Electric power data anomaly detection method and system combined with edge calculation

Abstract

The invention discloses an electric power data anomaly detection method and system combined with edge computing, and particularly relates to the field of electric power data anomaly detection.The electric power data anomaly detection method comprises the steps that a periodic abnormal behavior chain structure is established by extracting a continuous judgment result, actual physical feedback and a historical standard response track of edge nodes to electric power data; and a weighted residual trend path is generated, so that identification of a node judgment offset state and quantification of an abnormal trend are realized. By constructing a power data abnormal behavior chain and a multi-cycle residual error trend graph, an edge node misjudgment offset state is identified, and path suppression and channel structure correction are executed, so that the node safety judgment accuracy and the abnormal response stability in an edge computing environment are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power data anomaly detection, and more specifically, to a power data anomaly detection method and system combined with edge computing. Background Art

[0002] As power systems accelerate toward intelligent and distributed architectures, traditional data anomaly detection methods that rely primarily on central nodes are unable to effectively ensure node security in large-scale heterogeneous environments. Furthermore, they struggle to cope with the surge in the number of edge devices and their frequent state changes. In particular, response delays and recognition lags often occur when faced with sudden disturbances or early degradation signals.

[0003] To improve the ability to instantly perceive local anomalies and enhance the security and resilience of nodes in edge areas, power systems have gradually introduced edge computing architectures in practical applications. This shifts preliminary judgment and local response functions to edge nodes with sufficient computing power, enabling on-site analysis, rapid response, and network offloading.

[0004] Due to the limited computing resources of edge nodes, their node security relies on the accuracy of local models. Therefore, in practical applications, they often rely on lightweight models to perform anomaly identification tasks. Such models are prone to overfitting local features in long-term operation, causing the system to misidentify normal fluctuations as high-risk anomalies, thereby triggering unnecessary protection actions.

[0005] Once such misjudgments occur frequently, they will not only weaken the security robustness of nodes in the edge area, but will also cause linkage failure or control disorder among multiple edge nodes, and ultimately form a self-disturbing abnormal diffusion chain. Summary of the Invention

[0006] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a power data anomaly detection method and system combined with edge computing. By constructing a power data abnormal behavior chain and a multi-cycle residual trend map, it identifies the misjudgment offset state of edge nodes and performs path suppression and channel structure correction to improve the node safety judgment accuracy and abnormal response stability in the edge computing environment.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for detecting power data anomalies in combination with edge computing, comprising:

[0008] By extracting the continuous judgment results of edge nodes on power data, actual physical feedback and historical standard response trajectories, a periodic abnormal behavior chain structure is established, and a weighted residual trend path is generated from it, which enables the identification of node judgment offset status and quantification of abnormal trends.

[0009] For node behavior chains identified as misjudged deviation states, a response path suppression strategy group is constructed by combining response action history, misjudgment label trends, and the behavior patterns of adjacent nodes. This strategy also completes the generation of alternative target low-deviation paths and the unbinding of decision logic.

[0010] By summarizing the residual evolution trends of power data from multiple edge nodes and utilizing spatial location mapping and trend synergy modeling, we can extract the abnormal structure aggregation map within the region, and use this to determine the systematic model deviation phenomenon and the freezing level of the regulatory response.

[0011] Based on the residual comparison between the feedback path behavior and the judgment path, a channel-level correction tensor is constructed and model feature reconstruction is performed. The adjustment of the model version and the hierarchical tracking of the degradation state are realized through the structure evolution index sequence, so as to achieve continuous detection and correction of power data anomalies.

[0012] In a preferred embodiment, the model output label based on power data in each judgment cycle of the edge node, the triggered response category and its timing label are extracted and combined to form an abnormal judgment behavior chain unit; the response action called in the abnormal behavior chain unit is mapped to the feedback amount of voltage, current, frequency and phase angle change in the corresponding cycle to construct the power physical response signal flow;

[0013] Based on the state template preset by the power industry standard, the power physical response signal stream is subjected to piecewise normalization processing, and a residual vector between the signal stream and the standard response trajectory is generated;

[0014] The residual vectors are archived according to a single cycle to form a single cycle feedback residual indicator set corresponding to the power data, and the deviation magnitude between the judgment and feedback is marked.

[0015] In a preferred embodiment, the feedback residual indicator sets of multiple periods are spliced ​​in time series to establish a multi-period residual evolution matrix, and the first-order slope sequence, second-order fluctuation segment and adjacent disturbance transfer nodes are extracted therefrom;

[0016] The topological structure depth, load level, communication delay parameters and data refresh period of the area where the current edge node is located are introduced, and structural weighting is applied to the residual evolution matrix to generate a context-sensitive residual trend trajectory.

[0017] In a preferred embodiment, it is determined whether the residual trend trajectory meets at least two of the following conditions: the slope continues to rise for a long time, the local fluctuation amplitude exceeds the historical standard deviation threshold, the physical feedback occurs ahead of the judgment action, and the distance between the residual stable points decreases; if so, the node power data is marked as being in a misjudgment offset state.

[0018] In a preferred embodiment, the residual trend trajectory of the node behavior chain marked as misjudged offset state is input into the statistical feature analysis process to extract the trend duration, the number of abnormal dense segments, the reversal interval and the multi-scale residual envelope difference;

[0019] Based on the current node's power function type, historical misjudgment density, and adjacent node state synchronization rate, its response sensitivity matrix and misjudgment impact weight distribution map are calculated; power function types include grid-connected scheduling, load switching, and abnormal forwarding;

[0020] If the three joint conditions are met in the misjudgment impact weight distribution diagram, the response path suppression mechanism is triggered; the three joint conditions include the continuous growth of residual density within the cycle, the divergence of the collaborative behavior of adjacent nodes, and the increasing trend of misjudgment labels in the response action history.

[0021] In a preferred embodiment, a structural analysis is performed on the decision logic chain involved in the original response path to extract the decision instruction sequence and its execution order, and the model decision sub-channel on which each instruction depends and the mapping path relationship between it and the power feedback signal are calibrated;

[0022] Analyze the response behavior combination mode of the current judgment logic chain, and call the response suppression rule set including action-level delay strategy, feedback limiting rules and path substitution schemes to construct a set of candidate suppression paths, and perform node-level remapping and channel structure fusion with the current abnormal behavior chain structure;

[0023] The candidate suppression path set is compared with the pre-archived feedback quality stability verification path index table for residual similarity, and the path residual fit, feedback response matching and false trigger probability scores are jointly calculated. The optimal target alternative response path for the current cycle is screened and generated based on the weighted scoring function.

[0024] Perform a logical unbinding operation on the historical judgment channel structure bound to the optimal target alternative response path, and insert the alternative path into the current execution node of the abnormal behavior chain to complete the state connection and judgment sequence update after the path switching.

[0025] In a preferred embodiment, the context-weighted residual trend trajectories of all edge nodes in the region are aggregated to assemble a residual cooperative sequence set of multiple nodes in the region. Periodic realignment, node position spatial encoding, and load attribute cluster mapping are performed on the residual cooperative sequence set to generate a time-space joint offset matrix.

[0026] The trend coupling factor between node pairs is solved in the offset matrix, and a residual consensus graph between nodes is constructed based on the covariance kernel. A graph structure recognition mechanism is used to determine whether a target high-density coupled offset subgraph exists. If the following two conditions are met simultaneously: the node slope coordination degree continues to increase, the power factor in the system load partition fluctuates abnormally, and the behavioral structure clustering stability is weakened, it is considered a systemic misjudgment risk area.

[0027] Calculate the response freeze level distribution coefficient for nodes in the systemic misjudgment risk area, generate a control freeze priority vector table, apply the priority vector to the current regional control path set, construct a regional response control map, and clarify the main path intervention order and node de-throttling logic;

[0028] Report the regional response control map to the scheduling strategy control center to trigger the issuance of global model update strategy, path replacement instructions or control section isolation commands.

[0029] In a preferred embodiment, the response paths and corresponding physical feedback signals of the edge nodes for power data in each round are collected and summarized to form a consistent mapping set of paths and feedback states. Multi-dimensional residual calculations are performed on each path sample in the consistent mapping set to complete the solution of response offset values, quantify feedback delay intervals, and fit the judgment accuracy distribution curve, thereby generating a map of power data path behavior differences.

[0030] The path behavior difference map is input into the channel structure reconstruction process to determine the set of feature channels that need to be replaced and their corresponding channel weight update indexes. Based on the update index, the feature weights are reordered, the channel activation status is refreshed, and the path execution validity is detected to form the channel structure correction result.

[0031] In a preferred embodiment, the correction result is mapped into a node model structure map, and is bound to the corresponding power data determination version and then written into the structure evolution index sequence;

[0032] Perform response matching and residual fluctuation range statistics on each version in the structure evolution index sequence to determine whether its continuous evolution state is at the performance degradation boundary;

[0033] If the stability evaluation indicators of three consecutive versions are lower than the adaptive tolerance threshold, the set of difference factors between the current structure and the historical stable version is solved, and a degradation factor tensor graph is constructed, which serves as the configuration basis for the subsequent retraining of the power data anomaly detection structure.

[0034] An electric power data anomaly detection system combined with edge computing includes an offset identification module, a path correction module, a monitoring module, and a structure repair module;

[0035] The offset identification module extracts the edge node's continuous judgment results on power data, actual physical feedback, and historical standard response trajectories to establish a periodic abnormal behavior chain structure. From this, it generates a weighted residual trend path to identify the node's judgment offset state and quantify the abnormal trend.

[0036] The path correction module is used to construct a response path suppression strategy group for node behavior chains identified as misjudged deviation states, combining response action history, misjudgment label trends, and the behavior patterns of adjacent nodes. It also completes the generation of alternative target low-deviation paths and the unbinding of decision logic.

[0037] The monitoring module aggregates the residual evolution trends of power data from multiple edge nodes, uses spatial location mapping and trend synergy modeling to extract abnormal structure aggregation maps within the region, and uses this to determine systemic model deviation phenomena and the freezing level of regulatory response.

[0038] The structural repair module constructs a channel-level correction tensor and performs model feature reconstruction based on the residual comparison between the feedback path behavior and the judgment path. It adjusts the model version and hierarchically tracks the degradation state through the structural evolution index sequence to achieve continuous detection and correction of power data anomalies.

[0039] The technical effects and advantages of the present invention are as follows:

[0040] 1. By constructing abnormal behavior chains and residual trend paths for power data, we can identify misjudgment deviations and improve the timeliness and accuracy of edge node security assessments.

[0041] 2. By collaboratively modeling historical misjudgment density and adjacency status, we perform path suppression and replacement logic reconstruction to enhance node security's ability to self-suppress and dynamically adjust to abnormal misjudgments.

[0042] 3. Through regional multi-node residual coupling analysis and offset clustering identification, systematic abnormal propagation trends can be discovered in advance, improving the collaborative anti-proliferation capabilities of node security in regional regulation;

[0043] 4. Through channel-level residual back-injection and structural evolution index-driven model correction, misjudgment source channels are eliminated and the structure is reconstructed to ensure the continuous and stable operation of the node security model. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a flow chart of the misjudgment trend identification and path suppression phase of the present invention.

[0045] Figure 2 A flow chart for regional consensus identification and regulation map generation of the present invention.

[0046] Figure 3This is a flow chart of the path feedback re-injection and model structure evolution phase of the present invention.

[0047] Figure 4 The figure is a flow chart of the method steps of the present invention.

[0048] Figure 5 Schematic diagram of the system module of the present invention. DETAILED DESCRIPTION

[0049] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0050] Refer to the instruction manual Figure 1-5 According to an embodiment of the present invention, a method for detecting anomalies in power data combined with edge computing includes:

[0051] By extracting the continuous judgment results of edge nodes on power data, actual physical feedback and historical standard response trajectories, a periodic abnormal behavior chain structure is established, and a weighted residual trend path is generated from it, which enables the identification of node judgment offset status and quantification of abnormal trends.

[0052] For node behavior chains identified as misjudged deviation states, a response path suppression strategy group is constructed by combining response action history, misjudgment label trends, and the behavior patterns of adjacent nodes. This strategy also completes the generation of alternative target low-deviation paths and the unbinding of decision logic.

[0053] By summarizing the residual evolution trends of power data from multiple edge nodes and utilizing spatial location mapping and trend synergy modeling, we can extract the abnormal structure aggregation map within the region, and use this to determine the systematic model deviation phenomenon and the freezing level of the regulatory response.

[0054] Based on the residual comparison between the feedback path behavior and the judgment path, a channel-level correction tensor is constructed and model feature reconstruction is performed. The adjustment of the model version and the hierarchical tracking of the degradation state are realized through the structure evolution index sequence, so as to achieve continuous detection and correction of power data anomalies.

[0055] Extract the model output labels, triggered response categories, and their timing labels based on power data within each judgment cycle in the edge node, and combine them to form an abnormal judgment behavior chain unit. Map the response actions called in the abnormal behavior chain unit to the feedback quantities of voltage, current, frequency, and phase angle changes within the corresponding cycle to construct the power physical response signal flow.

[0056] Based on the state template preset by the power industry standard, the power physical response signal stream is subjected to piecewise normalization processing, and a residual vector between the signal stream and the standard response trajectory is generated;

[0057] The residual vectors are archived according to a single cycle to form a single cycle feedback residual indicator set corresponding to the power data, and the deviation magnitude between the judgment and feedback is marked.

[0058] The feedback residual indicator sets of multiple periods are spliced ​​in time series to establish a multi-period residual evolution matrix, and the first-order slope series, second-order fluctuation segments and adjacent disturbance transfer nodes are extracted;

[0059] The topological structure depth, load level, communication delay parameters and data refresh period of the area where the current edge node is located are introduced, and structural weighting is applied to the residual evolution matrix to generate a context-sensitive residual trend trajectory.

[0060] Determine whether the residual trend trajectory meets at least two of the following conditions: the slope continues to rise over a long period of time, the local fluctuation amplitude exceeds the historical standard deviation threshold, the physical feedback occurs ahead of the judgment action, and the distance between the residual stable points decreases; if so, the node power data is marked as being in a misjudgment offset state.

[0061] The residual trend trajectory of the node behavior chain marked as misjudged offset state is input into the statistical feature analysis process to extract the trend duration, the number of abnormal dense segments, the reversal interval and the multi-scale residual envelope difference;

[0062] Based on the current node's power function type, historical misjudgment density, and adjacent node state synchronization rate, its response sensitivity matrix and misjudgment impact weight distribution map are calculated; power function types include grid-connected scheduling, load switching, and abnormal forwarding;

[0063] If the three joint conditions are met in the misjudgment impact weight distribution diagram, the response path suppression mechanism is triggered; the three joint conditions include the continuous growth of residual density within the cycle, the divergence of the collaborative behavior of adjacent nodes, and the increasing trend of misjudgment labels in the response action history.

[0064] Perform structural analysis on the decision logic chain involved in the original response path, extract the decision instruction sequence it contains and its execution order, and calibrate the model decision sub-channel that each instruction depends on and the mapping path relationship between it and the power feedback signal;

[0065] Analyze the response behavior combination mode of the current judgment logic chain, and call the response suppression rule set including action-level delay strategy, feedback limiting rules and path substitution schemes to construct a set of candidate suppression paths, and perform node-level remapping and channel structure fusion with the current abnormal behavior chain structure;

[0066] The candidate suppression path set is compared with the pre-archived feedback quality stability verification path index table for residual similarity, and the path residual fit, feedback response matching and false trigger probability scores are jointly calculated. The optimal target alternative response path for the current cycle is screened and generated based on the weighted scoring function.

[0067] Perform a logical unbinding operation on the historical judgment channel structure bound to the optimal target alternative response path, and insert the alternative path into the current execution node of the abnormal behavior chain to complete the state connection and judgment sequence update after the path switching.

[0068] Summarize the context-weighted residual trend trajectories of all edge nodes in the region and assemble them into a residual cooperative sequence set of multiple nodes in the region. Perform period realignment, node position spatial encoding, and load attribute cluster mapping on the residual cooperative sequence set to generate a joint time and space offset matrix.

[0069] The trend coupling factor between node pairs is solved in the offset matrix, and a residual consensus graph between nodes is constructed based on the covariance kernel. A graph structure recognition mechanism is used to determine whether a target high-density coupled offset subgraph exists. If the following two conditions are met simultaneously: the node slope coordination degree continues to increase, the power factor in the system load partition fluctuates abnormally, and the behavioral structure clustering stability is weakened, it is considered a systemic misjudgment risk area.

[0070] Calculate the response freeze level distribution coefficient for nodes in the systemic misjudgment risk area, generate a control freeze priority vector table, apply the priority vector to the current regional control path set, construct a regional response control map, and clarify the main path intervention order and node de-throttling logic;

[0071] Report the regional response control map to the scheduling strategy control center to trigger the issuance of global model update strategy, path replacement instructions or control section isolation commands.

[0072] The response paths and corresponding physical feedback signals of the edge nodes in each round of power data are collected and summarized to form a consistent mapping set of paths and feedback states. Multi-dimensional residual calculations are performed on each path sample in the consistent mapping set to complete the solution of response offset values, quantify the feedback delay interval, and fit the judgment accuracy distribution curve, thereby generating a map of power data path behavior differences.

[0073] The path behavior difference map is input into the channel structure reconstruction process to determine the set of feature channels that need to be replaced and their corresponding channel weight update indexes. Based on the update index, the feature weights are reordered, the channel activation status is refreshed, and the path execution validity is detected to form the channel structure correction result.

[0074] Map the correction results into a node model structure graph, bind it with the corresponding power data judgment version, and write it into the structure evolution index sequence;

[0075] Perform response matching and residual fluctuation range statistics on each version in the structure evolution index sequence to determine whether its continuous evolution state is at the performance degradation boundary;

[0076] If the stability evaluation indicators of three consecutive versions are lower than the adaptive tolerance threshold, the set of difference factors between the current structure and the historical stable version is solved, and a degradation factor tensor graph is constructed, which serves as the configuration basis for the subsequent retraining of the power data anomaly detection structure.

[0077] An electric power data anomaly detection system combined with edge computing includes an offset identification module, a path correction module, a monitoring module, and a structure repair module;

[0078] The offset identification module extracts the edge node's continuous judgment results on power data, actual physical feedback, and historical standard response trajectories to establish a periodic abnormal behavior chain structure. From this, it generates a weighted residual trend path to identify the node's judgment offset state and quantify the abnormal trend.

[0079] The path correction module is used to construct a response path suppression strategy group for node behavior chains identified as misjudged deviation states, combining response action history, misjudgment label trends, and the behavior patterns of adjacent nodes. It also completes the generation of alternative target low-deviation paths and the unbinding of decision logic.

[0080] The monitoring module aggregates the residual evolution trends of power data from multiple edge nodes, uses spatial location mapping and trend synergy modeling to extract abnormal structure aggregation maps within the region, and uses this to determine systemic model deviation phenomena and the freezing level of regulatory response.

[0081] The structural repair module constructs a channel-level correction tensor and performs model feature reconstruction based on the residual comparison between the feedback path behavior and the judgment path. It adjusts the model version and hierarchically tracks the degradation state through the structural evolution index sequence to achieve continuous detection and correction of power data anomalies.

[0082] It should be further explained that in the formula structure involved in this scheme, dimensionless terms can serve as proportionality or structural adjustment factors. When combined with quantities with units, they only play a numerical scaling role and do not introduce new physical dimensions. Therefore, they will not change or confuse the overall unit system of expression. This combination of "dimensionless terms and units" can be understood as a composite structural expression commonly used in mathematical and physical modeling, conforming to the principle of dimensional consistency and having a clear physical interpretation basis.

[0083] Secondly, in the formula structure of this scheme, if multiple variables with different physical units are involved, including but not limited to time, mass, or energy variables, their joint appearance is to express the collaborative modeling relationship of multiple physical mechanisms. Each variable is formed into a unified structure through function mapping, ratio combination, or normalization adjustment. The units and meanings are clear, and the overall expression conforms to the principle of dimensional consistency and the common formula of engineering modeling.

[0084] The content of the basic steps for constructing the evolution of power data anomaly behavior chain and misjudgment deviation trend is to structurally analyze the power data anomaly judgment behavior chain, extract the judgment channel, physical feedback disturbance and structural propagation tension of each path segment in the edge node, and construct a complete misjudgment deviation trend map through nonlinear integral residual modeling and context parameter adjustment to achieve quantifiable identification of misjudgment status. Based on this, Model 1 is established:

[0085]

[0086] in is the total residual evolution trend strength index of power data in period t, which is used to measure whether the node enters the misjudgment offset state; N t represents the total number of path segments that constitute the behavior chain in period t. The total number of path segments of the behavior chain comes from the path segmentation of the actual power decision logic; is the decision channel coupling function of the i-th path in period t, which is used to reflect the adjustment ability of the current channel structure on the feedback of this segment; is the threshold deviation rate between the power feedback and the standard trajectory in the i-th path, and the threshold deviation rate is used to define the maximum absolute error function; is the channel hopping frequency factor of the i-th segment. The channel hopping frequency factor is the jump density that occurs in the continuous execution of the decision logic in this segment. The channel hopping frequency factor is used to measure the intensity of path switching; is the feedback signal trajectory function of path segment i in period t, the input is the time variable s, and the output is the change sequence of physical quantities (such as voltage, current, and frequency); are the start and end time points of the i-th path, respectively. The start and end time points come from the time boundaries of the path judgment labels in the behavior chain; is the node propagation tension coefficient corresponding to the i-th path in the edge topology. The node propagation tension coefficient is used to reflect the influence range of this segment. is the context adjustment item of period t, which is used to dynamically adjust the overall trend indicator by integrating the operation status of the node structure;

[0087] The composition of Model 1 needs to be further explained. The first item is Indicates the sensitivity of the adjustment ability of the current path segment judgment channel to the feedback error. If the offset is larger and the channel control response is more intense, the partial derivative value will be higher. is the hopping frequency of the decision structure in this path segment, which can be defined as: That is, the number of jumps in the channel structure per unit time, #(logicswitchesini) represents the time interval covered by path segment i The model determines the number of channel switches, and each switch from one channel to another (i.e., the "determination instruction jump" in the behavior chain) is recorded as a logic jump. The product of the two in the first term reflects the "response amplification effect" of the path segment to the error change, that is, whether the path is likely to amplify the error to the subsequent stage.

[0088] The second term of Model 1 is: This term is the integral term for the intensity of the nonlinear response to the disturbance of the power feedback signal in the path segment. The second-order derivative represents the "acceleration" or "curvature" of the feedback change. If there are spikes, sudden changes, or irregular shocks, the integral will increase. The absolute value of the second term in Model 1 is the disturbance intensity, and the integral covers the entire path segment, which is the dynamic structural basis for the residual.

[0089] The third term of Model 1 is: It represents the tension propagation capability of the edge node of the road section in the topology. For example, the value will be larger when it is connected to the main road, key forwarding node, or edge load node. It can be analyzed from the power topology diagram and is defined as: where deg out (v i ) is the node v i The output connectivity, κ(v i ) is its load coupling tension factor;

[0090] The fourth term in Model 1 is the context adjustment term, which can be expressed as: is the load level of the area where the node is located (numeric type); is the network topology depth of the node, that is, the path length from the main control center; It is the data refresh period of the node; its structure constitutes a nonlinear compression function, which produces an amplification effect when the topology is deep and the update is slow, and adjusts the total response of the behavior chain.

[0091] In the process of suppressing and correcting the power data response path under the misjudgment offset state, for the power data behavior chain identified as the misjudgment offset state, a set of candidate suppression paths is constructed through collaborative analysis of residual trajectory patterns, channel historical stability, and adjacent node states. The suppression paths are scored and ranked based on a multidimensional function coupling method, and the logical unbinding of the original response path and the access to the alternative path are completed to construct Model 2:

[0092]

[0093] where π t It represents the optimal response path replacement structure finally selected in period t. The optimal response path replacement structure is used to replace the current misjudged path. represents the candidate path set generated by the response suppression rule in the current cycle t, and the candidate path set includes action-level, feedback-level, and channel-level suppression combinations; π j is the jth alternative path in the candidate set, which has different path structures, channel calling orders, and response mechanisms; is π j and The residual fitting error function is used to quantify whether the path can effectively cover the misjudgment; α(π j ,Ξ t ) is the path π j Behavioral consistency scoring function in the stability of historical cycle feedback structure;Ξ t The historical feedback behavior trajectory matrix records the feedback timing and residual window of past execution results by path label index; is the path π j The behavior coordination separation function between the current neighboring node behavior state; is the behavioral state sequence of all adjacent nodes in period t, including action labels, feedback trends and structural path records;

[0094] The generation logic of each component in Model 2 is as follows:

[0095] Where as the path residual fitting term:

[0096] in is the path π j The predicted intensity of the feedback disturbance response in the i-th segment of the behavior chain; ω i The structural weight comes from the propagation tension of the path segment (which is related to Equivalent), used to control the influence of each segment matching error; the overall reflection path π j Whether its control output can effectively cover the original misjudgment offset segment;

[0097] Among them, as the path history feedback consistency item:

[0098] in is the path π j The sth feedback value in the kth execution of the historical period; represents the feedback center trajectory in the same period; ν s is the feedback response delay depth factor, which is the delay influence of the s-th feedback value relative to the path starting point; T path π j The total number of cycles that were actually executed in history; The length of the feedback response vector. The path history feedback consistency term is generally expressed as: whether the feedback behavior of the path is stable and whether the deviation is controllable in the historical execution, which is a score of the historical credibility of the path itself.

[0099] Among them, as the path adjacency state coordination item: in Represents the set of all adjacent nodes; is the path π j The corresponding response behavior structure vector; is the current response behavior structure vector of the adjacent node n; Sim(·,·) is the structural similarity function, which is calculated based on the dynamic time warping method of the path instruction sequence and the feedback segmentation trend; χ n is the output connectivity of the adjacent nodes in the power topology, which is used to measure the influence of its structure on the main path. The smaller the value of the path adjacency state coordination item, the stronger the path π j The less “conflicting” a neighboring behavior is, the more suitable it is as a local replacement structure under misjudgment.

[0100] In the process of integrating the residual trends of power data from multiple nodes in the region to build a consensus behavior graph and identify systemic offset risks, the power data residual trends of all edge nodes in the region are used to extract the trend coupling relationship between path segments, structural position adjacency, and behavioral structure consistency, and build a joint consensus graph. Through the graph structure aggregation index and offset propagation connectivity, the systemic offset risk is judged and a response freezing level graph is generated; Model 3 is constructed:

[0101]

[0102] Among them C t is the regional residual consensus coupling map strength index constructed under period t. The regional residual consensus coupling map strength index is used to evaluate the trend of systematic misjudgment; (i, j) represents a node pair consisting of any two edge nodes in the region; is the set of node pairs that have physical connectivity and participate in residual conduction under period t; is the structural connectivity weight between the node pair (i, j), which combines their topological adjacency and intercommunication load flow; is the residual trend synergy factor of the node pair (i, j) in period t; is the topological relative depth factor between nodes, quantifying its spatial conduction gradient; is the determination path consistency factor of the node behavior chain structure;

[0103] Structural connectivity weight of the connectivity weight item include: in Indicates the direct connection strength between node pairs in the topology graph (depending on path length and connectivity); Indicates the load interaction between the node pair in the cycle (such as the total amount of substation / branch current interaction); the connectivity weight term measures the degree of "residual conductivity" of the physical connection, that is, whether there is a propagation basis on the structural path;

[0104] As the residual trend cofactor

[0105]

[0106] in are the residual trend trajectory functions of nodes i and j, respectively. The difference in their first-order derivatives indicates whether the trend evolution directions between nodes are consistent, and the integral is the total difference in collaborative trends. The smaller the difference, the higher the trend synchronization, which is the basic indicator of consensus.

[0107] As the topological depth factor term is

[0108]

[0109] in is the structural depth of node i in the power control topology (e.g., the path length to the master control node); is the structural depth of node j in the power control topology. The smaller the depth gap, the more consistent the responses to the same type of interference. This topological depth factor ensures that consensus is formed only within the same logical layer.

[0110] As the factor for determining path consistency

[0111] Among them, P i is the set of path call sequences of node i; the numerator of the path consistency factor is the number of intersections of the path structures, and the denominator is the length of the longer path, which is used for normalization; a value close to 1 indicates that the behavior patterns of the two nodes are exactly the same and there is strong structural overlap;

[0112] C as an indicator of the overall structure diagram t , the synergy of each pair of nodes is quantified as the product of three factors, multiplied by the structural propagation weight; ∑ (i,j) Aggregate the weighted values ​​of the coordination degree of all node pairs to form the overall regional coordination deviation trend graph; when C t When a certain systematic deviation threshold is reached, it can be determined that systematic misjudgment propagation has been activated.

[0113] In establishing a power data path feedback and re-injection chain to implement structural correction and evolution tracking of lightweight models, a structured residual comparison is performed between the actual feedback of the misjudgment path and the execution result of the original judgment logic channel. The channel contribution and characteristic failure path are extracted, and the channel structure is dynamically reorganized. The previous correction status of the model structure is recorded in the evolution index chain to achieve traceable evolution of the power data anomaly judgment model. Based on this, Model 4 is constructed:

[0114]

[0115] in is the current decision model structure vector set before the kth feedback injection, including all channel activation states and decision orders; is the new model structure state after the k+1th update, which is the input model for the next cycle; Q (k) Eliminate the feature channels identified as invalid or high-risk in the k-th update; is the reconstruction set of alternative channels extracted in the kth reinjection cycle, constructed by the residual back-mapping; It is a structure encapsulation function, which means that the structure after channel reorganization is mapped into an actual executable decision graph structure;

[0116] The structural elimination of subset items in Model 4 include:

[0117] where c i is the i-th decision channel in the model structure; i ) is channel c i The misjudgment residual concentration function, channel c i The misclassification residual concentration function is defined by the following formula: represents the residual error between feedback and decision in the path p; σ p is the trigger frequency of path p; P t (c i ) indicates that channel c is included in the tth cycle i The set of all behavior path instances of η dropis the misjudgment rejection threshold, which is not a fixed constant and is dynamically estimated by the residual floating interval of the stable path set in the model; T k represents the total number of cycles executed in the model history up to the kth feedback injection cycle; the structural elimination subset item selects all channels that "repeatedly cause excessively large residuals in multiple cycles and are frequently triggered" as candidates for structural elimination;

[0118] As an alternative channel collection item include:

[0119]

[0120] in is the global channel candidate pool, which contains all the trained channel subgraphs of this node; Θ(c,c i ) is the channel structure function similarity distance function, which is used to measure the new channel c and the replaced channel c i The matching degree of the decision logic, input features and output response; Ω(c) is the residual risk constraint function of channel c, which is weighted by the residual integral caused by the channel in the historical feedback; each eliminated channel c in the replacement channel set i Corresponding to an optimal alternative channel Through the joint judgment of logical function equivalence and historical error stability; the alternative channel set item completes the insertion of alternative channels through judgment capability inheritance + residual safety selection;

[0121] As a model wrapper function The current structure vector is remapped into an executable model after elimination and completion; if the structure depth changes, the path graph is reconstructed and the node dependency table and behavior chain mapping table are updated; the new model output Will be used for the next cycle judgment. Supplementary explanation: The structure evolution index recording logic (not the main formula, but supporting the complete logic) is used to track the long-term trend of the model. k Represents the version snapshot of the model after the kth structural update, which is used to record the model structure, residual status and effectiveness score of this round;

[0122] where δ k is the k-th round residual mean square interval; γ k Score for structural validity after round k; all versions are stored in the structural evolution index chain in chronological order Used for subsequent judgment of degradation evolution risk or retraining startup.

[0123] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A power data anomaly detection method combined with edge computing, characterized in that: include: By extracting the continuous judgment results of edge nodes on power data, actual physical feedback and historical standard response trajectories, a periodic abnormal behavior chain structure is established, and a weighted residual trend path is generated from it, which enables the identification of node judgment offset status and quantification of abnormal trends. For the node behavior chain identified as misjudged deviation, a response path suppression strategy group is constructed by combining the response action history, misjudged label trend, and the behavior pattern of adjacent nodes. This strategy also completes the alternative generation of the target deviation path and the unbinding of the judgment logic. By summarizing the residual evolution trends of power data from multiple edge nodes and utilizing spatial location mapping and trend synergy modeling, we can extract the abnormal structure aggregation map within the region, and use this to determine the systematic model deviation phenomenon and the freezing level of the regulatory response. Based on the residual comparison between the feedback path behavior and the judgment path, a channel-level correction tensor is constructed and model feature reconstruction is performed. The adjustment of the model version and the hierarchical tracking of the degradation state are realized through the structure evolution index sequence, so as to achieve continuous detection and correction of power data anomalies.

2. The power data anomaly detection method combined with edge computing according to claim 1 is characterized by: Extract the model output labels, triggered response categories, and their timing labels based on power data within each judgment cycle in the edge node, and combine them to form an abnormal judgment behavior chain unit. Map the response actions called in the abnormal behavior chain unit to the feedback quantities of voltage, current, frequency, and phase angle changes within the corresponding cycle to construct the power physical response signal flow. Based on the state template preset by the power industry standard, the power physical response signal stream is subjected to piecewise normalization processing, and a residual vector between it and the standard response trajectory is generated; The residual vectors are archived according to a single cycle to form a single cycle feedback residual indicator set corresponding to the power data, and the deviation magnitude between the judgment and feedback is marked.

3. The power data anomaly detection method combined with edge computing according to claim 2 is characterized in that: The feedback residual indicator sets of multiple periods are spliced ​​in time series to establish a multi-period residual evolution matrix, and the first-order slope series, second-order fluctuation segments and adjacent disturbance transfer nodes are extracted; The topological structure depth, load level, communication delay parameters and data refresh period of the area where the current edge node is located are introduced, and structural weighting is applied to the residual evolution matrix to generate a context-sensitive residual trend trajectory.

4. The power data anomaly detection method combined with edge computing according to claim 3 is characterized by: Determine whether the residual trend trajectory meets at least two of the following conditions: the slope continues to rise over a long period of time, the local fluctuation amplitude exceeds the historical standard deviation threshold, the physical feedback occurs ahead of the judgment action, and the distance between the residual stable points decreases; if so, the node power data is marked as being in a misjudgment offset state.

5. The power data anomaly detection method combined with edge computing according to claim 4 is characterized in that: The residual trend trajectory of the node behavior chain marked as misjudged offset state is input into the statistical feature analysis process to extract the trend duration, the number of abnormal dense segments, the reversal interval and the multi-scale residual envelope difference; Based on the current node's power function type, historical misjudgment density, and adjacent node state synchronization rate, its response sensitivity matrix and misjudgment impact weight distribution map are calculated; power function types include grid-connected scheduling, load switching, and abnormal forwarding; If the three joint conditions are met in the misjudgment impact weight distribution diagram, the response path suppression mechanism is triggered; the three joint conditions include the continuous growth of residual density within the cycle, the divergence of the collaborative behavior of adjacent nodes, and the increasing trend of misjudgment labels in the response action history.

6. The power data anomaly detection method combined with edge computing according to claim 5 is characterized by: Perform structural analysis on the decision logic chain involved in the original response path, extract the decision instruction sequence it contains and its execution order, and calibrate the model decision sub-channel that each instruction depends on and the mapping path relationship between it and the power feedback signal; Analyze the response behavior combination mode of the current judgment logic chain, and call the response suppression rule set including action-level delay strategy, feedback limiting rules and path substitution schemes to construct a set of candidate suppression paths, and perform node-level remapping and channel structure fusion with the current abnormal behavior chain structure; The candidate suppression path set is compared with the pre-archived feedback quality stability verification path index table for residual similarity, and the path residual fit, feedback response matching and false trigger probability scores are jointly calculated. The target alternative response path for the current cycle is screened and generated based on the weighted scoring function. Perform a logical unbinding operation on the historical judgment channel structure bound to the target alternative response path, and insert the alternative path into the current execution node of the abnormal behavior chain to complete the state connection and judgment sequence update after the path switching.

7. The power data anomaly detection method combined with edge computing according to claim 6 is characterized by: Summarize the context-weighted residual trend trajectories of all edge nodes in the region and assemble them into a residual cooperative sequence set of multiple nodes in the region. Perform period realignment, node position spatial encoding, and load attribute cluster mapping on the residual cooperative sequence set to generate a joint time and space offset matrix. The trend coupling factor between node pairs is solved in the offset matrix, and a residual consensus graph between nodes is constructed based on the covariance kernel. A graph structure recognition mechanism is used to determine whether a target coupling offset subgraph exists. If the following two conditions are met simultaneously: the node slope coordination degree continues to increase, the power factor in the system load partition fluctuates abnormally, and the behavioral structure clustering stability is weakened, it is considered a systemic misjudgment risk area. Calculate the response freeze level distribution coefficient for nodes in the systemic misjudgment risk area, generate a control freeze priority vector table, apply the priority vector to the current regional control path set, construct a regional response control map, and clarify the main path intervention order and node de-throttling logic; Report the regional response control map to the scheduling strategy control center to trigger the issuance of global model update strategy, path replacement instructions or control section isolation commands.

8. The power data anomaly detection method combined with edge computing according to claim 7 is characterized in that: The response paths and corresponding physical feedback signals of the edge nodes in each round of power data are collected and summarized to form a consistent mapping set of paths and feedback states. Multi-dimensional residual calculations are performed on each path sample in the consistent mapping set to complete the solution of response offset values, quantify the feedback delay interval, and fit the judgment accuracy distribution curve, thereby generating a map of power data path behavior differences. The path behavior difference map is input into the channel structure reconstruction process to determine the set of feature channels that need to be replaced and their corresponding channel weight update indexes. Based on the update index, the feature weights are reordered, the channel activation status is refreshed, and the path execution validity is detected to form the channel structure correction result.

9. The power data anomaly detection method combined with edge computing according to claim 8, characterized in that: Map the correction results into a node model structure graph, bind it with the corresponding power data judgment version, and write it into the structure evolution index sequence; Perform response matching and residual fluctuation range statistics on each version in the structure evolution index sequence to determine whether its continuous evolution state is at the performance degradation boundary; If the stability evaluation indicators of three consecutive versions are lower than the adaptive tolerance threshold, the set of difference factors between the current structure and the historical stable version is solved, and a degradation factor tensor graph is constructed, which serves as the configuration basis for the subsequent retraining of the power data anomaly detection structure.

10. A power data anomaly detection system combined with edge computing, comprising an offset identification module, a path correction module, a monitoring module, and a structure repair module, characterized in that: The offset identification module extracts the edge node's continuous judgment results on power data, actual physical feedback, and historical standard response trajectories to establish a periodic abnormal behavior chain structure. From this, it generates a weighted residual trend path to identify the node's judgment offset state and quantify the abnormal trend. The path correction module is used to construct a response path suppression strategy group for the node behavior chain identified as misjudged deviation state, combining the response action history, misjudgment label trend, and the behavior pattern of adjacent nodes. It also completes the alternative generation of the target deviation path and the decoupling of the judgment logic. The monitoring module aggregates the residual evolution trends of power data from multiple edge nodes, uses spatial location mapping and trend synergy modeling to extract abnormal structure aggregation maps within the region, and uses this to determine systemic model deviation phenomena and the freezing level of regulatory response. The structural repair module constructs a channel-level correction tensor and performs model feature reconstruction based on the residual comparison between the feedback path behavior and the judgment path. It adjusts the model version and hierarchically tracks the degradation state through the structural evolution index sequence to achieve continuous detection and correction of power data anomalies.

Images