Analyzing pre-fault WMU current data to detect vegetation causing faults in electrical power delivery system

Abstract

A system and method predict whether a fault in an electrical power delivery system is caused by vegetation contact. Synchro-waveform data is received from a waveform measurement unit (WMU) that captures time-synchronized measurements indicative of electrical disturbances. Detected fault patterns are classified as line-to-ground or line-to-line events. Pre-fault data preceding the classified fault is extracted and processed through a feature transformation pipeline applying a direct quadrature zero (DQZ) transformation, a rolling root mean square (RMS) computation, and a Multiple Seasonal-Trend decomposition using LOESS (MSTL) to isolate signal components and extract trends. The resulting trend data is divided into sections, scaled for uniformity, and input to a gradient boosting decision tree model to predict vegetation contact as the fault cause. A notification identifying the fault as vegetation-caused and its estimated location is presented on a user interface with a visual representation of the waveform data.

Description

Docket No. 503461.70051ANALYZING PRE-FAULT WMU CURRENT DATA TO DETECT VEGETATION CAUSING FAULTS IN ELECTRICAL POWER DELIVERY SYSTEMRELATED APPLICATION

[0001] This application claims priority benefit of U.S. Provisional Patent Application No.63 / 777,619, filed March 25, 2025, which is hereby incorporated by reference in its entirety.TECHNICAL FIELD

[0002] This disclosure relates generally to power system diagnostics and, more particularly, to event type classification in electrical networks.BACKGROUND INFORMATION

[0003] In the IEEE Power & Energy Magazine, September / October 2023 edition (Vol. 21, No. 5), an article starting on page 68 by Hamed Mohsenian-Rad, Ph.D., and Wilsun Xu, Ph.D., titled “Synchro-Waveforms,” describes a power system measurement technology approach. This technology, referred to as a waveform measurement unit (WMU), obtains time-synchronized waveform measurements (i.e., synchro-waveforms) from different locations within a power system. WMUs can capture inconspicuous disturbances that are often overlooked by other types of time-synchronized sensors, such as phasor measurement units (PMUs). WMUs also monitor system dynamics at much higher frequencies, as well as much lower frequencies, than the fundamental components of voltage and current that are commonly monitored by PMUs.

[0004] Another paper by V. Barrera Nunez et al., titled “Feature Analysis and Classification Methodology for Overhead Distribution Fault Events,” presented in the IEEE / PES Transmission and Distribution Committee on July 25-29, 2010, describes the use of waveform data from power quality monitors (PQM) installed in substations. The paper presents an analysis of unique features and a classification method for identifying the root cause of overhead distribution fault events, which are multi-cycle events. In particular, it focuses on faults caused by animal contacts, tree contacts, and lightning-induced disturbances.

[0005] Additionally, there are a number of patents describing the calculation of fault locations. For example, U.S. Patent No. 8,941,387 of Kim describes a time-domain analysis for determining net fault electrical parameters, including inductive reactance.

[0006] FIG. 1 shows graphs from a paper titled, “Vegetation Conduction Ignition Tests,” by Marxsen Consulting Pty Ltd. The graphs show examples of how vegetation-caused faultsDocket No. 503461.70051in power systems are identified according to a specific sequence of four current change stages. These stages take over thirty seconds and include: development of full conductorvegetation contact; expulsion of moisture; progressive charring of bark extending from the thinner end of the vegetation sample; and flashover when flame bridges the high voltage conductors. Current change stage 1 begins with the development of power line-tree contact, indicated by a slow increase in current. This progresses to stage 2, where moisture expulsion (characterized by a rapid decrease in current when water, initially acting as a resistor, is removed) is observed. A fire risk is present in stage 3, where the outer layer of the branch progressively burns, potentially causing current fluctuations due to intermittent arcing when portions of the conduction path short-circuit. The final stage 4 is marked by a flashover, the main event where the fault becomes fully developed.SUMMARY OF THE DISCLOSURE

[0007] In light of the conventional methods outlined, which primarily focus on the gradual development of vegetation-caused faults as detailed in FIG. 1, there remains a need for automated detection of vegetation-caused faults that can leverage real-time data analysis to predict and mitigate such faults more efficiently. This disclosure, therefore, describes a system that integrates advanced machine learning algorithms and data processing techniques. The proposed system not only accelerates the detection of vegetation-caused faults but also enhances the accuracy and predictive power of these detections, transitioning from reactive to proactive management of power system integrity.

[0008] In some embodiments, a method for determining whether a fault in an electrical power delivery system is caused by vegetation contact includes receiving synchro-waveform data from a waveform measurement unit positioned within the electrical power delivery system, where the waveform measurement unit captures time-synchronized waveform measurements indicative of electrical disturbances. Patterns of current changes are detected from the synchro-waveform data to identify a fault, and identified fault patterns are classified into event types including line-to-ground faults and line-to-line faults. In response to a classified fault having a line-to-ground or line-to-line fault type, pre-fault data immediately preceding the classified fault is extracted, and feature data is generated from the pre-fault data for input to a machine learning model configured to determine whether vegetation contact caused the classified fault. In response to determining that vegetation contact is the cause of the classified fault, a notification is generated identifying the fault as vegetation-caused and indicating an estimated location of the fault within the electricalDocket No. 503461.70051power delivery system, and the notification is presented on a user interface alongside a visual representation of waveform data corresponding to the fault.

[0009] In some embodiments, generating the feature data includes applying a direct quadrature zero transformation to isolate reactive and zero-sequence components from the pre-fault data. In some embodiments, generating the feature data further includes applying a rolling root mean square computation to reduce noise and enhance detection of trends in current components derived from the pre-fault data. In some embodiments, the rolling root mean square computation uses a window size of 256 and a step size of 16. In some embodiments, generating the feature data includes processing data derived from the prefault data using Multiple Seasonal-Trend decomposition using LOESS to extract and separate trends and seasonal patterns from waveform data. In some embodiments, the prefault data comprises about ten seconds of synchro-waveform data immediately preceding the classified fault.

[0010] In some embodiments, generating the feature data includes deriving a fault phase current component, a reactive current component, and a zero-sequence current component from the pre-fault data. In some embodiments, trend data derived from the pre-fault data is divided into ten sections. In some embodiments, the ten sections correspond to equalduration intervals of the pre-fault data. In some embodiments, a section mean is calculated for each of the ten sections. In some embodiments, section values are scaled by dividing the section values by a mean of an initial set of sections, and in further embodiments the initial set comprises a first three sections. In some embodiments, the feature data includes thirty features comprising, for each of the ten sections, a scaled trend section mean of fault current, a scaled trend section mean of zero-sequence current, and a scaled trend section mean of reactive current. In some embodiments, the machine learning model comprises a gradient boosting decision tree model.

[0011] In some embodiments, cumulative reactance to the fault is determined and used to estimate the location of the fault within the electrical power delivery system. In some embodiments, the estimated location is presented on a geographic visualization. In some embodiments, the classified fault is grouped with one or more additional events based on shared waveform characteristics, location estimates, phases, or timing. In some embodiments, the notification is presented in an event overview timeline user interface. In some embodiments, the notification is presented together with an original signal user interface showing voltage waveforms, current waveforms, and a reactance-to-fault value corresponding to the classified fault. In some embodiments, synchro-waveform data is stored in one-hour periods before extracting the pre-fault data. In some embodiments, aDocket No. 503461.70051machine learning model is trained using a curated dataset of historical events grouped by event type and outage information. In some embodiments, the dataset includes only a first event for each outage. In some embodiments, a vegetation management action is initiated based on the estimated location.

[0012] In some embodiments, a system for determining whether a fault in an electrical power delivery system is caused by vegetation contact includes a waveform measurement unit positioned within the electrical power delivery system and configured to capture synchro-waveform data indicative of electrical disturbances, and a waveform measurement unit data analysis system including one or more processors and memory storing instructions that, when executed, cause the waveform measurement unit data analysis system to receive the synchro-waveform data, detect patterns of current changes that indicate a fault, classify identified fault patterns into event types including line-to-ground faults and line-to-line faults, extract pre-fault data from the synchro-waveform data immediately preceding a classified fault, generate feature data from the pre-fault data for input to a machine learning model configured to determine whether vegetation contact caused the classified fault, generate a notification identifying the fault as vegetation-caused and indicating an estimated location of the fault, and present the notification on a user interface alongside a visual representation of waveform data corresponding to the fault.

[0013] In some embodiments, a non-transitory computer-readable medium stores instructions that, when executed by one or more processors, cause the one or more processors to receive synchro-waveform data from a waveform measurement unit positioned within an electrical power delivery system, detect patterns of current changes that indicate a fault, classify identified fault patterns into event types including line-to-ground faults and line-to-line faults, extract pre-fault data from the synchro-waveform data immediately preceding a classified fault, generate feature data from the pre-fault data for input to a machine learning model configured to determine whether vegetation contact caused the classified fault, generate a notification identifying the fault as vegetation-caused and indicating an estimated location of the fault, and present the notification on a user interface alongside a visual representation of waveform data corresponding to the fault.

[0014] Additional aspects and advantages will be apparent from the following detailed description of embodiments, which proceeds with reference to the accompanying drawings.BRIEF DESCRIPTION OF THE SEVERAL VIEWS OF THE DRAWINGSDocket No. 503461.70051

[0015] To easily identify the discussion of any particular element or act, the most significant digit or digits in a reference number refer to the figure number in which that element is first introduced.

[0016] FIG. 1 shows example plots of electrical current exhibiting vegetation-caused fault in accordance with the prior art.

[0017] FIG. 2 is a block diagram of a WMU network architecture in accordance with one embodiment.

[0018] FIG. 3 is a block diagram showing an electrical power delivery system and WMU synchro-waveform event analysis in accordance with one embodiment.

[0019] FIG. 4 is an electrical schematic diagram of a single phase representation of the electrical power delivery system in FIG. 3, with WMUs showing an event in accordance with one embodiment.

[0020] FIG. 5 is a block diagram showing event classification in accordance with one embodiment.

[0021] FIG. 6 is a set of processed data preparatory to event classification in accordance with one embodiment.

[0022] FIG. 7 is a block diagram showing a dataset for event classification modeling and verification in accordance with one embodiment.

[0023] FIG. 8 is a block diagram showing how event type time and phasor domain characteristics data is input to machine learning algorithms for event type classification in accordance with one embodiment.

[0024] FIG. 9 is a set of waveforms showing an example of fault grouping in accordance with one embodiment.

[0025] FIG. 10 is a screenshot of a user interface for viewing grouped data in accordance with one embodiment.

[0026] FIG. 11 is a screenshot of an original signal user interface in accordance with one embodiment.

[0027] FIG. 12 is a flow diagram of a process for ingesting and processing data for determining whether a fault was caused by vegetation contact in accordance with one embodiment.

[0028] FIG. 13 A and FIG. 13B (collectively, FIG. 13) are a flow diagram of a process showing in greater detail the data processing of FIG. 12 for determining whether a fault was caused by vegetation contact in accordance with one embodiment.Docket No. 503461.70051

[0029] FIG. 14 is a set of waveforms for a vegetation contact prediction in accordance with one embodiment.

[0030] FIG. 15 is a data visualization showing a location of the vegetation contact prediction for the waveforms of FIG. 14.

[0031] FIG. 16 is a set of waveforms for a vegetation contact prediction in accordance with one embodiment.

[0032] FIG. 17 is a data visualization showing a location of the vegetation contact prediction for the waveforms of FIG. 16.

[0033] FIG. 18 is a block diagram showing components of a WMU or server in accordance with one embodiment.DETAILED DESCRIPTION OF EMBODIMENTS

[0034] This disclosure describes techniques for monitoring, analyzing, and responding to faults in electrical power delivery systems, including temporary faults (sometimes called “transitory” or “incipient” faults) and sustained faults (which lead to service outages). In many instances, the techniques also capture a broad range of grid disturbance events that may indicate developing problems in the system.

[0035] A temporary fault is a disturbance that triggers a protective mechanism (such as a recloser operation) but subsequently clears on its own. Because the underlying cause, such as a momentary line contact or minor insulation breakdown, is transient, the network can self-recover and continue operating. When these transient or incipient issues are detected early, proactive measures can prevent them from escalating into more serious conditions. In contrast, a sustained fault is one in which the protective mechanism remains open until the fault is physically removed or repaired, causing an outage or service disruption until corrective action is taken.

[0036] In some cases, smaller anomalies or precursors can be detected prior to either a more serious temporary or sustained fault. These minor disturbances are a type of temporary fault that are sometimes referred to as “incipient faults” or “precursor events.” These do not immediately cause a full interruption but can degrade conditions within the system and eventually evolve into a more significant fault. Examples of incipient faults include early-stage insulation breakdown in cables or intermittent contact caused by vegetation.Identifying and mitigating these precursors is central to effective predictive maintenance and helps avoid both minor and major disruptions.Docket No. 503461.70051

[0037] Additionally, the term PQ events encompasses disturbances such as voltage / current sags, swells, transients, harmonics, and frequency variations. These events can be transient (lasting milliseconds or seconds) or more prolonged, and they occasionally overlap with, precede, or result from either temporary or sustained faults. For instance, a voltage sag caused by a short but severe fault may resolve quickly, yet it still indicates a potentially vulnerable segment of the network. Electrical disturbance events include voltage or current sags, which are short-term reductions in magnitude often caused by fault conditions or sudden load changes and which may cause sensitive equipment to shut down or malfunction; voltage or current swells, which are temporary increases in magnitude caused by load changes or network failures; transients, which are short-duration, high-frequency disturbances caused by switching operations or lightning; harmonics, which are waveform distortions caused by nonlinear loads; and frequency variations, which are deviations from standard operating frequency caused by imbalances between generation and load.

[0038] Identifying and categorizing these events, whether they are “temporary” anomalies that self-clear or “sustained” faults leading to an outage, allows operators to coordinate more effectively with the broader predictive maintenance strategy and can significantly enhance system reliability. The data analysis techniques described herein process both continuously sampled waveforms (capturing early or subtle fault signatures) and discrete reports or alarms (indicating triggered protective devices or confirmed events) to locate the cause of an outage or incipient condition with precision.

[0039] FIG. 2 shows a WMU network architecture 200 for monitoring and analyzing faults within a power system, such as an electrical power delivery system 300 (FIG. 3). According to some embodiments, WMU network architecture 200 integrates subcomponents provided by a power transmission and distribution (T&D) company with infrastructure and software available from the applicant for this disclosure, Toumetis, Inc. of Boise, Idaho. This integrated system allows for comprehensive monitoring and analysis of the power grid, aiming to quickly and efficiently detect and respond to faults.

[0040] The main components shown in the example of FIG. 2 are a substation 202 and a WMU data analysis system 204, which are communicatively coupled via an internet connection 206. Internet connection 206 is available at substation 202 using cellular connection equipment 208 (e.g., LTE or 5G), fiber, or other communication devices. In some embodiments, WMU data analysis system 204 is a remotely located server (e.g., AWS server) such that the analysis is performed via cloud computing. In other embodiments, WMU data analysis system 204 may be located at an edge, such as within substation 202.Docket No. 503461.70051

[0041] In this example, substation 202 serves as the physical location where power system monitoring occurs. It contains equipment for controlling and monitoring the electricity flow in the power grid. Substation 202 also includes an internal LAN 210, which is communicatively coupled to cellular connection equipment 208 via wireless or ethernet connection 212 with a router 214.

[0042] LAN 210 interconnects all the monitoring and computing devices located in LAN 210. For instance, a substation server 216 includes local PC 218 with having an optional storage interface 220 for optional removable storage 222 and a wireless or wired router 214 connected via router interface 224 to local PC 218. Local PC 218 is configured with software processes 226 including a process 228 to get FTP data from devices and a process 230 push data to storage 232 in connection with cloud storage and processing system 234.

[0043] Cloud storage and processing system 234 is configured to read instructions from a machine-readable or computer-readable medium (e.g., a non-transitory machine-readable storage medium), including storage 232, and perform any one or more of the methods discussed herein (or portions thereof), such as discussed for process 1200 (FIG. 12), process 1300 (FIG. 13), or any portions of these embodiments.

[0044] Router 214 also connects to multiple WMU devices 236 via server connection 238. Multiple WMU devices 236 include monitoring device 240, which samples synchrowaveform data and provides it via an FTP interface 242 for process 228.

[0045] FIG. 3 illustrates an electrical power delivery system 300 monitored by WMU synchro-waveform event analysis 302. This monitoring and analysis capability uses continuously sampled voltage and current waveforms to detect, classify, and locate faults in electrical power delivery system 300, facilitating prompt responses to both temporary and sustained outages. For example, the system may classify various fault types, such as a tee connector fault, load break elbow fault, cable-to-ground fault, capacitor fault, or transformer fault, but other fault scenarios are possible as well.

[0046] Electrical power delivery system 300 includes a power generation plant 304 (e.g., a hydroelectric dam, wind farm, or thermal power plant), a step-up transmission substation 306, high-voltage transmission lines 308, a step-down transmission substation 310, a stepdown distribution substation 312, distribution network poles 314, and end users 316 (such as residential or commercial customers).

[0047] Within this system, WMU synchro-waveform event analysis 302 supports fault location estimation 318 by receiving and processing time-domain waveforms measured throughout the grid. Specifically, FIG. 3 shows a time-domain waveform 320 exhibiting aDocket No. 503461.70051temporary fault 322. The continuously sampled voltage and current measurements capture an event lasting less than a full cycle, which appears as a distinct spike or transient in the overlapping phase waveforms. By applying an analysis to this measurement data, the system produces fault location estimation 318. Here, the fault location is determined to lie on distribution line 324 between the step-down distribution substation 312 and the distribution network poles 314 or feeder lines 326.

[0048] Time-domain waveform 320 includes three overlapping waveforms, corresponding to multiple phase views of the electrical system. The transient voltage spikes on all three waveforms mark the precise, sub-cycle interval at which temporary fault 322 occurred. Additionally, techniques described in International Patent Application No.PCT / US2025 / 030681, titled “Phasor-Based Fault Location Estimation in Electrical Power Delivery System,” and incorporated by reference herein in its entirety, can be used in conjunction with the present disclosure. These techniques support multi-cycle or sub-cycle fault analysis, measuring waveform behavior at sub-cycle intervals to deliver high-resolution fault detection and location estimation.

[0049] By integrating both continuous waveform measurements and discrete fault monitoring inputs (e.g., event-driven alarms), the WMU synchro-waveform event analysis 302 can further pinpoint the source of an outage or transient event within electrical power delivery system 300 with greater speed and accuracy, thereby reducing system downtime and improving reliability.

[0050] FIG. 4 shows in greater detail a single phase 400 of electrical power delivery system 300 (FIG. 3). In this view, each node connection is monitored by a WMU. Skilled persons will appreciate, however, that one or more WMUs may be employed to monitor different locations along the circuit of single phase 400. For instance, a first WMU 402 measures an N1 node between a step-down distribution substation 312 (FIG. 3) and distribution network poles 314 (FIG. 3). Similarly, a second WMU 404 measures an N2 node between distribution network poles 314 and feeder lines 326. A third WMU 406 measures a node N3 between feeder lines 326 and residential or commercial users 316, and so forth. Examples of WMUs include Satec PM174 & PM180, Dranetz PQ3K, PQ5K, Encore, and PMI Phalanx.

[0051] In the scenario depicted in FIG. 4, a fault 408 is detected proximal to N2, subsequent to line fault Lfl. If this fault were a single line to ground (SLG), then the first WMU 402 at N1 might record a significant voltage drop coupled with a substantial increase in current. In contrast, the second WMU 404 at N2 would likely register a voltage dropDocket No. 503461.70051accompanied by a commensurate increase in current. These changes are analyzed to accurately detect and characterize the fault type.

[0052] Detection involves an assessment of electrical parameter changes captured by the WMUs, each set with specific voltage and current thresholds reflective of normal operating conditions for their respective network segments. When these thresholds are exceeded, such as a drop in voltage below 90% of typical levels or a rise in current above 120% of normal load, a fault alert is triggered. This initiates further analysis of the fault.

[0053] Accurate characterization of the event type reduces ambiguity in determining a fault location so as to reduce a number of candidate inspection sites in the electrical power delivery system for inspection by a maintenance team. Once data indicating a fault is captured, it is then used to determine the type of fault, which is also referred to as event classification.

[0054] FIG. 5 illustrates event type classification 500 used in an electrical power delivery system. Event type classification 500 captures synchro-waveform data 502, which includes real-time voltage and current measurements, from WMUs monitoring various segments of the power grid. Depending on operational needs, this data is either stored in a database 504 for later analysis or processed immediately in near real-time. The system identifies transitory electrical anomalies 506 within the data, characterized by their time-domain values such as magnitudes, rates of change, and the time of day. These anomalies are further analyzed to calculate phasor domain values for each half-cycle, including voltage and current phasors, phase angles, and impedances. This dual-domain analysis allows the WMU data analysis system to accurately determine the type of fault occurring, integrating both time-domain and phasor-domain characteristics. Additionally, the system classifies each event by type, including the persistence of the event (e.g., precursor, outage, or operational fault), inferred causes such as equipment failures, and the specific device type 508, such as cables, transformers, and capacitors.

[0055] FIG. 6 outlines the initial step in developing event type classification 500 depicted in previous figures. Here, both voltage and current data from three-phase waveforms are time-aligned and digitized, producing detailed time-series waveforms. These waveforms undergo analysis in both the time and phasor domains, focusing on characteristics such as magnitude, phase angle, frequency spectra, duration, and transient rate-of-change, achieved through advanced filtering and signal processing techniques. This processed data forms the foundation for training machine learning algorithms to enhance fault detection and classification.Docket No. 503461.70051

[0056] FIG. 7 demonstrates how a curated dataset is prepared using historical event data classified per the criteria established in FIG. 5. This dataset is grouped by specific event types and correlated outage log entries, ensuring a focus on relevant events for analysis. Subject matter experts conduct a thorough validation of this dataset to guarantee its accuracy and relevance. Subsequently, the dataset is bifurcated into training and validation subsets. The training subset, particularly emphasizing vegetation-contact data, undergoes iterative processing in a machine learning setup (shown in FIG. 8), enhancing the algorithm’s predictive accuracy and reliability.

[0057] FIG. 8 presents a flow diagram detailing a machine learning process 800. Input data, including waveform and signal-processed data 802, is utilized to train a neural network 804. This training involves adjusting model parameters, such as the weights of the neural network’s nodes, through iterative cycles until the network achieves specified performance metrics. Following training, a validation phase employs the validation subset of data to finetune the model, ensuring the algorithm meets standards of accuracy and reliability. Once trained and validated, neural network 804 is equipped to analyze new input data, effectively classifying event types and generating reliable model inferences 806. In this example, neural network 804 is used for event type classification as described with reference to FIG. 5-FIG.8; a separate machine learning model, gradient boosting decision tree model 1320 (FIG. 13), is used for the vegetation-cause prediction described below with reference to FIG. 12 and FIG. 13.

[0058] FIG. 9 shows how multiple transient or temporary events are grouped with a sustained fault based on shared attributes such as waveform characteristics, location estimates, phases, and timing. This grouping technique, elaborated in International Patent Application No. PCT / US2025 / 042658 titled “Grouping Electrical Disturbance Events,” utilizes waveform samples 902 which may represent momentary faults or incipient disturbances. In this example, a first event 904 might occur on one phase and subsequent events (906 and 908) on another. These events are grouped together if they share the same circuit, have similar phase configurations, are closely located, and occur within a predetermined time frame relative to an outage, enhancing diagnostic precision.

[0059] Building on this grouping technique, FIG. 10 and FIG. 11 showcase how a Cascadence system from Toumetis visualizes and utilizes the grouped data. Cascadence employs machine learning to derive analytic insights from data sourced from PQ meters, the Outage Management System (OMS), and external data like weather reports. It presents these insights through an intuitive GUI that supports proactive maintenance and operational decision-making. An event overview timeline user interface 1000, accessible via anDocket No. 503461.70051overview tab 1002, organizes data into a precursor timeline 1004 and an outage timeline 1006 for each day, offering detailed views 1012 of specific events, thereby aiding quick decision-making.

[0060] FIG. 11 shows original signal user interface 1100 for viewing the WMU data of a particular March 28, 2024 precursor event 1102. In this example, event 1102 is selected to show its original three-phase voltage signal waveforms 1104, an original three-phase current signal waveforms 1106, a reactance to fault 1108 (i.e., calculated as described in the ’681 application), and inferences 1110 that were determined using machine learning (ML) or artificial intelligence (Al) processes.

[0061] In the context of the document, Al and ML are terms that relate to the application of computer algorithms and systems to perform tasks that typically require human intelligence. Al is a general term that encompasses ML models as well as other techniques that enable machines to mimic human behavior and decision-making processes. ML, a subset of Al, specifically refers to algorithms that allow computers to learn from and make predictions or decisions based on data. In this document, ML is primarily mentioned in relation to the analysis of waveform data and fault detection in electrical power delivery systems. It involves training models on historical data to identify and predict fault types and causes, such as vegetation-induced faults. The term Al, however, references additional processes involved, including the automation of data analysis tasks that go beyond data-driven models to include adaptation and learning from new data, enhancing predictive accuracy and operational efficiency.

[0062] Inferences 1110 include an event persistence type 1112 (shown as a precursor event) and a fault type 1114 that is shown as a line-to-ground (LG) event.

[0063] Event persistence type 1112 includes precursor events, outages, and operational issues (e.g., schedule maintenance). This term conveys the various ongoing or transient states of events in the context of a power system, i.e., whether the events are persistent (ongoing or long-lasting in the case of planned outages) or transient (temporary or shortlived), which can be crucial for understanding and managing their impact on the power system.

[0064] Fault type 1114 includes single LG faults, line-to-line (LL) faults, line-to-line-to-ground (LLG) faults, and three-line faults (LLL).

[0065] In some embodiments, events can also be auto-labeled based on additional inferences. For example, labelling 1116 includes a cause 1118, a device 1120, and line placement 1122 (OH or UG).Docket No. 503461.70051

[0066] Causes 1118 indicate the cause of the event and can include any of the following: cap switch, cap switch re-strike, cap switch stuck contact, customer problem, deenergized for safety, dig in, equipment failed / damage, faulty meter, fire, foreign object in lines, improper installation, lightning, loose connection, meter power supply, non-standard construction, other, overloaded normal conditions, planned outage, power electronics switching, restoration switching event, severe weather, snapshot, switch (load off), switch (load on), switching error, switching event, sympathetic event, timing pulse, transformer inrush current, transmission event, tree / limb preventable, tree / limb unpreventable, under investigation, unknown, and vehicle contact.

[0067] Device 1120 indicates an electrical component that is determined to be involved with the event. Example electrical components include: conductor (OH), connector (OH), cross arm, customer equipment, cut out, disconnect, disconnect switch, down guy, elbow (L.B.), elbow (N.L.B.), electronic sectionalizer, feed through, fuse, fuse cabinet, fuse switch, ground bank, handhole, hotline clamp, insulator / pin, jumper, meter, miscellaneous substation, other equipment, pole, polymer arrester, pothead, probes, RA switch, RC-off (distribution only), recloser, regulator, relay, splice, step down transformer, switch, and T-connector.

[0068] In the example of FIG. 11, event 1102 is captured by a PQ meter (or equivalent device) and classified by Cascadence and its underlying machine learning algorithms: event type precursor, fault type LG, phases A and then the location algorithm provides a cumulative reactance to the fault of 10.104 ohms.

[0069] FIG. 12 details a process 1200 of ingesting and processing data within a WMU data analysis system 204 for prediction of vegetation caused faults. Process 1200 initiates with the collection of data from PQ monitors 1202, where real-time electrical parameters are recorded. This data is then stored in a Toumetis storage system 1204, serving as a central repository that facilitates subsequent data manipulation and analysis. The data within this storage is pre-processed and organized into one-hour periods 1206, standardizing the dataset for consistent analysis intervals.

[0070] An event detection 1208 analysis scrutinizes this structured data to identify anomalies, which are used for diagnosing potential issues within the electrical system.When an anomaly is detected, WMU data analysis system 204 triggers a vegetation model repository to extract ten seconds of data 1210 immediately preceding the anomaly. This targeted data extraction captures the moments before the disturbance, providing insights into the conditions leading up to the event. The extracted data is then provided into a featureDocket No. 503461.70051transformation pipeline 1212, where it is refined and transformed into features suitable for model input. This ensures that the data is formatted for predictive analysis.

[0071] The transformed data proceeds to a model inference stage 1214, where algorithms analyze the features to infer and predict potential causes of the detected anomalies. The culmination of this process is a prediction 1216 of the likelihood that a vegetation cause is responsible for the anomaly. Prediction 1216 aids in focusing maintenance efforts and managing vegetation to prevent future disturbances, enhancing the reliability and efficiency of the power delivery system.

[0072] FIG. 13 shows in greater detail an example of feature transformation pipeline 1212 (FIG. 12). In the example of FIG. 13, a process 1300 is for specific transformations of data used for fault analysis in an electrical power delivery system.

[0073] As explained previously, process 1300 begins with an hour-long data file 1302, which contains a comprehensive collection of waveform data from the power system.Initially, a detectable PQ event such as an SLG or LL fault (typically manifesting as current swells) is classified as explained above, enabling the analysis of pre-event data via pre-fault ML models.

[0074] Once the PQ event is classified, a pre-fault WMU current data buffer (e.g., about ten seconds of data) is retrieved for pre-fault analysis. Accordingly, from hour-long data file 1302, a ten-second segment 1304 that precedes a fault is isolated, focusing process 1300 on a relatively small amount of relevant data for fault detection.

[0075] This segment 1304 is processed using a direct quadrature zero (DQZ) transform 1306, which extracts reactive and zero-sequence components of the waveform, providing additional insight into fault dynamics. Thus, DQZ transform 1306 isolates the fault phase, in addition to iO and iq.

[0076] Following DQZ transform 1306, the data is processed through a rolling root mean square (RMS) computation 1308 to smooth out the noise and enhance the detection of significant trends and anomalies. In some embodiments, rolling RMS function 1308 is set with a window size of 256 and a step size of 16.

[0077] The smoothed data is further refined using multiple seasonal-trend decomposition using locally estimated scatterplot smoothing (LOESS) (MSTL) 1312. The seasonal periods tailored to effectively capture the dynamics of the waveform data include [2, 3, 4, 5, 32, 64, 128, 256, 512, 1024] for current phase settings and [4, 8, 64, 128] for DQZ phase settings, ensuring a comprehensive analysis across various frequencies and patterns. Each number in these lists represents the number of data points that complete one cycle of a seasonalDocket No. 503461.70051pattern. For instance, a seasonal period of two means the algorithm identifies patterns that repeat every two data points. Analyzing multiple seasonal periods allows MSTL to capture a range of underlying patterns that occur at different intervals, which is helpful in power systems exhibiting diverse cycles. This multi-period analysis is vital for isolating trends and seasonalities more precisely; shorter periods capture rapid fluctuations, while longer periods help identify broader, more persistent trends in the data.

[0078] LOESS is a non-parametric regression method that combines multiple regression models in a k-nearest-neighbor-based approach. It is used to create a smooth line through a scatterplot of data points. The method is particularly useful for datasets with a non-linear pattern since LOESS does not require a predetermined relationship model between the variables. LOESS fits a low-degree polynomial to subsets of the data, with the polynomial being fit using weighted least squares. Points nearest to the point of interest are given higher weight, which diminishes as the distance increases, making the regression model responsive to local changes in the dataset without being overly influenced by outliers or noise.

[0079] MSTL is an extension of a classical seasonal decomposition of time series data, adapted to handle data with multiple seasonal patterns. It is useful for complex time series datasets where multiple seasonal effects are present. The MSTL technique applies LOESS to decompose a time series into three components: (1) A trend component represents the longterm progression of the series (i.e., upwards, downwards). (2) Seasonal components are typically time series data and may exhibit multiple seasonal patterns (e.g., daily, weekly, yearly). MSTL can extract and separate these patterns, each with its own cyclic changes. (3) A residual component include the randomness or irregularities in the data that cannot be attributed to the trend or seasonal components.

[0080] The trends extracted are then converted into sections 1314, where the data is averaged over defined sections to simplify the data structure for subsequent analysis. In this example, there are ten sections (one per second). The reason for ten sections is that it balances predictive accuracy with use of computational resources.

[0081] These sections are scaled 1316 to ensure uniformity and comparability, making the data suitable for input into machine learning models. For instance, pre-event period trend sections are scaled by the event, by dividing all 10 sections by the mean of the first three. This is intended to remove substation / feeder specific information, while still keeping information about the size of the overall trend, which would be distorted by something like MinMax scaling.Docket No. 503461.70051

[0082] The final analytical step involves inputting this prepared data into a gradient boosting decision tree model 1320. There are 30 features provided into gradient boosting decision tree model 1320, i.e., three current features for each one of 10 sections. The three features for each section are scaled RMS trend section mean of fault current, scaled RMS trend section mean of zero-sequence current, and scaled RMS trend section mean of reactive current. These features are derived from a strategy where each event in the dataset is analyzed to maintain only the first event per outage, ensuring a focus on significant disturbances.

[0083] Gradient boosting decision tree model 1320 utilizes algorithms to predict fault types based on the processed waveform data. A gradient boosting decision tree (GBDT) model is an ensemble machine learning technique that enhances predictive accuracy by combining multiple decision trees. This model constructs a robust predictive system in stages, where each tree is built to correct the errors made by previous ones, thus forming a sequence of weak learners that cumulatively result in a strong predictive model. The essence of the GBDT approach lies in its use of gradient descent, a method typically used in optimizing continuous functions, to minimize a loss function that quantifies prediction errors. In this process, each new tree incrementally improves the model by moving in the direction that reduces the loss, effectively correcting the residual errors left by its predecessors.

[0084] Visual representations of the data transformations are provided in a graph 1310 and a graph 1318, which illustrate the appearance of the waveform data following the DQZ transform 1306 and MSTL trend extraction 1312, respectively. For instance, graph 1310 shows a fault phase, iO, and iq. Graph 1318 shows RMS, trend, and 10 section means.

[0085] The aforementioned pre-fault data is present for varying locations of WMUs relative to the fault. Different WMUs do provide varying cumulative reactance measurements, with the unit closest to the fault offering the smallest cumulative reactance. A fault location calculation can also be performed by incorporating source reactance and impedance calculations.

[0086] False positives in the event classification can be identified using additional models or by analysts trained to review model classifications. Curated datasets (see, e.g., FIG. 7) are developed that form the basis training and testing the ML programs. The characteristics the pre-fault model uses to detect vegetation contact are validated against investigations by a utility company following outages or via crew patrols, and resulting information is used to assess a model’s performance.Docket No. 503461.70051

[0087] When vegetation-induced faults are detected as explained above with reference to FIG. 12 and FIG. 13, they may be actively displayed in Cascadence (see, e.g., FIG. 14-FIG.17). For instance, in some embodiments, GIS-based maps (FIG. 15 and FIG. 17) are generated by integrating distribution network models from Synergi Electric or CYME Power Engineering Software with open-source geospatial data, such as OpenStreetMap. These network models contain GPS-based location data for substations, feeders, transformers, protection devices, and other components, as well as reactance values and electrical connectivity details. To generate the maps, the system first extracts the network topology and reactance data from Synergi or CYME and overlays this structured network model onto an open-source GIS base map. The GIS base map provides real-world geographical context, including roads, rivers, terrain, and urban infrastructure, which aids field crews in efficiently locating faulted sections.

[0088] In other embodiments, predictions are also communicated through periodic reports, meetings, and digital communications, forming part of grouped events that are likely due to similar causes. These events prompt actions such as sending out patrols to determine the fault’s cause, prioritizing tree trimming (e.g., a backbone issue has a larger customer impact), and removing overgrown vegetation. The predictions include location estimates to aid in prioritizing crew deployments and vegetation management initiatives. Additionally, FIG. 14-FIG. 17 show how results are presented as predictions of grouped waveform events, treating any group with a vegetation-caused event as indicative of the overall cause. This information is expected to enhance the client’s ability to manage vegetation and mitigate associated risks effectively. Finally, these figures also show how utilization of additional data types, such as geo / spatial data about vegetation, enhance the model’s functionality.

[0089] FIG. 18 is a block diagram illustrating components 1800, according to some example embodiments, configured to read instructions from a machine-readable or computer-readable medium (e.g., a non-transitory machine-readable storage medium) and perform any one or more of the methods discussed herein (or portions thereof) in connection with WMUs or WMU data analysis system 204 (FIG. 2), such as discussed for processes 1200, processes 1300, or any portions of these embodiments.

[0090] Specifically, FIG. 18 shows a diagrammatic representation of hardware resources 1802 including one or more processors 1804 (or processor cores), one or more memory / storage devices 1806, and one or more communication resources 1808, each of which may be communicatively coupled via a bus 1810. For embodiments where node virtualization (e.g., NFV) is utilized, a hypervisor 1812 may be executed to provide anDocket No. 503461.70051execution environment for one or more network slices / sub-slices to utilize hardware resources 1802.

[0091] Processors 1804 (e.g., a central processing unit (CPU), a reduced instruction set computing (RISC) processor, a complex instruction set computing (CISC) processor, a graphics processing unit (GPU), a digital signal processor (DSP) such as a baseband processor, an application specific integrated circuit (ASIC), another processor, or any suitable combination thereof) may include, for example, a processor 1814 and a processor 1816.

[0092] Memory / storage devices 1806 may include main memory, disk storage, or any suitable combination thereof. Memory / storage devices 1806 may include, but are not limited to, any type of volatile or non-volatile memory such as dynamic random-access memory (DRAM), static random-access memory (SRAM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), Flash memory, solid-state storage, etc.

[0093] Communication resources 1808 may include interconnection or network interface components or other suitable devices to communicate with one or more peripheral devices 1818 or one or more databases 1824 via a network 1820. For example, communication resources 1808 may include wired communication components (e.g., for coupling via a Universal Serial Bus (USB)), cellular communication components, NFC components, Bluetooth® components (e.g., Bluetooth® Low Energy), Wi-Fi® components, and other communication components.

[0094] Instructions 1822 may comprise software, a program, an application, an applet, an app, or other executable code for causing at least any of processors 1804 to perform any one or more of the methods discussed herein. Instructions 1822 may reside, completely or partially, within at least one of processors 1804 (e.g., within the processor’s cache memory), memory / storage devices 1806, or any suitable combination thereof. Furthermore, any portion of instructions 1822 may be transferred to hardware resources 1802 from any combination of peripheral devices 1818 or databases 1824. Accordingly, the memory of processors 1804, memory / storage devices 1806, peripheral devices 1818, and databases 1824 are examples of computer-readable and machine-readable media.

[0095] In light of this disclosure, skilled persons will appreciate that many changes may be made to the details of the above-described embodiments without departing from the underlying principles of the invention. The scope of the present invention should, therefore, be determined only by claims and equivalents.

Claims

Docket No. 503461.70051CLAIMSWhat is claimed is:

1. A method for determining whether a fault in an electrical power delivery system is caused by vegetation contact, comprising:receiving synchro-waveform data from a waveform measurement unit (WMU) positioned within the electrical power delivery system, in which the WMU captures time-synchronized waveform measurements indicative of electrical disturbances;detecting, from the received synchro-waveform data, patterns of current changes that indicate a fault;classifying identified fault patterns into event types including line-to-ground faults and line-to-line faults;in response to a classified fault having a line-to-ground or line-to-line fault type, extracting pre-fault data from the synchro-waveform data immediately preceding the classified fault;generating, from the pre-fault data, feature data for input to a machine learning model configured to determine whether vegetation contact caused the classified fault;in response to determining that vegetation contact is the cause of the classified fault, generating a notification about a detected vegetation-caused fault, which includes identification of the fault as vegetation-caused and provides an estimated location of the fault within the electrical power delivery system; andpresenting the notification on a user interface alongside a visual representation of waveform data corresponding to the fault, in which the user interface allows for user interaction to further analyze the waveform data.

2. The method of claim 1, in which the generating of the feature data includes applying a direct quadrature zero (DQZ) transformation to isolate reactive and zero-sequence components from the pre-fault data.

3. The method of claim 1, in which the generating of the feature data further includes applying a rolling root mean square (RMS) computation to reduce noise and enhance detection of trends in current components derived from the pre-fault data.

4. The method of claim 3, in which generating the feature data includes smoothing the prefault data using a rolling root mean square computation having a window size of 256 and a step size of 16.Docket No. 503461.700515. The method of claim 1, in which the generating of the feature data includes processing data derived from the pre-fault data using Multiple Seasonal-Trend decomposition using LOESS (MSTL) to extract and separate trends and seasonal patterns from waveform data.

6. The method of claim 5, in which the MSTL uses a plurality of seasonal periods to separate shorter-period fluctuations from longer-period trends in the pre-fault data.

7. The method of claim 1, in which the notification includes instructions for mitigating the vegetation-caused fault, including recommendations for tree trimming or other vegetation management practices near the estimated location.

8. The method of claim 1, further comprising storing historical fault data and associated machine learning model predictions in a database for trend analysis and predictive maintenance scheduling.

9. The method of claim 1, in which presenting the notification on the user interface includes displaying both historical and real-time data comparisons to provide analysis of a fault pattern over time.

10. The method of claim 1, in which the pre-fault data comprises about ten seconds of synchro-waveform data immediately preceding the classified fault.

11. The method of claim 1, in which the feature data is generated from a fault phase current component, a reactive current component, and a zero-sequence current component derived from the pre-fault data.

12. The method of claim 1, in which generating the feature data includes dividing trend data derived from the pre-fault data into ten sections.

13. The method of claim 12, in which the ten sections correspond to equal-duration intervals of the pre-fault data.

14. The method of claim 12, in which generating the feature data includes calculating a section mean for each of the ten sections.

15. The method of claim 14, in which the feature data comprises thirty features including, for each of the ten sections, a scaled trend section mean of fault current, a scaled trend section mean of zero-sequence current, and a scaled trend section mean of reactive current.Docket No. 503461.7005116. The method of claim 13, in which the generating of the feature data includes scaling trends within the equal duration intervals by normalizing section values relative to initial section values of the pre-fault data.

17. The method of claim 16, in which scaling includes dividing section means by a mean of an initial set of sections.

18. The method of claim 17, in which the initial set of sections comprises a first three sections.

19. The method of claim 1, in which the machine learning model comprises a gradient boosting decision tree model.

20. The method of claim 1, further comprising determining a cumulative reactance to the fault and using the cumulative reactance to estimate the location of the fault within the electrical power delivery system.

21. The method of claim 20, further comprising presenting the estimated location on a geographic visualization.

22. The method of claim 1, further comprising grouping the classified fault with one or more additional events based on shared waveform characteristics, location estimates, phases, or timing.

23. The method of claim 1, in which the notification is presented in an event overview timeline user interface.

24. The method of claim 1, in which the notification is presented together with an original signal user interface showing voltage waveforms, current waveforms, and a reactance-to-fault value corresponding to the classified fault.

25. The method of claim 1, further comprising storing the synchro-waveform data in one-hour periods before extracting the pre-fault data.

26. The method of claim 1, in which the machine learning model is trained using a curated dataset of historical events grouped by event type and outage information.

27. The method of claim 1, further comprising, in response to determining that vegetation contact caused the classified fault, initiating a vegetation management action based on the estimated location.Docket No. 503461.7005128. A system for determining whether a fault in an electrical power delivery system is caused by vegetation contact, comprising:a waveform measurement unit (WMU) positioned within the electrical power delivery system and configured to capture synchro-waveform data indicative of electrical disturbances; anda WMU data analysis system including a processor and memory storing instructions that, when executed by the processor, cause the WMU data analysis system to:receive the synchro-waveform data from the WMU;detect, from the synchro-waveform data, patterns of current changes that indicate a fault;classify identified fault patterns into event types including line-to-ground faults and line-to-line faults;in response to a classified fault having a line-to-ground or line-to-line fault type, extract pre-fault data from the synchro-waveform data immediately preceding the classified fault;generate, from the pre-fault data, feature data for input to a machine learning model configured to determine whether vegetation contact caused the classified fault;in response to determining that vegetation contact is the cause of the classified fault, generate a notification about a detected vegetation-caused fault, which includes identification of the fault as vegetation-caused and an estimated location of the fault within the electrical power delivery system; andpresent the notification on a user interface alongside a visual representation of waveform data corresponding to the fault.

29. The system of claim 28, in which the WMU data analysis system is configured to apply a direct quadrature zero (DQZ) transformation to isolate reactive and zero-sequence components from the pre-fault data.

30. The system of claim 28, in which the WMU data analysis system is configured to apply a rolling root mean square (RMS) computation to reduce noise and enhance detection of trends in current components derived from the pre-fault data.

31. The system of claim 30, in which the rolling root mean square computation uses a window size of 256 and a step size of 16.

32. The system of claim 28, in which the WMU data analysis system is configured to process data derived from the pre-fault data using Multiple Seasonal-Trend decompositionDocket No. 503461.70051using LOESS (MSTL) to extract and separate trends and seasonal patterns from waveform data.

33. The system of claim 28, in which the WMU data analysis system is configured to divide trend data derived from the pre-fault data into ten sections and calculate a section mean for each of the ten sections.

34. The system of claim 33, in which feature data provided to the machine learning model comprises thirty features including, for each of the ten sections, a scaled trend section mean of fault current, a scaled trend section mean of zero-sequence current, and a scaled trend section mean of reactive current.

35. The system of claim 28, in which the machine learning model comprises a gradient boosting decision tree model.

36. The system of claim 28, in which the WMU data analysis system is configured to determine a cumulative reactance to the fault and use the cumulative reactance to estimate the location of the fault within the electrical power delivery system.

37. The system of claim 36, in which the user interface is configured to present the estimated location on a geographic visualization.

38. A non-transitory computer-readable medium storing instructions that, when executed by a processor, cause the processor to perform operations comprising:receiving synchro-waveform data from a waveform measurement unit (WMU) positioned within an electrical power delivery system, in which the WMU captures time-synchronized waveform measurements indicative of electrical disturbances;detecting, from the synchro-waveform data, patterns of current changes that indicate a fault;classifying identified fault patterns into event types including line-to-ground faults and line-to-line faults;in response to a classified fault having a line-to-ground or line-to-line fault type, extracting pre-fault data from the synchro-waveform data immediately preceding the classified fault;generating, from the pre-fault data, feature data for input to a machine learning model configured to determine whether vegetation contact caused the classified fault;in response to determining that vegetation contact is the cause of the classified fault, generating a notification about a detected vegetation-caused fault, which includesDocket No. 503461.70051identification of the fault as vegetation-caused and an estimated location of the fault within the electrical power delivery system; andpresenting the notification on a user interface alongside a visual representation of waveform data corresponding to the fault.

39. The non-transitory computer-readable medium of claim 38, in which generating the feature data includes applying a direct quadrature zero (DQZ) transformation to isolate reactive and zero-sequence components from the pre-fault data.

40. The non-transitory computer-readable medium of claim 38, in which generating the feature data further includes applying a rolling root mean square (RMS) computation to reduce noise and enhance detection of trends in current components derived from the prefault data.

41. The non-transitory computer-readable medium of claim 38, in which generating the feature data includes processing data derived from the pre-fault data using Multiple Seasonal-Trend decomposition using LOESS (MSTL) to extract and separate trends and seasonal patterns from waveform data.

42. The non-transitory computer-readable medium of claim 38, in which generating the feature data includes dividing trend data derived from the pre-fault data into ten sections and calculating a section mean for each of the ten sections.

43. The non-transitory computer-readable medium of claim 42, in which the feature data comprises thirty features including, for each of the ten sections, a scaled trend section mean of fault current, a scaled trend section mean of zero-sequence current, and a scaled trend section mean of reactive current.

44. The non-transitory computer-readable medium of claim 38, in which the machine learning model comprises a gradient boosting decision tree model.

Images