A dynamic reasoning method for electrical fault of substation secondary circuit

Abstract

The application belongs to the technical field of data processing, and particularly relates to a dynamic deduction method of electrical fault of secondary circuit of a transformer substation. The method aligns multi-source data according to a unified time scale, adopts adaptive weighting fusion to obtain a comprehensive deviation degree, sets a threshold based on historical high quantile to screen an abnormal node set, in a reverse search stage, calculates an exponential decay base according to a comprehensive edge weight parameter to obtain a conduction attenuation degree, and obtains a path comprehensive conduction coefficient by multiplying along an effective branch, then selects a candidate starting point with a high conduction coefficient, performs a forward topological transfer according to a deterministic causal adjacency matrix to obtain a predicted node set, takes an asymmetric directional coverage rate as a coincidence coefficient, combines a normalized conduction coefficient and a coincidence coefficient to calculate a confidence degree, and outputs an abnormal trigger source as a fault root cause. The application effectively reduces the false alarm rate of multi-source data noise, and improves the speed of graph topological deduction and the reliability of data processing.

Description

Technical Field

[0001] This invention relates to the field of data processing technology. More specifically, this invention relates to a method for dynamically simulating electrical faults in the secondary circuits of substations. Background Technology

[0002] In the processing of secondary circuit status data and anomaly tracing in substations, existing inference methods typically handle step changes in single status data (such as discrete event parameters or continuous status parameters) and use graph search algorithms based on topological connections to locate the anomaly trigger source.

[0003] However, the multi-source data acquisition environment of substation secondary circuits is complex. External factors such as high-voltage switch switching and leakage flux of large transformers often introduce transient interference spikes and high-frequency state jitter in the acquired data. If only a single channel threshold is relied upon for triggering, such disturbances are easily misidentified as abnormal starting nodes. This will cause the graph search algorithm to spread along invalid paths, the branch size will grow exponentially, and it will seriously affect the real-time performance and computational efficiency of data processing and inference.

[0004] In the process of tracing back to the anomaly origin based on graph data, existing methods generally adopt a pure reverse graph accumulation search strategy, that is, tracing back and accumulating energy indicators according to topological adjacency relationships. This strategy does not take into account the attenuation constraints of parameters such as secondary cable length and series coil impedance on the correlation and transmission of state data, which results in the retention of a large number of paths that cannot effectively transmit anomalies, expanding the scope of invalid search. Especially when there is a multi-input single-output parallel structure in the cascaded control network, pure reverse accumulation is limited by topological symmetry, and often derives multiple paths with similar energy values ​​from symmetrical parallel branches, outputting multiple candidate anomaly origins. This causes multiple ambiguities in the deduction results, making it difficult to uniquely determine the real anomaly triggering source (root cause of the fault).

[0005] Therefore, a computer-implemented data processing method is needed to integrate multi-source heterogeneous state data to enhance anti-interference triggering capability, introduce conduction attenuation constraints in graph search algorithms, and use graph topology-based forward asymmetric verification to eliminate multiple solutions caused by parallel symmetry, thereby realizing dynamic and unique deduction of substation secondary circuit faults. Summary of the Invention

[0006] To address the shortcomings of the prior art, the present invention provides solutions in the following aspects.

[0007] In a first aspect, the present invention provides a dynamic deduction method for electrical faults in the secondary circuit of a substation, which adopts the following technical solution: collecting continuous state parameters and discrete event parameters of the acquisition nodes, aligning the data according to a unified time scale benchmark, using the deviation of continuous parameters and the jump of discrete event parameters as input, and using an adaptive weighting strategy based on data dispersion to obtain a comprehensive deviation. Based on the specified high confidence quantile of the comprehensive deviation historical samples, the activation threshold is adaptively set to filter out the set of abnormal nodes; During reverse search, the exponential decay base is calculated for the directed edge based on its comprehensive edge weight parameter, and the conduction attenuation is obtained by combining the comprehensive deviation of the pointing node. The comprehensive edge weight parameter is assigned a value based on the physical link transmission parameters or virtual node logical mapping parameters that can be obtained in the substation secondary circuit. Edges with conduction attenuation higher than the noise evaluation lower limit dynamically updated by a real-time sliding window are retained. The conduction attenuation of each edge is multiplied sequentially along the effective reverse branch to obtain the path comprehensive conduction coefficient. The tracing endpoint is used as the candidate starting point. Candidate starting points with large path integration transmission coefficients are selected, and forward topology propagation is performed based on the deterministic adjacency matrix that reflects the causal triggering relationship between network nodes to obtain the predicted node set; The asymmetric directional coverage rate is obtained by dividing the number of intersection nodes between the predicted node set and the actual abnormal node set by the total number of nodes in the actual abnormal node set, and used as the overlap coefficient. The path synthesis transmission coefficient is normalized and weighted with the coincidence coefficient to obtain the confidence level. Candidate starting points are arranged in descending order of confidence level. When there are multiple candidate starting points with confidence levels greater than the preset confidence threshold and topological distances between them exceeding the set interval, a multiple fault set containing multiple candidate starting points is obtained. Otherwise, the single candidate starting point with the highest confidence level is obtained as the abnormal triggering source, thus completing the electrical fault inference based on graph topology transmission.

[0008] Furthermore, the data alignment based on a unified timescale benchmark is to align heterogeneous sampled data to a unified timescale based on a time synchronization protocol.

[0009] Furthermore, the setting of the activation threshold includes: constructing a probability distribution model using nonparametric kernel density estimation, and extracting the probability value of the cumulative probability reaching a preset quantile as the activation threshold.

[0010] Furthermore, the comprehensive edge weight parameter is obtained from the hop count of the virtual logical node corresponding to the virtual terminal in the configuration file.

[0011] Furthermore, the step of multiplying the conduction attenuation of each side along the effective reverse branch to obtain the path comprehensive conduction coefficient includes: performing a product operation on the conduction attenuation corresponding to the continuously connected effective sides to reflect the overall signal retention capability of the entire reverse path.

[0012] Furthermore, the elements of the deterministic adjacency matrix are only 0 or 1, and do not contain probability transition semantics; Furthermore, the forward topology transfer is performed in the Boolean algebra domain, using Boolean logic AND-OR rules for iterative diffusion of the state vector.

[0013] Furthermore, the noise assessment lower limit based on real-time sliding window dynamic updates includes: calculating the root mean square value in real time using background environmental noise background data within a preset time window, and converting it into a dynamically updated conduction attenuation lower limit.

[0014] Furthermore, the step of selecting candidate starting points with large path comprehensive transmission coefficients includes: sorting all candidate starting points in descending order of their path comprehensive transmission coefficients, and truncating and retaining the top-ranked preset number of candidate starting points to enter the forward topology transmission stage.

[0015] Furthermore, obtaining a multiple fault set containing multiple candidate starting points includes: outputting multiple candidate starting points with confidence levels greater than a preset confidence threshold and topological distances between them exceeding a set interval in parallel to obtain a multiple fault set containing multiple candidate starting points.

[0016] The embodiments of the present invention have at least the following beneficial effects: 1. By using unified time-scale alignment and non-parametric high quantile threshold to trigger subsequent graph searches, the probability of erroneous results caused by heterogeneous data delays and misalignments and non-normal mixed data distributions is reduced. The size of the input nodes for reverse tracing is strictly constrained to the actual abnormal range, reducing the overhead of meaningless graph branch traversal caused by pseudo-source points.

[0017] 2. By injecting logical attenuation weights reflecting the tightness of coupling between nodes into the reverse directed edge traversal, and obtaining the path comprehensive transmission coefficient through multiplication along continuous effective branches, a reliable basis is provided for different tracing directions in symmetrical parallel topologies, thereby improving the reliability of the topology structure.

[0018] 3. Forward diffusion based on a 0 and 1 deterministic causal adjacency matrix is ​​performed on the reverse candidate anomaly starting point, which is consistent with the deterministic logic of the cascaded control network. Combined with asymmetric directional coverage and multiple fault-tolerant output mechanism, it alleviates the ambiguity caused by topological symmetry and improves robustness under complex working conditions. Attached Figure Description

[0019] Figure 1 The schematic diagram illustrates the steps of a dynamic simulation method for electrical faults in the secondary circuit of a substation according to the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] The following description, in conjunction with the accompanying drawings, details the specific implementation of the present invention using a substation secondary circuit as the target physical network as a specific application scenario.

[0022] S1: Collect the continuous state parameters and discrete event parameters of the acquisition node and then weight and fuse them to obtain the comprehensive deviation.

[0023] In the actual operation of the substation secondary circuit, the measurement and control device and the intelligent waveform recorder continuously collect multi-source heterogeneous time-series status data of each node with a sampling period of seconds or milliseconds. Each frame of data simultaneously contains the generalized measurement value of the continuous parameter channel and the status flag bit of the discrete parameter channel.

[0024] Transient interference spikes generated by high-voltage switch switching and high-frequency mechanical jitter caused by leakage flux of large transformers will inevitably mix background environmental noise into both types of time series data. Relying solely on continuous parameter deviations or discrete parameter jumps may misjudge noise disturbances as signs of abnormal conditions.

[0025] Therefore, this step processes the continuous and discrete parameter channels of the same node together. The system extracts the continuous parameter deviation and the discrete parameter jump indication respectively, and then weights and fuses them to obtain the comprehensive deviation, which serves as the original basis for judging whether the node has entered an abnormal state.

[0026] The data processing procedure for continuous parameter channels is as follows: the system pre-extracts the average value of the safety baseline state quantity from multiple samples collected by the node within a historical stable operating window. and historical statistical standard deviation The historical stable operation window is defined by personnel with relevant tool experience during the fault-free and stable normal operation phase, and the corresponding stable period window time is obtained as the historical stable operation window. It refers to the time interval corresponding to the system in the confirmed fault-free steady-state operation phase.

[0027] When a new sampling frame arrives, read the current continuous parameter value. This value generally refers to the physical state measurement values ​​such as voltage, current or power corresponding to the continuous parameter channel.

[0028] Next, calculate the absolute value of the difference between the current value and the benchmark mean, and divide it by the standard deviation and the small boundary value. The sum of these components yields the basic telemetry variance deviation component. .

[0029]

[0030] In the formula, This represents the degree of deviation of the current sampled frame from the steady-state reference on this continuous parametric channel.

[0031] This represents the value of a continuous parameter at the current moment.

[0032] The average value of the safety baseline state quantity. The historical statistical standard deviation reflects the normal fluctuation range of the continuous parameter channel under stable operating conditions.

[0033] To prevent the calculation from overflowing due to a zero denominator, this value is taken as an empirical value of 0.001, which implementers can adjust according to specific scenarios.

[0034] Normalizing the absolute deviation by standard deviation provides a comparable numerical scale for the deviation components generated by continuous parametric channels with different ranges and fluctuation ranges, which is a well-known technique.

[0035] For discrete parameter channels, the system sets up a high-frequency discrete state change detection window with timestamps to track the flipping events of discrete parameter state flag bits in a time sequence synchronized with the sampling period.

[0036] When a discrete parameter is detected to transition from a normal state to a non-preset state within the window, the Boolean discrete state transition indicator variable... Set to 1; if no state change occurs within the window or the direction of change does not meet the transition criterion, Set to 0.

[0037] This process directly utilizes the timestamps of discrete state changes, eliminating the need for additional detection delays, and thus enabling real-time capture of state abrupt changes in discrete acquisition channels. It is a dimensionless Boolean quantity.

[0038] get and Subsequently, to avoid transient feature misalignment caused by uncertain delays in network transmission between sampled values ​​and GOOSE (Generic Object Oriented Substation Event) messages, the system first aligns heterogeneous data to a unified microsecond-level time scale based on the IEEE 1588 high-precision time synchronization protocol.

[0039] The system then employs an adaptive weighting method based on information entropy to dynamically calculate the weighting factors of continuous parameters. Weighting factors for discrete parameters And calculate the overall deviation of the nodes. .

[0040]

[0041] In the formula, It represents the overall deviation of a node within the current sampling frame, comprehensively reflecting the strong coupling between the abnormal drift of continuous parameters and the sudden change of discrete parameters of the node.

[0042] As a continuous parameter weighting factor, an empirical value of 0.6 can be taken in a specific preferred embodiment to reduce computational overhead; As a discrete parameter weighting factor, in a specific preferred embodiment, an empirical value of 0.4 can be taken, wherein the empirical value... and All of these can be adjusted by the implementer according to the specific implementation scenario.

[0043] The fluctuations caused by background noise interference in continuous parametric channels are usually continuous small-amplitude disturbances, while the jumps caused by state jitter in discrete parametric channels are transient discrete pulses. Giving continuous parametric channels slightly higher weights can preserve more of the ability to reflect trend state anomalies in the fusion results. At the same time, the deterministic displacement information indicated by the jumps of discrete parametric channels can be used to supplement the deviation in real time.

[0044] After completing the above calculations, the comprehensive deviation of each node within each sampling period is obtained. The deviation value serves as a transient state indicator of a node under the fusion of multi-source heterogeneous data, and can be used to determine the adaptive activation threshold based on historical statistical distribution.

[0045] S2: Set an adaptive threshold based on historical high quantiles and filter out abnormal node sets.

[0046] To obtain the comprehensive deviation of each node in each sampling period Subsequently, the system faces the problem of distinguishing between actual anomalies and background environmental noise spurious sources from continuously changing deviation data. This is especially relevant when the target network operates under external transient disturbances (such as high-voltage switch switching). Short-term spikes may occur due to transient disturbance spikes.

[0047] Meanwhile, differences in the accuracy and physical link length of different data acquisition terminals lead to variations in the performance of each node under fault-free steady-state conditions. The normal fluctuation benchmark values ​​are not the same.

[0048] If a fixed threshold is used to uniformly judge all nodes, noise spikes are very likely to be mislabeled as abnormal nodes.

[0049] For each node, an adaptive activation threshold is independently established based on its own historical comprehensive deviation. And based on this, abnormal nodes are filtered in real time.

[0050] During the system commissioning initialization phase, the system checks whether there are any cascading control actions or abnormal alarm signals in the target network. If it is in a stable operating condition with no confirmed faults, the system continuously collects data from that node. Overall deviation per sampling period This constitutes the historical normal sample set. The empirical value of 1000 is used, and the implementer can adjust it according to the duration of steady state on site.

[0051] Based on a historical normal sample set, a probability distribution model of the comprehensive deviation is constructed using nonparametric kernel density estimation. The preset quantile with a cumulative probability of 99.9% is extracted and used as the adaptive activation threshold for that node. In the formula, The activation threshold for the node. This represents the 99.9th percentile of the overall deviation of the historical sample set. It is a preset empirical value that can be adjusted by the implementer according to the specific implementation scenario.

[0052] Considering that the combined deviation after the fusion of continuous and discrete parameters may exhibit a non-normal skewed or multimodal distribution, the normal distribution assumption is abandoned. This quantile threshold is robust to any distribution shape, does not require the data to follow a normal distribution, avoids the failure problem of the traditional 3σ criterion under non-normal mixed data, and isolates most normal random fluctuations below the threshold.

[0053] because It is derived entirely from the node's own historical operating data, and has the characteristic of following the actual operation of the node, thus enabling adaptive adjustment.

[0054] During operation, the current overall deviation of the node is obtained after each sampling cycle. Then, the system performs a comparison.

[0055] like If the node is not found, mark it as an anomaly at the current time and push its node identifier into the set of anomaly nodes. Otherwise, it will be considered a normal fluctuation and will not be processed.

[0056] To avoid false activation caused by isolated single pulses from switching operations, in a preferred embodiment, the node also needs to continuously... All sampling periods satisfy Only then was it finally confirmed that the node was an abnormal node. The empirical value is 3, which can be adjusted by the implementer according to the severity of the electromagnetic environment.

[0057] After completing the above discrimination, the system obtains the set of abnormal nodes selected in this cycle. This set serves as the starting node set to be traced in the reverse search step.

[0058] S3: Calculate the attenuation base based on the comprehensive edge weight parameters and obtain the conduction attenuation by combining the deviation.

[0059] After obtaining the set of abnormal nodes through adaptive threshold filtering, the system enters the reverse graph search algorithm stage based on edge weight constraints.

[0060] At this time, in the global directed graph topology of the target network corresponding to the secondary circuit of the substation, the deviation signal of the abnormal node can be propagated to the potential predecessor node along multiple directed edges. However, not every reverse edge reachable on the topology has the equivalent state data transmission capability in node association.

[0061] The secondary cable itself has a loss distribution parameter that increases with length, and the relay coil impedance connected in series in the circuit will also significantly attenuate the inductive reactance of the high-frequency disturbance component. If only the pure topological adjacency relationship is relied upon for reverse tracing, a large number of edges that cannot effectively conduct abnormal signals electrically will remain in the search space, causing the scale of the reverse search branch to expand exponentially with the topological depth.

[0062] Therefore, during the reverse search process, whenever the predecessor node of the current node is examined along the reverse direction of the directed edge, the system first determines the comprehensive edge weight parameter corresponding to that directed edge. .

[0063] In a hardware-based cabling implementation, this parameter can be determined by the length of the secondary cable. series coil impedance Weighted conversion.

[0064] In a logic topology-based implementation, this parameter can also be obtained from the number of virtual logical node hops corresponding to the virtual terminals in the IEC 61850 configuration file or the logic delay mapping of the IED device.

[0065] Those skilled in the art can flexibly set the integrated edge weight parameters according to the actual available data sources for the secondary loop. When the precise cable laying length or coil transient impedance cannot be obtained on-site, the theoretical laying path length in the secondary circuit design drawings shall be rounded down to the nearest integer. The value of the factory's power frequency AC resistance per unit is taken as the value. The value of .

[0066] However, due to and The dimensions are inconsistent, so the comprehensive edge weight parameters cannot be directly calculated. It is necessary to and Convert to the same dimension, and then calculate. In the formula, This indicates that the current travels along the cable's outgoing and return paths to form a complete circuit. For the resistivity of the cable conductor, Both the conductor cross-sectional area and the cable cross-sectional area can be directly obtained from the cable model parameters; here, we take... and The experience values ​​are respectively and It can be adjusted by the implementer according to the specific implementation scenario.

[0067] Furthermore, parameters The formula for weighted average conversion is: In the formula, This is the length conversion factor, taken as an empirical value. This reflects the degree of signal loss per unit length of cable; The impedance reduction factor is an empirical value. This reflects the degree to which the coil impedance attenuates the signal; and All values ​​are dimensionless, meaning that length and impedance contribute equally to λ, and can be calibrated and adjusted by the implementer based on the cable type and coil parameters on site.

[0068] Based on comprehensive edge weight parameters The system calculates the exponentially decaying base of the directed edge using a natural exponential function. In the formula, This is the exponentially decaying base, ranging from 0 to 1, reflecting the relative transmissibility of the directed edge to abnormal signals. The larger the value of the edge weight parameter, the greater the relative loss of the signal during transmission.

[0069] It should be clearly defined that the exponential decay base is a search ranking weight stored in the edge attributes of the directed graph. It is essentially used for path pruning and ranking in graph search and does not represent the actual decay value of electrical quantities.

[0070] The base of the exponential function causes the attenuation base to decrease smoothly with increasing length or impedance, and when Approaching zero It approaches 1, which aligns with the practical characteristic of electricalally short direct-drive paths having almost no losses.

[0071] After obtaining the exponential decay base of the directed edge, the system reads the overall deviation of the node pointed to by the directed edge. This deviation has already been calculated using weighted fusion in previous steps, and this node is precisely because... It is marked as abnormal if it exceeds its adaptive activation threshold.

[0072] Will and Multiplying them together gives the topological conduction attenuation of the directed edge. :

[0073] In the formula, The topological conduction attenuation represents the estimated effective signal strength that can be retained after an abnormal signal propagates backward from the pointing node along the directed edge to its predecessor node.

[0074] The overall deviation of the pointer node is derived from the weighted fusion calculation results of the previous steps.

[0075] By using multiplication, the degree of abnormal deviation of a node is coupled with the associated decay weight of an edge. Even if the edge weight decay base is high, if the deviation of the node is extremely weak, the decay of the reverse propagation will be correspondingly low, thus avoiding the graph search algorithm from performing ineffective depth tracing for low-intensity noise.

[0076] Thus far, for each directed edge currently examined in the reverse search, the system has calculated its corresponding topological conduction attenuation. It can be used to compare with the lower limit of the background noise baseline to determine whether the directed edge constitutes part of a valid reverse branch.

[0077] S4: Based on the lower limit of dynamic data noise, prune and multiply to calculate the comprehensive path transmission coefficient to determine the candidate starting point.

[0078] After completing the topological conduction attenuation of each reverse directed edge. After calculation, the system obtains a set of signal attenuation values ​​corresponding to each possible tracing direction.

[0079] In the multi-source data acquisition environment of the substation secondary circuit, there is always a background noise caused by high-voltage switch switching and transformer leakage flux.

[0080] This noise floor determines the lowest distinguishable intensity of state data transmitted along the actual physical link, when a directed edge... When the value is too low, the residual intensity of the abnormal signal after propagation in this direction will be completely submerged in the background noise, and it will be impossible to extract abnormal propagation information with data identifiability by continuing to trace back in this direction.

[0081] Therefore, the system introduces a noise evaluation lower bound based on real-time sliding window dynamic updates. The calculated directed edges are then filtered.

[0082] The system uses background noise data within a preset time window to calculate the root mean square value in real time, and converts it into a dynamically updated lower limit for conduction attenuation. .

[0083] This dynamic lower limit can adapt to the time-varying data acquisition environment of the target network, avoiding false alarms or missed alarms caused by fixed thresholds. The system will assign a lower limit to each directed edge. and Compare them one by one.

[0084] like If the strength of the abnormal signal that can be retained on the edge is higher than the lower bound of the noise, it has data distinguishability, and the directed edge is determined to be one of the constituent edges in the effective reverse branch and is retained.

[0085] like If the signal is obscured by noise, it indicates that the signal propagated along the edge is masked by noise. Continuing the reverse search will produce a noise-dominated pseudo-path. The system will prune and discard the edge and the reverse search branches following that direction.

[0086] This pruning operation strictly limits the expansion of the reverse search space to the set of edges whose signal-to-noise ratio satisfies the state data transmission condition, thus eliminating invalid tracing branches introduced by low-quality edges.

[0087] After pruning is completed, the system automatically retrieves the set of abnormal nodes. Starting from each abnormal node, trace backwards along the reverse branches formed by the retained valid edges. For each retained valid edge encountered, the corresponding edge is... Multiply the results into the cumulative path quantity.

[0088] If a reverse branch extending from an abnormal node to a higher-level node... The edges consist of a series of consecutive, connected valid edges, and are numbered from edge 1 to edge 2 in reverse tracing order. If a branch has a single edge, then the path comprehensive transmission coefficient of that branch is calculated. Calculate using the following formula.

[0089]

[0090] In the formula, This is the path integration conduction coefficient of the reverse branch, reflecting the overall signal retention capability of the entire reverse path.

[0091] The first in the reverse branch The topological conduction attenuation degree corresponding to each effective edge.

[0092] This represents the total number of valid edges that are continuously retained on the reverse branch.

[0093] Since conduction attenuation reflects the proportion of a signal remaining after passing through a single edge, according to the law of signal cascading transmission, the preferred accumulation method is multiplication. Multiplication can accurately reflect the retention characteristics of a signal in the process of multi-level transmission, avoiding the unreasonable situation that the longer the path, the larger the value, resulting from arithmetic addition.

[0094] When tracing back along a reverse branch to a certain node, if all incoming edges of that node in the topology no longer satisfy the aforementioned pruning conditions... If the node's in-degree is 0, meaning there is no upstream device in the secondary loop that can inject abnormal signals into the node, then the system determines the node as the tracing endpoint of the current reverse branch.

[0095] The system records this trace endpoint as a candidate anomaly starting point, and simultaneously records the path corresponding to... As the path energy evaluation value of the candidate starting point, reverse tracing initiated by different abnormal nodes may produce different tracing endpoints, or multiple paths may converge at the same endpoint.

[0096] The system collects all unique trace endpoints to form a candidate anomaly origin set. Each candidate starting point is associated with its corresponding Each traceable endpoint was used as a candidate anomaly starting point for this simulation.

[0097] S5: Select candidate starting points with high transmission coefficients and obtain the predicted node set by forward diffusion based on the deterministic adjacency matrix.

[0098] The set of candidate anomaly origins is obtained through reverse search. and their respective path comprehensive transmission coefficients Subsequently, the system faces the problem of distinguishing between the actual abnormal triggering source (root cause of failure) and the symmetric false source caused by the parallel structure from multiple candidate starting points that are possible on multiple topologies.

[0099] Because substation secondary circuits commonly feature multi-input, single-output parallel bus structures, the pure reverse accumulation algorithm, limited by topological symmetry, often results in similar outputs for multiple symmetrical branches. The value makes it impossible to determine the unique fault origin solely based on reverse energy.

[0100] Meanwhile, the reverse search only utilizes the one-way information of the abnormal signal tracing back from the current abnormal node, without examining the actual node range that the candidate starting point may affect along the normal causal propagation direction of the network state data. Therefore, it lacks verification of the forward graph topology propagation capability of the candidate starting point.

[0101] The system first examines the set of candidate anomaly origins. The overall transmission coefficient of each node in the process according to its associated path. Sort in descending order.

[0102] The larger the value, the higher the total effective signal strength retained along the entire reverse branch from the candidate starting point to a certain anomalous node, meaning the stronger the node correlation transitivity between the candidate starting point and the observed anomalous node.

[0103] To avoid computational latency caused by global matrix operations and to meet millisecond-level real-time requirements, only the path integration transmission coefficient is selected. Top The candidate starting point has entered the positive diffusion stage. The empirical value is 5, which can be adjusted by the implementer based on the hardware computing power.

[0104] The system retains the first few results in the sorting list. These candidate starting points constitute a high-energy candidate starting point set. The system then... Each high-energy candidate starting point is independently simulated for forward topological diffusion.

[0105] In the substation secondary circuit of this embodiment, the causal transmission direction of state data is consistent with the normal transmission direction of current and signal from the power supply side to the load side, that is, from the source node to the destination node; therefore, taking the candidate starting point as the diffusion origin, propagating step by step along the positive direction of the directed edge, the node range that the abnormal state may be affected by can be simulated based on the graph topology.

[0106] The forward diffusion process depends on the deterministic causal adjacency matrix of the globally connected topology. ,matrix for Square formation, in which The total number of nodes in the target network.

[0107] Matrix elements The definition is that if there exists a path from node Pointing to node Directed hardware or logic connections that can directly lead to nodes in cascading control logic. If a response is generated, then ,otherwise .

[0108] matrix The elements are only 0 or 1, and do not contain probabilistic semantics, in order to fit the deterministic logic of cascaded control actions.

[0109] When performing forward diffusion, the system constructs a dimensional initial diffusion state vector The element at the index corresponding to the candidate starting point has a value of 1, while the element at all other positions has a value of 0.

[0110] State vector This means that at the initial moment of diffusion, the abnormal states are completely concentrated at the candidate starting point. The iterative propagation of the state vector is carried out in the Boolean algebra field, and the state diffusion is performed using Boolean logic AND-OR rules. Therefore, after... After the second positive pass dimensional state vector In the formula, For the process After the second positive pass dimensional state vector, the first Each element value is the diffusion origin. Step-by-step propagation can reach nodes The cumulative impact state.

[0111] After being passed according to Boolean logic rules The node corresponding to the non-zero element is the first one. The nodes that can be reached by the diffusion step. The nodes corresponding to the zero elements in the non-zero elements are the nodes that can be directly affected by single-step propagation.

[0112] To simulate the propagation of abnormal states along serial branches, the system iteratively executes the above matrix multiplication process. step, The empirical value of 3 is adopted, and the implementer can adjust it according to the maximum series depth of the secondary circuit.

[0113] Finish After each iteration, the system calculates the state vector at each step. The results are combined to obtain the final predicted node set for the candidate starting point. .

[0114] The system traverses all Nodes, if node In the above If the element value in any of the state vectors is greater than 0, then the node is... Included in the prediction node set And also include the candidate starting point itself. .

[0115] This reflects the set of all nodes that may be affected by the abnormal state propagating in the positive direction of the graph topology, assuming that the candidate starting point is the actual source of the abnormality.

[0116] Through the above calculations, the system obtains an independent set of predicted nodes for each high-energy candidate starting point. And the predicted node set is compared with the candidate starting point. Used to match the actual abnormal activation node set Perform overlap calculations.

[0117] S6: Calculate the asymmetric directional coverage of the predicted node set over the abnormal node set as the overlap coefficient.

[0118] When the predicted node set of candidate starting points is obtained Then, the set of actual abnormal nodes previously filtered by adaptive thresholds is taken. .

[0119] For the current candidate starting points to be evaluated, the system compares... and Count the number of common nodes between two sets, and denote it as the intersection node count. .

[0120] Since the size of the node set affected by the forward diffusion prediction is usually much larger than the actual abnormal node set, the traditional crossover ratio index tends to become ineffective.

[0121] Therefore, asymmetric directional coverage is used as the overlap coefficient. The system directly calculates the ratio of the number of intersection nodes to the total number of actual abnormal nodes.

[0122]

[0123] In the formula, To cover the overlap coefficient, the value ranges from 0 to 1.

[0124] This represents the number of nodes in the intersection.

[0125] For abnormal node set The total number of nodes.

[0126] This index accurately reflects the complete explanatory power of candidate sources for observed anomalies, eliminating the ambiguity of multiple solutions caused by parallel symmetric topology.

[0127] When the forward diffusion prediction nodes completely cover the measured abnormal nodes The value is 1, and the value is 0 when there is no intersection.

[0128] For each candidate origin entering the high-energy candidate origin set, the system performs the above set operation to obtain the corresponding coverage overlap coefficient. .

[0129] S7: Normalize the conduction coefficients and calculate the confidence level by weighting them to output single or multiple anomaly trigger sources.

[0130] Based on high-energy candidate starting point set For each candidate starting point, the system has obtained its path synthesis transmission coefficient in the preceding processing. and coverage overlap coefficient .

[0131] because It is obtained by multiplying the conduction attenuation of multiple effective edges. Its absolute value is affected by the path length and attenuation base, and it lacks a closed upper bound.

[0132] and Strictly limited to the range of 0 to 1, this structural difference in the range of values ​​makes it possible to directly... and When adding linearly, Numerical fluctuations may overwhelm The provided topology coverage information makes it impossible for the fusion results to evenly reflect the evaluation contributions of both the reverse node association transitivity and the forward diffusion overlap.

[0133] Therefore, the system is currently All candidate starting points To perform minimum-to-maximum normalization, the system first retrieves all values ​​within the set. Determine the minimum value among the possible values. and maximum value .

[0134] like For the current candidate starting point, its normalized cumulative energy value is calculated using the following formula. ;like This indicates the origin of all candidate origins. Completely identical, at this point directly order .

[0135] In the formula, The normalized energy value is mapped to the 0 to 1 range after this transformation, which is a well-known technical method.

[0136] This is the path integration transmission coefficient of the current candidate starting point. This normalization process does not change the relationship between the candidate starting points. The relative size order is maintained, but the scale difference caused by the absolute span of the values ​​is eliminated.

[0137] After normalization, the system utilizes and Construct a two-way fusion confidence index for the current candidate starting point .

[0138]

[0139] In the formula, It is a two-way fusion confidence index that integrates reverse energy information and forward coverage information.

[0140] The coverage overlap coefficient is derived from the calculation of the asymmetric coverage of two sets.

[0141] As the inverse weighting factor, an empirical value of 0.5 is used; It is a positive weighting factor, with an empirical value of 0.5. Implementers can adjust it according to specific scenarios.

[0142] Through linear weighting, and Each contributes equally to the final confidence level; when both inverse energy and forward coverage are high... When both are low, the value is close to 1. Close to 0.

[0143] right After repeating the above calculation for each candidate starting point, the system obtains a set of confidence indices that correspond one-to-one with each candidate starting point. .

[0144] The system iterates through the set, sorting by confidence level. Before outputting in descending order These are candidate anomaly starting points, which are provided for maintenance personnel to make a comprehensive judgment based on on-site information. Take 3 experience points.

[0145] If there are multiple candidate starting points with confidence levels greater than the preset threshold and topological distances greater than the set value, they are determined to be multiple or chained state anomalies and output in parallel.

[0146] Furthermore, it effectively compensates for the limitations of the single root cause hypothesis in dealing with complex network conditions, and completes the electrical fault deduction based on graph topology transmission.

[0147] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make several improvements and substitutions without departing from the technical principles of the present invention, and these improvements and substitutions should also be considered within the scope of protection of the present invention.

Claims

1. A dynamic simulation method for electrical faults in the secondary circuits of a substation, characterized in that, include: After aligning the continuous state parameters and discrete event parameters of the acquisition nodes according to a unified time scale benchmark, the continuous parameter deviation and discrete event parameter jump are used as inputs, and a weighted fusion is performed using an adaptive weighting strategy based on data dispersion to obtain the comprehensive deviation. Based on the specified high confidence quantile of the comprehensive deviation historical samples, the activation threshold is adaptively set to filter out the set of abnormal nodes; During reverse search, the exponential decay base is calculated for the directed edge based on its comprehensive edge weight parameter, and the conduction attenuation is obtained by combining the comprehensive deviation of the pointing node. The comprehensive edge weight parameter is assigned a value based on the physical link transmission parameters or virtual node logical mapping parameters that can be obtained in the substation secondary circuit. Edges with conduction attenuation higher than the noise evaluation lower limit dynamically updated by a real-time sliding window are retained. The conduction attenuation of each edge is multiplied sequentially along the effective reverse branch to obtain the path comprehensive conduction coefficient. The tracing endpoint is used as the candidate starting point. Candidate starting points with large path integration transmission coefficients are selected, and forward topology propagation is performed based on the deterministic adjacency matrix that reflects the causal triggering relationship between network nodes to obtain the predicted node set; The asymmetric directional coverage rate is obtained by dividing the number of intersection nodes between the predicted node set and the actual abnormal node set by the total number of nodes in the actual abnormal node set, and used as the overlap coefficient. The path synthesis transmission coefficient is normalized and weighted with the coincidence coefficient to obtain the confidence level. Candidate starting points are arranged in descending order of confidence level. When there are multiple candidate starting points with confidence levels greater than the preset confidence threshold and topological distances between them exceeding the set interval, a multiple fault set containing multiple candidate starting points is obtained. Otherwise, the single candidate starting point with the highest confidence level is obtained as the abnormal triggering source, thus completing the electrical fault inference based on graph topology transmission.

2. The dynamic simulation method for electrical faults in the secondary circuit of a substation according to claim 1, characterized in that, The data alignment based on a unified time scale benchmark refers to aligning heterogeneous sampled data to a unified time scale based on a time synchronization protocol.

3. The dynamic deduction method for electrical faults in the secondary circuit of a substation according to claim 1, characterized in that, The setting of the activation threshold includes: constructing a probability distribution model using nonparametric kernel density estimation, and extracting the probability value of the cumulative probability reaching a preset quantile as the activation threshold.

4. The dynamic deduction method for electrical faults in the secondary circuit of a substation according to claim 1, characterized in that, The comprehensive edge weight parameter is obtained from the hop count of the virtual logical node corresponding to the virtual terminal in the configuration file.

5. The dynamic simulation method for electrical faults in the secondary circuit of a substation according to claim 1, characterized in that, The path-comprehensive conduction coefficient is obtained by multiplying the conduction attenuation of each side along the effective reverse branch, including: The conduction attenuation corresponding to the consecutively connected effective edges is multiplied to reflect the overall signal retention capability of the entire reverse path.

6. The dynamic deduction method for electrical faults in the secondary circuit of a substation according to claim 1, characterized in that, The elements of the deterministic adjacency matrix are only 0 or 1, and do not contain probability transition semantics.

7. The dynamic deduction method for electrical faults in the secondary circuit of a substation according to claim 1, characterized in that, The forward topology transfer is performed in the Boolean algebra field, using Boolean logic AND-OR rules for iterative diffusion of the state vector.

8. The dynamic deduction method for electrical faults in the secondary circuit of a substation according to claim 1, characterized in that, The noise assessment lower limit based on real-time sliding window dynamic updates includes: calculating the root mean square value in real time using background environmental noise background data within a preset time window, and converting it into a dynamically updated conduction attenuation lower limit.

9. The dynamic deduction method for electrical faults in the secondary circuit of a substation according to claim 1, characterized in that, The process of selecting candidate starting points with large path comprehensive transmission coefficients includes: sorting all candidate starting points in descending order of their path comprehensive transmission coefficients, and truncating and retaining the top-ranked preset number of candidate starting points to enter the forward topology transmission stage.

10. The dynamic deduction method for electrical faults in the secondary circuit of a substation according to claim 1, characterized in that, The process of obtaining a multiple fault set containing multiple candidate starting points includes: outputting multiple candidate starting points with confidence levels greater than a preset confidence threshold and topological distances between them exceeding a set interval in parallel to obtain a multiple fault set containing multiple candidate starting points.

Images