficient graph traversal and retrieval of spatiotemporally correlated anomaly information. This allows for the rapid location of anomaly clusters highly correlated with the current risk segment in massive heterogeneous data, effectively overcoming the problems of fuzzy positioning and response delays caused by data redundancy, fragmented marker information, and lack of indexing mechanisms in traditional backtracking methods. It achieves high-precision source tracing of risk segments and structured integration of multidimensional anomaly features, improving the accuracy of fault location and the timeliness of early warning response.

[0060] In this embodiment, "determining the location of risk zones" includes: based on the aforementioned constructed 3D view (with UTC time as the X-axis, WGS84 geographic coordinates as the YZ plane, and CIELAB color channels mapping voltage / current / power factor amplitude), identifying anomalous areas with significant color clustering in the visualized color heatmap, extracting their corresponding spatial node sets and time windows to form high-density spatiotemporal anomaly clusters; then, using the node ID and time range of this anomaly cluster as an index, retrieving the associated anomaly markers within ±30 minutes from the historical map database, obtaining information including mutation value timestamps, anomaly types, marker confidence levels, and sources. The algorithm identifies a structured set of outliers. A secondary spatiotemporal clustering is performed on this set, using a 300-meter spatial radius and a 5-minute time window as constraints. Dense sections consisting of ≥5 outliers with high confidence (≥0.8) are selected as high-risk core areas. Furthermore, based on the node topology connections within this section in the multidimensional graph structure, it is verified whether the spatial distribution of outliers exhibits a continuous and directional evolution trend along the feeder path. If satisfied, the physical feeder to which this section belongs and its corresponding geofence are identified, outputting the final risk section location information for the operation and maintenance system to achieve precise positioning and rapid response.

[0061] In this step of the embodiment, the risk probability distribution of high-risk fault types is used as input and mapped to a three-dimensional view of time, space, and electrical characteristics, realizing an intuitive representation of fault risk in multi-dimensional space. Subsequently, using this three-dimensional view and historical map data as input, a spatiotemporal correlation analysis method is used to cross-match the distribution patterns of high-risk faults in specific geographical areas and specific time windows, thereby establishing the correlation between spatial location and temporal evolution. On this basis, combined with the marked mutation values ​​and outliers in the multi-dimensional map structure, the concentrated clustering segments of these abnormal data in the spatiotemporal dimension are located. Finally, through the spatial distribution and temporal continuity analysis of the outliers, the specific risk segment locations with significant abnormal clustering characteristics are determined.

[0062] The technical solution of this embodiment collects and stores time-series electrical quantity data (voltage, current, power factor, etc.) containing timestamps and geographical coordinates in a hierarchical manner. It then constructs a dependency matrix between nodes based on the electrical network topology, calculates the dynamic characteristics of voltage and current in real time, and accurately identifies abrupt changes and outliers using an anomaly marking mechanism, forming a structured multidimensional map. This effectively suppresses misjudgments caused by high data noise and complex topological relationships. Furthermore, it extracts temporal, spatial, and electrical features, dynamically adjusting the feature weights of each dimension based on high-resistance faults or short-circuit anomaly modes to achieve adaptive fusion and expression optimization of multidimensional heterogeneous features, improving feature discriminability. Finally, it classifies and stores the weighted feature vectors through cloud-based distributed parallel processing, constructing a structured fault mode library. This method combines real-time streaming processing and principal component analysis to efficiently reduce the dimensionality of continuous electrical quantity data, obtaining optimized feature vectors and improving matching efficiency and accuracy. The optimized feature vectors are then matched with fault templates to automatically identify unknown patterns and store them in the analysis set, enabling continuous learning and updating of fault modes. Furthermore, a long short-term memory network is used for deep modeling of time-series data, accurately calculating the probability of high-resistance faults and short-circuit anomalies, achieving early warning of potential high-risk faults. Finally, the risk probability distribution is mapped into a three-dimensional view integrating time, space, and electrical features. Linking abrupt change points in the multidimensional map with historical map data, spatiotemporal correlation analysis is used to accurately locate multiple anomaly clusters, achieving visualization, localization, and retrospective confirmation of fault locations. These steps work synergistically to systematically solve the problems of inaccurate fault location and delayed early warning caused by strong electrical quantity noise interference, complex topology, heterogeneous multidimensional features, and low accuracy of dynamic pattern recognition in distribution network fault monitoring. This achieves the effects of improving fault identification accuracy, shortening early warning response time, and realizing precise location and intelligent source tracing of risk sections.

[0063] This embodiment also provides a distribution network fault monitoring system based on electrical quantity maps, which can implement the aforementioned distribution network fault monitoring method based on electrical quantity maps. The system includes a spatiotemporal data acquisition and hierarchical storage module, a multi-dimensional map construction module, a cross-dimensional feature fusion and weighting module, a cloud-based distributed feature processing module, a real-time streaming feature matching module, a deep time-series probability prediction module, and a three-dimensional visualization and risk location module.

[0064] The system comprises several modules: a spatiotemporal data acquisition and hierarchical storage module, and a cloud-based distributed feature processing module. The first module acquires time-series electrical quantity data containing timestamps and geographic coordinates, and stores it hierarchically. The second module collects time-series electrical quantity data including voltage, current, and power factor. The third module generates a multidimensional graph structure containing node associations and anomaly markers, based on the electrical network topology and uses dynamic feature extraction and anomaly labeling. The fourth module extracts temporal, spatial, and electrical features from the multidimensional graph structure and dynamically adjusts the feature weights based on high-resistance fault or short-circuit anomaly modes using an adaptive weight allocation mechanism, generating a weighted feature vector set. The fifth module performs parallel processing and aggregation of the weighted feature vector set using a cloud-based distributed computing framework, forming a structured fault feature vector set. This set is then categorized and stored according to high-resistance fault and short-circuit anomaly modes, constructing a fault mode library. The real-time streaming feature matching module segments continuous electrical quantity time-series data into time windows using a real-time streaming processing pipeline. After dimensionality reduction using principal component analysis, optimized feature vectors are obtained and matched against preset templates in the fault mode library. The deep time-series probability prediction module performs deep analysis on the continuous electrical quantity time-series data using a long short-term memory network, calculating the probability of high-resistance faults and short-circuit anomalies. When the probability exceeds a preset risk threshold, it is identified as a high-risk fault type. The 3D visualization and risk location module maps the risk probability distribution of high-risk fault types into a 3D view of time, space, and electrical characteristics. It combines spatiotemporal correlation analysis to match the spatial location and time window, and backtracks historical data to locate high-risk segments formed by clusters of multiple abrupt changes and outliers, outputting the final risk segment location information.

[0065] In this embodiment, the system achieves a fully intelligent closed-loop process for power distribution network fault monitoring, from "data acquisition" to "precise positioning," through a modular collaborative architecture. For the first time, the system organically integrates six major functions: spatiotemporal hierarchical storage, graph topology modeling, adaptive feature weighting, cloud-based parallel processing, real-time pattern matching, and 3D spatiotemporal positioning. This overcomes the limitations of traditional systems that prioritize alarms over positioning and single-point faults over correlations. Even in remote branch line scenarios with weak communication, it can still stably output high-precision, low-latency fault risk section information.

Claims

1. A method for monitoring distribution network faults based on electrical quantity spectra, characterized in that, Includes the following steps: Step S1: Collect electrical quantity time-series data containing timestamps and geographic coordinates, and store it in layers; wherein, the electrical quantity time-series data includes voltage, current, and power factor; Step S2: Construct a node dependency matrix based on the electrical network topology, and combine dynamic feature extraction and anomaly marking to generate a multi-dimensional graph structure containing node associations and anomaly markings; Step S3: Extract time, space and electrical features from the multidimensional graph structure, and dynamically adjust the feature weights of each dimension according to the high-resistance fault mode or short-circuit anomaly mode through an adaptive weight allocation mechanism to generate a weighted feature vector set. Step S4: Use a cloud-based distributed computing framework to process and summarize the weighted feature vector set in parallel, forming a structured fault feature vector set, and store it according to high-resistance fault and short-circuit anomaly modes to build a fault mode library. Step S5: The continuous electrical quantity time series data is segmented into time windows through a real-time streaming pipeline, and the optimized feature vector is obtained after dimensionality reduction by principal component analysis and matched with the preset template in the fault mode library. Step S6: Use a long short-term memory network to perform in-depth analysis on the continuous electrical quantity time series data to calculate the probability of high-resistance faults and short-circuit anomalies; when the probability exceeds a preset risk threshold, it is determined to be a high-risk fault type. Step S7: Map the risk probability distribution of high-risk fault types into a three-dimensional visualization view of time, space and electrical characteristics. Combine the spatiotemporal correlation analysis method to match its spatial location and time window, and backtrack the historical map data to locate the high-risk segment formed by the aggregation of multiple mutation values ​​and outliers, and output the final risk segment location information.

2. The method for monitoring distribution network faults based on electrical quantity spectra according to claim 1, characterized in that, The hierarchical storage in step S1 includes: dividing electrical quantity time series data into continuous time slices according to timestamps, dividing each time slice into several geographic grid units according to geographic coordinates, forming a spatiotemporal dual-dimensional data block, storing each data block independently and establishing an index mapping relationship, the index including timestamp interval, geographic grid code and data offset field.

3. The method for monitoring distribution network faults based on electrical quantity spectra according to claim 1, characterized in that, The dependency matrix construction in step S2 includes: quantifying the influence strength between nodes using a weighted graph model based on the electrical impedance coupling degree between nodes, the weight of energy transmission direction, and the topological adjacency relationship; And / or, the dynamic feature extraction in step S2 includes: real-time calculation of the voltage change rate, current harmonic distortion rate, power factor slope and time-varying characteristics of voltage and current phase difference of each node; And / or, the outlier marking in step S2 includes: using a sliding window statistical method to identify outliers in voltage or current data that deviate from the 3σ threshold, and combining the CUSUM algorithm to detect asymptotic abrupt changes in the voltage or current sequence, and the EWMA algorithm to detect small persistent deviations; when the CUSUM statistic exceeds the preset control limit or the EWMA value exceeds the ±2.5σ range for 5 consecutive sampling periods, it is marked as a transient abrupt change value.

4. The method for monitoring distribution network faults based on electrical quantity spectra according to claim 1, characterized in that, The adaptive weight allocation mechanism in step S3 includes: when a high-impedance fault mode is identified, the weight of time features is increased to 0.5~0.7, the weight of electrical features is decreased to 0.2~0.3, and the weight of spatial features is constant at 0.1~0.2; when a short-circuit anomaly mode is identified, the weight of electrical features is increased to 0.6~0.8, the weight of time features is decreased to 0.1~0.2, and the weight of spatial features is constant at 0.1~0.2; wherein, the time features include the rate of change of voltage, the slope of the power factor and the rate of change of phase difference, the electrical features include the current amplitude, the voltage amplitude, the harmonic distortion rate and the absolute value of the power factor, and the spatial features include the node geographic density and the topological centrality.

5. The method for monitoring distribution network faults based on electrical quantity spectra according to claim 4, characterized in that, In step S3, the adaptive weight allocation mechanism classifies the initial feature vectors using a shallow neural network.

6. The method for monitoring distribution network faults based on electrical quantity spectra according to claim 1, characterized in that, In step S4, the cloud-based distributed computing framework adopts the Spark architecture. The weighted feature vector set is divided into multiple sub-task units according to the feeder topology. Each sub-task unit corresponds to all nodes under a key node of the distribution network. Feature vector standardization, missing value imputation, and anomaly label verification are performed in parallel within the sub-task. When imputing missing values, a corresponding hierarchical imputation mechanism is used to complete the missing electrical features. After each sub-task completes local processing, the output is merged through the Reduce operation to form a structured and highly consistent fault feature vector set. And / or, the classification storage in step S4 includes: storing the fault feature vector set into the high-resistivity fault feature library and the short-circuit anomaly feature library according to the fault mode type, and attaching fault duration, mutation amplitude and spatial clustering density labels to each type of feature vector; wherein, the fault duration is calculated from the time difference between the first anomaly mark and the last anomaly mark, the mutation amplitude is the maximum absolute difference between the anomaly point and the previous steady state value, and the spatial clustering density is the ratio of the number of anomaly mark nodes in a unit geographic grid to the total number of grids.

7. The method for monitoring distribution network faults based on electrical quantity spectra according to claim 1, characterized in that, In step S5, the time window length of the real-time streaming pipeline is 10 to 60 seconds, and the specific window length is adjusted by the load prediction model. And / or, the principal component analysis method in step S5 includes: performing Z-score standardization on the original feature vectors of all key nodes in the distribution network within the time window, which consist of six dimensions: voltage amplitude, current amplitude, power factor, voltage change rate, current harmonic distortion rate, and phase difference change rate; constructing a covariance matrix in real time based on the standardized feature vector set, and performing eigenvalue decomposition on the covariance matrix to calculate each principal component and its corresponding variance contribution rate; sorting the cumulative variance contribution rates from high to low, and selecting the top 3 to 5 principal components with a cumulative contribution rate of not less than 95% as the optimized feature vectors; And / or, the preset templates in step S5 include: a high-resistance fault template and a short-circuit anomaly template; the high-resistance fault template includes a combination of a gradually decreasing voltage drop curve, a slowly rising current harmonic component, and a persistently low power factor; wherein, the gradually decreasing voltage drop curve is defined as a voltage amplitude monotonically decreasing by more than 8% within 10~120 seconds and a decrease rate of less than 0.1% / s, the slowly rising current harmonic distortion rate refers to the total harmonic distortion rate increasing from 5% to more than 12% within 10 minutes, and the persistently low power factor refers to a power factor below 0.7 for 30 consecutive minutes without periodic fluctuations; the short-circuit anomaly template includes a combination of instantaneous features including a voltage drop ≥30%, a current spike ≥5 times the rated value, and a phase change ≥15°; wherein, the voltage drop is determined by the voltage drop amplitude between sampling points, the current spike is determined by comparing the single-point sampling value with the average value of the previous 5 points, and the phase change is determined by the maximum value of the absolute value of the first-order difference of the instantaneous phase difference exceeding 15°, and all three must be satisfied simultaneously within 200ms; And / or, in step S5, the matching is characterized by the matching degree; wherein the matching degree is calculated using the Dynamic Time Warping (DTW) algorithm.

8. The method for monitoring distribution network faults based on electrical quantity spectra according to claim 1, characterized in that, The deep analysis in step S6 includes: continuously collecting electrical quantity time-series data at a preset sampling frequency; using a preset window length as an analysis window; extracting the latest time-series segment as the input of the LSTM; recursively processing through two layers of hidden units; extracting the nonlinear dependencies of electrical quantity time-series data in the time dimension layer by layer; and using the Sigmoid activation function in the final output layer of the network to independently generate risk probability values ​​for high-resistance fault probability, short-circuit anomaly probability, and normal state probability, and normalizing them. The initial hidden state of the LSTM directly reuses the final state at the end of the previous analysis window.

9. The method for monitoring distribution network faults based on electrical quantity spectra according to claim 1, characterized in that, In step S7, the three-dimensional view is mapped to the X-axis in the time dimension, the YZ plane in the spatial dimension, and the color heat map in the electrical features. And / or, the spatiotemporal correlation analysis method in step S7 includes: performing DBSCAN clustering on the abnormal clustered areas in the three-dimensional view, identifying the core node group with the highest spatial density after clustering, and matching its corresponding time window interval to form a spatiotemporal correlation cluster. Each cluster contains a set of node IDs, time start and end stamps, average voltage drop rate and current change amplitude. And / or, in step S7, the historical map data backtracking interface calls the multi-dimensional map structure data within ±30 minutes before and after the time window, extracts the nodes that overlap with the spatial location of the current risk segment and their associated anomaly marker set, the anomaly marker set including the mutation value timestamp, anomaly type, marker confidence and source algorithm identifier; And / or, determining the location of the risk segment in step S7 includes: extracting the corresponding spatial node set and time window based on the abnormal areas with significant color clustering in the 3D view to form a high-density spatiotemporal anomaly cluster; using the node ID and time window of the anomaly cluster as an index, retrieving the anomaly markers associated with it within ±30 minutes from the historical map database to obtain a structured set of anomaly points containing the mutation value timestamp, marker confidence, and source algorithm identifier; performing secondary spatiotemporal clustering on the anomaly point set to filter out dense segments composed of anomaly points; verifying whether the anomaly points in the dense segment exhibit spatial continuity and temporal directionality along the feeder path based on the electrical topology connection relationship in the multidimensional map; if satisfied, locking the physical feeder and geofence to which the dense segment belongs, and outputting the final risk segment location information.

10. A power distribution network fault monitoring system based on electrical quantity graphs, characterized in that, include: The system includes a spatiotemporal data acquisition and hierarchical storage module for acquiring and storing time-series electrical quantity data containing timestamps and geographic coordinates in hierarchical layers; the electrical quantity time-series data includes voltage, current, and power factor; a multidimensional graph construction module for generating a dependency matrix between nodes based on the electrical network topology, and generating a multidimensional graph structure containing node associations and anomaly markers by combining dynamic feature extraction and anomaly marking; a cross-dimensional feature fusion and weighting module for extracting temporal, spatial, and electrical features from the multidimensional graph structure, dynamically adjusting the feature weights of each dimension according to high-resistance fault mode or short-circuit anomaly mode through an adaptive weight allocation mechanism, and generating a weighted feature vector set; and a cloud-based distributed feature processing module for parallel processing and summarizing the weighted feature vector set using a cloud-based distributed computing framework to form a structured fault feature vector set, and categorizing it according to high-resistance fault and short-circuit anomaly modes. The system employs a classification and storage method to construct a fault mode library. A real-time streaming feature matching module segments continuous electrical quantity time-series data into time windows using a real-time streaming processing pipeline. Principal component analysis is used to reduce the dimensionality of the data, resulting in optimized feature vectors, which are then matched with preset templates in the fault mode library. A deep time-series probability prediction module uses a long short-term memory network to perform deep analysis on the continuous electrical quantity time-series data, calculating the probability of high-resistance faults and short-circuit anomalies. When the probability exceeds a preset risk threshold, it is identified as a high-risk fault type. A 3D visualization and risk location module maps the risk probability distribution of high-risk fault types into a 3D view of time, space, and electrical characteristics. It combines spatiotemporal correlation analysis to match the spatial location and time window, and backtracks historical data to locate high-risk segments formed by clusters of multiple abrupt changes and outliers, outputting the final risk segment location information.

Images