Fault detection and isolation method for intelligent power distribution system based on edge computing

Abstract

The present application belongs to the technical field of intelligent detection, and particularly relates to a method for fault detection and isolation of an intelligent power distribution system based on edge computing. The method is cooperatively executed by edge nodes and edge aggregation nodes, and comprises the following three steps: Step 1, data construction and graphization preprocessing on the edge node side are executed, wherein edge nodes are arranged at each segmented position of the intelligent power distribution system, and a switch state table and a controllable switch list corresponding to a graph signal frame are formed; Step 2, graph wavelet local anomaly focusing processing, BOCPD online confirmation processing and set member estimation processing are continuously executed on the edge aggregation node side to form a minimum suspicious segment set; and Step 3, isolation and power restoration control based on the minimum suspicious segment set is executed. The present application has the characteristics of high detection accuracy, fast response speed and strong fault tolerance, and can realize millisecond-level autonomous fault isolation and power restoration control under the conditions of limited communication and complex topology, thereby effectively improving the safety, reliability and recovery efficiency of the power distribution system.

Description

Technical Field

[0001] This invention belongs to the field of intelligent sensing system technology, and particularly relates to the field of intelligent detection technology, specifically a fault detection and isolation method for intelligent power distribution systems based on edge computing. Background Technology

[0002] As distribution network structures gradually evolve from unidirectional power supply to multi-source interconnection, distributed power generation penetration, and grid connection of electric loads, traditional distribution automation systems face new challenges in fault detection, section identification, and isolation control. Existing technologies typically rely on centralized monitoring master stations to periodically collect and analyze feeder operation data. For example, some urban distribution network monitoring master stations use methods based on synchronous sampling and waveform phasor analysis to determine fault sections through current mutations or phase angle jumps. However, such centralized methods require significant communication bandwidth and central computing resources, making it difficult to meet the real-time requirements when multiple feeders experience simultaneous disturbances. Furthermore, centralized computing platforms are prone to misjudgments or omissions under conditions of network latency, packet loss, or timestamp drift, leading to delayed issuance of isolation commands and extended power outage times for users.

[0003] Another existing approach introduces a fault location method based on wide-area measurement and distributed terminal collaboration. This method installs smart switches and segmented sampling units at feeder nodes, uploading the sampling results to the main control center. The central system then compares the voltage and current change trends of adjacent measuring points to identify the faulty section. While this approach alleviates the pressure of centralized computing to some extent, its essence remains "centralized reasoning and edge reporting." In the event of communication interruptions or data delays, fault diagnosis is forced to wait for the complete data packet to return to the main control center, resulting in detection and isolation times often exceeding hundreds of milliseconds. This does not meet the requirements of new flexible distribution networks for rapid isolation within tens of milliseconds.

[0004] Other studies have attempted to introduce AI-based fault identification models, such as using convolutional neural networks to extract features from waveform segments to distinguish between short circuits, grounding faults, or overloads. While these methods demonstrate high accuracy in experimental settings, they have significant limitations in practical applications: First, the models require a large number of labeled samples for training, while the number of real fault samples in distribution networks is limited; second, changes in waveform features are closely related to line length and power fluctuations of distributed generation sources, resulting in poor model transferability between different feeders; third, neural network models are sensitive to noise and susceptible to electromagnetic interference and communication jitter, thus reducing the reliability of online applications. Summary of the Invention

[0005] The main objective of this invention is to provide a fault detection and isolation method for intelligent power distribution systems based on edge computing. Through collaboration between edge nodes and edge aggregation nodes, an integrated autonomous process of fault detection, verification, and isolation control is achieved. Its core comprises three key steps: graph wavelet local anomaly focusing processing, BOCPD online verification processing, and set membership estimation processing. The former achieves multi-scale local energy focusing on the feeder topology map to quickly locate candidate anomaly regions; the latter identifies stability mutation points through a continuous sample verification mechanism using a running length table; and the latter calculates the minimum suspicious segment set by combining the propagation sequence and switch boundaries, and uses this to generate an isolation control sequence. This method can complete detection and control decisions at the edge without relying on a central master station, featuring high detection accuracy, fast response speed, and strong fault tolerance. It can achieve millisecond-level autonomous fault isolation and restoration control under conditions of limited communication and complex topology, effectively improving the safety, reliability, and recovery efficiency of the power distribution system.

[0006] To address the aforementioned technical problems, this invention provides a fault detection and isolation method for intelligent power distribution systems based on edge computing. This method is executed collaboratively by edge nodes and edge aggregation nodes, and includes the following three steps:

[0007] Step 1: Perform edge node side data construction and graph preprocessing. Edge nodes are set up at each segment of the intelligent power distribution system. Voltage, current, switch position and time stamp are obtained by the edge nodes. Feeder topology is generated based on the connection relationship between primary equipment and line. Synchronous sampling data within the same trigger event is organized into graph signal frames of uniform length and a switch status table and controllable switch list corresponding to the graph signal frames are formed.

[0008] Step 2: On one side of the edge convergence node, perform graph wavelet local anomaly focusing processing, BOCPD online verification processing and set membership estimation processing continuously to form the minimum suspicious segment set;

[0009] Step 3: Execute isolation and power restoration control based on the minimum suspicious segment set. The edge aggregation node generates an isolation control sequence based on the minimum suspicious segment set formed in Step 2. The isolation control sequence includes the opening and closing sequence of the controllable switches at both ends of each member in the minimum suspicious segment set, the activation sequence of the adjacent tie switches, and the coordination instructions with the distributed power sources involved. The edge aggregation node distributes the isolation control sequence to the relevant edge nodes for execution. The edge nodes complete the disconnection of the faulty segment, the reconstruction and power restoration of the healthy segment, and the status verification. The execution results are archived for reuse in subsequent similar events.

[0010] Furthermore, step two specifically includes: performing graph wavelet local anomaly focusing processing; performing BOCPD online verification processing; and performing set membership estimation processing to form a minimum suspicious segment set.

[0011] Furthermore, the graph wavelet local anomaly focusing processing specifically includes: generating three types of fixed neighborhoods—a first ring neighborhood, a second ring neighborhood, and a third ring neighborhood—centered on each node in the feeder topology; calculating the absolute difference between the sampled value of the center node and the arithmetic mean of each of the three types of fixed neighborhoods at each sampling time of the graph signal frame, denoted as the three types of local contrast values; applying the median and median absolute deviation methods to the three types of local contrast values ​​to obtain three types of thresholds; at any sampling time, if any node has a local contrast value higher than the corresponding threshold, this node is marked as a local anomaly node; expanding sequentially from the local anomaly node to the adjacent nodes on both sides along the connectivity of the feeder topology, retaining the adjacent node when its local contrast value is continuously higher than its corresponding threshold during the expansion process, stopping when the first boundary below the threshold appears, forming a candidate anomaly region composed of several adjacent edges; repeating the above process for all sampling times, merging candidate anomaly regions at the same position and recording the start and end sampling positions to obtain a set of candidate anomaly regions.

[0012] Furthermore, the BOCPD online verification process in step 2 specifically includes: taking the candidate abnormal region set as input, establishing a run length table according to the sampling order, with each item in the run length table recording the continuous sample count since the previous verification position; at each sampling time, for each candidate abnormal region appearing in the candidate abnormal region set, calculating the occurrence count of the corresponding local abnormal node in the candidate abnormal region at the current sampling time, and adding the occurrence count to the current item in the run length table.

[0013] Furthermore, when the occurrence count of the current item in the running length table reaches the preset upper limit, the current sampling time is marked as the change point position, and two time segments before and after the change point are segmented accordingly. A review is performed on a number of consecutive sampling times after the change point. The review rule is as follows: if the occurrence count of the candidate abnormal region continuously reaches or exceeds the preset lower limit and the running length table continuously increases within a number of consecutive sampling times, the continuous sample interval after the change point until the end of the review is marked as a confirmation segment. When the running length table is initialized after the review is completed, the next round of online confirmation begins. Within the same trigger event, one or more confirmation segments are obtained sequentially, and a one-to-one correspondence is established between each confirmation segment and the candidate abnormal region it covers.

[0014] Furthermore, the ensemble membership estimation process includes step A: propagation sequence generation: local anomalous nodes within the confirmed segment are arranged from early to late according to time stamps, and combined with the connectivity direction of the feeder topology, a propagation sequence from the power supply side to the load side is obtained.

[0015] Furthermore, the set membership estimation process also includes: Step B, temporary segment construction: starting from the first local anomaly node in the propagation sequence, a temporary segment is formed by merging into the first controllable switch along the connectivity direction; at the same time, another temporary segment is formed by merging into the adjacent controllable switch from the end local anomaly node in the propagation sequence along the connectivity direction. Step 3.3, boundary alignment: aligning the endpoints of the two temporary segments to the corresponding controllable switch positions respectively, resulting in two temporary segments with aligned boundaries.

[0016] Furthermore, the set membership estimation process also includes: Step C, overlap contraction: find the intersection of the temporary segments whose boundaries have been aligned. If the intersection is still a continuous edge set, replace the original temporary segment with the intersection. When the intersection is non-empty and there is a non-continuous edge set, split the non-continuous edge set into several continuous sub-segments according to connectivity and use them as new temporary segments respectively.

[0017] Furthermore, the set member estimation process also includes: Step D, Coverage Verification: Determine whether the new set of temporary segments covers all local abnormal nodes within the confirmed segment. If there are uncovered nodes, extend a temporary segment adjacent to the uncovered node along the connectivity direction to the next controllable switch and return to the boundary alignment step until coverage is completed.

[0018] Furthermore, the set membership estimation process also includes: Step E, Redundancy Removal: After coverage is completed, all temporary segments are checked in ascending order of segment length. If a temporary segment is completely contained by another temporary segment, the contained temporary segment is deleted; Step F, Connectivity Final Check: The connectivity of the remaining temporary segments is verified one by one. The verification rule is that each edge within the segment forms a single path on the feeder topology and both ends are bounded by controllable switches; All temporary segments obtained after the redundancy removal step and the connectivity final check step are the minimum suspicious segment set members corresponding to a confirmed segment; When there are multiple confirmed segments, the union of the set members obtained from each confirmed segment is taken, and the redundancy removal step and the connectivity final check step are executed again to obtain the final minimum suspicious segment set.

[0019] The edge computing-based intelligent power distribution system fault detection and isolation method of the present invention has the following beneficial effects: The present invention, through the collaboration of edge nodes and edge aggregation nodes, moves fault detection, verification, and control to the field, forming a closed-loop local autonomous process. The graph wavelet local anomaly focusing processing, oriented towards the spatial correlation of the feeder topology, performs multi-scale comparison of multi-node synchronous sampling for the same triggering event, rapidly amplifying the local changes caused by the fault in the graph structure, while suppressing scattered disturbances caused by random noise or load fluctuations, thus obtaining candidate anomaly regions faster and more stably. Subsequently, the BOCPD online verification processing, based on continuous samples during operation, performs change point verification and continuous review of candidate anomaly regions, effectively distinguishing between transient spikes and stable instability, avoiding false triggers and missed triggers. Afterwards, the ensemble membership estimation processing performs overlap shrinkage, coverage verification, and redundancy removal based on the propagation order and controllable switch boundaries, directly outputting the minimum suspicious segment set consistent with the executable boundary, making the detection results naturally correspond to the operational granularity of switching and reconnection. The isolation control sequence generated based on the minimum suspicious segment set unifies the disconnection sequence, the activation sequence of tie switches, and the coordination commands of distributed power sources into a single timing sequence. It also uses state verification and a stable window to drive gradual power restoration, reducing excessive isolation and repeated rollbacks. Compared to methods relying on centralized computation or single-point thresholds, this invention maintains topology consistency and selective action even in scenarios with limited communication, topology changes, and high distributed power source penetration. It shortens the links where isolation is detected, reduces bandwidth and transmission pressure, and improves the availability and recovery speed of intact segments. Simultaneously, the unified data structure and archived records enhance traceability and reuse efficiency, facilitating rapid migration and expansion between different feeders. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of the method flow for fault detection and isolation of an intelligent power distribution system based on edge computing, provided in an embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram of three-phase voltage sampling data of a graphical signal frame provided in an embodiment of the present invention;

[0023] Figure 3 This is a schematic diagram of the local abnormal node detection results provided in an embodiment of the present invention;

[0024] Figure 4This is a schematic diagram of the running length curve for BOCPD change point detection provided in an embodiment of the present invention;

[0025] Figure 5 This is a schematic diagram of the current change curves before and after fault isolation provided in an embodiment of the present invention;

[0026] Figure 6 This is a schematic diagram illustrating the principle of estimating the minimum suspicious segment set provided in an embodiment of the present invention. Detailed Implementation

[0027] The method of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0028] Example 1: Reference Figure 1 A fault detection and isolation method for intelligent power distribution systems based on edge computing is proposed. This method is executed collaboratively by edge nodes and edge aggregation nodes and includes the following three steps:

[0029] Step 1: Perform edge node side data construction and graph preprocessing. Edge nodes are set up at each segment of the intelligent power distribution system. Voltage, current, switch position and time stamp are obtained by the edge nodes. Feeder topology is generated based on the connection relationship between primary equipment and line. Synchronous sampling data within the same trigger event is organized into graph signal frames of uniform length and a switch status table and controllable switch list corresponding to the graph signal frames are formed.

[0030] Step 2: On one side of the edge convergence node, perform graph wavelet local anomaly focusing processing, BOCPD online verification processing and set membership estimation processing continuously to form the minimum suspicious segment set;

[0031] Step 3: Execute isolation and power restoration control based on the minimum suspicious segment set. The edge aggregation node generates an isolation control sequence based on the minimum suspicious segment set formed in Step 2. The isolation control sequence includes the opening and closing sequence of the controllable switches at both ends of each member in the minimum suspicious segment set, the activation sequence of the adjacent tie switches, and the coordination instructions with the distributed power sources involved. The edge aggregation node distributes the isolation control sequence to the relevant edge nodes for execution. The edge nodes complete the disconnection of the faulty segment, the reconstruction and power restoration of the healthy segment, and the status verification. The execution results are archived for reuse in subsequent similar events.

[0032] In one specific implementation of step one, using a hybrid overhead and cable feeder as the object, the method includes: setting up edge nodes at each segment of the intelligent power distribution system. Each edge node directly connects to the voltage and current sampling circuit and switch position acquisition channel of the corresponding segment and acquires a time stamp. To avoid misalignment caused by cross-site drift, each edge node is clock-aligned using a precise time protocol and uses the Global Positioning System (GPS) for timing redundancy. Each edge node performs synchronous sampling using a unified basic sampling arrangement. For example, each edge node continuously samples three-phase voltage and current at a rate of 5000 samples per second; when the operating system reports a trigger event, each edge node maintains synchronous sampling and records a time stamp within the time range covered by the trigger event. The reason for using a unified basic sampling arrangement is that the sampling amounts of multiple segments within the same time range need to be merged into the same signal frame. Only when the sampling cycle is consistent can the timing correspondence of each sampling point be single-valued and require no additional interpolation. To reduce the instantaneous impact of electromagnetic interference on site, each sampling point undergoes a fixed-length median calculation before being written into the temporary storage area of ​​the edge node. A fixed length of 5 sampling points is used, and the median value of 5 consecutive sampling points is taken as the representative value of the current sampling point. The reason for adopting this method is that median operation has a natural ability to suppress impulse interference and isolated spikes, does not destroy the step and abrupt edges of the waveform, and does not cause trailing in subsequent candidate event boundary identification.

[0033] A feeder topology map is generated based on the connection relationships between primary equipment and lines. Specifically, following the power supply direction from the power source to the load side, the connection information of the station bus, outgoing circuit breaker, line section switch, tie switch, and terminal load is read sequentially. Each physical location with independent sampling or status is established as a node, and the conductor or cable between two locations is established as an edge. A unique identifier is assigned to each node when generating the feeder topology map. For example, a three-segment naming convention of "station code – feeder number – section sequence number" is used, such as "AA01 – F12 – S03". This unique identifier is directly used for node indexing in the graph signal frame, avoiding ambiguity during subsequent multi-source data merging. For scenarios with ring network structures, a feeder topology map is still generated, but the positions of the controllable switches at both ends of each edge are recorded. This is done because it is necessary to derive the minimum suspicious segment set without changing the topology connectivity. Defining the physical boundaries of the edges allows edge convergence nodes to maintain consistent mapping between nodes and edges under different switching combinations. Each node is bound to the physical installation location of the edge node. Binding is accomplished by scanning a QR code or near-field identification tag on the edge node's shell, and the identification result and unique identifier are written to persistent storage inside the edge node. This method avoids manual secondary matching when replacing edge nodes during operation and maintenance, reducing the impact on time-stamped links.

[0034] The triggering event is determined on-site by the operating system. Upon receiving the start and end timestamps of the triggering event, the edge nodes retrieve synchronous sampling data and switch position change records within the corresponding time range from their local temporary storage. The reason for using event time range archiving instead of single-point triggering is that subsequent graph wavelet local anomaly focusing and online verification processing require observing background and change segments before and after the event; range archiving avoids missing the event's leading edge and decay segment. When the start timestamp of the triggering event is earlier than the buffer start time of any edge node, the edge node performs forward copy padding for the missing segment, that is, it fills in the missing segment with the first available set of valid sample data up to the start timestamp of the triggering event. Using forward copy padding instead of zero padding avoids introducing physically meaningless all-zero segments into the graph signal frame, thereby reducing misjudgments in subsequent processing. When the end timestamp of the triggering event is later than the buffer end time of any edge node, the edge node performs backward copy padding for the missing segment, that is, it fills in the missing segment with the last available set of valid sample data up to the end timestamp of the triggering event. This approach ensures consistency in frame length across different nodes and maintains stable alignment with event end boundaries.

[0035] Synchronous sampling data within the same trigger event is organized into graph signal frames of uniform length according to the node order of the feeder topology. The uniform length is determined by using the start time marker of the trigger event as the frame start point and the end time marker of the trigger event as the frame end point, calculating the number of sampling points between the two as the length of this graph signal frame. Using a uniform length ensures that subsequent sampling-time-by-sampling processing on the graph structure has a fixed number of iterations, facilitating real-time execution across devices. The graph signal frame is stored in a row-major format, with node order first and time order second. Each row corresponds to a node in the feeder topology, and each column corresponds to a sampling time within the trigger event. The reason for choosing row-major over column-major is that actual edge convergence nodes are often processed in parallel at the node level. Row-major allows continuous sampling quantities of the same node to maintain physical continuity in storage, thereby reducing random access overhead during retrieval. For multi-channel sampling of voltage and current, multiple graph signal frame groups are constructed using a channel-parallel approach. For example, three-phase voltage is grouped into one group, three-phase current into another, and a separate group is established for zero-sequence quantities. The reason for dividing different physical quantities into multiple groups is to ensure that the numerical range and statistical characteristics within each group are relatively consistent, which facilitates subsequent local comparisons within the same threshold system.

[0036] To ensure strict alignment of time stamps between different edge nodes, the relative error is calculated based on the time stamp attached to each sampling point before writing the graph signal frame. When the relative error of any edge node exceeds 50 microseconds, a linear interpolation alignment is performed. The interpolation is only performed between two adjacent sampling points and does not cross the boundary of switch position changes. Limiting the interpolation range is to avoid smoothing out the step transition caused by switch position changes and to maintain the edge nature of the event. For scenarios with slight amplitude drift, node-by-node detrending processing is performed without changing the event abrupt change. The method is to use the average value of several consecutive sampling points before the start of the trigger event as a baseline and subtract this baseline from the entire frame data. This processing can cancel out the differences in transformer bias and voltage level at different installation points, making it easier for the subsequent comparison-based focusing process to locate the relative change caused by the anomaly. The graph signal frame is bound to the metadata of the trigger event, which includes the trigger event number, frame length, frame start and end time stamps, number of participating nodes, and a list of unique node identifiers. The purpose of binding metadata is to support subsequent online verification processing for cross-frame cross-validation and to backtrack node context in set membership estimation processing.

[0037] Within the time frame of the triggered event, the position and time stamp of each perceptible switch are recorded, forming a switch position change record. These records are then compiled into a switch status table. The switch status table is arranged in chronological order, and each record includes a unique switch identifier, position status, time stamp, position source, and quality flag. The position source is used to distinguish whether the position is acquired from local contact data or from information from a distant upstream source, and the quality flag indicates whether short-term jitter has occurred. The purpose of setting the quality flag is to avoid mistaking jitter for actual position changes during subsequent framing and positioning processes. The controllable switch list lists all switches that can be issued control commands in the order of the feeder topology diagram. Each entry includes a unique switch identifier, switching direction, mapping to adjacent nodes, minimum allowable interval time, and the identifier of the adjacent tie switch. The reason for recording the switching direction and the mapping to adjacent nodes is to allow the subsequent isolation control sequence generation to directly locate the boundary on the graph, avoiding additional topology queries during the real-time phase. Edge nodes associate the switch status table and the controllable switch list with the graph signal frame, using version numbers of both in the metadata of the graph signal frame as the association method. Using version number association ensures that when maintenance personnel update switch names or wiring relationships on-site, the reproducibility of historical frames is not affected.

[0038] When a sampling loss occurs at an edge node within the trigger event range, the edge node marks the lost segment as an unusable segment and temporarily replaces it with the nearest available sampling point in the graph signal frame. Simultaneously, the start and end timestamps and replacement method of the unusable segment are recorded in the metadata. This is to maintain frame integrity, allowing subsequent processing to proceed first, and to reduce the confidence level of relevant conclusions based on metadata when necessary. When the switch status table and the graph signal frame show inconsistencies at the same time mark, the record obtained from the local contact point at the switch location source takes priority, and a "state conflict" is recorded in the metadata. Prioritizing local contact point acquisition is because this source has the most direct coupling with the physical contact, with less latency than distant information links, providing a more reliable boundary reference for subsequent candidate anomaly area boundary judgment.

[0039] In one optional implementation of step one, in a ring network power supply scenario with multiple power supply directions, the generation of the feeder topology graph, in addition to recording the connection relationships between nodes and edges, also records the switchable attributes of two power supply directions for each edge. When organizing the graph signal frames, they are still sorted by unique identifiers, and the index is not changed due to changes in power supply direction. This ensures that after a tie switch is activated in the operating system, the graph signal frames of two triggering events remain directly comparable. In this scenario, a "Direction Status" column is added to the switch status table. In events where a direction change occurs, the metadata of the graph signal frame includes a timestamp indicating the direction change. The significance of explicitly recording the direction status is that it allows subsequent set member estimation processing to identify the propagation sequence of the same edge in different directions.

[0040] In another optional implementation of step one, a fully cabled urban feeder is used as the object. In fully cabled feeders, the capacitance component is relatively large, resulting in a smoother rising edge of the trigger event. To avoid misinterpreting slow changes as missing data, the pre-buffer of the trigger event needs to be extended before framing. For example, 100 sampling points can be added before the start time marker of the trigger event as a preparatory segment. The introduction of the preparatory segment allows the complete preceding background to be seen in subsequent wavelet local anomaly focusing processing, thus more accurately defining the leading edge of the candidate anomaly region. Since there may be multiple reflections at cable joints, an endpoint protection process is added to the edge nodes after detrending processing. The method is as follows: identify the typical time window corresponding to the joint position on each node, and within this time window, do not perform linear interpolation and median calculations, but only make direct recordings. The purpose is to preserve subtle abrupt changes at the joints, allowing them to participate in local comparison in subsequent processing with their true form.

[0041] In one specific implementation of step two, a hybrid overhead and cable feeder is used as the object. First, based on the feeder topology, three types of fixed neighborhoods—a first ring neighborhood, a second ring neighborhood, and a third ring neighborhood—are generated for each node. Each type of fixed neighborhood is determined by its connectivity depth. For example, the first ring neighborhood contains the set of adjacent nodes directly connected to the node, the second ring neighborhood contains the set of adjacent nodes that cross one edge along the connectivity direction, and the third ring neighborhood contains the set of adjacent nodes that cross two edges along the connectivity direction. To ensure the stability of subsequent processing, the three types of fixed neighborhoods remain unchanged throughout the entire time range of the triggering event, and an index table is established using the node's unique identifier and the neighborhood sequence number. For each sampling moment in the graph signal frame, the nodes are processed sequentially: first, the sampled value of the node is taken; then, the arithmetic mean of the node in each of the three types of fixed neighborhoods is calculated; finally, the absolute value of the difference between the node's sampled value and the arithmetic mean of the three types is calculated to obtain the three types of local comparison values. The reason for using the absolute difference with the average value of the multi-ring neighborhood instead of the single-point difference is that the energy caused by the fault usually spreads from the node to the neighboring segment. Using the average value can reduce the influence of occasional noise at individual measurement points. At the same time, the three types of fixed neighborhoods form three observation scales from near to far, which can distinguish between local mutations and global slow fluctuations.

[0042] For each sampling time, a complete set of the three types of local contrast quantities is established, and the median and median absolute deviation of each are calculated. The median plus three times the median absolute deviation is used as the corresponding threshold for that sampling time. The combination of median and median absolute deviation is chosen because this combination is insensitive to a small number of extreme values ​​and can maintain the stability of the threshold in the presence of switching operations or pulse interference. The three types of local contrast quantities are compared with their respective thresholds. If any type of local contrast quantity is higher than the corresponding threshold, the node is marked as a local anomalous node at that sampling time. To avoid scattered single-point flicker causing false boundaries for subsequent expansion, the node is only confirmed as a local anomalous node at that sampling time if it is marked at least twice within three adjacent sampling times. For each sampling time, starting from each local anomalous node, the expansion proceeds sequentially towards both the power supply side and the load side along the connectivity of the feeder topology. The expansion rule is: if an adjacent node is still a local anomalous node at the current sampling time, it continues to be included, and the expansion stops at the first non-local anomalous node encountered. The set of continuous edges between the two stopping positions constitutes the candidate anomalous region for that sampling time. Candidate anomaly regions obtained at the same location at different sampling times are merged, and the start and end sampling times of each candidate anomaly region are recorded to form a set of candidate anomaly regions. The effect of this process is that it approximates the local energy response of graph wavelets by means of node differences, and transforms discrete node determinations into a continuous set of edges consistent with physical segments through neighborhood expansion, which facilitates subsequent confirmation in the time dimension.

[0043] The multiples for the three threshold types can be set to 3. The self-consistency constraint for local anomaly nodes can be set to at least two calibrations within three adjacent sampling times. The time interval tolerance for merging candidate anomaly regions can be set to no more than five sampling times. The reason for using the above values ​​is that, under the condition of 5000 samples per second, the time covered by three adjacent sampling times is less than 1 millisecond, which can filter out single-point jitter without sacrificing sensitivity; the time interval tolerance of five sampling times can accommodate extremely short sampling jitter without missegmenting two actually continuous regions.

[0044] Using a set of candidate anomaly regions as input, a run length table is built according to the sampling order. Each entry in the run length table records the continuous sample count since the last confirmed location. At each sampling time, for each candidate anomaly region appearing in the candidate anomaly region set, the number of nodes within that candidate anomaly region during that sampling time is counted, yielding an occurrence count, which is then added to the current entry in the run length table. The core consideration for this approach is that segmental anomalies caused by faults will manifest as a continuous accumulation of occurrence counts over time, while transient noise is mostly scattered and difficult to sustain. The run length table can transform this difference into a directly comparable increase in counts.

[0045] When the occurrence count of the current item in the running length table reaches the preset upper limit, the current sampling time is marked as the change point, and two time segments before and after the change point are segmented accordingly. Subsequently, a review is performed on a series of consecutive sampling times after the change point. The review rule is: if the occurrence count of the candidate abnormal region continuously reaches or exceeds the preset lower limit within these consecutive sampling times, and the running length table continues to increase continuously, the change point is deemed valid, and the continuous sample interval after the change point until the end of the review is marked as the confirmation segment. This two-stage strategy of triggering first and then reviewing separates instantaneous spikes from stable changes: the upper limit is responsible for triggering, and the review is responsible for confirming persistence, thus avoiding single-point false triggering.

[0046] Within the same triggering event, if multiple candidate anomaly regions pass the review, multiple confirmed segments are obtained sequentially. For each confirmed segment, the spatial and temporal ranges of the candidate anomaly regions it covers are recorded, establishing a one-to-one correspondence. Multiple confirmed segments may overlap temporally or be spatially adjacent, but the one-to-one correspondence ensures that subsequent ensemble membership estimation can proceed independently per segment, avoiding cross-segment interference.

[0047] The preset upper limit can be set to 12 occurrence counts, the number of consecutive sampling times during verification can be set to 24 sampling times, and the preset lower limit can be set to 18 occurrence counts. This combination of values ​​reflects the idea of ​​"trigger threshold lower than verification threshold": the change point location is identified as early as possible, and then verification is completed using stricter persistence conditions, thus balancing sensitivity and stability. At a sampling rate of 5000 times per second, 24 sampling times correspond to approximately 4.8 milliseconds, which is sufficient to cover the duration of common short-circuit or ground transients.

[0048] Locally anomalous nodes within the confirmed segment are arranged from earliest to latest according to their time stamps. Combined with the connectivity direction of the feeder topology, a propagation sequence from the power supply side to the load side is obtained. When multiple locally anomalous nodes appear at the same sampling time, the lexicographical order of the unique node identifiers is used as the parallel order to ensure the determinism of the propagation sequence. The reason for establishing a propagation sequence is that segmental faults exhibit a sequential response from the power supply side to the load side in the topology. By reconstructing the event progression order according to this sequence, the earliest and latest affected locations can be directly located during subsequent segment construction. Starting from the first locally anomalous node in the propagation sequence, a temporary segment is formed by merging along the connectivity direction to the first controllable switch; simultaneously, another temporary segment is formed by merging in the opposite direction from the last locally anomalous node in the propagation sequence to the adjacent controllable switch. This step transforms the node set into two continuous edge sets with the controllable switches as boundaries, allowing subsequent boundary operations to directly apply to the physical boundaries consistent with the execution control.

[0049] Align the endpoints of the two temporary segments to their corresponding controllable switch positions, including alignment to the power supply or load side endpoints of the switch, to maintain consistency with the actual fault boundary. Boundary alignment eliminates minor differences at the node level, ensuring that the final output edge set can be directly referenced by isolation and restoration control. Find the intersection of the two boundary-aligned temporary segments: if the intersection is a continuous edge set, replace the original temporary segment with the intersection to obtain a shorter edge set that more closely approximates the actual fault impact range; if the intersection is non-empty and contains a discontinuous edge set, split the discontinuous edge set into several continuous sub-segments based on connectivity, and treat each as a new temporary segment for subsequent processing; if the intersection is empty, retain both temporary segments and proceed to coverage verification. Overlap contraction is equivalent to approximating the impact areas derived from the early and late ends until only the overlapping core consistently identified at both ends is retained.

[0050] The process involves determining whether the new set of temporary segments covers all local anomaly nodes within the confirmed segment. If any local anomaly nodes remain uncovered, a temporary segment adjacent to the uncovered node is extended along the connectivity direction to the next controllable switch position, and boundary alignment and overlap contraction are performed again. Coverage verification reflects the principle of "constraints based on actual node measurements and means based on boundary approximation": a temporary segment is considered sufficient only when all local anomaly nodes are covered. Extensions are performed in steps using controllable switches, ensuring that each adjustment corresponds to a realistically operable segment granularity. After coverage is achieved, all temporary segments are checked from shortest to longest. If a temporary segment is completely contained by another, the contained segment is deleted. The remaining temporary segments are then verified for connectivity, with the verification rule being that each edge within the segment forms a single path on the feeder topology and both ends are bounded by controllable switches. The complete set of temporary segments obtained after redundancy removal and final connectivity checks constitutes the minimum suspicious segment set for the confirmed segment. If there are multiple confirmed segments, the union of the members of the set obtained from each confirmed segment is taken, and redundancy removal and connectivity final check are performed again to obtain the final minimum set of suspicious segments.

[0051] If the length of the temporary section after overlapping and contraction is less than two sides, it can directly proceed to the final connectivity check. The single extension step in the coverage check is one controllable switch position, and the maximum number of extensions can be set to three. Temporary sections that fail the final connectivity check are discarded and returned to the coverage check. The reason for limiting the maximum number of extensions to three is that in common distribution feeders with sectionalizing switch spacing of 500 to 1500 meters, extending three times to each side covers the area of ​​three to nine sides, which is usually sufficient to encompass the affected area of ​​a sectional fault.

[0052] In one optional implementation of step two, using a fully cabled urban feeder as an example, and considering the smoother cable leading edge and more pronounced reflections, the self-consistency constraint for local anomaly nodes in the graph wavelet local anomaly focusing processing is adjusted from at least two calibrations within three adjacent sampling times to at least three calibrations within five adjacent sampling times. This adjustment adds approximately 0.4 milliseconds of persistence judgment under a sampling rate of 5000 per second, ensuring that slow steps caused by capacitance effects are identified as stable changes rather than single-point noise. In the BOCPD online verification processing, the preset upper limit is set to 15 occurrence counts, the number of consecutive sampling times for verification is set to 30 sampling times, and the preset lower limit is set to 20 occurrence counts. This allows the verification segment to be more biased towards the stable range, reducing short-period fluctuations triggered by joint reflections. In the ensemble membership estimation processing, an "adjacent merging" strategy is added to the overlap contraction stage: when two continuous segments are separated by only one edge and both ends of that edge are controllable switch positions, that edge is directly merged to obtain a single continuous segment. This strategy is applicable to the "short-interval break points" commonly found in urban cables, which can reduce the number of subsequent isolation actions.

[0053] In one specific implementation of step three, a hybrid overhead and cable feeder is used as the object. After completing step two, the edge aggregation node obtains a minimum suspicious segment set. Each member in the minimum suspicious segment set is bounded by controllable switches at both ends. The edge aggregation node first reads the controllable switches at both ends of each member in a unique identifier order and establishes a "member and controllable switch list". When the minimum suspicious segment set contains multiple members, the edge aggregation node determines the relationship between members based on the feeder topology, classifying them into three categories: mutually independent, mutually inclusive, and sharing a boundary. Mutually independent members are grouped into the same batch, mutually inclusive members are executed first by the outermost member, and members sharing a boundary are merged into a single extended member and the controllable switch list at both ends is updated. The goal of this classification is to reduce repeated disconnection actions and keep the subsequent power restoration path singular.

[0054] For each member within the minimum suspected segment set, the edge convergence node determines the opening sequence of the controllable switches at both ends in the order of "power supply side first, load side second". The power supply side-first approach cuts off energy input from the power supply side in the earliest action, reducing the risk of arcing in subsequent actions. The opening interval of the controllable switches at both ends is set to 80 to 120 milliseconds. This interval covers the detection time required for a single opening feedback and current decay, while ensuring the overall isolation time is in the sub-second range. The edge convergence node adds a "pre-confirmation condition" to the second action in the opening sequence, including the status of the controllable switch on the power supply side being in place, the current detection value of the fault segment being less than 20 amps, and the voltage detection value of the fault segment being less than 5% of the rated value. This condition is added to ensure that the first action has already established a valid break point before opening the other end, thus avoiding repeated operations while the circuit is energized. The edge convergence node performs a shortest path search on both sides of the healthy segments of each member to find the tie switch closest to that healthy segment. If multiple available tie switches exist, the tie switch with the larger expected load recovery capacity is prioritized; if the capacities are the same, the tie switch with the shorter distance is prioritized. The activation order follows a "nearest to farthest" strategy: first activate the nearest tie switch, then activate the remaining unpowered sections in order of distance from nearest to farthest. This strategy maximizes the number of loads restored with the same number of actions. Before each tie switch is activated, the edge aggregation node requires the relevant edge nodes to complete three checks within a 40-millisecond window: the healthy section voltage is within 95% to 105% of the rated value, the healthy section frequency is within 49.5 Hz to 50.5 Hz, and the expected power flow direction is consistent with the power supply direction after activation. Only after passing the checks is the tie switch added to the isolation control sequence.

[0055] For distributed generation (DG) units falling within the minimum suspicious segment set, the edge aggregation node adds two coordination instructions before the disconnection sequence: first, it issues a DG disconnection instruction, followed by an instruction to reduce power to zero, requiring a status confirmation within 200 milliseconds. For DG units located in healthy segments, the edge aggregation node adds a "power smoothing boost" instruction before the tie switch is engaged, gradually increasing the rated power by 10% every 100 milliseconds to 50% of the target power, and then increasing it to the target power after the voltage and frequency have stabilized for 200 milliseconds. The "disconnect first, then disconnect" arrangement avoids backfeeding affecting the insulation recovery of the disconnecting contacts; the "power smoothing boost" arrangement significantly reduces inrush current at the moment of engagement.

[0056] After the isolation control sequence is formed, the edge aggregation node adds a time stamp and timeout setting to each action in the sequence. The time stamp is used to establish a unified rhythm among edge nodes, and the timeout setting is used to trigger status verification in case of on-site response delays. A commonly used single-action timeout setting is 300 milliseconds. The edge aggregation node broadcasts the isolation control sequence to the relevant edge nodes, and simultaneously sends a sequence summary for reconciliation. Upon receiving the instruction, the edge nodes execute the actions according to the time stamps and immediately send back the status, measurement value, and time stamp after each action. Using broadcast and summary reconciliation allows multiple edge nodes to obtain a consistent action list at the same time and corrects for out-of-order status feedback.

[0057] The process of disconnecting the faulty section, reconstructing and restoring power to the healthy section, and verifying the status includes: the controllable switches at both ends are executed sequentially according to the opening and closing order. After the first end is disconnected and confirmed by the aforementioned three steps, the second end is disconnected. Within 100 milliseconds after the disconnection at both ends is completed, the edge node performs a centralized recording of the voltage and current within the smallest suspicious segment set members and transmits the recording results back along with the status. This centralized recording is used to provide objective evidence of "residual energy dissipation after isolation". The edge convergence node is activated one by one according to the activation order of the tie switches. After each activation, a 200-millisecond stability window is observed. If the voltage, frequency, and power flow direction remain within the aforementioned verification range, the next activation is initiated; if an out-of-bounds condition occurs within the stability window, the tie switch is immediately disconnected and the process proceeds to the next selectable tie switch. Using the stability window as the timer for power restoration can distinguish between transient fluctuations and persistent anomalies, ensuring controllability during gradual power restoration. After all isolation control sequences have been executed, the edge aggregation node triggers a full network verification, which includes: all controllable switches at both ends of all members of the minimum suspicious segment set are in the open position; all activated tie switches are in the closed position; all involved distributed power sources are in the expected operating state; the voltage and frequency of all healthy segments are within the aforementioned range; and the power supply direction is consistent with the calculations before activation. If the verification passes, the edge aggregation node outputs the conclusion "Power restoration complete." If any item fails, the edge aggregation node regresses from the most recent activation and selects the next candidate tie switch to continue until the verification passes or all candidates are exhausted.

[0058] At the end of this triggered event, the edge aggregation node generates an archive record, which includes: a list of members of the minimum suspicious segment set, a complete list of isolation control sequences, a timestamp and on-site feedback for each action, measurement values ​​for each verification, a list of successfully activated tie switches, a set of nodes and edges in the final power supply topology, and the state changes of the involved distributed power sources. The archive file is named using timestamps and feeder numbers for easy reuse in subsequent similar events. To facilitate subsequent retrieval, the edge aggregation node writes three index fields into the archive record: total length of the faulty segment, total number of actions, and restored load capacity. This approach enables the rapid identification of structurally similar isolation control sequences for events of different scales in the future.

[0059] In one optional implementation of step three, a long-diameter feeder in rural areas is used as the target. Due to the longer segment spacing, an "intermediate check" action is added when generating the interruption sequence at the edge convergence node: after the first interruption, a current measurement is performed at the nearest intermediate measuring point for 60 milliseconds. If the measured value is lower than 20 amps, the second interruption is then executed. This arrangement can detect residual load coupling earlier on long-segment lines. The verification window before the tie switch is put into operation is adjusted to 60 milliseconds, and the stabilization window is adjusted to 150 milliseconds, using shorter verification and stabilization cycles to achieve faster power restoration. The coordination command of the distributed power source incorporates a "local load priority" strategy: when the target power reaches 50%, this power is maintained for 300 milliseconds, prioritizing the supply to healthy segments on the same side as the distributed power source, and then deciding whether to continue increasing based on the stability of voltage and frequency. This strategy can maintain the local voltage level even when the voltage drop is large on long lines. When the minimum suspicious segment set has shared boundary members, the edge convergence node performs boundary merging before generating the isolation control sequence. The merged members have new two-end controllable switches. The isolation control sequence performs only one end-to-end disconnection and one tie switch activation for the merged member, reducing the number of actions and making the power restoration sequence clearer. In scenarios with limited communication bandwidth, the isolation control sequence adopts a segmented delivery method: first, the opening and closing sequence of the controllable switches at both ends and the coordination instructions of the distributed power supply are delivered; after the edge nodes report back the completion of the opening and closing and the decoupling, the activation sequence of the first batch of tie switches is delivered; after the first batch of activations is completed and verified, the next batch is delivered. Segmented delivery can reduce the amount of data transmitted at one time and incorporate the verification results into the decision-making of subsequent actions in real time.

[0060] Example 2: A 10 kV feeder, 18 km long, uses a mixed configuration of segmented switches and tie switches. Sampling is arranged at 5000 samples per second. A single trigger event lasts 0.200 seconds, from 12:03:10.000 to 12:03:10.200. Edge nodes perform synchronous sampling, time stamping, and detrending at each segment location, generating graph signal frames, switch status tables, and a list of controllable switches. The segment locations, in order of power supply direction, are AA01–F12–S01 to AA01–F12–S09, corresponding to node sequence v1 to v9. Controllable switches are located at AA01–F12–S02, AA01–F12–S04, AA01–F12–S06, and AA01–F12–S08, while the tie switch is located at AA01–F12–Tie07. The distributed power source is located at adjacent nodes AA01–F12–S05, with a rated power of 800 kW. The following calculations are performed using the current path as an example (the voltage path is processed in parallel using the same method).

[0061] To facilitate subsequent calculations, we first define the graph structure and framed data: :in This represents the feeder topology diagram. Represents a set of nodes. Represents the set of edges. :in Indicates the number of nodes. :in The number of sampling moments in the signal frame is represented by the duration of the trigger event multiplied by the number of samples per second. . :in Represents the graph signal frame matrix. This indicates that the node index is And the sampling time index is The current sample value at the location is in amperes. :in This represents the image signal frame after median filtering and linear interpolation alignment. This represents the processed current sample value. :in This represents the detrended graph signal frame. This indicates the current sample value after subtracting the baseline from the mean of the preparatory segment before the trigger event. Detrending is performed using the following formula: ;in Represents a node The baseline value, This indicates the number of sampling points in the preparatory section, with a value of 100.

[0062] Graph wavelet local anomaly focusing processing: establishing three types of fixed neighborhoods for each node: :in Represents the first ring neighborhood, which is related to the node. The set of directly connected adjacent nodes. :in The second ring neighborhood is the set of adjacent nodes after crossing one edge along the connected direction. :in Let represent the neighborhood of the third ring, which is the set of adjacent nodes after crossing two edges along the connected direction. For each sampling time... Calculate the three types of local contrast quantities: ;in Represents a node At any moment Compared to the first The neighborhood mean of a fixed neighborhood. Indicates the first The number of nodes in a fixed neighborhood. ;in Indicates the first Local comparison quantity.

[0063] For each type of fixed neighborhood at time The complete set Find the absolute deviation of the median from the median: ;in This represents the median. This represents the absolute deviation from the median. The threshold is given. ;in Indicates the first Class threshold. If there exists any make Then the node At any moment The node is identified as a local anomalous node. A self-consistency constraint is applied: a node is confirmed as a local anomalous node only if it is identified at least twice within three consecutive sampling times.

[0064] Here is a calculation example for a specific time point (using the first ring neighborhood, time index). ):

[0065] Suppose that the local sample value (in amperes) on the detrended graph signal frame is: Then the mean of the first ring neighborhood and the local comparison value are:

[0066] ;

[0067] ;

[0068] ;

[0069] ;

[0070] ;

[0071] ;

[0072] ;

[0073] ;

[0074] .

[0075] set Sort by ,get Threshold is Therefore, at the first ring neighborhood scale, the node satisfy The index is calculated using the same method for the second and third ring neighborhoods at time [time]. A consistent result is reached. Further examination of adjacent time points is conducted. ,node A node is identified as a local anomalous node if it meets the threshold condition at least twice within three time points, thus satisfying the self-consistency constraint. Candidate anomalous region generation: Expanding along the connectivity direction from the local anomalous node. At time 1, the expansion proceeds from node 4 to both sides, incorporating the adjacent node 3, stopping at node 5 when the threshold is first lowered; similarly, the expansion proceeds from node 6 to both sides, incorporating node 7, stopping at node 5 when the threshold is first lowered. This results in two candidate anomaly regions: and Five locations were found to be below the threshold, causing the process to stop. Two candidate anomaly regions were identified: and As time progressed... Node 5 was identified as a local anomalous node at multiple time points, and the above two segments formed a single continuous edge set when merged in time. Record its start and end sampling times as follows: Let the set of edges be denoted as ;in This represents the candidate anomaly region within the time set.

[0076] BOCPD online verification processing includes: establishing a run length table and occurrence count: ;in This indicates the current item in the running length table at time [time]. Continuous sample count, This indicates the index of the first moment that triggered the event. ;in Indicates at time Candidate anomaly region The number of locally abnormal nodes within the region is counted. Indicates the indicator function. Sets the preset upper limit, preset lower limit, and review window: ;in This represents the cumulative occurrence threshold required to trigger the change point position. This indicates the threshold for the continuous occurrence of the count during the review phase. Indicates the continuous sampling length for verification. When the following conditions are met... At that time, The location is marked as a variable point, denoted as ; then in the interval Execution review: If any satisfy Then the interval The fragment was identified as a confirmatory fragment, in which Indicates the time range of the confirmed segment. In this example, within... Within, candidate anomaly region It persists, and the count of locally anomalous nodes at each time point is between 4 and 5, therefore in Reaching the preset upper limit, you will get .exist Within the verification window, the occurrence count at all times is not less than 4, and the run length table increases continuously, satisfying the condition. The cumulative requirement, therefore the confirmed fragment is .

[0077] Set membership estimation processing includes: propagation sequence generation: ;in Indicates the propagation sequence, This represents the local anomalous nodes within the confirmed segment, arranged chronologically from earliest to latest, along with the time they were first identified as local anomalous nodes. The sorting process yields... Temporary segment construction and boundary alignment: ;in This indicates a temporary segment that merges in from the beginning of the propagation sequence. This represents a temporary segment that merges in from the end of the propagation sequence. In a feeder, a node... The nearest controllable switch is the power supply side terminal of AA01-F12-S04, node The adjacent controllable switch is the load-side terminal of AA01-F12-S06. After boundary alignment, the following is obtained: Power supply side Overlapping contractions: in This represents the intersection of two temporary segments. After the two segments are approximated towards the center, in this example, the intersection becomes a set of continuous edges. If there are uncovered propagation sequence nodes, the extension is performed in steps using a controllable switch. In this example... Covered and through with Connectivity verification determines coverage is valid. Redundancy removal and final connectivity check: ;in This represents the smallest suspicious segment set, containing a single member, with controllable switches at both ends being AA01-F12-S04 and AA01-F12-S06.

[0078] The isolation control sequence generation includes: the switching order adopts "power supply side first, load side second", with a 100-millisecond interval between the two ends; the interlocking conditions include power supply side switching in place, section current below 20 amps, and section voltage below 5% of the rated value. The tie switch is activated according to the "nearest to farthest" strategy, prioritizing AA01–F12–Tie07. The coordination command for distributed power sources is: first disconnect, then reduce power to zero; for distributed power sources in healthy sections, a smooth power increase is adopted, increasing the rated power by 10% every 100 milliseconds until reaching 50% of the target power, and then increasing to the target power after the voltage and frequency stabilize for 200 milliseconds.

[0079] Execution process (time is measured relative to the triggering event): At 0 milliseconds, the distributed power source AA01–F12–S05 disconnection command and power reduction to zero command are issued, requiring a status confirmation return within 200 milliseconds. At 20 milliseconds, the AA01–F12–S04 disconnection command is issued; at 120 milliseconds, a status confirmation is received, and the segment measurement shows a current of 12 amps and a voltage of 450 volts. This meets the condition of "current below 20 amps and voltage below 5% of rated value". At 120 milliseconds, the AA01–F12–S06 disconnection command is issued; at 220 milliseconds, a status confirmation is received, and the segment measurement shows a current of 2 amps and a voltage of 120 volts. At 240 milliseconds, centralized recording of the faulty segment is triggered, with a recording window of 100 milliseconds, obtaining an average current of 1.5 amps and an average voltage of 100 volts within the segment, forming evidence of "residual energy dissipation after isolation". At 360 milliseconds, preparations are made to activate AA01–F12–Tie07, initiating a 40-millisecond verification window: the voltage in the healthy section is 10.2 kV, the frequency is 50.02 Hz, and the expected power flow direction is consistent with the power supply direction after activation; verification is passed. At 420 milliseconds, AA01–F12–Tie07 is activated; at 620 milliseconds, a 200-millisecond stability window observation is completed, with the voltage between 10.1 kV and 10.3 kV and the frequency between 49.98 Hz and 50.03 Hz, remaining stable. At 620 milliseconds, a smooth power boost is performed on the distributed power source in the healthy section: reaching 50% of the rated power at 720 milliseconds, maintaining stability for another 200 milliseconds, and then boosting to the target power at 920 milliseconds. At 1000 milliseconds, a full network check is performed: the controllable switches at both ends of the minimum suspicious segment set are in the open position, AA01–F12–Tie07 is in the closed position, the distributed power supply AA01–F12–S05 is in the disconnected state, and the voltage and frequency of the healthy segment meet the operating range of 10 kV and 50 Hz. The output is "Power restoration complete".

[0080] refer to Figure 2This figure illustrates the changes in sampled three-phase voltage data within the time frame of the trigger event. The horizontal axis represents time in milliseconds (ms), ranging from 0 to 12 ms; the vertical axis represents voltage amplitude in kilovolts (kV), ranging from -15 kV to 15 kV. The figure includes three curves: the solid line representing phase A voltage, the long dashed line representing phase B voltage, and the short dashed line representing phase C voltage. During the 0-6 ms timeframe, the three-phase voltages exhibit a normal sinusoidal waveform, with a 120-degree phase difference between each phase. At 6 ms, the fault occurrence is marked by a vertical dashed line. After the fault occurs, the three-phase voltage curves show significant distortion: phase A voltage drops sharply from its normal amplitude and oscillates, eventually stabilizing at approximately -5 kV; phase B voltage rises rapidly from the negative half-cycle to around 10 kV; and phase C voltage fluctuates slightly around -3 kV. The sampled data accurately reflects the voltage fluctuation characteristics when a single-phase ground fault occurs in the power distribution system, providing raw data support for subsequent local anomaly focusing processing.

[0081] refer to Figure 3 This figure shows the calculation results of the local contrast ratio of each node after performing graph wavelet local anomaly focusing processing on the graph signal frame. The horizontal axis represents the node number in the feeder topology diagram, from N1 to N7, a total of 7 nodes; the vertical axis represents the normalized value of the local contrast ratio, ranging from 0 to 1.2. The figure uses bar charts to display two types of local contrast ratios: the bars with solid borders represent the local contrast ratio of the first ring neighborhood, and the bars with dashed borders represent the local contrast ratio of the second ring neighborhood. Among them, the bars for nodes N4 and N5 are filled with black, indicating that these two nodes are judged as local anomaly nodes; the remaining nodes are filled with white, indicating normal nodes. The horizontal dashed line marks the threshold position, set at 0.8. The detection results show that the local comparison value of the first ring neighborhood of node N1 is approximately 0.2, that of node N2 is approximately 0.3, and that of node N3 is approximately 0.6, all below the threshold. The local comparison value of node N4 reaches 1.1, and that of node N5 reaches 1.0, both significantly exceeding the threshold, and are therefore identified as locally anomalous nodes. The local comparison values ​​of nodes N6 and N7 are approximately 0.4 and 0.1, respectively, within the normal range. These detection results accurately pinpoint the spatial extent of the fault's impact, providing a node-level basis for subsequent expansion of candidate anomalous regions.

[0082] refer to Figure 4This figure illustrates the variation of run length with sampling point number during Bayesian Online Change Point Detection (BOCPD). The horizontal axis represents the sampling point number, ranging from 0 to 700; the vertical axis represents the run length count, ranging from 0 to 30. The run length curve starts from sampling point 0 and shows a gradual upward trend. Within the range of sampling points 0 to 380, the run length gradually accumulates from 0 to approximately 12, indicating a continuous increase in the count of candidate anomaly regions during this period. At sampling point 420, the run length reaches a peak of approximately 28, marked with a vertical dashed line and labeled as the "change point location," indicating that the system detected a significant state change. After change point detection is triggered, the run length is reset and begins to accumulate again, forming a new run length sequence in the range of sampling points 440 to 800. The figure uses two horizontal dashed lines to mark the preset upper limit (corresponding to run length 12) and the preset lower limit (corresponding to run length 18), respectively. The upper limit is used to trigger change point detection, and the lower limit is used for verification. The interval between sampling points 440 and 800 is marked with a dashed rectangle as the "confirmation segment," indicating that this interval has passed continuous verification and has been determined as a valid fault segment. This detection method achieves accurate location of fault events through a two-stage strategy, avoiding false triggering caused by transient noise.

[0083] refer to Figure 5 This figure illustrates the current changes in the faulty and healthy sections during the execution of the isolation control sequence. The horizontal axis represents time in milliseconds (ms), ranging from 0 to 350 ms; the vertical axis represents current amplitude in amperes (A), ranging from 0 to 1200 A. The solid line represents the current curve of the faulty section: from 0 to 50 ms, the current remains at a normal operating level of approximately 400 A; at 50 ms, a fault occurs, and the current rises sharply to a short-circuit current level of approximately 1100 A, remaining high from 50 to 100 ms; at 100 ms, the first controllable switch S1 is opened, marked by a vertical dashed line, and the current begins to decrease; at 150 ms, the second controllable switch S2 is opened, also marked by a vertical dashed line, and the current further decreases; from 150 ms to 200 ms, the current gradually drops below 20 A; after 200 ms, the current in the faulty section returns to zero, indicating successful isolation. The long dashed line represents the current curve of the healthy section: throughout the process, it remains at a stable level of approximately 350 A. The 200ms mark indicates "Tie-up Engagement," signifying that the tie switch is closed, and the healthy section resumes power supply via transfer, maintaining a stable current. This curve verifies the effectiveness of the isolation control sequence and demonstrates the execution effect of the "isolate first, then restore power" control strategy.

[0084] refer to Figure 6This diagram illustrates the derivation process from the propagation sequence to the minimum suspicious segment in the ensemble membership estimation process. The upper part of the diagram is a linear representation of the feeder topology, containing five segmented switches S1 to S5, with the line segments between the switches representing feeder sections. On the feeder topology, the locations of five local anomaly nodes are marked with solid dots, distributed within the section between S2 and S4. The arrow sequence indicates the direction of anomaly propagation, pointing from the power supply side (left) to the load side (right), reflecting the propagation path of the fault energy. The middle of the diagram uses a long dashed rectangle to represent "Temporary Segment 1," which originates from the first local anomaly node in the propagation sequence and merges along the connectivity direction into the area between controllable switches S2 and S4, covering a length of 270 units. A short dashed rectangle represents "Temporary Segment 2," which merges in reverse from the end of the propagation sequence, covering a length of 280 units and overlapping with Temporary Segment 1. The gray-filled rectangle at the bottom of the diagram represents the intersection of the two temporary segments, known as the "minimum suspicious segment," with its ends aligned to the positions of controllable switches S3 and S4, respectively, and a length of 150 units. Vertical dashed lines mark the boundary alignment positions, and the arrows below indicate the process of aligning the boundaries to S3 and S4. This derivation process demonstrates the principle of gradually converging the initial, larger temporary segment to the minimum necessary isolation range through overlapping contraction and boundary alignment, ensuring the accuracy of isolation control and minimizing the power outage area.

[0085] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these specific embodiments are merely illustrative. Those skilled in the art can omit, substitute, and modify the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, combining the above method steps to perform substantially the same function and achieve substantially the same result according to substantially the same method falls within the scope of the present invention. Therefore, the scope of the present invention is defined only by the appended claims.

Claims

1. A fault detection and isolation method for an intelligent power distribution system based on edge computing, characterized in that, This method is executed collaboratively by edge nodes and edge aggregation nodes, and includes the following three steps: Step 1: Perform edge node side data construction and graph preprocessing. Edge nodes are set up at each segment of the intelligent power distribution system. Voltage, current, switch position and time stamp are obtained by the edge nodes. Feeder topology is generated based on the connection relationship between primary equipment and line. Synchronous sampling data within the same trigger event is organized into graph signal frames of uniform length and a switch status table and controllable switch list corresponding to the graph signal frames are formed. Step 2: On one side of the edge convergence node, perform graph wavelet local anomaly focusing processing, BOCPD online verification processing and set membership estimation processing continuously to form the minimum suspicious segment set; Step 3: Execute isolation and power restoration control based on the minimum suspicious segment set. The edge aggregation node generates an isolation control sequence based on the minimum suspicious segment set formed in Step 2. The isolation control sequence includes the opening and closing sequence of the controllable switches at both ends of each member in the minimum suspicious segment set, the activation sequence of the adjacent tie switches, and the coordination instructions with the distributed power sources involved. The edge aggregation node distributes the isolation control sequence to the relevant edge nodes for execution. The edge nodes complete the disconnection of the faulty segment, the reconstruction and power restoration of the healthy segment, and the status verification. The execution results are archived for reuse in subsequent similar events.

2. The method according to claim 1, characterized in that, Step two specifically includes: performing graph wavelet local anomaly focusing processing; performing BOCPD online verification processing; and performing set membership estimation processing to form a minimum suspicious segment set.

3. The method according to claim 2, characterized in that, The graph wavelet local anomaly focusing process specifically includes: on the feeder topology, generating three types of fixed neighborhoods—a first-ring neighborhood, a second-ring neighborhood, and a third-ring neighborhood—centered on each node; for each sampling time of the graph signal frame, calculating the absolute difference between the sampled value of the center node and the arithmetic mean of each of the three types of fixed neighborhoods, denoted as the three types of local contrast values; applying the median and median absolute deviation methods to the three types of local contrast values ​​to obtain three types of thresholds; at any sampling time, if any node has a local contrast value higher than the corresponding threshold, marking this node as a local anomaly node; sequentially expanding from the local anomaly node to adjacent nodes on both sides along the connectivity of the feeder topology, retaining the adjacent node when its local contrast value is continuously higher than its corresponding threshold during the expansion process, stopping when the first boundary below the threshold appears, forming a candidate anomaly region composed of several adjacent edges; repeating the above process for all sampling times, merging candidate anomaly regions at the same position and recording the start and end sampling positions to obtain a set of candidate anomaly regions.

4. The method according to claim 3, characterized in that, The BOCPD online verification process in step 2 specifically includes: taking the candidate anomaly region set as input, establishing a run length table according to the sampling order, with each entry in the run length table recording the continuous sample count since the previous verification position; at each sampling time, for each candidate anomaly region appearing in the candidate anomaly region set, calculating the occurrence count of the corresponding local anomaly node in the candidate anomaly region at the current sampling time, and adding the occurrence count to the current entry in the run length table.

5. The method according to claim 4, characterized in that, When the occurrence count of the current item in the running length table reaches the preset upper limit, the current sampling time is marked as the change point position, and the two time segments before and after the change point are divided accordingly. A review is performed on a series of consecutive sampling moments after the change point. The review rules are as follows: if the count of candidate abnormal regions continuously reaches or exceeds the preset lower limit and the running length table continuously increases within a series of consecutive sampling moments, the continuous sample interval after the change point until the end of the review is marked as a confirmation segment; when the running length table is initialized after the review is completed, the next round of online confirmation begins; within the same trigger event, one or more confirmation segments are obtained in sequence, and a one-to-one correspondence is established between each confirmation segment and the candidate abnormal region it covers.

6. The method according to claim 5, characterized in that, The ensemble member estimation process includes step A: propagation sequence generation: local anomalous nodes within the confirmed segment are arranged from early to late according to time stamps, and combined with the connectivity direction of the feeder topology, a propagation sequence from the power supply side to the load side is obtained.

7. The method according to claim 6, characterized in that, The set membership estimation process also includes: Step B, temporary segment construction: starting from the first local anomaly node in the propagation sequence, a temporary segment is formed by merging into the first controllable switch along the connectivity direction; at the same time, another temporary segment is formed by merging into the adjacent controllable switch from the end local anomaly node in the propagation sequence along the connectivity direction. Step 3.3, boundary alignment: align the endpoints of the two temporary segments to the corresponding controllable switch positions to obtain two temporary segments with aligned boundaries.

8. The method according to claim 7, characterized in that, The set membership estimation process also includes: Step C, overlap contraction: find the intersection of the temporary segments whose boundaries are aligned. If the intersection is still a continuous edge set, replace the original temporary segment with the intersection. When the intersection is non-empty and there is a non-continuous edge set, split the non-continuous edge set into several continuous sub-segments according to connectivity and use them as new temporary segments respectively.

9. The method according to claim 8, characterized in that, The set member estimation process also includes: Step D, Coverage Verification: Determine whether the new set of temporary segments covers all local abnormal nodes in the confirmed segment. If there are uncovered nodes, extend a temporary segment adjacent to the uncovered node along the connectivity direction to the next controllable switch and return to the boundary alignment step until coverage is completed.

10. The method according to claim 9, characterized in that, The set membership estimation process also includes: Step E, Redundancy Removal: After coverage is completed, all temporary segments are checked in ascending order of length. If a temporary segment is completely contained by another temporary segment, the contained temporary segment is deleted; Step F, Connectivity Final Check: The connectivity of the remaining temporary segments is verified one by one. The verification rule is that each edge within the segment forms a single path on the feeder topology and both ends are bounded by controllable switches; All temporary segments obtained after the redundancy removal step and the connectivity final check step are the minimum suspicious segment set members corresponding to a confirmed segment; When there are multiple confirmed segments, the set members obtained from each confirmed segment are combined, and the redundancy removal step and the connectivity final check step are executed again to obtain the final minimum suspicious segment set.

Images