An intelligent diagnosis model power quality disturbance tracing method based on a workflow platform

Abstract

This invention relates to the field of power quality monitoring technology and provides a power quality disturbance tracing method based on a workflow platform using an intelligent diagnostic model. The method includes: collecting three-phase voltage and current waveform data from each monitoring node; when a power quality indicator exceeds a preset threshold, the workflow platform performs feature extraction to obtain a disturbance feature vector; performs disturbance pattern discrimination and selects the corresponding tracing sub-algorithm branch based on the discrimination result to obtain branch tracing results; performs hierarchical event mapping to aggregate abnormal indicators of each monitoring node into disturbance events according to spatiotemporal clustering and constructs a directed acyclic graph-like disturbance propagation chain based on the branch tracing results; inputs the disturbance feature vector and disturbance propagation chain into a trained target hybrid model; outputs the corrected disturbance propagation chain and root cause node; and finally outputs the tracing results. This invention effectively improves the tracing accuracy and efficiency in complex disturbance scenarios.

Description

Technical Field

[0001] This invention relates to the field of power quality monitoring technology, and in particular to a method for tracing the source of power quality disturbances using an intelligent diagnostic model based on a workflow platform. Background Technology

[0002] In power system operation, power quality disturbances can affect power supply reliability and equipment safety, thus requiring accurate identification of disturbance types and location of disturbance sources. Existing technologies, after acquiring waveform data, extract waveform features before and after the disturbance occurrence, such as harmonic content, sag amplitude, and duration. These extracted features are then input into a pre-trained classification model, which outputs the disturbance type identification result. Combined with disturbance arrival time and amplitude attenuation information from multiple monitoring points, a simple propagation path analysis method is used to preliminarily determine the approximate direction of the disturbance source. However, due to the significant differences in waveform characteristics between different disturbance types, and the complex and variable propagation paths of disturbances in the power grid influenced by factors such as network topology and load changes, using a uniform feature extraction process and a fixed diagnostic model cannot adaptively adjust the source tracing strategy according to the disturbance type. This results in low accuracy and efficiency in tracing complex disturbances or complex propagation scenarios, making it difficult to accurately pinpoint the specific equipment where the disturbance source is located. Summary of the Invention

[0003] This invention provides a power quality disturbance tracing method based on a workflow platform using an intelligent diagnostic model. This method addresses the problems of low accuracy and efficiency in tracing the source of complex disturbances or complex propagation scenarios, making it difficult to accurately pinpoint the specific device where the disturbance source is located.

[0004] This invention provides a method for tracing the source of power quality disturbances using an intelligent diagnostic model based on a workflow platform, comprising:

[0005] By deploying power quality monitoring terminals at multiple monitoring nodes in the power grid, waveform data of each monitoring node is collected in real time. The waveform data includes three-phase voltage waveform data and three-phase current waveform data.

[0006] When the power quality index of any monitoring node exceeds a preset threshold based on the waveform data, the workflow platform executes the following steps:

[0007] Adaptive feature extraction is performed on the waveform data to obtain a disturbance feature vector;

[0008] Based on the perturbation feature vector, the perturbation mode is determined, and the corresponding source tracing sub-algorithm branch is selected according to the determination result. The source tracing sub-algorithm corresponding to the selected branch is then used for processing to obtain the branch source tracing result.

[0009] Hierarchical event mapping is performed, and the abnormal indicators of each monitoring node are aggregated into disturbance events by spatiotemporal clustering. The disturbance propagation chain in the form of a directed acyclic graph is constructed in combination with the branch tracing results.

[0010] The perturbation feature vector and the perturbation propagation chain are input into the trained target hybrid model, and the target hybrid model outputs the corrected perturbation propagation chain and root cause node, wherein the root cause node corresponds to the specific device where the perturbation source is located;

[0011] Output the source tracing results and add the current disturbance event and its corresponding source tracing results to the labeled dataset.

[0012] Furthermore, the adaptive feature extraction of the waveform data to obtain the disturbance feature vector includes:

[0013] The three-phase voltage waveform data and the three-phase current waveform data are respectively compared with each template in the preset standard disturbance template library to perform intersection point distribution analysis, and the degree of matching between each phase waveform data and the corresponding template in the time domain is determined.

[0014] Furthermore, based on the distribution type of intersection points of the waveform data for each phase, each phase waveform data is divided into four categories: complete matching, partial matching, no overlap but with intersection, and complete deviation; among which:

[0015] When the waveform data of a certain phase is divided into complete matches, the preset feature vector of the corresponding template is reused as the single-phase feature vector of that phase.

[0016] When a phase waveform data is divided into local matching, the difference between the phase waveform data and the template is extracted by a sliding window, and the incremental feature vector of the phase is generated based on the difference.

[0017] When a phase waveform data is divided into non-overlapping and overlapping sections, the slope change rate and distortion energy near the intersection point of the phase waveform data and the template are extracted to construct the cross feature vector of the phase.

[0018] When the waveform data of a certain phase is completely deviated, the waveform data of that phase is marked as a new category candidate, and the complete time-frequency domain statistical features of that phase are extracted as the single-phase feature vector of that phase;

[0019] The disturbance characteristic vector is obtained by concatenating the single-phase characteristic vectors of the three phases in phase order.

[0020] Furthermore, the step of determining the perturbation pattern based on the perturbation feature vector and selecting the corresponding source tracing sub-algorithm branch according to the determination result includes:

[0021] The transient sag depth, total harmonic distortion rate, and transient high-frequency energy in the disturbance feature vector are used as discriminant features.

[0022] When the sag depth exceeds the first threshold and the total harmonic distortion rate is lower than the second threshold, it is determined to be a sag-dominated type, and a propagation path deduction branch based on energy flow direction is selected.

[0023] When the total harmonic distortion rate exceeds the third threshold and the sag depth is lower than the fourth threshold, it is determined to be harmonic-dominated, and the source localization branch based on harmonic impedance calculation is selected.

[0024] When the transient high-frequency energy exceeds the fifth threshold and the transient sag duration is less than the sixth threshold, it is determined to be a transient oscillation type, and the traveling wave arrival time difference method based on waveform similarity is selected as the branch.

[0025] When the sag depth exceeds the first threshold and the total harmonic distortion rate exceeds the third threshold, it is determined to be a composite type, and the weighted fusion strategy branch is selected.

[0026] Furthermore, the first to sixth thresholds are adjusted adaptively:

[0027] The reference voltage and reference harmonic content of each monitoring node are calculated based on the current operation mode of the power grid, and the threshold is set as a preset percentage of the reference value; or the threshold is automatically set based on historical disturbance data through quantile statistics.

[0028] Furthermore, the step of processing the selected branch using the corresponding tracing sub-algorithm to obtain the branch tracing result includes:

[0029] When selecting a propagation path deduction branch based on energy flow direction, the direction and distance of the disturbance source are determined by deducing in reverse along the energy transmission based on the attenuation gradient of the transient amplitude at each monitoring point and the disturbance arrival time difference.

[0030] When selecting the source location branch based on harmonic impedance calculation, the harmonic impedance on the system side and the harmonic emission level on the user side are calculated using the harmonic voltage and harmonic current at each monitoring point, and the node with the highest harmonic emission level is identified as the harmonic source.

[0031] When the waveform similarity-based traveling wave time of arrival method branch is selected, the arrival time of the disturbance traveling wave is extracted, the distance between the disturbance source and the monitoring point is calculated through the multi-terminal ranging formula, and the optimal match is selected by combining waveform similarity.

[0032] When selecting the weighted fusion strategy branch, the source location confidence scores output by the algorithms corresponding to the propagation path deduction branch based on energy flow direction, the source location branch based on harmonic impedance calculation, and the traveling wave arrival time difference method branch based on waveform similarity are calculated respectively. The scores are then weighted and averaged according to preset weights to obtain the final disturbance source location.

[0033] Furthermore, the hierarchical event mapping, which aggregates the abnormal indicators of each monitoring node into disturbance events based on spatiotemporal clustering, includes:

[0034] Construct a four-layer exception event model, in which:

[0035] The L0 layer is the raw waveform data layer, which stores the three-phase voltage waveform data and three-phase current waveform data of each monitoring node;

[0036] The L1 layer is the feature index layer. The sag depth, total harmonic distortion rate, transient high-frequency energy and disturbance arrival time of each monitoring node are extracted from the L0 layer as anomaly indicators.

[0037] Layer L2 is the perturbation event layer, which aggregates the abnormal indicators in Layer L1 that meet the spatiotemporal clustering conditions into a perturbation event;

[0038] Layer L3 is the source localization layer, used to store the perturbation propagation chain and root cause nodes;

[0039] The spatiotemporal clustering conditions include that the anomaly indicators within the time window are continuous in time and that the anomaly indicators of spatially adjacent monitoring nodes have the same disturbance type.

[0040] Furthermore, the construction of a perturbation propagation chain in the form of a directed acyclic graph based on the branch tracing results includes:

[0041] Each perturbation event in the L2 layer is treated as a node, and directed edges between nodes are determined based on the temporal sequence of events, energy decay gradient, and waveform similarity.

[0042] Using the branch tracing results as the confidence weights of the edges, a perturbation propagation chain in the form of a weighted directed acyclic graph is obtained.

[0043] Furthermore, the target hybrid model includes:

[0044] The feature extraction part is composed of a multi-layer convolutional neural network, which is used to extract deep waveform features from the perturbation feature vector;

[0045] The propagation chain correction part is composed of a fully connected network, which is used to receive the graph embedding vector of the perturbation propagation chain and the depth waveform features, and output the correction weights of each edge in the perturbation propagation chain and the confidence of the root cause node.

[0046] Furthermore, the step of inputting the perturbation feature vector and the perturbation propagation chain into the trained target mixture model, and having the target mixture model output the corrected perturbation propagation chain and root cause node, includes:

[0047] The perturbation feature vector is input into the multi-layer convolutional neural network to obtain deep waveform features;

[0048] The disturbance propagation chain is encoded as a graph embedding vector;

[0049] The graph embedding vector and the depth waveform feature are concatenated and then input into the fully connected network.

[0050] The fully connected network outputs corrected weights for each edge in the perturbation propagation chain;

[0051] Adjust the confidence level of each edge according to the corrected weight, delete the edges with confidence levels lower than the preset threshold, and obtain the corrected perturbation propagation chain;

[0052] Nodes with an in-degree of zero are selected from the corrected disturbance propagation chain as root cause nodes, and the root cause nodes correspond to the specific devices where the disturbance source is located.

[0053] As can be seen from the above technical solutions, the present invention has the following advantages:

[0054] This invention collects three-phase voltage and current waveform data of each monitoring node in real time through power quality monitoring terminals deployed at multiple monitoring nodes in the power grid. When the power quality index exceeds a preset threshold, the workflow platform automatically performs adaptive feature extraction to obtain a disturbance feature vector. Then, based on the disturbance feature vector, disturbance mode discrimination is performed, and the corresponding source tracing sub-algorithm branch is selected for processing according to the discrimination result to obtain the branch source tracing result. Then, hierarchical event mapping is performed to aggregate the abnormal indicators of each monitoring node into disturbance events according to spatiotemporal clustering, and a disturbance propagation chain in the form of a directed acyclic graph is constructed by combining the branch source tracing result. Next, the disturbance feature vector and the disturbance propagation chain are input into the trained target mixture model, and the target mixture model outputs the corrected disturbance propagation chain and root cause node. Finally, the source tracing result is output, and the current disturbance event and the corresponding source tracing result are added to the labeled dataset. This invention implements a branched tracing strategy for different disturbance types through adaptive feature extraction and disturbance pattern discrimination, significantly improving the tracing accuracy in complex disturbance scenarios. By constructing hierarchical event mapping and disturbance propagation chains, it explicitly models the spatiotemporal correlation information of multiple monitoring points, enhancing the interpretability of the tracing results. By correcting the propagation chain through a target hybrid model, it further improves the accuracy of root cause localization. At the same time, the fully automated execution of the workflow platform and the continuous feedback of labeled datasets improve the efficiency of tracing. Attached Figure Description

[0055] Figure 1 This is a schematic flowchart of an embodiment of a power quality disturbance tracing method based on a workflow platform in this invention;

[0056] Figure 2 This is a flowchart of the adaptive feature extraction and branch tracing of power quality disturbances in this invention. Detailed Implementation

[0057] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “corresponding to,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0058] Example

[0059] The implementation method in this embodiment can be implemented in a system, on a server, or on a terminal; no specific limitation is made. The method in this application will be described below from the perspective of system implementation. Please refer to... Figure 1 The method provided in this application includes the following steps:

[0060] S1. Through power quality monitoring terminals deployed at multiple monitoring nodes in the power grid, waveform data of each monitoring node is collected in real time. The waveform data includes three-phase voltage waveform data and three-phase current waveform data.

[0061] Multiple monitoring nodes are selected based on key power grid buses, important load connection points, distributed power generation grid connection points, and branch nodes in the network topology to ensure that the monitoring coverage reflects the power quality status of major areas of the power grid. In actual implementation, node locations are determined by equidistant arrangement or topology vulnerability analysis, based on the power grid's voltage level, short-circuit capacity, and frequency of historical disturbance events. For example, monitoring nodes are set up on the 10kV or 35kV outgoing lines of 110kV and above substations, the incoming lines of large industrial users, and the grid connection points of new energy power plants. The power quality monitoring terminal uses a standard-compliant three-phase power quality monitoring device, which incorporates high-precision voltage transformers and current transformers, enabling simultaneous acquisition of instantaneous values ​​of three-phase voltage and three-phase current.

[0062] Specifically, the monitoring terminal continuously samples at a sampling rate of 256 or 512 points per cycle, generating discrete three-phase voltage and three-phase current time series, with each sampling point accompanied by a precise timestamp. The acquired raw waveform data is uploaded in real time to the workflow platform's data access service via a 4G wireless communication module, using the common format for power system transient data exchange. After receiving the waveform data from each monitoring node, the workflow platform segments and stores the data according to node number and timestamp, and maintains a fixed-length circular buffer for each monitoring node to cache waveform data for the current and several historical cycles, enabling rapid extraction of waveform segments before and after disturbances when a threshold is triggered.

[0063] S2. When the power quality index of any monitoring node exceeds the preset threshold based on waveform data, the workflow platform executes the following steps:

[0064] The preset thresholds are determined in advance based on the power quality statistical characteristics under normal grid operating conditions and industry regulations. The power quality indicators here include the effective voltage value reflecting voltage sags, the total harmonic distortion (THD) rate reflecting harmonic distortion, and the high-frequency energy reflecting transient oscillations. For voltage sags, the preset threshold is set to 90% of the nominal voltage. A sag event is triggered when the effective voltage value at a monitoring node drops below 90% of the nominal voltage for more than half a cycle. For harmonic indicators, the preset threshold is set according to GB / T 14549 standard, with the THD limit set to 2% to 5%, specifically selected based on the grid voltage level. For transient oscillation indicators, the preset threshold is dynamically set by adding three standard deviations to the average high-frequency energy of historical normal waveforms, with a typical range of 3 to 5 times the normal high-frequency energy. The thresholds for each monitoring node can be configured independently, allowing for adjustments by experts based on field conditions, or automatic calculation by the workflow platform using quantile statistics based on historical data.

[0065] The workflow platform is a visual, orchestratable, automated operation and maintenance management software system running on a central server or cloud server. It provides a drag-and-drop component orchestration interface, a task scheduling engine, and data storage and processing services. The platform receives waveform data uploaded from various monitoring terminals and calculates power quality indicators in real time. This involves extracting the effective voltage value, total harmonic distortion rate (THD), and high-frequency energy from the waveform data, comparing each indicator with corresponding preset thresholds, and automatically triggering a series of processing steps defined in steps S21 to S25 when any indicator on any node continuously exceeds the threshold for a preset number of times. Internally, the platform adopts a microservice architecture. The data access service is responsible for waveform data parsing and indicator calculation, the rule engine is responsible for threshold comparison and event triggering, and the workflow engine is responsible for calling subsequent components such as feature extraction, pattern discrimination, and event mapping according to the visually orchestrated flowchart. Each component is encapsulated as an independent containerized service, and data is passed between components via message queues, supporting parallel execution and retrying. The workflow platform also provides a graphical user interface where operation and maintenance personnel can view real-time waveforms, threshold configurations, trigger records, and execution status logs for each step. The above configuration ensures the reasonableness of the threshold settings and the reliability of the workflow platform's execution steps. The workflow platform executes the following steps:

[0066] S21. Perform adaptive feature extraction on the waveform data to obtain the disturbance feature vector;

[0067] In this embodiment, the three-phase voltage waveform data and the three-phase current waveform data are respectively compared with each template in the preset standard disturbance template library to perform intersection point distribution analysis, thereby determining the degree of matching between each phase waveform data and the corresponding template in the time domain.

[0068] The pre-defined standard disturbance template library is a collection of waveform samples constructed by extracting features and labeling waveform data from typical power quality disturbance events. Specifically, it is constructed by collecting historical waveform data of typical disturbance events occurring in actual power system operation, such as voltage sags, harmonic distortions, transient oscillations, and pulses. Each disturbance event is labeled by experts to confirm its disturbance type and key parameters. For each labeled waveform, waveform segments of ten cycles before and after the disturbance are extracted, their amplitudes are normalized to the per-unit value of the nominal voltage, and Fourier transform and wavelet transform are used to extract time-domain and frequency-domain features to form a standard template. Each type of disturbance in the template library contains multiple waveform templates with different severity and propagation distances. For example, for voltage sags, templates include sag depths of 90%, 80%, and 50%, and durations of half a cycle, ten cycles, and thirty cycles. The template library is stored as a structured file on the workflow platform. Each template contains a template identifier, disturbance type label, feature vector, waveform sampling point sequence, and sampling rate.

[0069] During intersection point distribution analysis, the workflow platform processes the waveform data for each phase separately: For phase A voltage waveform data, its sampling point sequence is aligned with the sampling point sequence of the same type of template in the template library along the time axis, and the amplitude difference is calculated point by point. If the amplitude difference is less than 2% of the nominal voltage, the point is determined to be an intersection point. The number of intersection points, the intersection point distribution interval, and the length of the continuous interval without intersection points on the entire waveform are counted to determine the degree of matching. For phase B and phase C waveform data, the exact same method is used for independent analysis. The quantitative indicators of the degree of matching include intersection point coverage (the proportion of intersection points to the total number of points in the waveform), the maximum continuous matching length, and the average amplitude deviation. For example, when the intersection coverage exceeds 95% and the maximum continuous matching length covers the entire waveform, it is determined to be a complete match; when the intersection coverage is between 50% and 95%, and there are alternating continuous matching segments and deviation segments, it is determined to be a partial match; when the intersection coverage is less than 50% but the waveform and the template intersect in amplitude, it is determined to be non-overlapping but intersecting; when the intersection coverage is close to 0 and the overall waveform deviates from the template amplitude by more than 30%, it is determined to be a complete deviation. Please refer to [link to relevant documentation]. Figure 2 The details are as follows:

[0070] In this embodiment, based on the intersection distribution type of each phase waveform data, each phase waveform data is divided into completely matched, partially matched, non-overlapping but intersecting, and completely offset; wherein:

[0071] 1. When the waveform data of a certain phase is divided into perfectly matched phases, the preset feature vector of the corresponding template is reused as the single-phase feature vector of that phase;

[0072] 2. When the waveform data of a certain phase is divided into local matching, the difference between the waveform data of that phase and the template is extracted by a sliding window, and the incremental feature vector of that phase is generated based on the difference.

[0073] 3. When a phase waveform data is divided into non-overlapping and overlapping sections, extract the slope change rate and distortion energy near the intersection point of the phase waveform data and the template, and construct the intersection feature vector of that phase;

[0074] 4. When the waveform data of a certain phase is completely deviated, the waveform data of that phase is marked as a new category candidate, and the complete time-frequency domain statistical features of that phase are extracted as the single-phase feature vector of that phase;

[0075] 5. Concatenate the single-phase characteristic vectors of the three phases according to the phase sequence to obtain the disturbance characteristic vector.

[0076] Specifically, perfect matching occurs when a phase waveform data highly overlaps with a standard template in the time domain, with an intersection coverage exceeding 95%, and the amplitude deviation of all intersection points is within a set tolerance range. In this case, the disturbance characteristics reflected by the phase waveform are completely consistent with the template, so the pre-stored feature vector in the template is directly reused. This feature vector contains key indicators such as sag depth, harmonic content, and duration. Partial matching occurs when a portion of the waveform data overlaps with the template, while another portion deviates significantly. For example, the first three cycles of the waveform match the sag template, while the last two cycles deviate due to the superposition of harmonics. In this case, a sliding window with a length of half a cycle or one cycle is used, moving one-quarter of a cycle each time. The residual signal between the waveform within the window and the corresponding template window is calculated, and a Fourier transform is performed on the residual signal to obtain the incremental spectrum. The dominant frequency component and energy in the incremental spectrum are extracted as the incremental feature vector. "No overlap but with intersection" means that the waveform and the template cross each other in amplitude, but there are no long-term overlapping sections. For example, a waveform with amplitude oscillation decay and a steady-state sag template. The waveform curve crosses the template curve multiple times. In this case, each intersection point (the point where the waveform amplitude equals the template amplitude) is located, and the waveform slope of five sampling points before and after the intersection point is calculated. The average of the slopes is taken as the slope change rate. At the same time, the ratio of the waveform energy to the template energy of ten sampling points near the intersection point is calculated as the distortion energy. These parameters are combined into the intersection feature vector. "Complete deviation" means that the waveform and the template have no intersection points, and the overall amplitude of the waveform deviates from the template amplitude by more than 30%. In this case, the current disturbance is considered not to belong to any type already existing in the template library. Therefore, the waveform is marked as a new category candidate. At the same time, the complete time-frequency domain statistical features of the waveform are extracted, including the voltage RMS value sequence, total harmonic distortion rate sequence, high-frequency energy envelope, waveform kurtosis, skewness, etc., as the single-phase feature vector of this phase.

[0077] After extracting the single-phase feature vectors for each of the three phases, the three vectors are concatenated into a long vector in the order of phase A, phase B, and phase C. This long vector is the perturbation feature vector. If the lengths of the single-phase feature vectors of a certain phase are inconsistent, linear interpolation is used to unify them to a set length to ensure that the concatenated feature vector has a fixed dimension, which is convenient for subsequent input into the target hybrid model.

[0078] S22. Based on the perturbation feature vector, perform perturbation pattern discrimination, select the corresponding source tracing sub-algorithm branch according to the discrimination result, and use the source tracing sub-algorithm corresponding to the selected branch for processing to obtain the branch source tracing result;

[0079] This step automatically identifies the main type of the current disturbance based on the numerical relationships of different indicators in the disturbance feature vector, and matches the most suitable tracing algorithm branch for each type, thereby improving the accuracy and efficiency of tracing. Specifically, it includes the following:

[0080] 1. The transient sag depth, total harmonic distortion rate, and transient high-frequency energy in the disturbance feature vector are used as discriminant features;

[0081] Sag depth is defined as the ratio of the minimum effective value of the monitored node voltage during a disturbance to the nominal voltage, expressed as a percentage. For example, a sag depth of 0.9 indicates that the voltage has dropped to 90% of the nominal voltage. Sag depth can be directly extracted from the disturbance feature vector; a smaller value indicates a more severe voltage drop. Total harmonic distortion (THD) is defined as the ratio of the sum of the effective values ​​of all harmonic components to the effective value of the fundamental frequency, reflecting the overall degree of waveform distortion, expressed as a percentage. A higher THD indicates more severe harmonic pollution. Transient high-frequency energy is calculated by integrating the energy of the filtered signal over the duration of the disturbance after high-pass filtering of the waveform. It characterizes the intensity of high-frequency disturbance components such as transient oscillations and spikes. These three discriminant features correspond to the core physical quantities of the three basic disturbance types: sag, harmonics, and transient oscillations, effectively distinguishing different types of disturbance events.

[0082] 2. When the sag depth exceeds the first threshold and the total harmonic distortion rate is lower than the second threshold, it is determined to be a sag-dominated type, and a propagation path deduction branch based on energy flow direction is selected;

[0083] 3. When the total harmonic distortion rate exceeds the third threshold and the sag depth is lower than the fourth threshold, it is determined to be harmonic-dominated, and the source localization branch based on harmonic impedance calculation is selected.

[0084] 4. When the transient high-frequency energy exceeds the fifth threshold and the transient sag duration is less than the sixth threshold, it is determined to be a transient oscillation type, and the traveling wave arrival time difference method branch based on waveform similarity is selected;

[0085] 5. When the sag depth exceeds the first threshold and the total harmonic distortion rate exceeds the third threshold, it is determined to be a composite type, and the weighted fusion strategy branch is selected.

[0086] In this embodiment, the first to sixth thresholds are adjusted adaptively:

[0087] The reference voltage and reference harmonic content of each monitoring node are calculated based on the current operation mode of the power grid, and the threshold is set as a preset percentage of the reference value; or the threshold is automatically set based on historical disturbance data through quantile statistics.

[0088] Specifically, the first threshold is used to determine whether the sag depth is significant, with a typical value range of 80% to 90% of the nominal voltage, and a preferred value of 85%. When the sag depth is greater than 85%, it indicates a significant voltage drop, constituting a sag event. The second threshold is used to determine whether the total harmonic distortion (THD) is low, with a typical value range of 1% to 3%, and a preferred value of 2%. When the THD is below 2%, it indicates that harmonic interference is negligible, and the disturbance is mainly caused by sags. The third threshold is used to determine whether the THD is high, with a typical value range of 3% to 5%, and a preferred value of 4%. When the THD exceeds 4%, it indicates significant harmonics. The fourth threshold is used to determine whether the sag depth is low, with a typical value range the same as or slightly lower than the first threshold, and a preferred value of 80%. When the sag depth is below 80%, the sag itself is not significant, and the disturbance is mainly caused by harmonics. The fifth threshold is used to determine whether transient high-frequency energy is significant. This threshold is obtained by adding three times the standard deviation to the average high-frequency energy during historical periods without disturbances, with a typical value being three to five times the normal high-frequency energy. The sixth threshold is used to determine whether the duration of a sag is short, with a typical value ranging from half a cycle to three cycles, and a preferred value of two cycles. When the duration of a sag is less than two cycles, it indicates that the disturbance is transient. All of the above thresholds can be adaptively adjusted according to the actual operation mode of the power grid and historical data. For example, when the power grid operation mode changes, the workflow platform recalculates the reference voltage and reference harmonic content of each monitoring node, and then updates the first and fourth thresholds by a preset percentage of the reference value; or, the workflow platform collects historical disturbance data of all triggered events in the past year, performs quantile statistics on the characteristic indicators of each disturbance type, and uses the 75th percentile as the corresponding threshold. Through the above threshold settings, it is possible to accurately distinguish between four modes: sag-dominated, harmonic-dominated, transient oscillation, and composite, and select the corresponding tracing sub-algorithm branch for each mode.

[0089] In this embodiment, the source tracing sub-algorithm corresponding to the selected branch is used for processing to obtain the branch source tracing result, including the following:

[0090] 1. When selecting a propagation path deduction branch based on energy flow direction, the direction and distance of the disturbance source are determined by deducing in reverse along the energy transmission based on the attenuation gradient of the transient amplitude at each monitoring point and the arrival time difference of the disturbance.

[0091] The propagation path deduction branch based on energy flow direction is a method to reverse-engineer the disturbance source by utilizing the physical laws of amplitude attenuation and arrival time delay during voltage sag propagation in the power grid. When a sag-dominated disturbance is identified, the workflow platform collects data from all nodes that have detected the sag event, obtaining the sag depth and disturbance arrival time for each node. The attenuation gradient of the sag amplitude is the difference in sag depth between two adjacent monitoring nodes divided by the electrical distance between the two nodes; this gradient reflects the energy loss rate of the sag along the transmission path. The disturbance arrival time difference is the time difference between the first detection of the same sag event at different nodes; nodes closer to the disturbance source detect the sag earlier.

[0092] In practical implementation, the workflow platform first establishes an undirected graph based on the power grid topology (obtaining switch states and line parameters from the scheduling system). Nodes are monitoring points, and edges are line or transformer branches, with the impedance and length of each edge known. Then, for each node detecting a sag, its sag depth and arrival time are used as attribute values. A breadth-first search algorithm is employed, starting from the node with the deepest sag depth and comparing the decay trend of sag depth to its neighboring nodes: if along an edge from node A to node B, the sag depth of node B is less than that of node A, and the arrival time of node B is later than that of node A, then it is determined that energy flows from A to B, and the disturbance source may be located on the A side. During reverse deduction, the path is traced backward along the direction with the slowest decay gradient, because the gradient of sag depth change is smaller near the disturbance source. Simultaneously, using the arrival time difference, the distance between the disturbance source and the two monitoring points is calculated using the two-end ranging principle: let the arrival times of nodes A and B be... and The length of the line between the two nodes is The speed of travel wave propagation is Then the distance from the disturbance source to node A is The calculation results from multiple neighboring nodes are combined, and the average value is taken as the distance from the disturbance source to the reference node, thus determining the direction of the disturbance source. The final output branch tracing result is a result object containing the direction of the disturbance source and the location confidence level.

[0093] 2. When selecting the source location branch based on harmonic impedance calculation, the harmonic impedance on the system side and the harmonic emission level on the user side are calculated using the harmonic voltage and harmonic current at each monitoring point, and the node with the highest harmonic emission level is identified as the harmonic source.

[0094] The source localization branch based on harmonic impedance calculation identifies the main harmonic sources by measuring harmonic voltage and current and using the circuit equivalence principle to calculate the contribution of each node to system harmonic pollution. When a harmonic-dominant disturbance is identified, the workflow platform extracts the three-phase harmonic voltage and current at each monitoring point during the disturbance, and simultaneously obtains the fundamental voltage and current during steady-state operation before the disturbance. For each monitoring node, the system-side harmonic impedance is calculated using bilinear regression: the change in harmonic voltage over a certain period is selected. Harmonic current variation Establish a linear regression equation ,in This is an estimate of the system-side harmonic impedance. After obtaining the system-side harmonic impedance, the user-side harmonic emission level is then calculated: User-side harmonic voltage emission level = Measured harmonic voltage at monitoring point - Measured harmonic current. This emission level reflects the contribution of downstream (user-side) equipment to harmonics at this node. The workflow platform repeats the above calculation for all monitored nodes to obtain the harmonic emission level of each node. The node with the highest harmonic emission level is identified as the main harmonic source, and the harmonic emission level value of this node and the contribution percentage of each harmonic are output. To improve the reliability of the location, the calculation results of multiple harmonic orders can be combined. If multiple orders all point to the same node, the confidence level is higher. The branch tracing results include the identification of the equipment or feeder where the harmonic source is located, the emission level of each harmonic, and the overall confidence level.

[0095] 3. When selecting the traveling wave time of arrival method branch based on waveform similarity, the arrival time of the disturbance traveling wave is extracted, the distance between the disturbance source and the monitoring point is calculated using the multi-terminal ranging formula, and the optimal match is selected by combining waveform similarity.

[0096] The waveform similarity-based traveling wave arrival time difference method utilizes the time difference between the arrival times of transient traveling waves generated by disturbances along the line to different monitoring points, combined with waveform correlation, to accurately locate the disturbance source. This method is particularly suitable for transient oscillatory disturbances. When a transient oscillatory disturbance is identified, the workflow platform performs high-frequency filtering on the voltage or current waveforms recorded at each monitoring node to extract the transient traveling wave component. Then, the arrival time of the traveling wave is detected using the wavelet transform modulus maxima method: a discrete wavelet transform is performed on the filtered waveform, and the modulus maxima point is found in the detail coefficients from level 1 to level 3; the time corresponding to this point is the arrival time of the traveling wave. For at least three monitoring nodes, the location of the disturbance source is calculated using a multi-terminal ranging formula. Taking three nodes A, B, and C as an example, let the geographical coordinates of node A be... The coordinates of node B are The coordinates of node C are The arrival times of the traveling waves are respectively , , The speed of travel wave propagation is The time of the disturbance is Then the coordinates of the disturbance source. The following system of equations must be satisfied:

[0097]

[0098] in, , The coordinates of monitoring node A on a two-dimensional plane are represented by the horizontal and vertical coordinates, which are obtained by conversion based on the power grid geographic information system or the line tower number. For unknown quantities, and These together constitute three unknowns, and the three equations are solvable. The coordinates of the disturbance source are obtained by solving this system of nonlinear equations using the least squares method. The distance from the disturbance source to each monitoring point is then calculated. Since noise and measurement errors can lead to inconsistent ranging results, waveform similarity is used to select the optimal match. Specifically, the theoretical traveling wave waveform at each monitoring point is derived from the obtained disturbance source coordinates and cross-correlated with the actual detected traveling wave waveform to obtain the cross-correlation coefficient. A higher cross-correlation coefficient (closer to 1) indicates a more reliable ranging result. For multiple possible ranging results, the combination with the highest average cross-correlation coefficient is selected as the final result. The branch-source tracing results include the specific coordinates of the disturbance source or its distance relative to the reference point, as well as a matching score based on waveform similarity.

[0099] 4. When selecting the weighted fusion strategy branch, calculate the source location confidence scores output by the algorithms corresponding to the propagation path deduction branch based on energy flow direction, the source location branch based on harmonic impedance calculation, and the traveling wave arrival time difference method branch based on waveform similarity, respectively. Then, perform a weighted average according to the preset weights to obtain the final disturbance source location.

[0100] The weighted fusion strategy branch is used to handle complex disturbances. When multiple characteristics such as sags, harmonics, or transient oscillations coexist, a single source tracing algorithm may be inaccurate, requiring the integration of results from multiple algorithms. When a complex disturbance is identified, the workflow platform calls the three source tracing sub-algorithms in parallel, obtaining the source location candidate set and the confidence level of each candidate location from the output of each algorithm. The confidence level is automatically calculated internally by each algorithm based on factors such as data quality, the number of nodes involved in the calculation, and measurement errors, with a value ranging from 0 to 1. Preset weights are determined based on actual power grid operation experience or offline training: for example, for complex disturbances dominated by sags, the weight of the energy flow algorithm is set to 0.5, the harmonic impedance algorithm to 0.3, and the traveling wave time difference method to 0.2; for complex disturbances dominated by harmonics, the harmonic impedance algorithm is set to 0.6, and the others to 0.2; for general cases that are difficult to distinguish, equal weights can be used. The specific weight values ​​are adjusted by maintenance personnel based on the on-site effects in the workflow platform's visualization interface, and automatic learning through algorithm optimization based on historical data is also supported.

[0101] The weighted average is calculated as follows: for each candidate location, its confidence score across all algorithms is multiplied by the corresponding algorithm's weight, and then summed to obtain the overall score for that location. For example, if a candidate location has a confidence score of 0.8 in the energy flow algorithm, 0.6 in the harmonic impedance algorithm, and 0.4 in the traveling wave time difference method, with preset weights of 0.4, 0.3, and 0.3 respectively, then its overall score is 0.62. All candidate locations are sorted from highest to lowest overall score, and the location with the highest score is selected as the final disturbance source location, and its overall confidence score is output. The branch tracing results include a description of the final disturbance source location, intermediate results from each sub-algorithm, and the fused confidence score.

[0102] S23. Perform hierarchical event mapping, aggregate the abnormal indicators of each monitoring node into disturbance events according to spatiotemporal clustering, and construct a disturbance propagation chain in the form of a directed acyclic graph based on the branch tracing results;

[0103] In this embodiment, hierarchical event mapping is performed, and the abnormal indicators of each monitoring node are aggregated into disturbance events according to spatiotemporal clustering, including the following:

[0104] Construct a four-layer exception event model, in which:

[0105] The L0 layer is the raw waveform data layer, which stores the three-phase voltage waveform data and three-phase current waveform data of each monitoring node;

[0106] The L1 layer is the feature index layer. The sag depth, total harmonic distortion rate, transient high-frequency energy and disturbance arrival time of each monitoring node are extracted from the L0 layer as anomaly indicators.

[0107] Layer L2 is the perturbation event layer, which aggregates the abnormal indicators in Layer L1 that meet the spatiotemporal clustering conditions into a perturbation event;

[0108] Layer L3 is the source localization layer, used to store the perturbation propagation chain and root cause nodes;

[0109] The spatiotemporal clustering conditions here include that the anomalous indicators within the time window are continuous in time and that the anomalous indicators of spatially adjacent monitoring nodes have the same disturbance type.

[0110] Specifically, the four-layer anomaly event model is a hierarchical abstract representation of power quality disturbances from raw data to final location results. Layer L0 is the raw waveform data layer, where the workflow platform stores the three-phase voltage and current sample values ​​uploaded by each monitoring terminal in chronological order. Each sample point includes a timestamp and amplitude, preserving complete waveform information. Layer L1 is the feature index layer, where the workflow platform calculates the raw waveforms from Layer L0, extracting the sag depth, total harmonic distortion rate, transient high-frequency energy, and disturbance arrival time for each monitoring node; these anomaly indicators are the foundational data for subsequent aggregation and source tracing. Layer L2 is the disturbance event layer, where the workflow platform aggregates the anomaly indicators of all monitoring nodes in Layer L1 according to spatiotemporal clustering conditions. In this spatiotemporal clustering condition, the anomaly indicators within the time window are continuous in time, meaning that the anomaly indicators of the same monitoring node or multiple adjacent monitoring nodes appear consecutively in time or at intervals shorter than the set time window. The anomaly indicators of spatially adjacent monitoring nodes have the same disturbance type, meaning that the two monitoring nodes are directly connected in the power grid topology and, based on the disturbance mode discrimination result, are both determined to be of the same disturbance type. When the above conditions are met, the workflow platform classifies these anomaly indicators into the same disturbance event. For example, if a sag depth exceeding 85% is detected on the 10 kV outgoing line side of substation A and at node B 2 kilometers downstream, and the disturbance mode discrimination result is sag-dominant in both cases within five milliseconds, then the anomaly indicators of nodes A and B are aggregated into the same disturbance event and assigned a unique event identifier. Layer L3 is the source localization layer, used to store the disturbance propagation chain generated after source tracing analysis and the finally determined root cause node information, providing data support for the result output.

[0111] In this embodiment, the perturbation propagation chain in the form of a directed acyclic graph is constructed as follows:

[0112] 1. Treat each L2 layer perturbation event as a node, and determine the directed edges between nodes based on the temporal sequence of events, energy decay gradient, and waveform similarity;

[0113] 2. Using the branch tracing results as the confidence weights of the edges, we obtain a perturbation propagation chain in the form of a weighted directed acyclic graph.

[0114] The purpose of constructing a disturbance propagation chain is to organize individual disturbance events into a directed acyclic graph (DAG) according to their spatiotemporal relationships, thereby visually displaying the propagation path of disturbances in the power grid and locating the source. The workflow platform first treats each L2-level disturbance event aggregated in step S23 as a node. Node attributes include event occurrence time, disturbance type, a list of affected monitoring nodes, sag depth, or harmonic content, among other characteristic values. Then, for any two disturbance event nodes... and Determine whether it is necessary to follow these three criteria. and Establish a directed edge between them and the direction of the edge:

[0115] Timing criterion: Compare the arrival times of two disturbance events; if the event... The arrival time was earlier than the event If the arrival time of the two is within the preset propagation time window, then a propagation connection is established from... point to A directed edge represents a perturbation from... spread to .

[0116] Energy attenuation gradient criterion: Compare the sag depth or harmonic content amplitude of two disturbance events at the same monitoring node or adjacent nodes. If the event The descent depth at the corresponding monitoring node is greater than the event. If the descent depth at the corresponding monitoring node is greater than 0.02 per unit, then a system is established from... point to A directed edge represents the attenuation propagation of energy from a node with a high amplitude to a node with a low amplitude. For harmonic-dominated disturbances, comparing the total harmonic distortion rate, nodes with high amplitudes point towards nodes with low amplitudes.

[0117] Waveform similarity criterion: Calculate the cross-correlation coefficient of waveforms of two disturbance events at overlapping monitoring nodes. Event extraction. and events Cross-correlation calculations are performed on waveform segments at common monitoring nodes to obtain cross-correlation coefficients. .if If the similarity exceeds the threshold of 0.85, then a similarity record is established from... point to A directed edge indicates that two events have the same waveform characteristics and belong to the same propagation chain.

[0118] Assume there are three monitoring nodes N1, N2, and N3 in the power grid, and a sag event occurs. The workflow platform aggregates three disturbance events: Event A occurs at N1, time 10:23:45.123, sag depth 0.85; Event B occurs at N2, time 10:23:45.135, sag depth 0.80; Event C occurs at N3, time 10:23:45.140, sag depth 0.78. Based on the time sequence, Event A precedes Event B, which precedes Event C; therefore, directed edges A→B and B→C are established. Based on the energy decay gradient, the sag depth decreases by 0.05 from A to B and by 0.02 from B to C, both greater than the gradient threshold of 0.01, further confirming the propagation direction. Waveform similarity is calculated; the cross-correlation coefficient between Event A and Event B on N1 and N2 is 0.92, and the cross-correlation coefficient between Event B and Event C is 0.89, both greater than 0.85, further confirming the propagation direction. The final constructed directed graph path is A→B→C, and there are no reverse edges, forming a directed acyclic graph. The output of the branch-source algorithm in step S22 is used as the confidence weight of each edge. For example, the weight of edge A→B is set to 0.95, and the weight of edge B→C is set to 0.90, resulting in a perturbation propagation chain in the form of a weighted directed acyclic graph. In this propagation chain, node A with an in-degree of zero is the root cause node, corresponding to the device or bus where the perturbation source is located.

[0119] S24. Input the perturbation feature vector and perturbation propagation chain into the trained target mixture model, and output the corrected perturbation propagation chain and root cause node from the target mixture model. The root cause node corresponds to the specific device where the perturbation source is located.

[0120] In this embodiment, the target hybrid model includes:

[0121] The feature extraction part consists of a multi-layer convolutional neural network, which is used to extract deep waveform features from the perturbation feature vector;

[0122] The propagation chain correction part, consisting of a fully connected network, is used to receive the graph embedding vector and depth waveform features of the perturbation propagation chain, and output the correction weights of each edge in the perturbation propagation chain and the confidence of the root cause node.

[0123] The target hybrid model here is a pre-trained deep learning model with an architecture consisting of two parts. The feature extraction part employs a three-layer convolutional neural network. Each layer contains a 3×3 convolutional layer, a batch normalization layer, a linear rectified unit activation function, and a max-pooling layer with a stride of 2. The input is the perturbation feature vector obtained in step S21, which has a length of 768 dimensions (256 dimensions for each of the three phases). The convolutional neural network extracts local features through sliding convolutional kernels, abstracting layer by layer, and finally outputs a 128-dimensional depth waveform feature vector through a global average pooling layer. The propagation chain correction part uses a three-layer fully connected network. The first layer has an input dimension of 128 (depth waveform features) plus the dimension of the graph embedding vector, totaling 192 dimensions. The number of hidden layer nodes is 128 and 64 respectively. The number of output layer nodes equals the number of edges in the perturbation propagation chain plus the number of root cause node classifications. The fully connected network uses random deactivation regularization to prevent overfitting. During model training, the labeled dataset accumulated in step S25 is used, and the sum of cross-entropy loss and edge correction loss is used as the total loss function for optimization. After training, the model parameters are fixed and deployed in the inference service of the workflow platform. When online inference is required, the workflow platform calls the model, inputs the feature vector and propagation chain of the current perturbation event, and the model performs forward computation to obtain the output.

[0124] In this embodiment, the target mixture model outputs the corrected perturbation propagation chain and root cause node, specifically including the following steps:

[0125] 1. Input the perturbation feature vector into a multi-layer convolutional neural network to obtain deep waveform features;

[0126] The multilayer convolutional neural network has been described in the model architecture. After inputting a perturbation feature vector, it outputs a 128-dimensional depth waveform feature vector through forward computation.

[0127] 2. Encode the perturbation propagation chain as a graph embedding vector;

[0128] The graph embedding vector is the process of converting the perturbation propagation chain (directed acyclic graph) constructed in step S23 into a fixed-length vector.

[0129] 3. The graph embedding vector and depth waveform features are concatenated and then input into the fully connected network;

[0130] The node features and adjacency matrix of the graph are input into a two-layer graph convolutional network, with each layer outputting 32 dimensions. Finally, a 64-dimensional graph embedding vector is obtained through global summation. The 128-dimensional depth waveform features and the 64-dimensional graph embedding vector are concatenated in dimensions to form a 192-dimensional combined feature vector.

[0131] 4. The fully connected network outputs corrected weights for each edge in the perturbation propagation chain;

[0132] The fully connected network receives the combined feature vector, performs computation through three fully connected layers, and outputs a vector of length N, where N equals the number of directed edges in the perturbation propagation chain. Each element of this vector has a value between 0 and 1, representing the corrected weight of the corresponding edge.

[0133] 5. Adjust the confidence level of each edge according to the corrected weights, and delete the edges with confidence levels lower than the preset threshold to obtain the corrected perturbation propagation chain;

[0134] The workflow platform reads the corrected weights output by the fully connected network and multiplies them by the original confidence weights of the edges obtained in step S23 to obtain the updated edge weights. A preset threshold of 0.5 is set. When the updated edge weight is lower than 0.5, the edge is considered unreliable and is removed from the perturbation propagation chain. The preset threshold of 0.5 is obtained based on model performance tuning on the validation set. A threshold below 0.5 introduces more false positive edges, while a threshold above 0.5 loses weak but true edges. Experiments have verified that 0.5 is the optimal balance. After deleting low-confidence edges, the corrected perturbation propagation chain is obtained.

[0135] 6. Select nodes with an in-degree of zero from the corrected disturbance propagation chain as root cause nodes. The root cause node corresponds to the specific device where the disturbance source is located.

[0136] In the corrected directed acyclic graph, the in-degree (the number of edges pointing to that node) of each node is calculated. A node with an in-degree of zero indicates that no other node points to it, which is the starting point of the disturbance propagation. Therefore, the device or feeder corresponding to that node is identified as the specific device where the disturbance source is located. For example, if the in-degree of node A in the corrected propagation chain is 0, and the in-degrees of nodes B and C are 1, then node A is the root cause node, and the name of its associated device is output.

[0137] Through the above steps, the target hybrid model utilizes deep learning and graph embedding techniques to automatically correct the propagation chain constructed based on physical rules, thereby improving the robustness and accuracy of root cause localization.

[0138] S25. Output the source tracing results and add the current disturbance event and its corresponding source tracing results to the labeled dataset.

[0139] After completing the above steps, the workflow platform outputs the tracing results to operations and maintenance personnel in the form of a visual report. The tracing results include at least the following: a unique identifier for the disturbance event, the occurrence time, the disturbance type, the specific device or feeder name corresponding to the root cause node, a directed acyclic graph of the disturbance propagation chain, and a comprehensive confidence level. Output formats include pop-up prompts on the workflow platform's user interface, automatically generated PDF reports, and email pushes to designated operations and maintenance personnel. Simultaneously, the workflow platform adds this disturbance event and its corresponding tracing results to the labeled dataset. The labeled dataset is a structured database table stored in the workflow platform's relational database or data lake. Each record contains the following fields: event identifier, disturbance occurrence time, original waveform data file path for each monitoring node, extracted disturbance feature vector, branch tracing results, graph structure of the disturbance propagation chain, corrected propagation chain and root cause node output by the target hybrid model, and the final label after manual review.

[0140] Through the above methods, the present invention realizes a complete process from data collection, feature extraction, pattern discrimination, source tracing analysis to result feedback, and the source tracing performance is continuously improved as data accumulates.

[0141] It is understood that those skilled in the art can combine various implementation methods in the above embodiments under the guidance of the above examples to obtain technical solutions with multiple implementation methods.

[0142] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for tracing the source of power quality disturbances using an intelligent diagnostic model based on a workflow platform, characterized in that, include: By deploying power quality monitoring terminals at multiple monitoring nodes in the power grid, waveform data of each monitoring node is collected in real time. The waveform data includes three-phase voltage waveform data and three-phase current waveform data. When the power quality index of any monitoring node exceeds a preset threshold based on the waveform data, the workflow platform executes the following steps: Adaptive feature extraction is performed on the waveform data to obtain a disturbance feature vector; Based on the perturbation feature vector, the perturbation mode is determined, and the corresponding source tracing sub-algorithm branch is selected according to the determination result. The source tracing sub-algorithm corresponding to the selected branch is then used for processing to obtain the branch source tracing result. Hierarchical event mapping is performed, and abnormal indicators of each monitoring node are aggregated into perturbation events by spatiotemporal clustering. A perturbation propagation chain in the form of a directed acyclic graph is constructed based on the branch tracing results. This includes: Construct a four-layer exception event model, in which: The L0 layer is the raw waveform data layer, which stores the three-phase voltage waveform data and three-phase current waveform data of each monitoring node; The L1 layer is the feature index layer. The sag depth, total harmonic distortion rate, transient high-frequency energy and disturbance arrival time of each monitoring node are extracted from the L0 layer as anomaly indicators. Layer L2 is the perturbation event layer, which aggregates the abnormal indicators in Layer L1 that meet the spatiotemporal clustering conditions into a perturbation event; Layer L3 is the source localization layer, used to store the perturbation propagation chain and root cause nodes; The spatiotemporal clustering conditions include that the abnormal indicators within the time window are continuous in time and that the abnormal indicators of spatially adjacent monitoring nodes have the same disturbance type. The perturbation feature vector and the perturbation propagation chain are input into the trained target hybrid model, and the target hybrid model outputs the corrected perturbation propagation chain and root cause node, wherein the root cause node corresponds to the specific device where the perturbation source is located; Output the source tracing results and add the current disturbance event and its corresponding source tracing results to the labeled dataset.

2. The intelligent diagnostic model-based power quality disturbance tracing method based on a workflow platform according to claim 1, characterized in that, The adaptive feature extraction of the waveform data to obtain the disturbance feature vector includes: The three-phase voltage waveform data and the three-phase current waveform data are respectively compared with each template in the preset standard disturbance template library to perform intersection point distribution analysis, and the degree of matching between each phase waveform data and the corresponding template in the time domain is determined.

3. The intelligent diagnostic model-based power quality disturbance tracing method based on a workflow platform according to claim 2, characterized in that, Based on the intersection distribution type of the waveform data of each phase, the waveform data of each phase is divided into complete matching, partial matching, no overlap but with intersection, and complete deviation; among which: When the waveform data of a certain phase is divided into complete matches, the preset feature vector of the corresponding template is reused as the single-phase feature vector of that phase. When a phase waveform data is divided into local matching, the difference between the phase waveform data and the template is extracted by a sliding window, and the incremental feature vector of the phase is generated based on the difference. When a phase waveform data is divided into non-overlapping and overlapping sections, the slope change rate and distortion energy near the intersection point of the phase waveform data and the template are extracted to construct the cross feature vector of the phase. When the waveform data of a certain phase is completely deviated, the waveform data of that phase is marked as a new category candidate, and the complete time-frequency domain statistical features of that phase are extracted as the single-phase feature vector of that phase; The disturbance characteristic vector is obtained by concatenating the single-phase characteristic vectors of the three phases in phase order.

4. The intelligent diagnostic model-based power quality disturbance tracing method based on a workflow platform according to claim 1, characterized in that, The step of determining the perturbation pattern based on the perturbation feature vector and selecting the corresponding source tracing sub-algorithm branch according to the determination result includes: The transient sag depth, total harmonic distortion rate, and transient high-frequency energy in the disturbance feature vector are used as discriminant features. When the sag depth exceeds the first threshold and the total harmonic distortion rate is lower than the second threshold, it is determined to be a sag-dominated type, and a propagation path deduction branch based on energy flow direction is selected. When the total harmonic distortion rate exceeds the third threshold and the sag depth is lower than the fourth threshold, it is determined to be harmonic-dominated, and the source localization branch based on harmonic impedance calculation is selected. When the transient high-frequency energy exceeds the fifth threshold and the transient sag duration is less than the sixth threshold, it is determined to be a transient oscillation type, and the traveling wave arrival time difference method based on waveform similarity is selected as the branch. When the sag depth exceeds the first threshold and the total harmonic distortion exceeds the third threshold, it is determined to be a composite type, and the weighted fusion strategy branch is selected. Wherein, the first threshold is greater than the fourth threshold, and the second threshold is less than the third threshold.

5. The intelligent diagnostic model-based power quality disturbance tracing method based on a workflow platform according to claim 4, characterized in that, The first to sixth thresholds are adjusted adaptively. Thresholds are automatically set based on historical disturbance data using quantile statistics.

6. The intelligent diagnostic model-based power quality disturbance tracing method based on a workflow platform according to claim 4, characterized in that, The process of using the source tracing sub-algorithm corresponding to the selected branch to obtain the branch tracing result includes: When selecting a propagation path deduction branch based on energy flow direction, the direction and distance of the disturbance source are determined by deducing in reverse along the energy transmission based on the attenuation gradient of the transient amplitude at each monitoring point and the disturbance arrival time difference. When selecting the source location branch based on harmonic impedance calculation, the harmonic impedance on the system side and the harmonic emission level on the user side are calculated using the harmonic voltage and harmonic current at each monitoring point, and the node with the highest harmonic emission level is identified as the harmonic source. When the waveform similarity-based traveling wave time of arrival method branch is selected, the arrival time of the disturbance traveling wave is extracted, the distance between the disturbance source and the monitoring point is calculated through the multi-terminal ranging formula, and the optimal match is selected by combining waveform similarity. When selecting the weighted fusion strategy branch, the source location confidence scores output by the algorithms corresponding to the propagation path deduction branch based on energy flow direction, the source location branch based on harmonic impedance calculation, and the traveling wave arrival time difference method branch based on waveform similarity are calculated respectively. The scores are then weighted and averaged according to preset weights to obtain the final disturbance source location.

7. The intelligent diagnostic model-based power quality disturbance tracing method based on a workflow platform according to claim 1, characterized in that, The construction of a perturbation propagation chain in the form of a directed acyclic graph based on the branch tracing results includes: Each perturbation event in the L2 layer is treated as a node, and the directed edges between nodes are determined based on the temporal sequence of events, energy decay gradient, and waveform similarity. Using the branch tracing results as the confidence weights of the edges, a perturbation propagation chain in the form of a weighted directed acyclic graph is obtained.

8. The method for tracing the source of power quality disturbances using an intelligent diagnostic model based on a workflow platform according to claim 1, characterized in that, The target hybrid model includes: The feature extraction part is composed of a multi-layer convolutional neural network, which is used to extract deep waveform features from the perturbation feature vector; The propagation chain correction part is composed of a fully connected network, which is used to receive the graph embedding vector of the perturbation propagation chain and the depth waveform features, and output the correction weights of each edge in the perturbation propagation chain and the confidence of the root cause node.

9. The intelligent diagnostic model-based power quality disturbance tracing method based on a workflow platform according to claim 8, characterized in that, The step of inputting the perturbation feature vector and the perturbation propagation chain into the trained target mixture model, and having the target mixture model output the corrected perturbation propagation chain and root cause node, includes: The perturbation feature vector is input into the multi-layer convolutional neural network to obtain deep waveform features; The disturbance propagation chain is encoded as a graph embedding vector; The graph embedding vector and the depth waveform feature are concatenated and then input into the fully connected network. The fully connected network outputs corrected weights for each edge in the perturbation propagation chain; The confidence level of each edge is adjusted according to the corrected weight, and edges with confidence levels lower than a preset threshold are deleted to obtain the corrected perturbation propagation chain. Nodes with an in-degree of zero are selected from the corrected disturbance propagation chain as root cause nodes, and the root cause nodes correspond to the specific devices where the disturbance source is located.

Images