Method and system for determining an electric power grid condition, and measurement units and computer programs for use in the method

Abstract

This invention discloses a method and apparatus for determining the state of a power grid. A plurality of measurement units are connected to the power grid. Each measurement unit is configured to perform measurements of at least one characteristic of the electricity flowing in the power grid to obtain characteristic data. The method includes performing a classification procedure at each measurement unit to determine at least one piece of respective classification data based on the characteristic data. Each piece of respective classification data is related to the state of the power grid and indicates one or more respective categories of that state. Each respective category includes one of a predetermined plurality of categories of the state of the electricity flowing in the power grid. The method includes transmitting the determined classification data from each measurement unit to a monitoring system.

Description

Determine the condition of the power grid This invention relates to a method, apparatus, and computer software for determining the state of a power grid. Monitoring the condition of the power grid is useful for maintaining a stable power supply. Conditions occurring on the power grid (such as anomalies including generator tripping and interconnection tripping) can have undesirable consequences for consumers. Therefore, it is desirable to identify any conditions occurring on the power grid. It is also desirable to identify these conditions quickly so that any faults can be swiftly corrected or compensated for. Existing methods for determining the condition of a power grid include those involving periodically measuring electrical characteristics using a plurality of measuring units at their respective locations within the power grid. The obtained characteristic data is then sent to a monitoring system, which analyzes the data to determine, for example, whether a power generation facility has tripped. However, in such systems, there is a trade-off between accurate and rapid determination of the power grid's condition. To improve accuracy, measurements can be performed more frequently. But more frequent measurements require transmitting a larger amount of data to the monitoring system, which in turn introduces latency. According to a first aspect of the present invention, a first method for use in a power grid, in which electricity flows, wherein a plurality of measurement units are connected to the power grid, each of the measurement units being configured to perform measurement of at least one characteristic of the electricity to obtain characteristic data, the method comprising: performing a classification procedure at each measurement unit to determine at least one piece of respective classification data based on the characteristic data, each piece of respective classification data being related to the state of the power grid and indicating one or more respective categories of the state of the power grid, each respective category including one of a predetermined plurality of categories of the state of the electricity flowing in the power grid; transmitting the determined classification data from each measurement unit to a monitoring system; and determining a consensus classification of the state of the power grid at the monitoring system based on the classification data, wherein the classification procedure comprises: performing a machine learning-based procedure based on the characteristic data using a machine learning-based system of measurement devices held at the measurement units; and determining the classification data based on the output of the machine learning-based system of the measurement devices. Depending on the circumstances, the consensus classification may include: selecting the state of existence as the one with the highest number of such measurement units. Depending on the circumstances, the consensus classification may include: inputting the received classification data into a consensus classification system based on machine learning. According to a second embodiment, a method for use in a power grid, in which electricity flows, is provided, wherein a plurality of measurement units are connected to the power grid, each of the measurement units being configured to perform measurement of at least one characteristic of the electricity to obtain characteristic data, the method comprising: performing a classification procedure at each measurement unit to determine at least one piece of respective classification data based on the characteristic data, each piece of respective classification data being related to the state of the power grid and indicating one or more respective categories of the state of the power grid, each respective category including one of a predetermined plurality of categories of the state of the electricity flowing in the power grid; and transmitting the determined classification data from each measurement unit to a monitoring system, wherein the classification procedure includes: performing a machine learning-based procedure based on the characteristic data using a machine learning-based system of measurement devices held at the measurement unit; and determining the classification data based on the output of the machine learning-based system of the measurement devices. In the first and / or second state, depending on the circumstances, the method includes receiving parameters of the measurement device based on a machine learning system from the monitoring system by each of the measurement units, and implementing the measurement device based on the machine learning system according to the received parameters. In the first and / or second sample, the classification procedure may, as appropriate, include preprocessing the feature data to obtain preprocessed data, the preprocessing including applying a transform to the feature data, the transform including at least one of wavelet transform and fast Fourier transform, and the classification procedure includes inputting the preprocessed data into a machine learning-based system of the measurement device. In the first and / or second state, depending on the situation, the method includes performing a training procedure to train the measurement device into a machine learning-based system using training data, the training data including training feature data indicating the at least one feature. In the first and / or second sample, the training data may, as appropriate, include baseline real data of one or more of the predetermined plurality of categories indicating the condition. In the first and / or second state samples, depending on the situation, the measurement device of each measurement unit is based on a different machine learning system, and the training program for training each of the respective measurement device's machine learning system is based on training characteristic data indicating at least one characteristic of the power at the respective location of the measurement unit. In the first and / or second state, the second method may, as appropriate, include: at the first measurement unit, performing a training procedure based on characteristic data obtained at the first measurement unit to determine an update to the machine learning-based system of the measurement device; transmitting data indicating the updates; and receiving the updated machine learning-based system of the measurement device at the plurality of measurement units, wherein the updated machine learning-based system of the measurement device has been generated based on the data indicating the updates, and the classification procedure includes inputting the characteristic data into the updated machine learning-based system of the measurement device. In the first and / or second state samples, the plurality of categories of conditions may include a plurality of categories of anomalies in the power flow, depending on the circumstances. In the first and / or second state, depending on the circumstances, the plurality of categories of anomalies include at least one of the following: power generation facility tripping, interconnector tripping, interconnector starting, transmission line tripping, demand load tripping, transformer tripping, rewiring of the power flow, converter tripping, and amplitude of voltage harmonics that meet the criteria. In the first and / or second state samples, the plurality of categories of the state may, as appropriate, include categories indicating normal power flow and categories indicating abnormal power flow. In the first and / or second state samples, the training characteristic data may include data representing normal power flow, depending on the circumstances. In the first and / or second state samples, depending on the circumstances, the training characteristic data is essentially composed of the data representing the normal power flow. In the first and / or second state samples, the at least one characteristic may, as appropriate, include frequency, the derivative of frequency with respect to time, inertia, voltage, effective power, reactive power, and / or current. In the first and / or second state samples, the characteristic data may have a sampling rate of more than 120 Hz, depending on the circumstances. In the first and / or second sample, the classification data may, as appropriate, indicate the probability of the presence or absence of one of the predetermined plurality of categories. In the first and / or second state samples, depending on the circumstances, the characteristic data is obtained during the measurement time interval, and the condition of the power grid is related to a time interval later than the measurement time interval. According to the third state sample, a plurality of measurement units are provided, which are configured to perform the method according to the second state sample. According to a fourth state sample, a system is provided, comprising: a plurality of measurement units; and a monitoring system configured to perform a method according to a first state sample. According to the fifth state, a plurality of computer programs are provided, wherein when the respective processor of the plurality of measurement units executes one of the computer programs, the plurality of measurement units execute the method according to the second state. Further features and advantages of the invention will become apparent from the following description of preferred embodiments of the invention, given only by way of example and with reference to the accompanying drawings. Power supply from suppliers such as power plants to consumers such as homes and businesses is typically delivered via power grid 100. Figure 1 illustrates an exemplary power grid 100 in which embodiments of the present invention may be implemented, including transmission grid 102 and distribution grid 104. Transmission network 102 is connected to power generation facility 106, which may be a nuclear power plant or a gas-fired power plant. For example, transmission network 102 transmits a large amount of electrical energy at very high voltage (typically around several hundred kV) from power generation facility 106 to distribution network 104 via power lines such as overhead power lines. Transmission network 102 is connected to distribution network 104 via transformer 108, which converts the power supply to a lower voltage (typically around 50 kV) for distribution in distribution network 104. Distribution network 104 is connected to a regional network via substation 110, which includes further transformers for converting to lower voltages. This regional network supplies power to power-consuming devices connected to power grid 100. The regional network may include networks for residential consumers, such as urban networks 112, which supply power to household appliances within private residences 113, which draw a relatively small amount of power, approximately a few kW. Private residences 113 may also use photovoltaic systems 117 to provide a relatively small amount of power for household appliances or to supply power to power grid 100. The regional network may also include industrial sites such as factories 114, where larger appliances operating in the industrial site draw larger amounts of power, approximately a few kW to MW. The regional network may also include networks of smaller generators (such as wind farms 116) that supply power to power grid 100. Although only one transmission network 102 and one distribution network 104 are shown in Figure 1 for the sake of simplicity, in practice, a typical transmission network 102 supplies power to multiple distribution networks 104, and a transmission network 102 can also be interconnected with one or more other transmission networks 102. Electricity flows as alternating current (AC) in the power grid 100 at the system frequency (which may be referred to as the grid frequency, typically 50 Hz or 60 Hz depending on the country). The power grid 100 operates at a synchronous frequency, such that the frequency is substantially the same at all points in the power grid 100. Power grid 100 may include one or more DC interconnects 119 providing direct current (DC) connections between power grid 100 and other power grids. Typically, the DC interconnects 119 are connected to the high-voltage transmission network 102 of power grid 100. The DC interconnects 119 provide DC links between various power grids, allowing power grid 100 to define areas operating at a given synchronous grid frequency and unaffected by variations in the grid frequencies of other power grids. For example, the UK transmission network is connected to the synchronous grid of continental Europe via HVDC cross-channel interconnects. The power grid 100 also includes a plurality of measurement units 118 (such as measurement units 118a, 118b) and a monitoring system 120. Figure 2a is a flowchart illustrating a consensus classification method 200 for determining the state of power grid 100. Method 200 includes an "edge method 210" performed by each of a plurality of measurement units 118 and a "central method 220" performed by the monitoring system 120. Figures 2b and 2c are flowcharts of edge method 210 and central method 220 in more detail, respectively. A schematic diagram of this measurement unit 118a is shown in Figure 5a, and a schematic diagram of the monitoring system 120 is shown in Figure 5b. Each measurement unit 118 is connected to the power grid 100 at its respective location. The measurement units 118 may be located in the distribution network 104 or the transmission network 102, or at any other location in the power grid 100. Although only seven measurement units 118 are shown in Figure 1 for simplicity, it should be understood that in practice, the power grid 100 may include hundreds or thousands of such devices. Each measurement unit 118 is configured to perform measurements of at least one characteristic of the power supply to obtain characteristic data (e.g., measurement data), and may include at least one sensor 510 for this purpose. For example, the measurement unit 118 may be configured to measure one or more of the following: the frequency of the power flowing in the power grid 100, the derivative of the frequency with respect to time, inertia, voltage, active power, reactive power, and / or current. The characteristic data may additionally or alternatively include data related to harmonics of voltage or other characteristics. The characteristic data may have a sampling rate (number of measurements per unit time) exceeding 120 Hz, 1 kHz, 10 kHz, 40 kHz, or 400 kHz. Generally, the higher the sampling rate, the more reliable and accurate the measurement unit 118 can be in determining the condition of the power grid (e.g., whether an abnormal event has occurred on the power grid 100, and, where applicable, what type of event has occurred). Each measurement unit 118 stores characteristic data recorded at measurement time intervals in memory 504. The measurement time interval can be the time interval ending with the most recently recorded measurement and beginning at a previous time. The previous time can be a fixed point in time, such as January 1, 2020, and in this case, the characteristic data accumulates gradually over time. Alternatively, the previous time can be a fixed time interval earlier than the most recently recorded measurement (such as 10 seconds before the most recently recorded measurement), such that the characteristic data consists of data from the most recently recorded 10 seconds. In the latter case, characteristic data recorded before the measurement time interval is erased when it is no longer within the measurement time interval, to efficiently utilize space in memory 504. In any case, the characteristic data includes a time series of measurements of at least one characteristic of the power. For example, the characteristic data may include a time series of measurements of the AC frequency of the power grid 100. Edge method 210 includes, in step 212, performing a classification procedure to determine at least one individual classification record based on characteristic data. Each individual classification record is related to the condition of power grid 100. Here, "condition of power grid 100" refers to the grid condition, such as (for example) conditions arising from events occurring on power grid 100 (such as power consumption from the grid or sudden changes in power supply to the grid), and may include the condition of components of power grid 100 (such as transmission lines, generating facilities, converters, or interconnectors). Examples of events may include anomalies such as generating facility tripping, interconnector tripping, interconnector startup, transmission line tripping, demand load tripping, transformer tripping, and changes in system state (such as rewiring of power flows). Events are not necessarily anomalous. For example, in the case of scheduled power plant startup or shutdown, the event may be a planned event. In other instances, the grid condition is not related to a specific identified event. For example, grid conditions can be related to distortion levels (e.g., harmonic amplitudes in voltage) introduced by converters connected to the power grid (e.g., converters used in solar farms). In this case, grid conditions can indicate normal distortion levels or high distortion levels (e.g., distortion levels exceeding thresholds, such as harmonic amplitudes). The detected distortion can be constant; that is, grid conditions can be independent of changes in distortion levels but related to the current measurement of the distortion level. As another example, grid conditions can indicate a regular increase in the total demand load (electricity consumed by consumers) of power grid 100 during the early morning. In other instances, grid conditions may include oscillations in power flow characteristics for which no specific cause has been identified. The state of power grid 100 is sometimes normal; that is, the grid conditions are normal. The examples in this document are primarily described in relation to events and more specifically, anomalies, but the methods can also be applied to determine grid conditions not associated with specific events. Each piece of classification data indicates one or more individual categories of power grid conditions. For example, as shown in Figure 2e, various categories of conditions labeled 1 to 8 are displayed, and each of the plurality of measurement units 118a, 118b, and 118c indicates the presence of each of the stated conditions. Each individual category includes one of a predetermined plurality of categories of the condition of electricity flowing in the power grid 100. In the example shown in Figure 2e, the predetermined plurality of categories of conditions includes condition categories 1 to 8. In this example, the categories indicated by the classification data are one of these categories. In some instances, categorical data is associated with the presence of a category of grid condition (e.g., anomaly) during the measurement interval. Such instances are appropriate when it is desirable to determine the state of the power grid 100 during the measurement interval that has been obtained. For example, categorical data may include data indicating timestamps or time intervals associated with the condition. For instance, a categorical data entry might indicate that a generating facility tripped at 12:36:01 on October 11, 2023. Alternatively, the categorical data might indicate that the generating facility was not functioning properly at 12:36:04 on October 11, 2023 (i.e., it tripped and has not yet restarted). Figure 2d illustrates an example of frequency versus time for electricity in the case where a generating facility tripped at 12:36:01 and was still not functioning properly at 12:36:04. As another example, classification data may indicate that a power generation facility will restart at a certain point in time during the time interval from 12:48:00 to 12:49:00 on October 11, 2023, without further specifying the restart time. Measurement unit 118 may store a table representing categories of conditions identified as existing during each of a plurality of time intervals, similar to the table shown in Figure 3a for training purposes. In some instances, the time associated with the condition of the power grid is the time within the measurement time interval; that is, the classification data is associated with the existence or non-existence of this category of condition of the power grid 100 during the measurement time interval. In some instances, classification procedures are used to predict classification data related to the grid conditions following a time interval in which characteristic data was acquired. For example, by observing that the damping ratio of periodic oscillations in the grid frequency signal increases over time, some transmission line and interconnect trips can be predicted up to 20 to 30 seconds before tripping. This phenomenon can be detected by analyzing the frequency signal using Yule-Walkers or Prony algorithms to determine the damping ratio. The rate of change of the damping ratio can be used to determine when the oscillation will become unstable and when the transmission line or interconnect experiencing such oscillation will trip. In situations where it is desirable to predict the future state of the power grid 100, a predictive procedure is appropriate. For example, a fault in a nuclear power plant (such as a decrease in output power) can develop gradually; early identification of this fault is desirable. The measurement unit 118 can maintain the measurement device based on a machine learning system (also referred to herein as an edge ML system), which is trained to predict the state of the power grid 100, or can generate frequency predictions to determine whether the frequency is expected to exceed a threshold; two examples are further described in detail below. The categories of conditions indicate events (such as anomalies) that cause a particular type of condition in the power flow. For example, a category might indicate one of the anomalies described above, such as generator tripping, interconnector tripping, interconnector starting, transmission line tripping, demand load tripping, transformer tripping, power flow redistribution, or converter tripping. These categories of anomalies have specific effects on various characteristics of the power flow. For example, interconnector tripping typically results in a momentary decrease in effective power. It can also lead to a decrease in the frequency of the power flow because the resulting imbalance between generation and power consumption causes a self-rotating generator to continuously draw kinetic energy, thus reducing the generator's rotational frequency and the frequency of the power flow. However, some faults in generator facilities result in a more gradual decrease in effective power, such as some faults in nuclear power generators. Furthermore, the frequency decreases at a lower rate after a generator facility trip compared to an interconnector trip. Additionally, transmission line tripping can cause sudden changes in frequency, accompanied by frequency oscillations. Therefore, the rate of frequency change can be used additionally or alternatively to identify more specific faults. Demand load tripping has a similar effect to the opposite effect of power generation facility tripping, as it leads to an increase in frequency. The increase in total demand load offset by the increase in power generation can be detected by measuring the inertia of the power grid 100. When inertial contributing generators are turned on to offset the increase in demand load, the inertia of the power grid 100 may increase. The multiple categories of conditions may include categories indicating that the power grid 100 is in a normal state. For example, when the frequency of the power grid 100 is stable and at or near its nominal level (e.g., 50 Hz in the UK), the measurement unit 118 determines that the power is in a normal state during a first time interval. The frequency versus time graph shown in Figure 2d illustrates an example of frequency behavior up to 12:36:01 when the power grid 100 is in a normal state. In some instances, categorical data indicates the probability of the presence of one of a predetermined plurality of categories. For example, categorical data might indicate a 70% probability of transformer tripping during the measurement interval. The probability of each category's presence can be determined by the measurement unit 118. The probabilities can be output by a machine learning-based system of the measurement device based on characteristic data, as described in further detail below. Alternatively, each piece of classification data may include a binary number indicating the presence or absence of a specific category. For example, as shown in Figure 2e, each possible category is represented by a label, and in response to determining that a specific category exists, measurement unit 118 may set the corresponding label to 1. If measurement unit 118 determines that a specific category does not exist, it may set the label to 0. The binary number may be determined by, for example, comparing the probability mentioned above with a threshold value (e.g., 0.5), and setting the corresponding label to 1 if the probability is greater than the threshold value and setting the corresponding label to 0 if the probability is less than the threshold value. Compared to indicating probabilities, using this binary data reduces the amount of data that measurement unit 118 transmits to monitoring system 120. Measurement unit 118 can identify the existence of multiple categories. For example, measurement unit 118 can identify the existence of interconnect tripping and frequency oscillation over time based on characteristic data. Measurement unit 118 may include the measurement device mentioned above based on a machine learning system (also referred to herein as an "edge ML system"), which includes a single neural network for identifying which of the multiple categories of conditions exists at a given time, or a neural network for each condition category. The training of such neural networks is described below. In some instances, the classification data does not specify any particular condition category, but only whether a condition has occurred on power grid 100, or whether the power grid is in a normal state. For example, measurement unit 118 can determine that no abnormality has occurred in power grid 100 based on characteristic data. An example of characteristic data indicating no abnormality in power grid 100 is shown in the frequency diagram before 12:36:01 in Figure 2d. In some instances, the classification process includes preprocessing the feature data to obtain preprocessed data, and then inputting the preprocessed data into an edge ML system. Preprocessing may include applying transforms (such as discrete wavelet transform or fast Fourier transform) to the feature data. More generally, preprocessing may include any feature extraction operations performed on the characteristic data. For example, preprocessing may include obtaining the derivative of the transformed data, obtaining the variance of the characteristic data, and / or obtaining a measure describing the transitions (such as voltage transitions) that occur in the characteristic data. In any case, preprocessing results in input data being fed into the edge ML system. The input data may include vectors on which the edge ML system operates, where each component of the vector represents a value obtained from the preprocessing of a time series of frequency measurements; for example, frequency data acquired at time intervals of 1 second, 2 seconds, 3 seconds, 4 seconds, 5 seconds, 10 seconds, 1 minute, or 10 minutes for a given wave frequency and time. Alternatively, in instances where no preprocessing is performed, the components represent one of the original frequency measurements. By applying preprocessing to the data, the complexity of patterns occurring in the unprocessed data can be reduced, thereby reducing the burden on edge ML systems when recognizing patterns. Therefore, applying preprocessing to the data allows for a reduction in the complexity of edge ML systems and / or their training procedures. In this example, edge method 210 includes inputting characteristic data (which may have undergone the preprocessing described above) into a machine learning-based procedure using an edge ML system held at measurement unit 118, and determining classification data based on the output from the edge ML system. Before using edge method 210, the edge ML system is trained based on (as preprocessed) characteristic data to determine at least one piece of individual classification data, each piece of individual classification data relating to the state of the power grid and indicating one or more individual categories of the state of the power grid, each individual category including one of a predetermined plurality of categories of the state of electricity flowing in the power grid 100. The manner in which the edge ML system can determine classification data from characteristic data is described below with reference to the training section. In some instances, the edge ML system may output the probability of the existence of a certain category of power condition (e.g., the probability that an interconnector has tripped). Using machine learning-based systems allows for reliable and / or accurate determination of the presence or absence of categories of power conditions. Furthermore, compared to Supervisory Control and Data Acquisition (SCADA) based systems that monitor the status of individual power grid components, embodiments of the present invention detect power grid events and identify their types by measuring at locations within the power grid (e.g., at power lines, which can be remotely located relative to individual power grid components) without monitoring individual power grid components. In addition, the measurement unit 118 can further input time data representing the time of day, week, month, and / or year for which the measurement is performed into the edge ML system, and further determine classification data based on the time data. In these examples, the edge ML system may be able to identify typical patterns that occur periodically (e.g., daily or weekly) in the power grid 100, and distinguish these patterns from patterns representing, for example, anomalies in the power grid 100. In some instances, measurement unit 118 self-monitors system 120 to receive parameters from the edge ML system. Measurement unit 118 implements the edge ML system based on the received parameters. Self-monitoring system 120 receiving parameters and transmitting classified data to the same monitoring system 120 allows for a simplified network architecture. The machine learning-based systems described herein may include, for example, feedforward neural networks, such as convolutional neural networks or recursive neural networks. The parameters received by the self-monitoring system 120 may include the weights of this neural network. In some instances, all measurement units 118 use the same edge ML system; that is, all edge ML systems used by measurement units 118 have the same parameters. However, in other instances, measurement units 118 may use edge ML systems with different parameters. Illustrative methods for training such edge ML systems are described in detail below. In some instances, non-ML methods are used to determine the classification data. For example, the classification procedure may include comparing characteristic data with one or more threshold values. For example, the rate of change of frequency may be compared with a positive threshold value. If the rate of change of frequency exceeds the threshold value, measurement unit 118 may determine that a demand load has tripped. If the rate of change of frequency is less than the negative threshold value, measurement unit 118 may determine that a generating facility has tripped. Alternatively, if the measured characteristic of the power is frequency and measurement unit 118 is located on the power grid (e.g., in the UK, with a nominal grid frequency of 50 Hz), the two such threshold values ​​may be 49.5 Hz and 50.5 Hz. If measurement unit 118 determines that the frequency is less than 49.5 Hz, measurement unit 118 may determine that a generating facility has tripped. If measurement unit 118 determines that the frequency is greater than 50.5 Hz, measurement unit 118 may determine that a demand load has tripped. In one example where the measurement unit 118 is configured to predict the future condition of the power grid 100, the measurement unit 118 may generate a prediction of the frequency of the power grid by fitting a function (e.g., a polynomial function) to a frequency measurement that has been obtained in the most recent period (e.g., the last 5 seconds of the measurement that has been obtained) and extrapolating the function to determine whether the frequency will exceed one of the aforementioned threshold values ​​in the future; an example of this procedure is further described in the applicant's patent application WO 2015 / 158525 A1. Edge method 210 includes, in step 214, transmitting determined classification data from each measurement unit 118 to the monitoring system 120. For example, measurement unit 118 may include an output interface 508 for communicating with monitoring system 120, through which classification data can be transmitted. Communication between monitoring system 120 and measurement unit 118 may include wired communication (e.g., using Ethernet wide area network) or wireless communication (which may include one or more of Global System for Mobile Communications (GSM), Universal Mobile Telecommunications System (UMTS), Long Term Evolution (LTE), 5G New Radio, fixed wireless access (such as IEEE 802.16 WiMax), and wireless network connections (such as IEEE 802.11 WiFi). Measurement units 118 do not need to transmit characteristic data to monitoring system 120 because the monitoring system alternatively uses classification data to determine a consensus classification of the state of power grid 100. Measurement unit 118 can individually transmit feature data and / or classification data to database 122. For example, feature data can be used for training the machine learning-based system described below. Classification data can be transmitted at regular intervals; for example, once classification data has been determined, classification data determined from the characteristic data recorded 1 second ago can be transmitted. This can be useful for the daily monitoring of power grid 100. Additionally or alternatively, classification data can be transmitted when measurement unit 118 has identified a new power grid event. For example, measurement unit 118 can detect an interconnector trip and, in response to detecting the interconnector trip, transmit classification data indicating that an interconnector trip has occurred. In addition to classification data, the measurement unit 118 can also transmit data representing signal quality, such as the estimated level of noise occurring in the characteristic data. As described above, method 200 includes the central method 220 performed by monitoring system 120 as shown in FIG2c. Central method 220 includes receiving classification data from each of the plurality of measurement units 118 in step 222. Monitoring system 120 includes an input interface 606 for communicating with the plurality of measurement units 118, through which classification data can be received. Central method 220 includes, in step 224, determining a consensus classification of the state of power grid 100 based on received classification data. The consensus classification may include one or more of a predetermined plurality of categories of state. For example, the consensus classification may be "normal" or "interconnector tripped". As described above, the classification data transmitted from a given measurement unit 118 may be related to the existence or non-existence of a certain category of condition in the power grid 100 at a specific time, and may include data indicating the time associated with the condition (e.g., determining the time when the condition occurred). The monitoring system 120 may process the classification data based on timestamps. For example, the monitoring system 120 may receive classification data associated with the time interval 12:36:03 to 12:36:04 from several measurement units 118, and form a set of classification data associated with that time interval received from different measurement units 118. Classification data related to this time interval can be received from different measurement units 118 at different times; for example, if an interconnect trip occurs at 12:36:01, one measurement unit 118 (e.g., the measurement unit 118 near the interconnect) can transmit classification data indicating the interconnect trip at 12:36:03:00, while another measurement unit 118 (e.g., the measurement unit 118 far from the interconnect) can transmit classification data indicating the interconnect trip at 12:36:03:01 (due to the limited propagation speed of changes in power characteristics). The monitoring system 120 can use timestamps to form a set of received classification data related to the time interval 12:36:03 to 12:36:04, and determine a consensus classification based on this set of classification data. The monitoring system 120 can periodically execute this procedure (e.g., once per second) to determine the consensus classification for each 1-second time interval. In this example, determining the consensus classification involves selecting the state as being classified by the highest number of measurement units 118. As shown in Figure 2e, measurement units 118a and 118b can determine the state as being "interconnector tripped," while measurement unit 118c alternatively determines the state as being "normal" (i.e., no abnormality has occurred in the power grid 100). The monitoring system 120, having received classification data indicating these determined categories from measurement units 118a, 118b, and 118c, determines that the category "interconnector tripped" has been identified as being present by more measurement units than any other category, and in response to this determination, selects the category "interconnector tripped" as the consensus classification. Selecting the state as being classified by the highest number of measurement units 118 provides a simple and / or computationally efficient method for determining the state of the power grid 100. In an example where the classification data received from measurement unit 118 indicates the probability of each category of condition, for each measurement unit 118, the monitoring system 120 can identify the category of condition associated with the highest probability. For example, if the classification data received from the first measurement unit 118a indicates that the probability of interconnect tripping is 20%, the probability of power generation facility tripping is 70%, and the probability of each other condition is less than 10%, then the monitoring system 120 can identify the condition identified by the first measurement unit 118a and associated with the highest probability as "power generation facility tripping". If the classification data received from the second measurement unit 118b indicates that the probability of interconnect tripping is 60%, the probability of power generation facility tripping is 40%, and the probability of each other condition is less than 10%, then the monitoring system 120 can identify the condition identified by the second measurement unit 118b and associated with the highest probability as "interconnect tripping". The monitoring system 120 can perform the same operation for all other measurement units 118. Next, the monitoring system 120 can generate classification data for each condition, indicating the number of measurement units 118 with the highest probability for that condition. The monitoring system 120 can select the condition with the highest count as the consensus classification of the state of the power grid 100. Before performing any of the probability-based calculations mentioned above, monitoring system 120 or measurement unit 118 may apply their respective weights to the probabilities. For example, when monitoring system 120 receives data representing signal quality from each of measurement units 118, monitoring system 120 may determine the weighted probabilities of each category of the situation, for example, by multiplying the received probability by a function of signal quality (e.g., a sigmoid function). As another example, monitoring system 120 may determine the weights of measurement unit 118 based on the historical accuracy of the classification data received from measurement unit 118. For example, monitoring system 120 may determine the weights as the ratio of correct classifications received from measurement unit 118. A correct classification may be defined, for example, as an instance where the consensus classification matches the category of the situation assigned the highest probability by measurement unit 118. Additionally or alternatively, the monitoring system 120 may, in response to the identification of a power grid component causing an anomaly (such as a tripped generating facility), determine a consensus classification based on the known failure rate of that power grid component. Power grid components known to have failed multiple times are generally more likely to fail again than those known to have failed fewer times. In addition to received classification data, the monitoring system 120 may also determine a consensus classification, for example, using a previously determined consensus classification or alternatively using publicly available data regarding anomalies observed in the power grid 100. As another example, generating facilities (especially gas-fired generating facilities) are more likely to trip shortly after startup, and interconnectors are more likely to trip while operating at their maximum power capacity. In addition to received classification data, the monitoring system 120 may also use data representing this information. In instances where categorical data only indicates whether characteristic data represents an anomaly, determining a consensus classification may include comparing the number of measurement units 118 whose classification indicates an anomaly with a threshold value. The threshold value may be, for example, 30% of the measurement units 118. If the threshold value is exceeded, the consensus classification may indicate the presence of an anomaly; if the threshold value is not exceeded, the consensus classification may indicate the absence of an anomaly. In other instances, determining a consensus classification involves: inputting received classification data into a consensus classification-based machine learning system, and determining the consensus classification based on the output of the consensus classification-based machine learning system. The consensus classification-based machine learning system is trained before use to generate an output representing the consensus classification based on the received classification data. The training of the consensus classification-based machine learning system is described in detail below. Using a consensus classification-based machine learning system allows for reliable and / or accurate determination of the state of power grid 100. We now describe exemplary methods for training edge ML systems and consensus classification machine learning-based systems based on training data and (where appropriate) benchmark real data. In the examples below, the training data used for training may be historical characteristic data acquired by measurement unit 118 during historical time intervals. This historical characteristic data may be acquired by measurement unit 118 and transferred to database 122 for storage. The training data may be preprocessed according to the preprocessing described above before it is used for training. Benchmark real data may include data indicating at least one of one or more predetermined plurality of categories indicating the presence or absence of power during the historical time interval. Human operators may obtain reports of known detected events (such as anomalies or a regular increase in total demand load during the morning) (such as interconnect tripping) that occurred in the power grid 100 during a specific time interval. Historical characteristic data related to the power grid 100, obtained during the historical time intervals in which events occurred, can be retrieved from database 122, for example, by searching database 122 using the time of the events, and used as training data. Human operators can set baseline true data based on detected events. As an example, baseline true data may include classification data indicating that the event occurred during the time interval, such as the label "1" for the category "power generation facility tripping". Training data can be grouped into sets, each set including historical characteristic data related to its respective historical event (or indicating the power grid 100 when no event occurred), and baseline true data indicating that its respective historical event has occurred and, depending on the event type (or indicating that no event occurred). The time matching of historical characteristic data and baseline true data, and how this can be used to train a machine learning-based system for predicting power grid conditions, are described below with reference to Figures 3a and 3b. Figure 2d illustrates an example of historical characteristic data, specifically representing frequency versus time data of electricity, which is related to power generation facility tripping. In some instances, for each category of electricity condition, the training data includes at least one (or at least 10, 100, 1,000, or 10,000) historical characteristic datasets for each category of condition. In some instances, supervised training procedures are performed to train the edge ML system. In others, baseline real-world data only indicates whether the grid condition is normal. In supervised learning, the edge ML system "guesses" whether a specific category of power condition exists during a given time interval (or whether the grid condition is normal depending on the baseline real-world data) based on characteristic data associated with that time interval. The loss function is used to modify the parameters of the edge ML system, where the loss function is high if the "guess" does not match the baseline real-world data, and low if they do. For example, if the baseline real-world data indicates that the correct category of an anomaly existing during a given time interval is "generator tripping," the loss function might penalize the incorrect classification of "interconnector tripping," or penalize the incorrect classification of an anomaly even if no anomaly actually exists. In instances where the edge ML system outputs probabilities rather than binary classifications, the loss function can be a continuous function of the output probabilities. Therefore, if the edge ML system outputs a probability of 0.2 for "power generation facility tripping", and the baseline real data indicates that the correct category of the anomaly is power generation facility tripping, the value of the loss function will be higher than the case where the output probability is, for example, 0.8. The loss function may include terms for categories of power conditions, and in such cases, the edge ML system is configured to output data (binary output or probability) representing the presence or absence of each category of condition during training, such that the edge ML system can be trained to output the presence or absence of each of multiple categories of condition. Using a single loss function containing many such terms simplifies the architecture of the edge ML system. However, as an alternative, the edge ML system may include multiple neural networks, each trained separately to determine the presence or absence of its respective condition (e.g., anomaly). The same training data can be used to train each neural network. In this case, the loss function of each neural network only includes terms that penalize the neural network for incorrect classification of its respective condition. As an example, one neural network may be trained to classify power generation facility tripping, while another neural network may be trained to classify interconnector tripping. In any of the supervised learning procedures described herein, backpropagation can be used to adjust the parameters of the marginal ML system and find parameters that minimize the loss function with respect to the training data. Minimizing the loss function with respect to the training data is effectively equivalent to finding the parameter set of the marginal ML system that most successfully determines classification data from feature data, at least with respect to the training dataset used. By using baseline real data indicating the presence or absence of at least one of one or more predetermined plurality of categories, measurement unit 118 can output more reliable data indicating the presence or absence of such conditions when implementing marginal method 210. Whether supervised or unsupervised learning is used to perform training, at least one characteristic of the electrical power used in training is generally the same characteristic used in step 212 of the marginal method 210 to determine the classification data. However, a pattern (such as oscillation) in one characteristic of the electrical power can also be observed in another characteristic. Therefore, in some instances, different characteristics can be used. For example, voltage amplitude can be used in the training procedure, while current amplitude can be used in the marginal method 210. The characteristics of the electricity (e.g., frequency, current, etc.) can vary depending on the location of the measuring unit 118. For example, a measuring unit 118 placed near a solar power field can observe more significant changes in the current due to changes in cloud cover, while a measuring unit 118 placed near a wind farm 116 can observe changes due to changes in wind speed. As another example, a measuring unit 118 placed near an interconnector (such as an HVDC cross-channel interconnector between the European power grid and the UK power grid) will observe large frequency changes generally within a very short time after the interconnector has tripped, while a measuring unit 118 further away from the interconnector will observe smaller changes at a later time. Therefore, in some instances, the edge ML systems of each measurement unit 118 are distinct machine learning-based systems, and the training procedures for each edge ML system are based on characteristic data indicating at least one characteristic (e.g., frequency) of the power at the location of the measurement unit 118. The supervised training procedures described above can be used to train different machine learning-based systems separately. In this case, the training data used to train the machine learning-based system used by a given measurement unit 118 may include characteristic data related to the location of that measurement unit 118, but not characteristic data obtained at other locations. For example, the training data used by a given measurement unit 118 to train its respective machine learning-based system may have already been obtained by that measurement unit 118. The monitoring system 120 may provide a consensus classification for a given time interval to the measurement unit 118. The provided consensus classification may be obtained from the reporting of events as described above. Next, measurement unit 118 can use the data it acquired during the time interval as training data and use the received consensus classification as benchmark real data to train its machine learning-based system locally. Monitoring system 120 can provide the consensus classification to multiple measurement units 118, each of which uses its own characteristic data acquired during the time interval for training. Therefore, the measurement units learn how the local behavior of the power grid 100 (such as that captured by characteristic data) can be used to determine which conditions exist in the power grid 100. By training using characteristic data of at least one characteristic (e.g., frequency) of the power at the location of the measurement unit 118, the measurement unit 118 can more reliably determine the presence of power flow. For example, a measurement unit 118 placed near a solar field will typically be less sensitive to sudden changes in power flow caused by variations in cloud cover due to its training; that is, it is less likely to determine the presence of anomalies for a given magnitude of change in power flow compared to another measurement unit not placed near a solar field. Meanwhile, using the training procedure described above, a measurement unit 118 placed in an area of ​​the power grid 100 that typically has a more stable power flow (such as an area of ​​the power grid 100 with higher inertia per unit area) can be trained to be more sensitive to changes in power flow. In cases where the edge ML systems of each measurement unit 118 are different machine learning-based systems, as described above, in some instances, the training data used by a given measurement unit 118 includes a combination of feature data related to the location of that measurement unit 118 and feature data related to the location of(s) other measurement units 118 experiencing similar power grid conditions. For example, if a measurement unit 118 is located near a specific type of power grid component (e.g., a power generation facility), it can use data obtained by measurement units 118 near other power grid components of the same type (e.g., measurement units 118 near other power generation facilities) in its training, in addition to the data obtained by the measurement unit 118 itself. This allows a larger training dataset to be used to train the various machine learning-based systems, thereby improving the accuracy of the machine learning-based systems in classifying power grid conditions. To further increase the amount of training data used, feature data related to the location of(s) other measurement units 118 experiencing different power grid conditions can be used, if necessary. In one instance, unsupervised learning is used, for example, as a discriminator in a Generative Adversarial Network (GAN), to train an edge ML system. Unsupervised learning does not use the baseline real data described above. For example, in the case mentioned above regarding Figures 2a and 2b, where the classification data only specifies the presence or absence of anomalies, characteristic data representing the power flow during time intervals when no anomalies exist in the power supply could be used to train the edge ML system. In unsupervised learning, the generator of the GAN, which includes the edge ML system, generates artificial characteristic data. If the artificial characteristic data deviates more from the training characteristic data, a higher loss function is used to modify the parameters of the edge ML system. The discriminator of the GAN (the discriminator being the edge ML system) can be trained in parallel with the generator using GAN learning to distinguish between the baseline real training characteristic data and the artificial characteristic data generated by the generator. Since the training data represents the power flow during time intervals when no anomalies exist in the power supply, this training allows the discriminator to distinguish between normal power flow (represented by the training data) and power flow that is not typical of the normal state of power grid 100. In this way, the discriminator can be trained to determine whether the characteristic data indicates the absence of power anomalies or power anomalies. Unsupervised learning methods like this can produce machine learning-based systems that are robust to anomalous behavior in at least one characteristic of power and can identify anomalous behavior caused by previously unobserved or unknown anomalies. Another type of edge computing system that can be trained using unsupervised learning is the autoencoder. In this case, feature data representing non-anomalies is used to jointly train the encoder and decoder. The encoder learns how to efficiently represent the feature data, and the decoder attempts to reconstruct the feature data from the efficient representation. In this case, the loss function has a higher value if the difference between the original feature data and the reconstructed feature data is large. In this case, the difference can be defined, for example, as the cosine distance between the vector representing the original feature data and the vector representing the reconstructed feature data. Once trained, this autoencoder can be used to determine whether the target feature data represents anomalies by comparing the value of the loss function between the target feature data and the reconstructed target feature data with a threshold value. If the value of the loss function exceeds the threshold value, it means that the autoencoder cannot accurately reconstruct the feature data using the decoder, which in turn means that the feature data may represent a pattern (such as anomalies) not observed in the training data. In the above examples of unsupervised learning, the training data consists purely of power characteristic data representing time intervals without power anomalies, and the machine learning-based system is configured to determine whether the characteristic data indicates anomalies. However, in some instances, the training data also includes characteristic data representing abnormal behavior of the power grid 100. In this case, the machine learning-based system will be able to identify previously unrecorded anomaly types (anomaly types not present in the training data) through training using this training data. A consensus classification machine learning system can be trained, for example, by using a pre-trained edge ML system at each measurement unit 118 to determine training classification data for historical time intervals associated with known detected events. Then, the consensus classification machine learning system can be trained using the training classification data and consensus baseline real data indicating the state of the power grid 100. As described above regarding the supervised training procedure for training the edge ML system, consensus baseline real data can be obtained from event reports. In this way, the consensus classification machine learning system can be trained using classification data representing what measurement unit 118 will actually output during the execution of edge method 210. This allows the consensus classification machine learning system to be robust to failures in identifying the category of the actual state during the first time interval by one or more of the plurality of measurement units 118. For all of the above supervised training procedures, whether the machine learning-based system (of the measurement unit 118 or monitoring system 120) predicts the future behavior of the power grid 100 or classifies power grid events that have already occurred depends on how the training feature data matches the baseline real data over time, as will now be described. Figures 3a and 3b illustrate exemplary tables of training data. The training data includes training characteristic data and baseline real-world data. The training data represents the characteristics of the power supply (e.g., frequency) during each of the six time intervals. The baseline real-world data indicates the power grid status during each of these time intervals (in this case, "normal," "interconnector tripped," or "interconnector started"). In this example, the interconnector tripped during time interval 14:01 to 14:02, remained tripped during time interval 14:02 to 14:03, restarted during time interval 14:03 to 14:04, and otherwise the power grid status was normal. In an instance where the machine learning-based system of measurement unit 118 and monitoring system 120 aims to determine whether an event has occurred / what type of event has occurred during a measurement time interval, the training feature data for a given time interval can be matched with the baseline real data for that time interval, such that a single training dataset contains both the training feature data for the given time interval and the baseline real data for that time interval. Figure 3a illustrates the matching of baseline real data and training feature data in this case, where one training dataset is indicated by element symbol 306. Next, the loss function is evaluated by comparing the output of the machine learning-based system (inputting training feature data from a given time interval) with the baseline real data for that time interval. This can be repeated for hundreds, thousands, or millions of time intervals. In this way, the machine learning-based system will be trained to use feature data acquired during the time interval to determine the presence / type of events occurring in that time interval. However, in instances where machine learning-based systems are trained to predict future events, training feature data for a given time interval can be matched with baseline real data for a later time interval, such that a single training dataset contains training feature data for the measurement time interval and baseline real data for the later time interval. The later time interval can begin, for example, 1 millisecond, 1 second, 2 seconds, 3 seconds, 4 seconds, 5 seconds, 10 seconds, 20 seconds, or 30 seconds after the end of the measurement time interval. The later time interval can also begin at a fixed time interval after the end of the measurement time interval, which is the same for all training datasets. Figure 3b illustrates the matching of baseline real data and training feature data in this case, where one training dataset is indicated by element symbol 311. In this case, for a given training dataset, the baseline real data is offset by one minute from the training feature data. Next, the loss function is evaluated by comparing the output of the machine learning-based system (inputting feature data from a given time interval) with baseline real data from a later time interval. In this way, the machine learning-based system is trained to predict the presence / type of events occurring in later time intervals using feature data acquired during the time interval. For example, the training feature data M1 could indicate an early warning sign of an impending interconnect trip, such as the pattern occurring in frequency data. In this case, the loss function would reward the machine learning-based system for outputting classification data based on feature data M1, or a consensus classification of "interconnect trip". In some instances, the training procedures described above for edge ML systems and consensus classification systems based on machine learning are executed, for example, "offline" on a general-purpose computing device. Then, before use (e.g., in the case of measurement unit 118), the trained ML system(s) can be stored in the respective memory 504 of the measurement unit(s) 118 and / or monitoring system 120; this can be done before installation. In other instances, the relevant training procedures can be executed on the measurement unit(s) 118 and / or monitoring system 120. In other instances, the measurement unit 118 may be equipped with an untrained edge ML system (e.g., a neural network system with a set of preset weights), and the weights are subsequently transmitted to the measurement unit 118, for example, via a communication link from the monitoring system 120. Regardless of how the edge ML system is initially trained, it can subsequently be updated. This can take the form of supervised learning. For example, after classification data is provided to monitoring system 120, a given measurement unit 118 can receive feedback from monitoring system 120 indicating the results of the consensus classification procedure, according to a reinforcement learning method. This result can then serve as baseline real data, against which the edge ML system can update its internal parameters (e.g., neural network weights). For example, if a given measurement unit 118 provides classification data indicating a "normal" condition, and the consensus classification is "interconnector activated," this can trigger an adjustment of the internal parameters of measurement unit 118. In some cases, it may be appropriate for different measurement units 118 to provide different classification results, for example, where they are located at geographically dispersed locations within the power grid and / or, for other reasons, typically experience significantly different power grid conditions. Therefore, in some instances, measurement units 118 are triggered to update their edge ML systems depending on the fulfillment of certain conditions. For example, an update is only performed if the measurement unit 118 is located in a similar position within the power grid 100. Additionally or alternatively, unsupervised learning according to the techniques described above can be used to perform updates, wherein measurement unit 118 uses data recorded locally, which is accumulated as training data in the unsupervised learning process to update the parameters of the edge ML system. In some examples which we will now describe with reference to Figure 4, the training of the edge ML system is performed collaboratively by the measurement unit 118 and the monitoring system 120 in a joint learning method. In step 404, according to one of the methods described above, measurement unit 118 executes a local training procedure to perform initial training (or update its respective edge ML systems) on its respective edge ML systems. This determines the ML parameter set of each edge ML system. In step 406, each measurement unit 118 transmits its respective ML parameter set defining its own edge ML system to the monitoring system 120. In step 408, the monitoring system 120 summarizes the ML parameter set received from the measurement unit 118 to generate summarized ML parameters. Any suitable summarization method can be used, such as average summarization or momentum summarization. In step 402a, the monitoring system 120 transmits the aggregated ML parameters to each of the measurement units 118. Then, the measurement units 118 can implement the edge ML system based on the aggregated ML parameters to perform the edge method 210 described above. In step 404a, measurement unit 118 performs further local training to update the edge ML system. The further training procedure may be based at least in part on the feature data accumulated since the previous local training procedure in step 404 (and, where appropriate, additional baseline real data). Steps 406, 408, 402a, and 404a can be repeated repeatedly, for example, at regular intervals. Each repetition can be triggered, for example, by a timer or a message received by the self-monitoring system 120. By training the edge ML system with data used at the measurement unit itself and aggregating the training results using feature data obtained from other locations by other measurement units, the machine learning-based system can determine the existence or non-existence of both local anomalies (trained at the local end) and anomalies affecting a larger area or the entire power grid 100 (trained by aggregation). Figure 5a shows a schematic diagram of a measurement unit 118a according to an example. Only one measurement unit 118a is shown in Figure 5a, but other measurement units 118 may be similar or identical. The measurement unit 118a is configured to perform the edge method 210 described above with respect to Figures 2a and 2b and / or the training method 400 described above with respect to Figure 4. The measurement unit 118a includes a processor 502, memory 504, an input interface 506, an output interface 508, and a sensor 510. The processor 502 may include a CPU, a GPU, or an AI hardware accelerator. The memory 504 is a non-transitory computer-readable medium storing a computer program that, when executed by the processor 502, causes the processor 502 to perform the edge method 210 described above with respect to Figures 2a and 2b and / or the training method 400 described above with respect to Figure 4. In an instance where measurement unit 118 performs only edge method 210 without performing training method 400, and where measurement unit 118 uses an edge ML system (e.g., where its self-monitoring system 120 receives parameters from the edge ML system), measurement unit 118 uses TensorFlow Lite to determine the classification data. In an instance where measurement unit 118 performs training method 400, measurement unit 118 uses TensorFlow to perform training method 400 and / or edge method 210. Figure 5b shows a schematic diagram of the monitoring system 120 according to the example. The monitoring system 120 is configured to execute the central method 220 described above with respect to Figures 2a and 2c and / or the training method 400 described above with respect to Figure 4. The monitoring system 120 includes a processor 602, memory 604, an input interface 606, and an output interface 608. The processor 602 may include a CPU, GPU, or AI hardware accelerator. The memory 604 is a non-transitory computer-readable medium storing a computer program that, when executed by the processor 602, causes the processor 602 to execute the central method 220 described above with respect to Figures 2a and 2c and / or the training method 400 described above with respect to Figure 4. According to the present invention, the plurality of measurement units 118 described above with respect to FIG. 5a are configured in conjunction with the monitoring system 120 described above with respect to FIG. 5b to perform a consensus classification method 200 for determining the state of the power grid 100. For example, each of the plurality of measurement units 118 may be configured to use its respective sensors 510 to perform measurements of at least one characteristic (e.g., frequency) of the power at its respective location to obtain respective characteristic data. Each measurement unit 118 is configured to perform step 212 of edge method 210 on its respective characteristic data to obtain at least one piece of its own classification data. Each measurement unit 118 is configured to transmit the determined classification data to the monitoring system 120 described above with respect to FIG. 5b. Due to this transmission, the monitoring system 120 receives classification data from each of the plurality of measurement units 118. The monitoring system 120 determines a consensus classification of the state of the power grid 100 based on the received classification data. Embodiments of the present invention offer various advantages. For example, by sending classification data from measurement unit 118 to monitoring system 120 instead of sending characteristic data itself, a smaller amount of data can be sent while still enabling monitoring system 120 to determine a consensus classification of the state of power grid 100. This, in turn, reduces the latency of data transmission. In cases where it is desirable to determine a consensus classification of the state of power grid 100 over a time interval associated with characteristic data, time interval-related classification data can be received by monitoring system 120 from measurement unit 118 earlier than characteristic data due to the lower latency. In existing methods where characteristic data is transmitted instead, there is a trade-off between capturing a larger amount of characteristic data (leading to more reliable consensus classification by monitoring system 120) and capturing a smaller amount of characteristic data (reducing the latency of sending characteristic data to monitoring system 120 and thus allowing monitoring system 120 to determine consensus classification earlier). The present invention resolves this trade-off by determining classification data at the measurement unit 118 that is related to the presence or absence of one or more categories of conditions, which means that a relatively small amount of classification data can be provided in place of the original characteristic data. It should be understood that any feature described with respect to any embodiment may be used alone or in combination with other described features, and may also be used in combination with one or more features of any other embodiment or any combination of any other embodiment. Furthermore, equivalents and modifications not described above may be employed without departing from the scope of the invention as defined in the appended claims. This description includes the following numbered items: 1. A method for use in a power grid, in which electricity flows, wherein a plurality of measurement units are connected to the power grid, each of the measurement units being configured to perform measurement of at least one characteristic of the electricity to obtain characteristic data, the method comprising: performing a classification procedure at each measurement unit to determine at least one piece of respective classification data based on the characteristic data, each piece of respective classification data being related to the state of the power grid and indicating one or more respective categories of the state of the power grid, each respective category including one of a predetermined plurality of categories of the state of the electricity flowing in the power grid; and transmitting the determined classification data from each measurement unit to a monitoring system. 2. The method of item 1, wherein the classification procedure comprises: performing a machine learning-based procedure based on the characteristic data using a machine learning-based system of measurement devices held at the measurement unit; and determining the classification data based on the output of the machine learning-based system of the measurement devices. 3. The method of clause 2, comprising receiving parameters of the measurement device based on a machine learning system from the monitoring system by each of the measurement units, and implementing the measurement device based on the received parameters. 4. The method of clause 2 or 3, wherein the classification procedure comprises preprocessing the feature data to obtain preprocessed data, the preprocessing comprising applying a transform to the feature data, the transform comprising at least one of wavelet transform and fast Fourier transform, and the classification procedure comprising inputting the preprocessed data into the measurement device based on a machine learning system. 5. A method for use in a power grid, in which electricity flows, wherein a plurality of measurement units are connected to the power grid, each of the measurement units being configured to perform measurement of at least one characteristic of the electricity to obtain characteristic data, each of the measurement units being configured to transmit at least one piece of respective classification data to a monitoring system, each piece of respective classification data relating to the state of the power grid and indicating one or more respective categories of the state of the power grid, each respective category including one of a predetermined plurality of categories of the state of the electricity flowing in the power grid, the method comprising: receiving the classification data from each of the plurality of measurement units at the monitoring system; and determining a consensus classification of the state of the power grid at the monitoring system based on the received classification data. 6. The method of claim 5, wherein determining the consensus classification comprises: selecting the state classified as present by the highest number of the measurement units. 7. The method of claim 5 or claim 6, wherein determining the consensus classification comprises: inputting the received classification data into a consensus classification-based machine learning system.8. The method of any of clauses 5 to 7, wherein each of the measurement units maintains a measurement device based on a machine learning system and is configured to: perform a classification procedure to determine classification data based on the characteristic data, the classification procedure including inputting the characteristic data to the measurement device based on the machine learning system; and determine the classification data based on the output from the measurement device based on the machine learning system. 9. The method of any of clauses 8 or 2 to 4, comprising performing a training procedure to train the measurement device based on the machine learning system using training data, the training data including characteristic data indicating the at least one characteristic and baseline real data indicating one or more of a predetermined plurality of categories. 10. The method of clause 9, wherein the measurement device based on the machine learning system of each measurement unit is a different machine learning system, and the training procedure for training each of the respective measurement device based on the characteristic data indicating the at least one characteristic of the power at the respective location of the measurement unit. 11. The method of claim 9 or 10, comprising: at the first measurement unit, performing a training procedure based on characteristic data obtained at the first measurement unit to determine an update of the machine learning-based system of the measurement device; transmitting data indicating the updates; receiving the updated machine learning-based system of the measurement device at the plurality of measurement units, wherein the updated machine learning-based system of the measurement device has been generated based on the data indicating the updates, and the classification procedure comprising inputting the characteristic data into the updated machine learning-based system of the measurement device. 12. The method of any one of claims 1 to 11, wherein the plurality of categories of conditions includes a plurality of categories of anomalies in the power flow. 13. The method of claim 12, wherein the plurality of categories of anomalies includes at least one of: power generation facility tripping, interconnector tripping, interconnector starting, transmission line tripping, demand load tripping, transformer tripping, redistribution of the power flow, converter tripping, and voltage harmonic amplitude meeting criteria. 14. The method of any of clauses 1 to 13, wherein the at least one characteristic includes frequency, the derivative of frequency with respect to time, inertia, voltage, active power, reactive power, and / or current. 15. The method of any of clauses 1 to 14, wherein the characteristic data has a sampling rate exceeding 120 Hz. 16. The method of any of clauses 1 to 15, wherein the classification data indicates the probability of the presence or absence of one of the predetermined plurality of categories. 17. The method of any of clauses 1 to 16, wherein the characteristic data is obtained during a measurement time interval, and the condition of the power grid is related to a time interval later than the measurement time interval. 18. A plurality of measurement units configured to perform any of the methods of clauses 1 to 4 or any of clauses 9 to 17 attached to clause 1.19. A monitoring system configured to perform a method as described in any of clauses 5 to 8 or as described in any of clauses 9 to 17 attached to clause 5. 20. A system comprising: a plurality of measurement units as described in clause 18; and a monitoring system as described in clause 19. 21. A computer program, when executed by a processor of the monitoring system, causes the monitoring system to perform a method as described in clause 5 or as described in any of clauses 6 to 17 attached to clause 5. 22. A plurality of computer programs, wherein when a respective processor of a plurality of measurement units executes one of the computer programs, the plurality of measurement units perform a method as described in any of clauses 1 to 4 or as described in any of clauses 9 to 17 attached to clause 1. 100: Power grid; 102: Transmission network; 104: Distribution network; 106: Power generation facilities; 108: Transformer; 110: Substation; 112: Urban network; 113: Private residence; 114: Factory; 116: Wind farm; 117: Photovoltaic equipment; 118: Measurement unit; 118a: Measurement unit / First measurement unit; 118b: Measurement unit / Second measurement unit; 118c: Measurement unit; 119: DC interconnect; 120: Monitoring system; 122: Database; 200: Square Method 210: Edge Method 212: Step 214: Step 220: Center Method 222: Step 224: Step 306: Training Data Set 311: Training Data Set 400: Training Method 402a: Step 404: Step 404a: Step 406: Step 408: Step 502: Processor 504: Memory 506: Input Interface 508: Output Interface 510: Sensor 602: Processor 604: Memory 606: Input Interface 608: Output Interface Figure 1 is a schematic diagram of the power grid according to the example; Figure 2a is a flowchart of the method performed by multiple measurement units and a monitoring system according to the example; Figure 2b is a flowchart of the edge method performed by the measurement units according to the example; Figure 2c is a flowchart of the center method performed by the monitoring system according to the example; Figure 2d is a graph of the time-varying frequency of electricity according to the example; Figure 2e is a schematic diagram of the method performed by multiple measurement units and a monitoring system according to the example; Figure 3a is a table of training data for the classification procedure according to the example; Figure 3b is a table of training data for the prediction procedure according to the example; Figure 4 is a flowchart of the training method according to the example; Figure 5a is a schematic diagram of the measurement units according to the example; and Figure 5b is a schematic diagram of the monitoring system according to the example. 100: Power Grid 118a: Measurement Unit / First Measurement Unit 118b: Measurement Unit / Second Measurement Unit 122: Database 200: Method 212: Steps 224: Steps

Claims

1. A method for use in a power grid, wherein electricity flows in the power grid, wherein a plurality of measurement units are connected to the power grid, each of the measurement units being configured to perform measurement of at least one characteristic of the electricity to obtain characteristic data, the method comprising: A classification procedure is performed at each measurement unit to determine at least one piece of individual classification data based on the characteristic data, each piece of individual classification data being related to the condition of the power grid and indicating one or more individual categories of the condition of the power grid, each individual category including one of a predetermined plurality of categories of the condition of the electricity flowing in the power grid; and the determined classification data is transmitted from each measurement unit to a monitoring system, wherein the classification procedure includes: performing a machine learning-based procedure based on the characteristic data using a machine learning-based system of the respective measurement devices held at the measurement unit; and determining the classification data based on the output from the machine learning-based system of the respective measurement devices, wherein each of the respective machine learning-based system of the respective measurement devices has its own different set of machine learning parameters.

2. The method of claim 1, comprising receiving, by each of the measurement units, the parameters of the respective measurement device's machine learning-based system from the monitoring system, and implementing the respective measurement device's machine learning-based system based on the received parameters.

3. The method of claim 1, wherein the classification procedure includes preprocessing the feature data to obtain preprocessed data, the preprocessing including applying a transform to the feature data, the transform including at least one of wavelet transform and fast Fourier transform, and the classification procedure includes inputting the preprocessed data into a machine learning-based system of the respective measurement device.

4. The method of claim 1, comprising performing a training procedure to train the respective measurement device into a machine learning-based system using training data, the training data including training feature data indicating the at least one feature.

5. The method of claim 4, wherein the training data includes baseline real data of one or more of the predetermined plurality of categories of the indicated conditions.

6. The method of claim 4, wherein the training procedure for training each of the respective measurement devices in a machine learning-based system is based on training characteristic data indicating at least one characteristic of the power at the respective location of the measurement unit.

7. The method of any of claims 1 to 6, comprising: At the first measurement unit, a training procedure is performed based on feature data obtained at the first measurement unit to determine updates to the machine learning-based system of the respective measurement device; data indicating the updates is transmitted; the updated machine learning-based system of the measurement device is received at the plurality of measurement units, wherein the updated machine learning-based system of the measurement device has been generated based on the data indicating the updates, and the classification procedure includes inputting the feature data into the updated machine learning-based system of the measurement device.

8. The method of any of the requests 1 to 6, wherein the plurality of categories of conditions includes a plurality of categories of anomalies in the power flow.

9. The method of claim 8, wherein the plurality of categories of anomalies includes at least one of the following: power generation facility tripping, interconnector tripping, interconnector starting, transmission line tripping, demand load tripping, transformer tripping, rewiring of the power flow, converter tripping, and amplitude of voltage harmonics that meet the criteria.

10. The method of any one of claims 1 to 6, wherein the plurality of categories of conditions includes a category indicating normal power flow and a category indicating abnormal power flow.

11. The method of claim 4, wherein the plurality of categories of conditions includes categories indicating normal power flow and categories indicating abnormal power flow, and wherein the training characteristic data includes data representing normal power flow.

12. The method of claim 11, wherein the training characteristic data substantially consists of the data representing normal power flow.

13. The method of any one of claims 1 to 6, wherein the at least one characteristic includes frequency, the derivative of frequency with respect to time, inertia, voltage, effective power, reactive power and / or current.

14. The method of any one of requests 1 to 6, wherein the characteristic data has a sampling rate of more than 120 Hz.

15. The method of any one of claims 1 to 6, wherein the classification data indicates the probability of the presence or absence of one of the predetermined plurality of categories.

16. The method of any one of claims 1 to 6, wherein the characteristic data is obtained during the measurement time interval, and the condition of the power grid is related to a time interval later than the measurement time interval.

17. The method of any of claims 1 to 6, which includes a consensus classification at the monitoring system based on the classification data to determine the state of the power grid.

18. The method of claim 17, wherein determining the consensus classification includes selecting the highest number of such measurement units as the state of existence.

19. The method of claim 17, wherein determining the consensus classification includes inputting the received classification data into a consensus classification system based on machine learning.

20. A plurality of measurement units configured to perform any one of the methods requested in items 1 to 19.

21. A system for determining the condition of a power grid, comprising: Multiple measurement units; and a monitoring system configured to perform the methods described in any of requests 1 to 19.

22. A plurality of computer programs, wherein when a processor of a plurality of measurement units executes one of the computer programs, the plurality of measurement units perform the method of any one of claims 1 to 19.

Images