A method and system for automatic monitoring and early warning of power distribution based on voltage measurement

Abstract

This invention proposes a method and system for automatic monitoring and early warning of power distribution based on voltage measurement, belonging to the field of power distribution monitoring. Wavelet decomposition is performed on the voltage signals of each node in the spatiotemporal matrix of the power distribution network topology to obtain the frequency band components at each level and calculate their energy entropy. Each frequency band component is treated as an independent entity, and the energy entropy of the lowest level is integrated from bottom to top to form individual characteristics. Based on mutual information analysis, the correlation between individuals is analyzed, and individual pairs are selected as independent states of nodes. All states are integrated to obtain node characteristics. Based on the node characteristics, the correlation between any two nodes is calculated using an improved dynamic time warping distance, which serves as the voltage fluctuation correlation coefficient. A dynamic fault propagation graph is constructed using the voltage fluctuation correlation coefficient as edges. A pre-trained multi-head attention neural network is used to analyze the node characteristics to obtain the node risk coefficient, and the total risk coefficient of the connected path is calculated for early warning, thus realizing automatic monitoring and early warning of the power distribution network.

Description

Technical Field

[0001] This invention relates to the field of power distribution monitoring technology, and in particular to a power distribution automatic monitoring and early warning method and system based on voltage measurement. Background Technology

[0002] As the distribution network structure becomes increasingly complex and the scale of distributed energy access continues to expand, the safety risks during its operation are constantly rising. Problems such as abnormal voltage fluctuations, electrical faults, and equipment aging occur frequently, placing higher demands on the stability, power supply reliability, and intelligent management level of the power system.

[0003] Traditional power distribution network monitoring relies heavily on manual inspections, periodic maintenance, or single-point monitoring. This approach is insufficient to reflect the real-time and comprehensive operation status of the power grid under complex interconnections of multiple nodes. In particular, in large-scale wide-area power distribution network scenarios, the information collection and processing capabilities are significantly limited, failing to meet the needs of efficient monitoring.

[0004] In actual operation, the electrical connections between distribution network nodes exhibit significant spatiotemporal dynamic characteristics. Affected by factors such as load fluctuations and wiring structure adjustments, node voltage signals show nonlinear and multi-scale variation characteristics. Existing monitoring systems struggle to efficiently and synchronously acquire and process these dynamic changes, easily leading to delayed or missed fault identification. Furthermore, most monitoring methods focus on signals from single nodes or local areas, neglecting the topological connections and dynamic transmission relationships between nodes, making it impossible to effectively trace fault propagation paths and resulting in difficulties in timely prediction and intervention of anomaly propagation risks.

[0005] In the feature extraction and anomaly detection stages, mainstream solutions rely on conventional signal processing techniques such as Fourier transform and short-time energy analysis, which have weak anti-noise interference capabilities and are difficult to deal with multi-frequency band anomaly information. The Pearson correlation coefficient and Euclidean distance methods commonly used for node correlation calculation cannot reveal the deep dynamic connections between nodes at different frequencies and time periods, resulting in low sensitivity and accuracy of anomaly signal identification.

[0006] In addition, the operation of the power distribution network is affected by multiple factors such as the external environment and equipment parameters. The electrical coupling relationship between nodes is dynamically adjusted. Traditional static or semi-static monitoring models cannot reflect the time evolution of the power grid structure and signal characteristics, making it difficult to accurately depict fault propagation and risk areas, which restricts the development of intelligent power distribution operation and maintenance. Summary of the Invention

[0007] To address the aforementioned issues, this invention proposes a distribution network automatic monitoring and early warning method and system based on voltage measurement. The method constructs a spatiotemporal matrix by collecting voltage signals from associated nodes based on the distribution network topology. Multi-state superposition node features are obtained through wavelet decomposition, energy entropy calculation, and mutual information filtering. An improved dynamic time warping distance is used to calculate the correlation coefficient of voltage fluctuations between nodes to construct a dynamic fault propagation graph. A pre-trained multi-head attention neural network is then used to analyze node features and obtain risk coefficients. Combined with the propagation graph, the total risk coefficient of connected paths is calculated, and early warnings are issued for paths exceeding preset values, thus achieving automatic monitoring and early warning of the distribution network.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] In a first aspect, the present invention provides a power distribution automatic monitoring and early warning method based on voltage measurement, comprising:

[0010] Based on the distribution network topology, voltage signals of associated nodes are collected to construct a spatiotemporal matrix;

[0011] Wavelet decomposition is performed on the voltage signals of each node in the spatiotemporal matrix to obtain the frequency band components of each layer and calculate their energy entropy; each frequency band component is regarded as an independent individual, and the energy entropy of the lowest layer is integrated from bottom to top to form individual features; based on mutual information analysis, the correlation between individuals is analyzed, and individual pairs are selected as independent states of nodes according to preset rules, and all states are integrated to obtain the node features of multi-state superposition.

[0012] Based on the node characteristics, the correlation between any two nodes is calculated using the improved dynamic time warping distance, which serves as the voltage fluctuation correlation coefficient.

[0013] The edge relationships are determined based on the topological associations between nodes, and a dynamic fault propagation graph is constructed using the voltage fluctuation correlation coefficient as the edge. The node risk coefficient is obtained by analyzing the node characteristics using a pre-trained multi-head attention neural network, and the total risk coefficient of the connected path is calculated by combining the dynamic fault propagation graph. Paths with a total risk coefficient exceeding a preset value are given an early warning.

[0014] Secondly, the present invention provides a power distribution automatic monitoring and early warning system based on voltage measurement, comprising:

[0015] The matrix construction module is configured to collect voltage signals from associated nodes based on the distribution network topology and construct a spatiotemporal matrix.

[0016] The feature extraction module is configured to perform wavelet decomposition on the voltage signals of each node in the spatiotemporal matrix to obtain the frequency band components of each layer and calculate their energy entropy; treat each frequency band component as an independent individual, integrate the energy entropy of the lowest layer from bottom to top to form individual features; analyze the correlation between individuals based on mutual information, select individual pairs as independent states of nodes according to preset rules, and integrate all states to obtain node features with multiple states superimposed.

[0017] The correlation calculation module is configured to calculate the correlation between any two nodes based on the node characteristics and using the improved dynamic time warping distance as the voltage fluctuation correlation coefficient.

[0018] The risk warning module is configured to determine the edge relationship based on the topological association between nodes, construct a dynamic fault propagation graph with the voltage fluctuation correlation coefficient as the edge, analyze node features using a pre-trained multi-head attention neural network to obtain node risk coefficients, calculate the total risk coefficient of connected paths in combination with the dynamic fault propagation graph, and issue a warning for paths with a total risk coefficient exceeding a preset value.

[0019] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the power distribution automatic monitoring and early warning method based on voltage measurement described in the first aspect.

[0020] Fourthly, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the power distribution automatic monitoring and early warning method based on voltage measurement described in the first aspect.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0022] This invention collects voltage signals based on the distribution network topology and constructs a spatiotemporal matrix. Combining wavelet decomposition and energy entropy calculation, it can completely preserve the multi-scale frequency domain characteristics of voltage signals, avoiding the information loss of single-node analysis in traditional methods. By screening individual pairs of integrated multi-state node features through mutual information, redundant information is eliminated while enhancing feature representativeness, providing a high-quality data foundation for subsequent analysis. The improved dynamic time warping distance can accurately characterize the correlation of voltage fluctuations between nodes, and the constructed dynamic fault propagation map can reflect the fault association path in real time. By using a multi-head attention neural network to calculate the node risk coefficient and the total path risk, it can achieve accurate identification of anomalies and risk warning, effectively improving the timeliness of distribution network monitoring, the accuracy of fault location, and the reliability of power supply, and adapting to the intelligent operation and maintenance needs of complex distribution networks.

[0023] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0024] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute a limitation thereof.

[0025] Figure 1 The main flowchart of a power distribution automatic monitoring and early warning method based on voltage measurement provided in an embodiment of the present invention is shown below. Detailed Implementation

[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0027] Example 1

[0028] like Figure 1 As shown in the figure, this embodiment discloses a power distribution automatic monitoring and early warning method based on voltage measurement, including the following steps:

[0029] S1: Collect voltage signals from associated nodes based on the distribution network topology and construct a spatiotemporal matrix;

[0030] S2: Perform wavelet decomposition on the voltage signals of each node in the spatiotemporal matrix to obtain the frequency band components of each layer and calculate their energy entropy; treat each frequency band component as an independent individual, integrate the energy entropy of the lowest layer from bottom to top to form individual features; analyze the correlation between individuals based on mutual information, select individual pairs as independent states of nodes according to preset rules, and integrate all states to obtain node features with multiple states superimposed.

[0031] S3: Based on the node characteristics, the correlation between any two nodes is calculated using the improved dynamic time warping distance, which serves as the voltage fluctuation correlation coefficient.

[0032] S4: Determine the edge relationships based on the topological associations between nodes, and construct a dynamic fault propagation graph using the voltage fluctuation correlation coefficient as the edge; use a pre-trained multi-head attention neural network to analyze node characteristics to obtain node risk coefficients, and combine the dynamic fault propagation graph to calculate the total risk coefficient of connected paths, and issue early warnings for paths with total risk coefficients exceeding preset values.

[0033] Next, combined Figure 1 This embodiment provides a detailed description of a power distribution automatic monitoring and early warning method based on voltage measurement.

[0034] In S1, the power distribution network typically exhibits a complex topology, interconnected by various transmission and distribution lines to form a vast and orderly power transmission network. Within this network, nodes, as key elements, represent power plants, substations, load centers, and network connection points; they are abstract convergence points where electrical energy is input, output, and converted.

[0035] There are direct or indirect relationships between nodes. Among them, the voltage signal changes of directly connected nodes can intuitively reflect the power transmission status and potential faults on the line; while the voltage signal changes of indirectly connected nodes, such as nodes connected through multiple levels of lines, contain key information such as the large-scale power flow fluctuations of the entire distribution network, the propagation path of potential faults, and the coupling effects between different areas.

[0036] In-depth analysis of these voltage signals helps to accurately identify abnormal states in the distribution network and predict potential faults in advance.

[0037] Therefore, in this embodiment, nodes in the distribution network that are interconnected by lines and have direct or indirect relationships are considered as topology-related nodes. By setting a high-frequency sampling rate (not less than 10kHz), the voltage signals of N topology-related nodes in the distribution network are captured, providing comprehensive data support for subsequent analysis.

[0038] By setting a high-frequency sampling rate, it is possible to effectively capture fine-grained changes in voltage signals at the microsecond level, thus preserving key information for subsequent processing.

[0039] Furthermore, for the captured voltage signal, a spatiotemporal matrix is ​​constructed according to the method of row corresponding to time sequence and column corresponding to node position.

[0040] In this embodiment, a spatiotemporal matrix is ​​constructed to regularize discrete voltage signals based on time and location information, allowing the signals to simultaneously carry temporal evolution and node location information, thus avoiding data confusion. For subsequent steps, this matrix can directly support the layered processing of time-series signals of each node by wavelet packet decomposition, and also provides a structured data foundation for calculating correlation coefficients and analyzing the voltage fluctuation correlation of different nodes within the same time window. This enables the subsequent construction of dynamic fault propagation maps to accurately combine the spatiotemporal correlation of nodes, improving the efficiency and accuracy of fault analysis and early warning.

[0041] In S2, the correlation relationship between different nodes is constructed by using the joint analysis of wavelet packet decomposition and the spatiotemporal matrix.

[0042] Using the db4 wavelet function, a 6-level wavelet packet decomposition is performed on the voltage signals of each node in the spatiotemporal matrix. This decomposes the voltage signals of each node from the time domain to the frequency domain, and then further decomposes them level by level to obtain 2 6 =64 non-overlapping frequency bands; calculate the energy of each frequency band separately, and calculate the energy entropy of each frequency band after normalizing the energy of each frequency band.

[0043] In this embodiment, by performing 6-level db4 wavelet packet decomposition on the voltage signal of each node, the original signal can be decomposed into sub-signals of different scales and frequency bands. This meticulously captures various features in the signal, such as low-frequency trends, harmonic components, and transient anomalies, effectively improving the ability to distinguish between minute fluctuations and complex disturbances in the signal. Then, by normalizing the energy of each frequency band and calculating the energy entropy, the distribution and uncertainty of energy in each frequency band can be quantified, revealing the local and global dynamic change characteristics of the node signal. This not only enhances the sensitivity and accuracy of anomaly detection but also provides a rich and structured feature foundation for subsequent multi-node correlation analysis and fault propagation path modeling.

[0044] Furthermore, a sliding window is preset, for example, set to a sliding time window of not less than 200ms. Based on the sliding time window, the correlation of each frequency band of a single node in time sequence between different decomposition layers is analyzed, and then the node features of each node in multiple states superimposed within the sliding window are generated.

[0045] The construction process for node features with multiple states superimposed includes:

[0046] Treating each frequency band in each layer as an individual, the j-th individual in the i-th layer for any node n is represented as: .

[0047] It should be understood that this embodiment employs a 6-level wavelet decomposition, where the level number represents the depth of the recursive decomposition. After the 6-level decomposition, a total of 64 frequency bands are finally obtained. In order to fully explore the multi-scale features of the voltage signal at different decomposition levels and improve the ability to characterize abnormal node states, this embodiment not only uses the 64 frequency bands obtained from the final 6-level decomposition but also introduces the frequency bands of each intermediate layer (layer 1 to layer 5). By integrating the frequency domain information at different resolutions contained in each layer of frequency bands, a more discriminative multi-state superimposed node feature is constructed to adapt to the analysis needs of complex and ever-changing operating conditions and fault scenarios in the distribution network.

[0048] The energy entropy of an individual (frequency band) is used as an individual attribute. All the individual attributes of the lower-level individuals contained in the current layer are integrated to form a sequence containing the energy entropy of each lowest-level frequency band up to the energy entropy of the current individual, which is used as the individual feature of the current individual.

[0049] Considering that a single entity can only reflect its own attributes, while an entity can accurately characterize the correlation and interaction between two features, this aligns with the characteristic that distribution network faults are often caused by multi-frequency band effects. Furthermore, it can transcend the inter-level boundaries of wavelet packet decomposition, uncovering the intrinsic connections between frequency bands at different scales and comprehensively showcasing the multi-scale characteristics of voltage signals. Further, based on individual features, mutual information is used to represent the correlation between two entities, and then individuals with high information content are selected and integrated into node features based on this correlation.

[0050] Specifically, the mutual information between any two individuals is analyzed, and each pair of individuals is traversed to filter individual pairs. The filtering rule is as follows: Let there be two pairs of individuals, individuals a and individuals b as individual pair 1, and individuals c and individuals d as individual pair 2;

[0051] Rule 1: When individuals a and b can respectively contain the frequency bands of individuals c and d (a contains c, b contains d; or a contains d, b contains c), or individuals c and d can respectively contain the frequency bands of individuals a and b (the opposite of the above containment relationship), then the pair of individuals with higher mutual information is selected as the result of this screening.

[0052] Rule 1 simplifies data relationships, effectively reducing feature redundancy and improving the independence and representativeness of node feature representation. This mechanism filters out pairs of individuals with greater information content and stronger correlation in the multi-scale decomposition structure, eliminates low-contribution features with inclusion relationships, and reduces the interference of duplicate and invalid information on the overall analysis results. This provides a more refined and efficient data foundation for subsequent node state integration, anomaly identification, and correlation analysis.

[0053] Rule 2: When the inclusion relationship in Rule 1 is not satisfied, individual pair 2 and individual pair 1 are treated as two individual pairs respectively, and the current individual pair is retained.

[0054] Rule 3: If there is no complete containment relationship between the two individuals in each pair, the current pair is retained.

[0055] After the filtering is completed, each individual pair will be obtained as a state of a node, and each state is independent of the others. By integrating all the states of the nodes, the filtered energy entropy is aggregated into a time series, thus obtaining the node features.

[0056] In this embodiment, by treating the wavelet frequency bands in each layer as independent entities and integrating the energy entropy of the lowest-level frequency band from bottom to top, all characteristic information of each frequency band of the signal at different decomposition levels can be systematically preserved, avoiding feature loss. Introducing mutual information as a metric for measuring the correlation between entities helps to discover deep connections hidden in the multi-scale decomposition structure.

[0057] Using only the last layer's frequency band would result in the loss of features from intermediate layers (such as layers 1-5) at different decomposition depths. Directly superimposing all layer frequency bands indiscriminately would lead to excessive feature quantity and redundancy, drastically increasing computational complexity and introducing noise interference. However, by using hierarchical analysis and mutual information filtering, we can preserve the multi-scale features of each layer's frequency bands while effectively filtering out highly informative, representative, and independent pairs of individuals, avoiding information redundancy and feature conflicts. This reduces computational complexity while ensuring the integrity of multi-scale features, making node features more concise and efficient, and better suited to subsequent analysis needs.

[0058] This process not only simplifies the coupling relationship between node features, but also improves the distinguishability of features in multiple states, laying a solid multidimensional feature foundation for subsequent correlation analysis between nodes, anomaly detection, and dynamic propagation graph construction.

[0059] In S3, based on the node characteristics obtained in S2, the correlation between different nodes is calculated using the improved dynamic time warping distance, and the dynamic time warping distance between two nodes is denoted as the voltage fluctuation correlation coefficient ρ.

[0060] The specific calculation process is as follows:

[0061] Obtain the time series sequences of two nodes A and B: and .

[0062] in, This represents the energy entropy of the u-th frequency band in the node frequency band sequence F; The energy entropy of the m-th frequency band in the node frequency band sequence E.

[0063] The sample points of A and B are paired step by step. During the pairing, one-to-one and one-to-many matching are supported, thereby automatically aligning temporal differences.

[0064] After each pairing is completed, the Euclidean distance of the pairing result is calculated. After the Euclidean distance of all pairing results is calculated, the pairing results are weighted according to the node characteristics. After weighting, the total distance is calculated to minimize the total distance.

[0065] The feature weighting process includes: obtaining the frequency bands contained in the node feature data pairs and marking the mutual information of the corresponding frequency bands; and assigning standard mutual information to each frequency band. Then, all data pairs are traversed, and the mutual information of each frequency band is accumulated. The accumulated mutual information of each frequency band is normalized to obtain the attention weight of each frequency band, i.e., the voltage fluctuation correlation coefficient ρ.

[0066] In this embodiment, by introducing an improved Dynamic Time Warping Distance (DTW), not only can one-to-one and one-to-many sample point pairing be flexibly supported, automatically aligning asynchronous and length differences in timing between nodes and accurately capturing the true trajectory of signal changes, but it can also effectively reflect the dynamic coupling characteristics of node voltage signals under actual distribution network conditions. After calculating the pairing distance based on Euclidean distance, by combining node characteristics to weight the pairing results, the focus on representative and important features can be highlighted, while suppressing interference from noise and irrelevant information, making the final correlation coefficient ρ more physically meaningful and capable of fault detection. This correlation analysis method based on dynamic alignment and feature weighting provides a solid and scientific foundation for the detailed characterization of dynamic fault propagation relationships and the identification of abnormal paths among multiple nodes in a distribution network.

[0067] Dynamic time warping distance (VTW) is used to analyze correlations through time alignment. This means it's a long-term correlation analysis and therefore may not overemphasize the characteristics of a single frequency band. Utilizing the temporal correlation of nodes enhances the analysis of temporal correlations between different nodes. By labeling the frequency bands involved in each node's features with mutual information tags, and accumulating and normalizing the frequency band mutual information of all data pairs based on standard mutual information, the actual role of each frequency band in multi-node signal correlation can be effectively quantified. The resulting attention weights highlight key frequency bands that play a major role in long-term time-series correlation analysis, while suppressing redundant or noisy features that have a weak impact on overall correlation. Combining VTW with the temporal alignment of voltage signals between nodes not only accurately captures the dynamic coupling relationships between multiple nodes at different time scales but also avoids information loss or biased judgments caused by over-reliance on a single frequency band, thus enhancing the reliability and discriminative ability of node temporal correlation analysis.

[0068] In S4, based on the direct or indirect relationship between nodes, it is determined whether there is an edge relationship between two nodes. Between two nodes with an edge relationship, the voltage fluctuation correlation coefficient is used as an edge to obtain a dynamic fault propagation graph, thereby identifying the risk status of the connected path.

[0069] Directly related nodes include those directly connected by a circuit. Indirectly related nodes include those that are required to be synchronized, or those where adjusting one node will cause another node to change synchronously.

[0070] Specifically, risk identification involves first using mutual information to analyze the risk of each node, and then using edge relationships to analyze the total risk coefficient of the connected paths formed by nodes with direct or indirect connections.

[0071] In each node, a pre-trained multi-head attention neural network is used, with each attention head focusing on individual pairs of features in a node. Anomaly identification is performed using Transformer as the basic architecture, and the output anomaly probability is used as the risk coefficient of the node.

[0072] In this embodiment, when modeling the multi-state features of nodes, the multi-scale features of each node (such as combinations of energy entropy from different wavelet decomposition layers and different frequency bands) are first paired to form "individual pairs" as the smallest analysis unit. Then, using the multi-head attention mechanism in the Transformer architecture, an independent attention head is assigned to each "individual pair". Each attention head focuses only on its corresponding individual pair, and through adaptive learning of the correlation and importance between the features of the pair, the unique contribution of the feature combination in node anomaly identification is extracted.

[0073] The attention calculation is as follows: For the h-th attention head (assuming there are H heads in total), its input is the feature vector of a specific individual pair (a, b). This head maps the input to a query Q, key K, and value V vector through a linear transformation and calculates the attention score. The attention weight of each head is the softmax normalized weight score, representing the importance of that individual pair in the overall node features. The outputs of all heads are concatenated or weighted and fused in the final stage for anomaly probability discrimination.

[0074] In this way, each attention head focuses on only one "individual pair," and its attention weight reflects the degree of influence of that pair of features on the anomaly detection result. The entire multi-head structure can capture the interactions and contributions of different levels and scales in the multidimensional features of nodes in parallel, which greatly improves the model's ability to identify and interpret anomalies in complex states.

[0075] Furthermore, G connected nodes are obtained, and the average risk coefficient of these G nodes is used as the total risk coefficient. If the total risk coefficient is greater than a preset value, then the G nodes are determined to be high-risk connected paths (G is the maximum value). In fault propagation analysis, G represents the number of nodes within a connected path. Taking the maximum value of G means that in the dynamic fault propagation graph, for each connected path, the complete connected component containing the most nodes is selected for risk coefficient calculation and determination.

[0076] Specifically, in the high-risk connectivity path, the anomaly coefficient of each node is analyzed, and each node is arranged according to the magnitude of the anomaly coefficient to obtain the prediction result of the anomaly node. For node K, the anomaly coefficient is calculated as follows: the risk coefficient of other nodes is multiplied by the ρ of the edges directly or indirectly connected to node K, and then multiplied sequentially (if there are two or more edges connecting other nodes Y and node K, then the risk coefficient of node Y needs to be multiplied by these multiple edges), and then multiplied by the risk coefficient of node K. If there is more than one link between other nodes Y and node K, then the calculation results of different links are summed.

[0077] By conducting in-depth analysis of identified high-risk connectivity paths and calculating node anomaly coefficients, the risk impact of each node within a connectivity path is quantified by combining its correlation coefficient ρ with the edges it connects to. This not only fully reflects the node's own risk but also comprehensively considers the strength of risk transmission between the node and its surrounding nodes, as well as the complexity of the path structure. By multiplying the risk coefficients of other nodes layer by layer with ρ on all paths to node K, the transmission and superposition effects of risk in the topology are accurately reproduced, thus more accurately identifying key nodes in the fault propagation chain where risk accumulates and is prone to causing anomaly propagation. By sorting nodes by their anomaly coefficients, high-risk locations can be quickly identified, facilitating precise, tiered, and zoned early warning and targeted maintenance intervention, thereby improving the overall efficiency of accident response and risk prevention in the distribution network.

[0078] After the early warning is issued, further analysis based on the dynamic fault propagation graph can be performed on the distribution, connectivity and changing trends of abnormal nodes, enabling the tracing of the source of the fault and the tracking of its spread path.

[0079] In this embodiment, a dynamic fault propagation graph constructed based on correlation relationships can reflect the real-time coupling and risk diffusion characteristics of electrical states between nodes within the distribution network. Furthermore, by combining mutual information and a multi-head attention mechanism, fine-grained risk assessment is performed on each node, enabling the node risk coefficient to comprehensively consider the complex relationships between multi-state characteristics and features at different scales. A multi-head attention neural network structure with Transformer as its core is adopted, allowing each attention head to focus on different feature pairs, significantly improving the ability and sensitivity of abnormal state identification. By comprehensively evaluating the risk coefficients of multiple nodes within a path, high-risk connected paths are identified, which not only reflects the aggregation and diffusion trends of local risks in real time but also provides a scientific basis for fault early warning, operation and maintenance intervention, and accident prevention. This overcomes the shortcomings of traditional methods in dynamic risk propagation and intelligent anomaly identification between nodes, laying a solid foundation for high-reliability operation and intelligent safety management of the distribution network.

[0080] This invention achieves holographic acquisition, in-depth analysis, and precise anomaly early warning of voltage signals among multiple nodes in a distribution network through multi-channel high-frequency synchronous acquisition, multi-scale feature extraction based on wavelet packet decomposition and energy entropy, construction of multi-state node features within a sliding time window, correlation analysis based on dynamic time warping and feature weighting, and node risk assessment based on mutual information and multi-head attention mechanisms. This invention effectively captures the complex spatiotemporal coupling characteristics of the distribution network, improving the ability to identify minor disturbances and nonlinear anomalies. By using dynamic fault propagation graphs and high-risk connectivity path identification, it dynamically models the risk transmission and anomaly aggregation processes between nodes, supporting priority ranking and targeted early warning of key risk nodes.

[0081] Compared to existing technologies, this invention not only overcomes the shortcomings of traditional methods, such as low spatiotemporal resolution, coarse correlation calculation, and incomplete risk identification, but also provides theoretical and technical support for intelligent operation and maintenance and rapid fault response in power distribution networks. Overall, the invention has advantages such as data-driven approach, adaptive structure, traceable risks, and interpretable results, and can be widely applied to the fields of safety monitoring and intelligent early warning in modern power distribution networks.

[0082] Example 2

[0083] This embodiment provides a power distribution automatic monitoring and early warning system based on voltage measurement, including:

[0084] The matrix construction module is configured to collect voltage signals from associated nodes based on the distribution network topology and construct a spatiotemporal matrix.

[0085] The feature extraction module is configured to perform wavelet decomposition on the voltage signals of each node in the spatiotemporal matrix to obtain the frequency band components of each layer and calculate their energy entropy; treat each frequency band component as an independent individual, integrate the energy entropy of the lowest layer from bottom to top to form individual features; analyze the correlation between individuals based on mutual information, select individual pairs as independent states of nodes according to preset rules, and integrate all states to obtain node features with multiple states superimposed.

[0086] The correlation calculation module is configured to calculate the correlation between any two nodes based on the node characteristics and using the improved dynamic time warping distance as the voltage fluctuation correlation coefficient.

[0087] The risk warning module is configured to determine the edge relationship based on the topological association between nodes, construct a dynamic fault propagation graph with the voltage fluctuation correlation coefficient as the edge, analyze node features using a pre-trained multi-head attention neural network to obtain node risk coefficients, calculate the total risk coefficient of connected paths in combination with the dynamic fault propagation graph, and issue a warning for paths with a total risk coefficient exceeding a preset value.

[0088] Example 3

[0089] This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the power distribution automatic monitoring and early warning method based on voltage measurement as described in Embodiment 1 above.

[0090] Example 4

[0091] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the power distribution automatic monitoring and early warning method based on voltage measurement as described in Embodiment 1 above.

[0092] The steps or modules involved in Embodiments 2 to 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0093] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A power distribution automatic monitoring and early warning method based on voltage measurement, characterized in that, The method comprises the following steps: Collecting voltage signals of associated nodes based on the power distribution network topology structure, and constructing a space-time matrix; Wavelet decomposition is performed on the voltage signals of each node in the space-time matrix to obtain band components at each layer and calculate their energy entropy; Each layer of band components is regarded as an independent individual, and the energy entropy of the lowest layer is integrated from bottom to top to form an individual feature; Based on mutual information, the correlation between individuals is analyzed, and individuals are selected as the independent state of the node according to the preset rule; Specifically, the energy entropy of the independent individual is taken as the individual attribute, and all lower layer individual attributes contained in the individual at the current layer are integrated to form a sequence containing the energy entropy of each lowest layer band to the energy entropy of the current individual, which is taken as the individual feature of the current individual; Any two independent individuals are taken as an individual pair to obtain a first individual pair and a second individual pair, and the mutual information is calculated respectively; If the individuals in the first individual pair respectively contain the band part of the individuals in the second individual pair, the individual pair with larger mutual information is retained as the screening result of this time; otherwise, the first individual pair and the second individual pair are retained respectively; if there is no completely contained relationship between the first individual pair and the second individual pair, the two individual pairs are retained; all states are integrated to obtain a node feature with multiple state superpositions; Based on the node feature, the correlation between any two nodes is calculated using an improved dynamic time warping distance as the voltage fluctuation correlation coefficient; Specifically, based on the node features of any two nodes, the time sequence sequence of the band energy entropy is obtained; The two sequence sample points are paired step by step and support one-to-one and one-to-many matching to align the time sequence difference; After calculating the Euclidean distance of each pairing, the node feature is marked based on the mutual information of each band, and each band is given a standard mutual information and all data pairs are traversed to obtain the attention weight of each band; The attention weight is used to weight the Euclidean distance, and the minimum total distance is calculated as the voltage fluctuation correlation coefficient of the two nodes; According to the topological correlation between nodes, the edge relationship is determined, and the voltage fluctuation correlation coefficient is taken as the edge to construct a dynamic fault propagation graph; The node risk coefficient is obtained by analyzing the node feature using a pre-trained multi-head attention neural network, and the total risk coefficient of the connected path is calculated based on the dynamic fault propagation graph, and the path with a total risk coefficient exceeding a preset value is warned.

2. The power distribution automatic monitoring and early warning method based on voltage measurement according to claim 1, characterized in that, The rows of the space-time matrix correspond to time sequences, and the columns correspond to node positions.

3. The method of claim 1, wherein the method comprises: The wavelet decomposition is performed on the voltage signals of each node in the space-time matrix to obtain band components at each layer and calculate their energy entropy, which specifically comprises: Wavelet decomposition is performed on the voltage signals of each node in the space-time matrix to obtain multi-layer wavelet packet decomposition coefficients, each layer containing multiple band components; The energy of each band is calculated, and the energy entropy of each band is calculated after normalizing the energy of each band.

4. The method of claim 1, wherein the method comprises: The node risk coefficient is obtained by analyzing the node feature using a pre-trained multi-head attention neural network, and the total risk coefficient of the connected path is calculated based on the dynamic fault propagation graph, which specifically comprises: A multi-head attention neural network is constructed based on Transformer, each attention head paying attention to an individual pair in the node feature, and the node anomaly probability is output through the network as the node risk coefficient; In the dynamic fault propagation graph, G nodes with direct or indirect connections are extracted, and the average of the risk coefficients of the G nodes is calculated as the total risk coefficient of the connected path, where G is the maximum value of the number of nodes in the connected path.

5. A voltage measurement based automatic monitoring and early warning system for power distribution, characterized in that, The method comprises the following steps: The matrix construction module is configured to collect voltage signals of associated nodes based on the power distribution network topology to construct a space-time matrix. The feature extraction module is configured to perform wavelet decomposition on the voltage signals of each node in the space-time matrix to obtain band components at each layer and calculate the energy entropy thereof. Each layer of band components is regarded as an independent individual, and the energy entropy of the lowest layer is integrated from the bottom up to form an individual feature. The correlation between individuals is analyzed based on mutual information, and individual pairs are selected as the independent state of the node according to a preset rule; specifically, the energy entropy of the independent individual is taken as the individual attribute, and all lower layer individual attributes contained in the individual at the current layer are integrated to form a sequence containing the energy entropy of each lowest layer band to the energy entropy of the current individual, which is taken as the individual feature of the current individual. Any two independent individuals are taken as individual pairs to obtain a first individual pair and a second individual pair, and mutual information is calculated. If the individuals in the first individual pair respectively contain the band part of the individuals in the second individual pair, the individual pair with larger mutual information is retained as the screening result; otherwise, the first individual pair and the second individual pair are retained respectively; if there is no completely contained relationship between the first individual pair and the second individual pair, the two individual pairs are retained; and the node feature of the multi-state superposition is obtained by integrating all states. The correlation calculation module is configured to calculate the correlation between any two nodes based on the node feature using an improved dynamic time warping distance as the voltage fluctuation correlation coefficient; specifically, based on the node features of any two nodes, the time sequence sequence of the band energy entropy is obtained; the two sequence sample points are paired step by step and one-to-one and one-to-many matching is supported to align the time sequence difference; After calculating the Euclidean distance of each pairing, the mutual information of each band is marked based on the node feature, a standard mutual information is given to each band, and all data pairs are traversed to obtain the attention weight of each band; The Euclidean distance is weighted using the attention weight to calculate the minimum total distance as the voltage fluctuation correlation coefficient of the two nodes. The risk warning module is configured to determine the edge relationship according to the topological correlation between nodes, construct a dynamic fault propagation graph with the voltage fluctuation correlation coefficient as the edge, analyze the node features using a pre-trained multi-head attention neural network to obtain a node risk coefficient, and calculate a total risk coefficient of a connected path in combination with the dynamic fault propagation graph to perform early warning on the path with a total risk coefficient exceeding a preset value.

6. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the steps of the power distribution automatic monitoring and early warning method based on voltage measurement in any one of claims 1-4.

7. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the steps of the power distribution automatic monitoring and early warning method based on voltage measurement in any one of claims 1-4.

Images