Method and apparatus for estimating the cause of power system failures

Abstract

To estimate a cause of a failure in a power system with high accuracy by using not only a voltage waveform of an existing measurement device that measures voltages but also weather information and land usage information.SOLUTION: A power system failure cause estimation device 1 includes: a feature value calculation unit 12 that calculates a feature value on the basis of data from a power measurement signal input unit 25; a classification unit 13 based on machine learning for estimating a failure cause for a power system on the basis of the feature value; and a failure cause output unit 14 that outputs the estimated failure cause. The feature value calculation unit 12 calculates a first feature value from a measurement signal at the time of occurrence of a failure, calculates a second feature value by acquiring weather information and land usage information corresponding to a power transmission line section containing a location of occurrence of the failure, and estimates the failure cause with high accuracy by performing machine learning with the first and second feature values.SELECTED DRAWING: Figure 3

Description

Technical Field

[0001] The present invention relates to a method and apparatus for estimating the cause of a power system failure, and particularly to accurately analyzing the factors causing a failure when a failure occurs in a high-voltage transmission line by using AI technology.

Background Art

[0002] A power system is a large-scale system that is constructed and operated by combining many devices and control methods for stable power supply. To transmit the alternating current power generated at a power plant to the demand area, a high-voltage or extra-high-voltage power transmission network is used. The state of a power system can be represented by physical electrical quantities such as voltage, current, power, and frequency, and these states are monitored in real time by a power system monitoring device. The state of a power system has a spatial spread created by the system configuration and also involves temporal changes. Also, the system state changes greatly depending on the power generation and load connected to the system.

[0003] Conventionally, when an accident or failure that causes a disruption to power supply occurs, the power supply company quickly reduces the power outage area and shortens the restoration time. These measures are essential requirements for improving the power supply reliability. Among the many causes of power system failures, in order to achieve prompt accident restoration, it is required to quickly identify the cause of the failure (accident cause).

[0004] When a power system failure occurs, the cause is analyzed using oscilloscope waveform information at the time of the failure. Monitors (such as engineers) then consider weather information around the transmission lines and use past experience to make a comprehensive judgment and estimate the cause of the failure. However, a challenge with this method of cause estimation is that it is time-consuming because it depends on the skill of the engineer making the judgment. Furthermore, since this engineer's skill is tacit knowledge and highly individual, there is a challenge in that it is difficult to pass on this judgment technique. To solve these problems, a method for classifying the causes of transmission line failures using AI technology has been proposed. For example, if the cause of the failure can be mechanically estimated using signal processing with measured signals and collected data, it will be possible to make a quick judgment that does not depend on the engineer's skill. In addition, by mechanically memorizing and learning the engineer's tacit knowledge, the individuality of the knowledge can be eliminated, and the challenge of technology transfer can be solved.

[0005] Patent documents 1 and 2 describe methods for estimating the cause of power transmission line faults using AI technology (learning models). In Patent Documents 1 and 2, when a ground fault occurs in the power transmission network, the cause of the fault is estimated by analyzing the zero-sequence current waveform and zero-sequence voltage waveform of the wiring detected at the time of the ground fault. In the technology described in Patent Document 3, the ground fault current waveform of the power line is spectrally analyzed, and the cause of the ground fault is determined based on the power spectral value. [Prior art documents] [Patent Documents]

[0006] [Patent Document 1] Japanese Patent Publication No. 2020-10505 [Patent Document 2] Japanese Patent Publication No. 2019-208317 [Patent Document 3] Japanese Patent Application Publication No. 6-289086 [Overview of the Initiative] [Problems that the invention aims to solve]

[0007] Patent Document 1 selects potential power system failure causes based on the waveform characteristics at the time of the fault included in the measured signal, calculates the probability of the failure occurring based on the input surrounding information, and ranks the failure causes by combining the failure cause candidates and failure cause probabilities. In order to estimate the power system failure causes, signal processing is performed by combining a failure cause candidate selection unit that selects potential failure causes and a failure cause probability calculation unit that calculates the probability of the failure occurring. Therefore, if the failure cause candidate selection unit misjudges a failure cause candidate and the correct failure cause is excluded from the candidates, the failure cause probability of the correct failure cause will not be output in subsequent failure cause estimation, and the ranked screen display will not be shown. As a result, if the correct failure cause is among the failure cause candidates excluded due to misjudgment, information such as the probability of the correct failure cause may not be obtained.

[0008] Patent Document 2 describes how to acquire zero-sequence data showing the temporal change in the zero-sequence potential of a three-phase distribution line to determine the ground fault period during which a ground fault occurs. The estimation unit inputs the zero-sequence data corresponding to the ground fault period determined by the determination unit into a trained model that has learned the relationship between the waveform pattern that appears due to the ground fault and the ground fault cause, and estimates the ground fault cause for each predetermined period included in the ground fault period based on the classification results obtained. However, the waveform at the time of fault included in the measurement signal analyzes the zero-sequence current and zero-sequence voltage of the power system, and the type of fault targeted is only ground faults. There is a problem that it is not possible to estimate the cause of faults other than ground faults, such as short-circuit faults and open-circuit faults.

[0009] Patent Document 3 describes spectral analysis of ground fault current waveforms in power lines. Since ground fault current waveforms exhibit unique patterns depending on the cause of the fault, it is possible to estimate the cause of the fault with considerable accuracy. However, it is not possible to estimate the cause of faults that do not appear as patterns in the ground fault current waveform.

[0010] The present invention was made to solve conventional problems, and its objective is to realize a power system fault cause estimation method and apparatus that can accurately estimate the cause of faults, including not only ground faults but also short-circuit faults, by utilizing not only the voltage waveform of existing measuring devices that measure voltage, but also weather information and land use information that can be obtained from external sources. Another object of the present invention is to realize a power system failure cause estimation method and apparatus that eliminates the subjectivity of the features used for cause estimation by inputting weather information and land use information to be considered when estimating the cause of a power system failure, selecting and processing the input information, and calculating features used in machine learning. Another object of the present invention is to realize a power system failure cause estimation method and apparatus that allows arbitrary modification of the features input to a machine learning model. [Means for solving the problem]

[0011] In this invention, a power system failure cause estimation device is provided, comprising: a power measurement signal input unit that inputs measurement signals of the three-phase voltage and zero-phase voltage of a power system and power measurement signals that can be obtained by an existing measuring device; a feature calculation unit that calculates a first feature quantity from the data input to the power measurement signal input unit; a failure cause estimation unit that estimates the cause of a power system failure based on the first feature quantity; and a failure cause output unit that outputs the estimated failure cause. In this device, a peripheral information input unit is provided for inputting peripheral information related to the power system (weather information, land use information, etc.), and the feature calculation unit calculates a second feature quantity from the data input from the peripheral information input unit, and the device is configured to estimate the cause of a failure using the first and second feature quantities. The items calculated as this second feature quantity can be arbitrarily set, changed, or deleted by setting the feature quantity item input unit.

[0012] Data input by the power measurement signal input unit is used to create a first feature vector through CNN image recognition and digital signal processing. The failure cause estimation device integrates this data with information acquired from existing measurement devices, externally or internally, weather information, land use information, etc., to create table data. The table data is stored in a memory device. The failure cause estimation device includes a CPU (Central Processing Unit) that executes a computer program. [Effects of the Invention]

[0013] In this invention, in addition to using the phase voltages and zero-sequence voltages of three-phase AC power detected from the voltage waveform at the time of an accident, obtained from existing voltage measuring devices that are standardly installed in most substations, as input data, a machine learning model is used that also uses weather information and land use information, which can be obtained externally as open data, as input data. As a result, it is possible to estimate the probability of most fault causes with high accuracy without missing the correct fault cause. In addition to ground faults, it is also possible to estimate the fault causes of short-circuit faults and open-circuit faults. Furthermore, since the existing power system fault cause estimation device has been improved by adding a feature calculation program that uses open data as additional data, no new capital investment in hardware is required, and a power system fault cause system can be constructed at low cost. Moreover, by using weather information and land use information and performing machine learning on a computer, the method of determining the fault cause, which was previously the tacit knowledge of engineers, can be inherited regardless of the engineer's skill level. [Brief explanation of the drawing]

[0014] [Figure 1] This is a diagram illustrating the procedure for estimating the cause of a power system failure according to an embodiment of the present invention. [Figure 2] This figure shows the input data for the power system failure cause estimation device 1 according to an embodiment of the present invention. [Figure 3] This is a system configuration diagram of the power system failure cause estimation device 1 according to an embodiment of the present invention. [Figure 4]A list of feature quantities input to the classification unit 13 in FIG. 3. [Figure 5] A list of items output as failure causes by the classification unit 13. [Figure 6] FIG. 4 is a diagram for explaining a method of quantifying features of weather observation data (No. 16, 17, 19 to 24). [Figure 7] FIG. 4 is a diagram for explaining a method of quantifying features of sunshine duration (No. 18). [Figure 8] FIG. 4 is a diagram for explaining a method of quantifying features of land use information. [Figure 9] A flowchart for explaining the power system failure cause estimation procedure of the present invention. [Figure 10] FIG. 3 is a diagram showing a display screen of an estimated result of a failure cause output to the output device 16.

Embodiments for Carrying Out the Invention

[0015] FIG. 1 is a diagram for explaining the power system failure cause estimation procedure. When a failure 310 such as a short circuit, a ground fault, or a disconnection occurs in the transmission line 300 wired via the tower 320 as shown in FIG. 1(a) for some reason, the power system failure cause estimation device 1 of the present invention performs machine learning based on various collectible information and estimates the failure cause by executing a computer program. The estimated failure cause is output to the output device 16 of the monitor 180. In FIG. 1(a), in order to easily assume the image of the power system failure 310, a state in which a short circuit or a disconnection has occurred due to the contact of the loading platform of the large truck 350 with the transmission line 300 is schematically illustrated. However, since the high-voltage transmission line 300 is usually strung by a tower 320 having a sufficient height, it is impossible for the loading platform of the large truck 350 to come into contact. The actual possible causes of failure include lightning strikes, snow damage, bird contact, bird droppings, salt damage, entanglement of vines, contact with motor vehicles, contact with metal objects, etc.

[0016] When a failure occurs in the power transmission line 300, the power company immediately identifies the location where the failure occurred. The method for identifying the location of the failure is well-known, and various methods have been proposed. For example, if the failure occurs at a location between specific relay points of the power transmission line 300, for example, between point A with a substation and point B with another substation, it is identified within a line section. Also, in some cases, it may be determined as a surface area, such as the interior of the area surrounded by three points A, B, and C in the power transmission network.

[0017] Next, the power system fault cause estimation device 1 collects the necessary information that can be used to identify the fault cause as shown in Fig. 1(b) and performs its analysis. As conventionally available information, for example, there is power monitoring data 40 detected by various sensors in the power system, such as measurement signals collected by current transformers and transformers installed in substations, and waveform values such as the voltage of each phase of three-phase alternating current, zero-phase voltage, the current of each phase of three-phase alternating current, zero-phase current, etc. The power monitoring data 40 is information that can be obtained by existing measurement devices. These various data are monitored in real time, and the measurement data is recorded in a storage device in a power monitoring device (not shown). Also, in order to detect a phenomenon that is a precursor before detecting a fault, the power monitoring device (not shown) uses not only the power monitoring data 40 at the time of fault detection but also the past measurement data before the fault detection, and sets the power measurement signal before the fault and the measurement signal after the fault as comparison targets.

[0018] The conventional power system fault cause estimation device 1 estimated the fault cause using only the power monitoring data 40. Therefore, while the monitor 180 was looking at the screen of the output device 16 that displayed the fault cause using only the power monitoring data 40, a determination of attribution was made while considering the weather information at the time of the fault occurrence and the surrounding environment information of the power transmission line. In contrast, in this embodiment, the weather information 60 at the time of the fault occurrence and the surrounding environment information 70 of the power transmission line are also input to the power system fault cause estimation device 1, and the fault cause is estimated by AI technology using these data 40, 60, and 70. As a result, it has become unnecessary for the monitor 180 to add a determination of attribution while considering natural phenomena, etc., to the fault cause result determined and output by the power system fault cause estimation device as was conventionally done.

[0019] Figure 2 is a diagram illustrating the input data of the power system fault cause estimation device 1 according to an embodiment of the present invention, and the machine learning-based determination process using that data. The power monitoring data 40 acquired by the power company includes various types of data. In Figure 2, from this data, the recorded information 41 at the time of fault occurrence, the oscilloscope data A44 at the time of fault occurrence, and the oscilloscope data B47 at the time of fault occurrence are extracted as training data 150 for machine learning.

[0020] The recorded information 41 at the time of the fault is existing measurement data 42 indicating the time of the fault, and includes measurement signals collected by sensors installed in power system sensor-equipped switches, etc., and information that can be obtained by existing measuring devices. The measurement signals collected by the sensors include the phase voltages and zero-sequence voltages of the three-phase AC. Alternatively, the phase currents and zero-sequence currents of the three-phase AC, and active power calculated from the measured values ​​may also be used. From the measurement data 42, the date, time, and voltage of the transmission line at the time of the fault are extracted and used as feature quantities 43, and as shown by arrow 81, feature quantities 43 are integrated into the training data 150. The oscilloscope data A44 at the time of the fault is waveform data included in the power monitoring data 40 acquired by the power company. The oscilloscope data A44 is subjected to digital signal processing using a known AI model. For example, the transient response time and fault characteristics of the waveform shown in the oscilloscope data 45 are quantified by digital signal processing and used as feature quantities 46. The detected features 46 are integrated into the training data 150 as shown by arrow 82, and this training data 150 becomes the input to the machine learning-based classification unit 13.

[0021] The oscilloscope data B47 at the time of failure is another oscilloscope data (second oscilloscope data) 48 detected at the time of failure, and the probability of pulse waveform inclusion is detected using this oscilloscope data 48. This detection is performed using digital signal processing 49 with a known AI model. For example, pulse waveforms commonly seen during bird contact are recognized using image recognition technology, and their inclusion probability is used as a feature. The features detected in this way are integrated into the training data 150 as shown by arrow 83. Here, the data shown by the white arrows 81 to 83 correspond to the first feature of the present invention.

[0022] The weather and land use information 140 is additional information in this invention. This includes weather data 60 and land use information 70 that can be used as open data. For example, as weather information, weather data obtainable from the Japan Meteorological Agency's website can be used. As land use information, national land numerical information (land use subdivision mesh (raster version) data) obtainable from the Ministry of Land, Infrastructure, Transport and Tourism's website can be used. When a failure is detected and the failure location is identified, the obtained weather data and land use information are extracted as feature quantities 61 and 71 related to the failure location and included in the training data 150 as shown by the arrow 84. The data shown by this black arrow 84 corresponds to the second feature quantity of this invention.

[0023] In this way, the features 81-84 extracted from multiple data (41, 44, 47, 140) are integrated as training data and input to the machine learning-based classification unit (failure cause estimation unit) 13, as shown by arrow 85, to perform cause classification of the failure.

[0024] Figure 3 is a system configuration diagram of a power system failure cause estimation device 1 according to an embodiment of the present invention. The power system failure cause estimation device 1 can be implemented using a general-purpose computer device. The power system failure cause estimation device 1 is configured to include a CPU 10, to which an input device 15, an output device 16, and a storage device 20 are connected via a data bus 5. In addition, although not shown in Figure 3, it may have internal devices of known computer devices or known external connection devices.

[0025] The power system fault cause estimation device 1 is equipped with a power measurement signal input unit 25 for receiving power measurement signals 31 from power measurement devices 30 on the power transmission lines. The number and type of power measurement devices 30 can be arbitrary, and existing equipment can be used. The power measurement signals 31 include the measurement data 42 and oscilloscope data 45 and 48 shown in Figure 2. The power measurement signals 31 input from the power measurement signal input unit 25 are transmitted to the CPU 10 via the data bus 5.

[0026] The output from the power measurement device 30 may be configured to be input to the power measurement signal input unit 25 as an analysis result analyzed by an external device before being input to the power system fault cause estimation device 1, or the output from the power measurement device 30 (power measurement signal 31) may be directly input to the power system fault cause estimation device 1, and the CPU 10 may process the power measurement signal 31 by executing an analysis program (not shown). Regardless of whether a fault has occurred or not, the collected power measurement signal 31 is stored in the data storage unit 22 of the storage device 20. By recording past power measurement signals 31 in the storage device 20 in this way, it is possible to compare the power measurement signal 31 collected at the time of the fault with past power measurement signals 31, enabling effective estimation of the cause of the fault. Furthermore, in order to detect phenomena that may precede the detection of a fault, it is important to continuously monitor the power measurement signal 31 before fault detection.

[0027] The power system failure cause estimation device 1 is equipped with a peripheral information input unit 26. The peripheral information input unit 26 can be implemented using a network interface that can connect to an open network 100 such as the Internet or an in-house LAN. An example of peripheral information is information related to factors that cause power system failures, such as weather information such as temperature, wind conditions (wind direction, wind speed), and temperature. Weather information can be obtained from an external weather information server 110, such as the Japan Meteorological Agency website. Sunrise and sunset times can also be obtained from the weather information server 110 provided by the National Astronomical Observatory of Japan website. It is preferable to obtain this weather information via the Internet at intervals when the open data of the weather information server 110 is updated, and record the data in the data storage unit 22 of the storage device 20.

[0028] Another example of surrounding information is land use data provided by the land use information server 120, such as the Ministry of Land, Infrastructure, Transport and Tourism website. There is a lot of other information similar to land use, so it is not limited to data provided by the Ministry of Land, Infrastructure, Transport and Tourism. Data acquired from the weather information server 110 and the land use information server 120 is stored in the data storage unit 22 of the storage device 20. Although this information is collected independently of the power measurement signal 31, it is relevant to the cause of the failure, so in this embodiment, it is used for machine learning-based classification to estimate the cause of the failure by combining the measurement signal with weather information and surrounding information.

[0029] The CPU 10 of the failure cause estimation device 1 implements specific functions for estimating the cause of failure using software by executing a computer program (not shown). The computer program (not shown) that is executed includes a program to implement the functions of the feature item input unit 11, a program to implement the functions of the feature calculation unit 12, and a program to implement the functions of the machine learning-based classification unit 13, and these are stored in the program storage unit 21 of the storage device 20.

[0030] The feature item input unit 11 is an input unit for setting the feature items calculated from the inputs from the power measurement signal input unit 25 and the peripheral information input unit 26. In the example shown in Figure 4, 26 feature items are set, and the content of each feature is listed in the explanatory variable 87. Here, items No. 1 to 15 (42) are information that can be obtained from the power measurement device 30 and are also information that is recorded when a power transmission line fails. Items No. 16 to 26 are newly added information (61) in the power system failure cause estimation device 1 according to this embodiment and are information that can be obtained from the weather information provision server 110. Item No. 27 is newly added information (71) in the power system failure cause estimation device 1 according to this embodiment and is information that can be obtained from the land use information provision server 110. The newly added information (61, 71) is information related to the second feature.

[0031] The feature item input unit 11 allows for changes to the feature items once they have been set. These changes are made at the direction of the monitor 180. For example, it is possible to exclude the month of occurrence (No. 1 in Figure 4) from the features used in the failure cause estimation process, or to add new features. By selecting (setting, changing, deleting) appropriate feature items based on the monitor 180's judgment, the accuracy of failure cause estimation can be improved.

[0032] The feature calculation unit 12 calculates first and second features to be input to the machine learning-based classification unit 13 by processing input data from the power measurement signal input unit 25 and the surrounding information input unit 26 according to the feature items set by the feature item input unit 11. An example of this process is, for example, the pulse waveform inclusion probability shown in No. 15 of Figure 4, which involves detecting pulse waveforms seen in bird / animal contact from waveform images contained in power measurement signals 31 measured when a power system accident occurs, using CNN image recognition, and extracting waveform features. A CNN is a feedforward neural network that includes special layers called convolutional layers and pooling layers, and is known to be suitable for image recognition.

[0033] The machine learning-based classification unit 13 classifies the cause of failure using machine learning with the features calculated by the feature calculation unit 12 described above, and corresponds to the failure cause estimation unit of the present invention. For example, Figure 5 is a diagram showing an example of the target variable 89 output as the cause of failure by the machine learning-based classification unit 13. As shown in Figure 5, the target variable 89 indicating the cause of failure was selected to classify 12 types of causes that are assumed to be the main causes of power system failures: lightning, bird droppings, metal, bird contact, motor vehicles, salt damage, animals, snow damage, snakes, flying objects, vines, and tree contact. The setting of these causes can be arbitrarily increased or decreased by the operator's instructions, by increasing or decreasing the number of item numbers 88.

[0034] To classify the causes of power system failures, the system uses decision trees, which represent data classification in a tree structure, gradient boosting decision trees (GBDTs), which are extensions of decision trees capable of high accuracy and complex branching, and random forest machine learning for classification. In this embodiment, the power system failure cause estimation device uses a machine learning model to classify the causes of failures using AI, thereby enabling classification that is not dependent on human judgment.

[0035] The failure cause output unit 14 selects the content to be displayed on the output device 16 from among the multiple failure causes calculated by the feature calculation unit 12. For example, it selects the top few results with high probability of matching from among the multiple classified estimates and outputs them to the output device 16.

[0036] The input device 15 is a device used by an operator to input operation requests or changes to setting values ​​to the power system fault cause estimation system. Examples of such devices include keyboards, mice, and touch-type input devices. Alternatively, an external personal computer or smartphone may be used as the input device 15.

[0037] The output device 16 includes a display device such as a liquid crystal display and presents the fault cause estimated by the CPU 10 to the operator. For example, the fault causes classified as described above are displayed on the screen in a way that is visually understandable to the operator. It is preferable to output not just one fault cause, but multiple fault causes ranked in order of importance. The output device 16 may also output the probability of occurrence of multiple fault causes. In addition to output to the output device 16, various output formats that allow the supervisor to recognize the information can be considered, such as output to paper, transmission via email, illumination of indicator lights, and output of a buzzer sound.

[0038] Figure 6 is a diagram illustrating how to extract features to be input to the machine learning-based classification unit 13 from the meteorological observation data shown in Figures 4, Nos. 16 to 24. Figure 6(a) is a map 62 plotting the area where the power lines are installed, the locations of the power lines 300 arranged from point A to point B, and meteorological data observation points near the power lines 300. In the area shown in map 62, when some kind of fault occurs in the power lines 300, a known fault location identification method detects that the fault occurred between point A (e.g., a substation) and point B (e.g., a substation) of the power lines 300. In reality, the approximate locations of points A and B (fault location 300a) are known, so the power system fault cause estimation device according to this embodiment searches for the meteorological observation point closest to the fault location 300a of the power lines 300 and selects the meteorological data of "Takayama Observatory 65 in XX Prefecture" as the meteorological observation point 64.

[0039] Next, the feature calculation unit 12 obtains the latest weather data from the weather information server 110 immediately after the failure occurs and extracts data for items No. 16 to 22 in Figure 4. Figure 6(b) shows the extracted weather data for "Takayama Observatory, XX Prefecture". Here, the weather observation data for "Takayama Observatory, XX Prefecture" includes temperature 66a, precipitation 66b, sunshine duration 66c, wind speed 66d, wind direction 66e, snowfall 66f, and snow depth 66g. In addition, the integrated snowfall value 66h for the past 10 days and the maximum snowfall value 66i for the past 10 days are calculated by the feature calculation unit 12 after collecting weather data for the past 10 days from the date of the failure (stored in the data storage unit 22 of the memory device 20). The calculated features are input to the machine learning-based classification unit 13. In this way, by preparing weather data in advance in the storage device 20, when a power transmission line 300 accident occurs and the affected section (in this case, section A to section B) is identified, the feature quantities shown in No. 16 to 24 of Figure 4 can be immediately calculated as weather information related to the fault location 300a, and these feature quantities can be input into the machine learning-based classification unit 13. Furthermore, even if the network 100 goes down for some reason, it is possible to estimate the cause of the accident with high accuracy using the acquired weather information. Note that if the weather information server 110 provides the past 10-day snowfall integral value 66h and the past 10-day snowfall maximum value 66i shown in Figure 6, the calculation of the integral value on the fault cause estimation device 1 side is unnecessary.

[0040] Figure 7 is a diagram illustrating the method for feature extraction related to sunshine duration in Figure 4. Unlike the meteorological observation data used in Figure 6, the calendar data can be obtained from the National Astronomical Observatory's weather information server 110. Here, when a failure occurs, data on the sunrise and sunset times on the day of the failure in the prefecture to which the maintenance location of the faulty power transmission line belongs is used. It is preferable to obtain this calendar data in advance and store it in the data storage unit 22 in the memory device 20. The feature calculation unit 12 calculates items No. 25 and 26 in Figure 4 by calculating the difference between the acquired sunrise and sunset time data and the failure time. In the example in Figure 7, the failure time 67 is 8:09 in 24-hour format, the sunrise time 68a is 5:39, and the sunset time 68b is 18:14. Therefore, the feature calculation unit 12 can calculate the time from sunrise time 68a to failure time 67 as failure time - sunrise time = 150 [minutes]. Similarly, the feature calculation unit 12 can calculate that the time from the time of failure to the time of sunset is time of failure - time of sunset = -605 [minutes]. Therefore, the time from sunrise to the time of failure (No. 25 in Figure 4) and the time from the time of failure to the time of sunset (No. 26 in Figure 4) are input to the machine learning classification unit 13. In this specification, sunrise and sunset times are also considered as a type of weather information, and are included in the "weather information" as defined in the claims of the present invention.

[0041] Figure 8 is a diagram illustrating the method of feature-quantifying the land use information in Figure 4. Figure 8(a) is statistical data showing land use data 71 in an area where power lines are laid. The divided grid (mesh) is a figure composed of cells that divide the target area into equally spaced grids, with each equally spaced grid point containing geographic information indicating how the land is used. Here, an example is shown where the area is divided into grids X1 to X13 in the horizontal direction (east-west) and grids Y1 to Y9 in the vertical direction (north-south). In this figure, a total of 117 grids are shown as an example for the sake of explanation, but in actual statistical data, grids with sides of several hundred meters are defined for the target land, so the total number of grids is enormous. All grids are classified by land use categories that indicate how the land is used, by applying colors, patterns, etc. In Figure 8(a), the land use categories are divided into fields 72, rivers 73, building sites 74, paddy fields 75, and forests 76, and patterns are applied to the corresponding grids. In actual land use data,71 land use is often displayed using color rather than patterns. Note that in the example in Figure 8(a), for ease of understanding, the number of land use classifications is shown as five, and an example with blank grids where no land use is indicated is also shown. However, in the actual statistical data provided by the Ministry of Land, Infrastructure, Transport and Tourism, numerous classifications are used for land use, such as wasteland, beaches, other agricultural land, roads, railways, lakes and marshes, and golf courses, and a land use classification is assigned to every grid.

[0042] In this embodiment, the land use classifications of the grids through which the power transmission line 300 passes from point A to point B, i.e., both ends of the power transmission line 300 where a failure has occurred, are counted. For example, pattern 74 indicates building land, and here the power transmission line 300 passes through four grids: (X5, Y3), (X5, Y4), (X6, Y4), and (X6, Y5), so the count for building land 74 is 4. Also, pattern 73 indicates a river, which is a river with a certain width or more, and in this land use data 71, it indicates that the river flows as follows: (X6, Y9~Y6), (X7, Y5), (X8, Y4)... Here the power transmission line 300 passes through the (X7, Y5) grid, so the count for river 73 is 1. The number of grids through which building site 74 and river 73 pass can be calculated in the same way, and as a result, a table showing the land use status from point A to point B of power transmission line 300 can be created, as shown in Figure 8(b).

[0043] In Figure 8(b), the line name 77a is information indicating a section of the power transmission line 300, and is stored as "AB line" as shown in 77b. The right side of item 78a stores the names of the counted land use categories. Here, they are stored in the order of field 72, river 73, building site 74, paddy field 75, and forest 76. The grid count 78b is the count value of the items corresponding to each item 72 to 76. For example, it indicates that within the AB line there are 3 grids for field 72, 1 grid for river 73, 4 grids for building site 74, 5 grids for paddy field 75, and 4 grids for forest 76. The total value of these grid counts 78b is 17. The ratio 78c stores the ratio value of the item to the total value of 17 for each item. For example, since there are 3 grids for field 72, the ratio of 0.18 to the total of 17 is calculated and stored in the table. In this way, the calculated land use status is compiled into a table as the proportion of land use categories and used as features. These are input to the machine learning-based classification unit 13 as features shown in No. 27 of Figure 4. In the example of Figure 8(a), the land use categories of the two-terminal section from point A to point B were aggregated, but the land use categories of a three-terminal area such as points A, B, and C, or more areas, may also be aggregated and used as features. Even in the case of three or more terminal areas, the same features as the table in Figure 8(b) can be calculated by counting all the land use categories of the grid within the area enclosed by them.

[0044] The count of land use classifications for another power transmission line 300 (not shown) leading up to point A, and the count of land use classifications for the grid through which another power transmission line 300 (not shown) extending from point B passes, can be counted in the same way as shown in Figure 8. These land use classification counts should be aggregated for each power transmission line when the data on the land use information server 120 is updated and stored in the data storage unit 22 of the storage device 20. If the land use classifications are prepared in advance in this way, when an accident occurs on the power transmission line 300, if the section where the accident occurred (in this case, section A to section B) can be identified, features can be obtained from a table like the one shown in Figure 8(b) and input into the machine learning-based classification unit 13.

[0045] Next, the procedure for estimating the cause of a power system failure according to this embodiment will be explained using the flowchart in Figure 9. The procedure shown in this flowchart can be implemented in software by having the CPU 10 shown in Figure 3 execute a computer program (power system failure cause estimation program). First, as preparation before executing this flowchart, the power system failure cause estimation device 1 selects the types of weather information to be used as new features and land use information, obtains the information from an external weather information server 110 and a land use information server 120, and records it in the data storage unit 22 of the storage device 20. Since the current weather information is updated at regular intervals, it may be acquired before a failure occurs in accordance with the update timing, or the current weather information may be acquired immediately after detecting a failure. Note that the weather information and land use information data acquired from external sources are not limited to paid or free open data, but may also be special data provided by specific companies, or other weather information, land use information, land management information, land development information, wildlife habitat information, etc. Furthermore, the means of acquiring weather information and land use information data are not limited to acquiring them via the network 100, but may also include methods of inputting them via a storage medium or manually inputting them via the input device 15.

[0046] Next, the monitor 180 of the power system fault cause estimation device 1 individually sets which items to extract from the weather information to be used as new features. In this embodiment, these extracted items for use as features correspond to items No. 16 to 27 in Figure 4. Next, the monitor 180 sets when to acquire the features No. 16 to 27 and which items to extract as features. For example, the items to be acquired and used from an external source can be selected in advance as data No. 16 to 24 in Figure 4. When monitoring power lines in areas where snowfall is not expected, such as from May to September, it is possible to set it so that explanatory variables No. 21 to 24 in Figure 4 are not used.

[0047] Regarding land use information for No. 27 in Figure 4, it is advisable to obtain statistical data, such as the land use data 71 shown in Figure 8(a), from an external source and store the statistical data for the area corresponding to all power transmission lines in advance in a storage device. Feature extraction is performed immediately after the failure occurs using the method shown in Figure 8. Once the preliminary preparations for starting to use the power system failure cause estimation device 1 are complete, the CPU 10 starts executing the power system failure cause estimation program.

[0048] In the flowchart of Figure 9, first, the power system fault cause estimation device 1 performs conventional power monitoring, that is, monitoring whether there are any abnormalities or faults in the transmission line 300 (step 201). The power monitoring includes monitoring the measurement signals of the three-phase voltage and zero-phase voltage of the power system, and monitoring information that can be obtained by existing measuring devices. These monitoring techniques are conventionally known, so their explanation is omitted here. Next, the CPU 10 of the power system fault cause estimation device 1 determines whether or not a fault has occurred in the power system based on the results of the power monitoring (step 202). If no fault has occurred, the process returns to step 201 and the power monitoring continues.

[0049] If a fault is detected in step 202, the CPU 10 identifies the location of the fault (step 203). This identification of the fault location is performed using a known method. The entity that identifies the fault location may be the power system fault cause estimation device 1, or the CPU 10 may be configured to obtain fault detection results from a monitoring system (not shown) separate from the power system fault cause estimation device 1. The fault location can be identified, for example, between points A and B on the power transmission line 300, or a specific narrowed-down location between points A and B. It may also be identified as an area encompassed by multiple points such as points A, B, and C.

[0050] Next, the CPU 10 extracts the power monitor values ​​at the point of failure, i.e., the recorded information 42 at the time of failure shown in Figure 2, the oscilloscope data A44 at the time of failure, the oscilloscope data B47 at the time of failure, etc., as training data 150 for machine learning (step 204). The extraction work for machine learning using these oscilloscope data A44 and oscilloscope data B47 utilizes conventionally known methods.

[0051] Next, the CPU 10 obtains the latest weather information for the vicinity of the failure location from the weather information server 110 via the network 100. It also reads past weather information from the data storage unit 22 in the storage device 20 or obtains it from the weather information server 110 via the network 100. Next, it performs extraction and aggregation from the acquired information as shown in Figures 6 and 7 to calculate feature quantities (No. 16 to 26 in Figure 4) for input into the machine learning-based classification unit 13 (step 205).

[0052] Next, the CPU 10 reads the land use information near the fault location stored in the data storage unit 22 of the memory device 20, performs extraction and aggregation as explained in Figure 8, and calculates feature quantities to be input to the machine learning-based classification unit 13 as shown in Figure 8(b) (step 206). Alternatively, the feature extraction work shown in Figure 8 may be performed for each section of the power transmission line, aggregated according to pre-selected items, stored in the data storage unit 22, and the data for that section may be read immediately after fault detection.

[0053] Using the features prepared in steps 204-206 above, the machine learning-based classification unit 13 performs fault cause estimation using AI technology (step 207). In this embodiment, the machine learning-based classification unit 13 classifies 12 types of faults in the power transmission line 300, which are assumed to be the main causes of failure: lightning, bird droppings, metal objects, bird contact, vehicles, salt damage, animals, snow damage, snakes, flying objects, vines, and tree contact. It performs classification using two types of AI: a decision tree that represents data classification in a tree structure using 27 explanatory variables (see Figure 4), including weather and land use information added in this embodiment, and a gradient boosting decision tree that is an extension of the decision tree capable of high accuracy and complex branching.

[0054] The failure cause estimated using AI is displayed on the operator's output device 16 shown in Figure 1 (step 208). In this case, the system may be configured to provide the information not only to the output device 16 of the failure cause estimation device 1, but also to relevant departments within the company and external parties via the network. The above procedure completes one cause estimation method associated with the current failure. The program that executes the procedure in Figure 9 is executed repeatedly and continuously, and once the processing in step 208 is completed, it starts the processing in step 201 in order to repeat the next cause estimation process.

[0055] Figure 10 shows the display screen 161, which displays the estimated failure cause results output to the output device 16 in Figure 3. On the display screen 161, the failure cause items are listed on the left, and the probability of each failure cause (a value estimated using AI) is displayed on the right. Here, the top three estimated failure causes are shown, indicating that the first cause 162 is salt damage with a 60% probability, the second cause 163 is the influence of trees with a 20% probability, and the third cause 164 is the influence of bird droppings with a 15.5% probability. Based on the estimated top failure causes and their probabilities, the supervisor 180 who secures the failure recovery personnel can recognize the details of the accident in advance before going to the site, and can prepare personnel and equipment corresponding to the failure cause, thus enabling faster recovery with reduced variability.

[0056] As described above, this embodiment actively utilizes weather and land use information, which were not previously input, in addition to oscilloscope data and information estimated using oscilloscope data that power system management companies have conventionally acquired, and performs machine learning using an AI model, thereby achieving even greater accuracy than conventional fault cause classification. In this invention, since a machine learning model based on factual data from past accidents is used, there is no need for prior assumptions such as probability distributions showing the relationship between surrounding information and fault causes by the monitor 180, and there is an advantage that classification without subjectivity is possible. Furthermore, in this invention, surrounding information such as weather information and map information is used as data for machine learning that links past accidents, and there is no need for humans to set the probability distribution of fault causes in advance. Since the probability of fault causes is analyzed by the machine learning model, there is an advantage that the estimation accuracy improves as more surrounding information is added, and the amount of human work does not change regardless of the amount of information added. Moreover, since the method of this embodiment uses a machine learning model that treats surrounding information input as equivalent to measurement signal input, it is possible to obtain the fault cause probability for all fault causes without missing the correct fault cause during processing.

[0057] Although the present invention has been described above using examples, these examples are provided as illustrations and are not intended to limit the scope of the invention. These examples can be implemented in various other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These examples and their variations are included in the scope and spirit of the invention, as well as in the claims and their equivalents. [Explanation of Symbols]

[0058] 1 Power system failure cause estimation device 10 CPU 11 Feature Item Input Section 12 Feature Calculation Unit 13. Classification Unit using Machine Learning 14. Output section for fault cause 15 Input device 16 Output device 20 Storage device 21 Program Storage Unit 22 Data Storage Unit 25 Power measurement signal input section 26 Peripheral Information Input Section 30 Power measuring device 31 Power measurement signal 100 Networks 110 Weather Information Server 120 Land Use Information Provision Server 180 Observer 300 power transmission lines

Claims

1. A power measurement signal input unit that inputs measurement signals of power transmitted by the power grid, and a feature quantity calculation unit that calculates a first feature quantity from the data input to the power measurement signal input unit, A failure cause estimation unit that estimates the cause of failure in the power system based on the first characteristic quantity, A power system failure cause estimation device having a failure cause output unit that outputs the estimated failure cause, A peripheral information input unit is provided for inputting peripheral information related to the aforementioned power system. The feature calculation unit calculates a second feature from the data input from the peripheral information input unit, The failure cause estimation unit estimates the cause of failure using the first and second feature quantities. A power system failure cause estimation device characterized by having a feature item input unit that allows the items calculated as the second feature quantity to be arbitrarily changed.

2. The measurement signal input to the power measurement signal input unit includes measurement signals of three-phase voltage and zero-phase voltage acquired by sensors provided in the power system. The power system fault cause estimation device according to claim 1, characterized in that the feature calculation unit calculates the first feature by image recognition of the pulse waveform of the input measurement signal.

3. The feature calculation unit, As a process for creating the first feature quantity, the measurement signal of the voltage at the time of the accident, which is included in the measurement signal, is input, and digital signal processing is performed. The power grid failure cause estimation device according to claim 2, characterized in that, as a process for creating the second feature quantity, weather information and land use information are integrated to create table data.

4. The failure cause estimation unit calculates the probability of each of the failure causes using a nonlinear classification model based on a gradient boosting decision tree or a random forest, or a classification model achieved through ensemble learning using multiple models combining these. The power system failure cause estimation device according to claim 3, characterized in that the failure cause output unit displays the calculated failure causes and probabilities in descending order of probability on the output device.

5. The power system failure cause estimation device according to claim 3, characterized in that the feature calculation unit calculates the second feature after the failure occurs by selecting the weather information corresponding to the power transmission line section including the failure location.

6. The power system failure cause estimation device according to claim 5, characterized in that the feature calculation unit collects land use information of points through which the power transmission line passes in advance and stores it in a storage device.

7. A power measurement signal input unit that receives measurement signals for power transmitted by the power grid, A processor that implements the functions of a feature calculation unit that calculates a first feature quantity from data input to the power measurement signal input unit by executing a computer program, a failure cause estimation unit that estimates the cause of failure in the power system based on the feature quantity, and a failure cause output unit that outputs the estimated cause of failure. A method for estimating the cause of a power system failure in a power system failure cause estimation device having a memory device, A peripheral information input unit is provided to acquire peripheral information related to the power system from an external device. The aforementioned processor, The power measurement signal input unit detects whether or not a power system failure has occurred from the measurement signal input to the power measurement signal input unit. The first feature quantity is calculated from the measurement signal at the time the fault occurred. Using the weather information corresponding to the power transmission line section including the location of the fault, obtained from the surrounding information input unit, and the land use information of the power transmission line section, a second feature quantity is calculated. A method for estimating the cause of a power system failure, characterized by inputting the first feature quantity and the second feature quantity into a failure cause estimation unit to estimate the cause of the power system failure.

8. The aforementioned weather information is open data obtained from an external source via a network. The aforementioned land use information is external data that classifies the land use status of the land on which the power transmission lines are installed, The failure cause estimation method according to claim 7, characterized in that the processor pre-stores the acquired land use information in the storage device.

9. The failure cause estimation unit calculates the probability of each of the failure causes from the input first feature and second feature using a nonlinear classification model using a gradient boosting decision tree or a random forest, or a classification model obtained by ensemble learning of multiple models combining these, The failure cause estimation method according to claim 8, characterized in that the failure cause output unit displays the calculated plurality of failure causes and the estimated probability of those failure causes on an output device.

10. As the aforementioned weather information, observation data from observation stations near the location where the malfunction occurred is used, including information such as temperature, precipitation, wind speed, wind direction, snowfall, and snow depth at the time of the malfunction. The aforementioned processor, The weather information, which can be obtained from an external source at regular intervals, is stored in the storage device. The failure cause estimation method according to claim 9, characterized in that when the failure is detected, the latest weather information is obtained and the second feature quantity is calculated in combination with past weather information stored in the storage device.

11. The aforementioned land use information is a land use subdivision mesh composed of cells that divide the target area into equally spaced grids, and each equally spaced grid point contains geographic information indicating how the land is used. The aforementioned processor, The storage device pre-stores the land use subdivision mesh that covers the entire land on which the power transmission line is installed, The method for estimating the cause of a failure according to claim 9 or 10, characterized in that, when the failure is detected, the second feature quantity is calculated from the land use information shown in the land use subdivision mesh through which the power transmission line, including the location of the failure, passes.

Images