COMPUTER-IMPLEMENTED METHOD FOR POWER LINE PROTECTION, INTELLIGENT ELECTRONIC DEVICE AND POWER SYSTEM - Patent application

Abstract

A technique for distance protection of a transmission line (11) includes determining a fault onset time from voltage and / or current, determining rate of change sample values ​​indicative of a rate of change of voltage and / or current with respect to at least one sample time that is dependent on the fault onset time, and using the rate of change sample values ​​to generate a phase classifier for fault classification or a zone classifier for faulted zone identification.

Description

[Technical field]

[0001] FIELD OF THEINVENTION The present invention relates to power transmission systems and devices and methods for protecting power transmission lines, and in particular to methods and devices operative to perform at least one of fault classification, zone classification, and / or distance protection. [Background technology]

[0002] 2. Background of the Invention The power grid is undergoing a significant change in its generation mix, from synchronous AC rotating machines to inverter-based resource (IBR) technologies. As a result of this trend, renewable energy is deployed more in both transmission and distribution grids. In addition to environmental benefits, the introduction of renewable energy sources changes the operation of the power system, e.g., reduced inertia, reduced stability margins, and increased unpredictability. Since most of these renewable technologies are interfaced with inverters, their behavior under fault conditions differs from traditional rotating machines. This creates new challenges for power system protection. One issue with the high penetration of inverter-based resources, such as wind and photovoltaic (PV) generation, is the reduction in fault current levels and short-circuit strength of the power grid.

[0003] Furthermore, the power system stability margin depends on the inertia of the system, which determines the desired speed of the protection scheme. IBR systems often have less inertia and therefore less stability margin. A reduced stability margin means a reduced critical clearing time. Faults must be cleared faster than the critical clearing time or the system may lose transient stability, which leads to power system outages.

[0004] Distance relays are protective devices designed to provide primary and backup protection for transmission lines. Traditional distance relays do not auto-adjust as systems change. Routine manual adjustment and automated adaptability can be provided to some degree in protective relays.

[0005] WO 2005 / 076428 A9 discloses a system and method for detecting high impedance faults (HIF) in power lines using an artificial neural network (ANN). The method is wavelet transform based and therefore may be computationally intensive for practical applications. A detection rate of 70.83% with a false alarm rate of 22.06% may not be a satisfactory performance for protection applications. The use of a spectrum of 3 cycle windows of data may not be desirable for protection applications.

[0006] US Patent No. 7,720,619 B2 discloses a method for detecting high impedance faults. Artificial intelligence (AI) methods of classification and pattern recognition, such as neural networks, expert systems or decision trees, can be used to distinguish HIF conditions from other system conditions, such as switching activity and noisy loads.

[0007] US6405184B1 discloses a method for generating a fault classification signal to identify a faulty loop in case of a fault. A neural network is used that is trained with input variables that simulate a faulty loop. Measurements are derived from the current and voltage of the loop of the energy supply system, and in case of a fault, a fault classification signal is generated by the trained model.

[0008] WO 1995 / 009465 A1 discloses a method for generating a directional signal indicative of the direction of a short circuit current. The directional signal indicates whether a short circuit has occurred in the forward direction from a measurement point. A neural network is used to generate the directional signal, which is formed using normalized samples of the current.

[0009] An artificial neural network (ANN) based approach for three-zone distance protection of transmission lines is presented in A. Feilat and K. Al-Tallaq, "A new approach for distance protection using artificial neural network", 39th International Universities Power Engineering Conference, 2004.UPEC2004, Bristol, UK, 2004, pp. 473-477, vol. 1. This technique deals only with the fault detection and classification aspects of distance protection. The input features of the neural network are fundamental frequency voltage and current magnitudes extracted by discrete Fourier transform. Hence, this solution is not expected to improve the speed of existing phasor-based distance protection algorithms.

[0010] JR de Carvalho et al., "Development of detection and classification stages for a new distance protection approach based on cumulants and neural networks," 2011 IEEE Power and Energy Society General Meeting, Detroit, MI, USA, 2011, pp. 1-7, proposes the use of artificial intelligence (AI) for distance relaying. Fault detection and classification capabilities can be implemented.

[0011] "Transmission line distance relaying using machine intelligence technique" by SR Samantaray et al., IET Generation, Transmission & Distribution, Vol. 2, No. 1, pp. 53-61, January 2008, describes a technique for distance relaying of transmission lines using machine intelligence techniques such as Support Vector Machine (SVM). Post-fault current and voltage samples are used as input to the SVM. The total time taken for fault phase selection and ground detection is 10 ms (half cycle) from fault initiation.

[0012] Neural network based impedance protection for power transmission lines to eliminate the undesirable effect of fault resistance on conventional distance relay operation is presented in P. Vaidya and PAVenikar, "ANN based distance protection of long transmission lines by considering the effect of fault resistance", IEEE-International Conference On Advances In Engineering, Science And Management (ICAESM-2012), Nagapattinam, Tamil Nadu, 2012, pp. 590-594, 2012. The algorithm uses the magnitude of resistance and reactance as input. However, the pre-processing done to reduce the neural network size involves DFT to calculate the fundamental frequency components. Hence, the speed of the phasor based distance protection algorithm is not significantly improved.

[0013] F. Rafique et al., "An accurate digital distance protection scheme using neural network in UHV transmission line," Proceedings of 2014 IEEE Chinese Guidance, Navigation and Control Conference, Yantai, pp. 1949-1954, 2014, discloses an impedance relay system combined with an ANN block that uses the zero-sequence component of the fault current. The fast Fourier transform (FFT) to extract the fundamental component introduces limitations in terms of speed. Also, the use of the zero-sequence component introduces further limitations when the fault is a symmetrical fault or does not involve ground, and no zero-sequence quantity exists.

[0014] M. Oleskovicz et al., "A complete scheme for fault detection, classification and location in transmission lines using neural networks," Proceedings of the IEEE 7th International Conference on Developments in Power System Protection (DSP'01), pp. 335-338, 2001, presents an ANN approach for a complete scheme of distance protection. The inputs of the presented solution are three-phase voltage and current samples in pre-fault and fault conditions. However, the average response time is 13 ms with an accuracy of 98%.

[0015] There is a continuing need for improved techniques for transmission line protection. Summary of the Invention [Means for solving the problem]

[0016] overview The addition of inverter-based resources (IBRs) to the power grid poses challenges for transmission line protection. Fault currents generated by IBRs are often low in magnitude and often do not contain sufficient levels or negative quantities to apply traditional protection schemes. The lack of these negative quantities can result in incorrect phase and zone classification of line-distance intelligent electronic devices (IEDs). Lower fault current magnitudes can affect distance relay performance, causing malfunction or non-operation of distance protection elements. Techniques that mitigate these challenges through human expert-based relay settings may not be able to completely eliminate protection issues.

[0017] There is a need for improved techniques for power system protection. There is a need for improved techniques adapted for power line protection. There is also a need for methods, devices, and systems that can provide protection functions (e.g., autonomous distance protection methods) that can efficiently protect systems having IBRs. In particular, there is a need for such methods, devices, and systems that can generate phase classifications, zone classifications, and / or protection functions using a data-driven approach. There is also a need for such methods, devices, and systems that can determine phase classifications, zone classifications, and / or protection functions in a self-configuring manner. In particular, there is a need for such methods, devices, and systems that provide fast and safe fault and zone classifications for power systems having IBRs.

[0018] Methods, devices, and systems according to embodiments may operate to provide artificial intelligence (AI) / machine learning (ML) based implementation of functions for power system protection and / or monitoring. Methods, devices, and systems according to embodiments may operate to provide AI / ML based phase identification (for fault classification) and zone identification for transmission line systems. The methods, devices, and systems are particularly suited for power systems with high IBR penetration.

[0019] Methods, devices, and systems according to embodiments may be operable to perform fault onset time detection using analytical techniques: The methods, devices, and systems may be operable to determine a sliding-window moving average and a sliding-window standard deviation of at least one power system characteristic (such as current and / or voltage) and compare the absolute value of the deviation from the sliding-window moving average to a threshold that depends on the sliding-window standard deviation, thereby determining the fault onset time.

[0020] The sliding window moving average may be an average of at least one power system characteristic (eg, voltage and / or current or several voltages and / or currents) over a sliding window.

[0021] The sliding window standard deviation may be the standard deviation of at least one power system characteristic (eg, voltage and / or current or several voltages and / or currents) over a sliding window.

[0022] Methods, devices, and systems according to embodiments may operate to perform smart training window selection to improve accuracy. Methods, devices, and systems may operate to determine a rate of change of current and / or voltage based on the number of samples per cycle used for a moving average filter.

[0023] Methods, devices, and systems according to embodiments provide system-agnostic feature selection for training and building artificial intelligence (AI) / machine learning (ML) models.

[0024] Methods, devices, and systems according to embodiments may be to generate and / or use interpretable code from trained AI / ML models for implementation in IEDs.

[0025] Methods, devices, and systems according to embodiments may operate to continuously update phase classification, zone classification, and / or protection functions during field operation of a transmission system.

[0026] The method, device, and system according to the embodiments can operate to provide power line protection in power systems with high penetration of inverter-based (renewable energy) resources. The machine learning model obtained by the method is independent of system parameters. It can be applied to any system without the need for additional training.

[0027] In the method, device, and system according to the embodiments, the features used as inputs for the AI / ML model training and / or used as inputs for the trained classifier or protection function may be specifically selected. That is, features indicating the rate of change of voltage for the three phases and the rate of change of current for the three phases of the transmission line may be used as inputs. This may be applied both during training, where random forest (RF) training or other AI / ML model training may be performed, and during the use of the created decision logic.

[0028] The input layer of the AI / ML model may receive rate of change values ​​indicative of the rate of change of one or several power system characteristics (such as the rate of change of voltage and / or current).

[0029] The output layer of the AI / ML model may provide a signal indicative of a fault and / or a type of fault and / or a zone of fault.

[0030] In accordance with embodiments, a method, device, and system may generate a first random forest having a first plurality of decision trees to operate as a phase classifier that receives as input the rate of change values ​​of voltage in three phases and the rate of change values ​​of current in three phases, performs fault classification, and outputs information regarding the fault type. In response to receiving as input the rate of change values ​​of voltage in three phases and the rate of change values ​​of current in three phases, the phase classifier may operate to classify the fault into one of types ABC, Ag, BC, CA-g.

[0031] In methods, devices, and systems according to embodiments, a second random forest having a second plurality of decision trees may be generated to operate as a zone classifier that receives as input the rate of change values ​​of the voltage in the three phases and the rate of change values ​​of the current in the three phases, performs zone classification, and outputs information regarding the zone type. In response to receiving as input the rate of change values ​​of the voltage in the three phases and the rate of change values ​​of the current in the three phases, the zone classifier may operate to classify the fault into one of a plurality of distinct zones (Zone 1, Zone 2, etc.).

[0032] A computer-implemented method according to one embodiment includes calculating a feature for use in generating decision logic operative to generate a signal operative for use with an electric power system, where calculating the feature includes determining a fault onset time from at least one electrical characteristic of sampled data for a given type of electric power system, and determining a rate of change value indicative of a rate of change of the at least one electrical characteristic with respect to at least one sample time, where the at least one sample time is dependent on the fault onset time; and using the feature to generate the decision logic.

[0033] The method may further include deploying decision logic in an intelligent electronic device (IED) to generate the signal.

[0034] The method may further include executing, by an intelligent electronic device (IED), a decision logic to generate the signal.

[0035] The method may be a method for power system protection and / or monitoring. The decision logic may be a protection logic.

[0036] The decision logic may be or may implement a distance protection function.

[0037] A given type of power system may be a transmission line. The decision logic may be deployed to protect and / or monitor multiple power systems of a given type.

[0038] Power systems of a given type may be transmission lines. The multiple transmission lines may include transmission lines whose installation locations are distinct from one another.

[0039] The multiple transmission lines may include transmission lines having distinct line lengths. The plurality of transmission lines may include transmission lines having distinct conductor diameters.

[0040] The method may be a method for transmission line protection. The decision logic may be or may comprise at least one of a protection function, a phase classifier for fault classification, and a zone classifier for faulty zone identification.

[0041] The signal may be or may include at least one of a circuit breaker control signal, a switch control signal, an alarm, a warning, status information, an output for output via a human machine interface (HMI).

[0042] Determining the fault onset time may include determining a sliding-window moving average and a sliding-window standard deviation of the at least one electrical characteristic and comparing an absolute value of the deviation of the at least one electrical characteristic from the sliding-window moving average to at least one threshold value.

[0043] At least one threshold may be dependent on a sliding window standard deviation. The fault onset time can be identified as the earlier of the time when the absolute value of the deviation of the voltage from a sliding-window moving average of the voltage reaches or exceeds a phase-specific voltage threshold and the time when the absolute value of the deviation of the current from a sliding-window moving average of the current reaches or exceeds a phase-specific current threshold.

[0044] Determining the rate of change value may include, for each phase, determining a filtered voltage, determining a filtered current, and calculating the rate of change value from the filtered voltage and the filtered current.

[0045] The filtered voltage may be determined by averaging N>1 sample values ​​of the voltage, and the filtered current is determined by averaging N>1 sample values ​​of the current.

[0046] Calculating the rate of change value may include determining a voltage difference between the filtered voltage sample value at time k and the filtered voltage sample value at time kN, and a current difference between the filtered current sample value at time k and the filtered current sample value at time kN.

[0047] The rate of change value of the voltage may be determined as the change in the voltage difference between a sample time and a previous sample time.

[0048] The rate of change value of the current may be determined as the change in the current difference between a sample time and a previous sample time.

[0049] Deploying the decision logic may include deploying a phase classifier and / or a zone classifier for execution by the IED for distance protection.

[0050] Generating the decision logic may include using the features to generate one or both of a phase classifier and a zone classifier.

[0051] Using the features to generate decision logic may include using ensemble machine learning (ML) methods to generate a phase classifier and to generate a zone classifier.

[0052] Generating the phase classifier and generating the zone classifier can include random forest training using the features to generate a random forest.

[0053] The features of the nodes in the input layer of the random forest decision tree can be selected from the rate of change values.

[0054] The features of the nodes in the input layer of the random forest decision tree can be selected from at least m × n rate of change values ​​indicating the rate of change of voltage and m × n rate of change values ​​indicating the rate of change of current, where m ≧ 2 indicates the number of phases and n ≧ 1 indicates the number of sample times.

[0055] The features of the nodes in the input layer of the random forest decision tree can be selected from at least m × n rate of change values ​​indicating the rate of change of voltage and m × n rate of change values ​​indicating the rate of change of current, where m ≥ 2 indicates the number of phases and n ≥ 2 indicates the number of sample times depending on the fault onset time.

[0056] The features of the nodes in the input layer of the random forest decision tree can be selected from at least m × n rate of change values ​​indicating the rate of change of voltage and m × n rate of change values ​​indicating the rate of change of current, where m = 3 indicates the number of phases and n ≥ 2 indicates the number of sample times depending on the fault onset time.

[0057] The features of the nodes in the input layer of the random forest decision tree can be selected from sample times taken from up to a quarter cycle after the fault onset time.

[0058] The features of the nodes in the input layer of the random forest decision tree can be selected from sample times that are not obtained from the time one-quarter cycle after the fault onset time.

[0059] Using the rate of change values ​​to generate one or both of the phase classifier and the zone classifier can include training a first machine learning model with the features to generate the phase classifier.

[0060] Using the rate of change values ​​to generate one or both of the phase classifier and the zone classifier can include training a second machine learning model with the features to generate the zone classifier.

[0061] The phase classifier can receive as input the rate of change of voltage for three phases and the rate of change of current for three phases and output a fault type.

[0062] The zone classifier can receive as input the rate of change of voltage for three phases and the rate of change of current for three phases and output a zone.

[0063] The phase classifiers and / or zone classifiers may be generated in a self-configuring manner. The decision logic may be a decision logic for transmission line protection.

[0064] The transmission line protection may include high impedance fault protection. Generating the phase classifier and / or zone classifier can be performed using a data set that includes data for a number of distinct source-line impedances.

[0065] A computing system according to one embodiment has an interface or memory operative to receive or store training data indicative of power system characteristics (such as voltage and / or current of a faulted line). The computing system has at least one integrated circuit operative to calculate features for use in generating decision logic for power system protection and / or to calculate features for use in generating decision logic operative to generate a signal operative for use with the power system, where calculating the features includes determining a fault onset time from at least one electrical characteristic of sampled data for a given type of power system, and determining a rate of change value indicative of a rate of change of the at least one electrical characteristic relative to at least one sample time, where the at least one sample time is dependent on the fault onset time. The at least one integrated circuit is operative to use the features to generate the decision logic.

[0066] The computing system may have an output interface for deploying the decision logic to an intelligent electronic device (IED) to generate a signal.

[0067] The computing system may be equipped with an intelligent electronic device (IED) that operates to execute decision logic to generate the signal.

[0068] The computing system may be a computing system for power system protection and / or monitoring.

[0069] The decision logic may be a protection logic. The decision logic may be or may implement a distance protection function.

[0070] A given type of power system may be a transmission line. The computing system is operable to deploy decision logic for protecting and / or monitoring multiple power systems of a given type.

[0071] Power systems of a given type may be transmission lines. The multiple transmission lines may include transmission lines whose installation locations are distinct from one another.

[0072] The multiple transmission lines may include transmission lines having distinct line lengths. The plurality of transmission lines may include transmission lines having distinct conductor diameters.

[0073] The computing system is operable to generate decision logic for transmission line protection.

[0074] The computing system may be operable to generate decision logic that may be or may comprise at least one of a protection function, a phase classifier for fault classification, and a zone classifier for faulty zone identification.

[0075] The signal may be or may include at least one of a circuit breaker control signal, a switch control signal, an alarm, a warning, status information, an output for output via a human machine interface (HMI).

[0076] The computing system can be operable such that determining the fault onset time can include determining a sliding-window moving average and a sliding-window standard deviation of the at least one electrical characteristic, and comparing an absolute value of the deviation of the at least one electrical characteristic from the sliding-window moving average to at least one threshold value.

[0077] The computing system may be operable such that at least one threshold may be dependent on a sliding window standard deviation.

[0078] The computing system can be operated such that the fault onset time can be identified as the earlier of the time when the absolute value of the deviation of voltage from a sliding-window moving average of voltage reaches or exceeds a phase-specific voltage threshold and the time when the absolute value of the deviation of current from a sliding-window moving average of current reaches or exceeds a phase-specific current threshold.

[0079] The computing system may be operable such that determining the rate of change value may include, for each phase, determining a filtered voltage, determining a filtered current, and calculating the rate of change value from the filtered voltage and the filtered current.

[0080] The computing system may be operable such that the filtered voltage is determined by averaging N>1 sample values ​​of the voltage, and the filtered current is determined by averaging N>1 sample values ​​of the current.

[0081] The computing system may be operable such that calculating the rate of change value may include determining a voltage difference between the filtered voltage sample value at time k and the filtered voltage sample value at time kN, and a current difference between the filtered current sample value at time k and the filtered current sample value at time kN.

[0082] The computing system may be operable such that a rate of change value of the voltage may be determined as the change in voltage difference between a sample time and a previous sample time.

[0083] The computing system may be operable such that a rate of change value of the current may be determined as the change in the current difference between a sample time and a previous sample time.

[0084] The computing system is operable to deploy decision logic for execution by an intelligent electronic device (IED).

[0085] The computing system can be operable such that deploying the decision logic can include deploying a phase classifier and / or a zone classifier for execution by the IED for distance protection.

[0086] The computing system may be operable such that generating the decision logic may include using the features to generate one or both of a phase classifier and a zone classifier.

[0087] The computing system can be operable such that using the features to generate the decision logic can include using an ensemble machine learning (ML) method to generate a phase classifier and to generate a zone classifier.

[0088] The computing system can be operable such that generating the phase classifier and generating the zone classifier can include random forest training using the features to generate a random forest.

[0089] The computing system may be operable such that features of an input node of a random forest decision tree may be selected from rate of change values.

[0090] The computing system may be operable such that the features of an input node of the random forest decision tree may be selected from at least m×n rate of change values ​​indicative of a rate of change of voltage and m×n rate of change values ​​indicative of a rate of change of current, where m≧2 indicates a number of phases and n≧1 indicates a number of sample times.

[0091] The computing system may be operable such that the features of an input node of the random forest decision tree may be selected from at least m×n rate of change values ​​indicative of a rate of change of voltage and m×n rate of change values ​​indicative of a rate of change of current, where m≧2 indicates a number of phases and n≧2 indicates a number of sample times that is dependent on the fault onset time.

[0092] The computing system may be operable such that the features of an input node of the random forest decision tree may be selected from at least m×n rate of change values ​​indicative of a rate of change of voltage and m×n rate of change values ​​indicative of a rate of change of current, where m=3 indicates a number of phases and n≧2 indicates a number of sample times that is dependent on the fault onset time.

[0093] The computing system can be operable such that features for an input node of the random forest decision tree can be selected from sample times taken from up to one-quarter cycle after the fault onset time.

[0094] The computing system can be operable such that features for an input node of the random forest decision tree can be selected from sample times not taken from times one-quarter cycle after the fault onset time.

[0095] The computing system can be operable such that using the rate of change values ​​to generate one or both of the phase classifier and the zone classifier can include training a first machine learning model with the features to generate the phase classifier.

[0096] The computing system can be operable such that using the rate of change values ​​to generate one or both of the phase classifier and the zone classifier can include training a second machine learning model using the features to generate the zone classifier.

[0097] The computing system may be operable to enable the phase classifier to receive as input the rate of change of voltage for three phases and the rate of change of current for three phases and output a fault type.

[0098] The computing system may be operable to enable the zone classifier to receive as input the rate of change of voltage for three phases and the rate of change of current for three phases, and output a zone.

[0099] The computing system may be operable to generate the phase classifiers and / or zone classifiers in a self-configuring manner.

[0100] The computing system may be operable such that the decision logic may be decision logic for transmission line protection.

[0101] The computing system may be operable such that the transmission line protection may include high impedance fault protection.

[0102] The computing system may be operable such that generating the phase classifier and / or the zone classifier may be performed using a data set that includes data for a number of distinct source-line impedances.

[0103] According to an embodiment that can be used in combination with or independently of the above-mentioned embodiments, a method for power system protection, in particular transmission line protection, includes determining a fault onset time from a voltage and / or a current. The method includes determining a rate of change value indicative of a rate of change of the voltage and / or a rate of change of the current with respect to at least one sample time, the at least one sample time being dependent on the fault onset time. The method includes using the rate of change value to generate a protection function (such as a protection logic), a phase classifier for fault classification, and / or a zone classifier for faulty zone identification.

[0104] The method may be a computer-implemented method. The method may be performed automatically by at least one integrated circuit.

[0105] The at least one integrated circuit may be at least one integrated circuit of a computing device. The at least one integrated circuit may be at least one integrated circuit of an intelligent electronic device (IED).

[0106] The method may further include deploying a phase classifier and / or a zone classifier for execution by the intelligent electronic device IED for distance protection.

[0107] The method may further include generating an intelligent electronic device (IED) decision logic based on the phase classifier and the zone classifier, and deploying the IED decision logic for execution.

[0108] The IED decision logic may be a decision tree. The IED decision logic may be an interpretable IED decision logic.

[0109] The IED may be a distance relay. Generating the phase classifier and generating the zone classifier may include ensemble machine learning methods.

[0110] The method can include deriving features for training a machine learning model from training data including voltage and current.

[0111] Deriving the features may include determining a rate of change of voltages of multiple phases of the transmission line and a rate of change of currents of multiple phases of the transmission line at one or more times that are dependent on the fault onset time.

[0112] Generating the phase classifier and generating the zone classifier may include random forest training to generate a random forest.

[0113] The features of the nodes in the input layer of the random forest decision tree can be selected from the rate of change values.

[0114] The features of the nodes in the input layer of the random forest decision tree can be selected from at least m × n rate of change values ​​indicating the rate of change of voltage and m × n rate of change values ​​indicating the rate of change of current, where m ≧ 2 indicates the number of phases and n ≧ 1 indicates the number of sample times.

[0115] The features of the nodes in the input layer of the random forest decision tree can be selected from at least m × n rate of change values ​​indicating the rate of change of voltage and m × n rate of change values ​​indicating the rate of change of current, where m ≥ 2 indicates the number of phases and n ≥ 2 indicates the number of sample times depending on the fault onset time.

[0116] The features of the nodes in the input layer of the random forest decision tree can be selected from at least m × n rate of change values ​​indicating the rate of change of voltage and m × n rate of change values ​​indicating the rate of change of current, where m = 3 indicates the number of phases and n ≥ 2 indicates the number of sample times depending on the fault onset time.

[0117] The features of the nodes in the input layer of the random forest decision tree can be selected from sample times taken from up to a quarter cycle after the fault onset time.

[0118] The features of the nodes in the input layer of the random forest decision tree can be selected from sample times that are not obtained from the time one-quarter cycle after the fault onset time.

[0119] Random forest training may be performed using a data set of L>1 faulty lines, each containing voltages and currents for three phases.

[0120] Using the rate of change values ​​to generate the phase classifier and / or the zone classifier may include training a first machine learning model to generate the phase classifier and training a second machine learning model to generate the zone classifier.

[0121] The first machine learning model can comprise a first random forest having a first plurality of decision trees, and the second machine learning model comprises a second random forest having a second plurality of decision trees different from the first plurality of decision trees.

[0122] The number of decision trees in the second plurality may exceed the number of decision trees in the first plurality. The first number of variables available for splitting at each tree node in the first random forest is greater than the second number of variables available for splitting at each tree node in the second random forest.

[0123] The first machine learning model may be a first random forest classifier and the second machine learning model may be a second random forest classifier.

[0124] Determining the failure onset time may include determining a sliding-window moving average and a sliding-window standard deviation.

[0125] Determining the failure onset time can include comparing an absolute value of the deviation from a sliding window moving average to a threshold value.

[0126] The threshold may depend on the sliding window standard deviation. Determining the fault onset time may include performing, for each phase, a comparison of the absolute value of the deviation of the voltage from a sliding-window moving average of the voltage to a phase-specific voltage threshold that depends on the sliding-window standard deviation of the voltage for that phase, and a comparison of the absolute value of the deviation of the current from a sliding-window moving average of the current to a phase-specific current threshold that depends on the sliding-window standard deviation of the current for that phase, respectively. For any phase, and for either the voltage or the current, the earliest time at which the threshold is reached may be determined to be the fault onset time.

[0127] Determining the rate of change value may include, for each phase, determining a filtered voltage, determining a filtered current, and calculating the rate of change value from the filtered voltage and the filtered current.

[0128] The filtering may include applying a moving average filter. The phase classifier can receive as input the rate of change of voltage for three phases and the rate of change of current for three phases and output a fault type.

[0129] The disorder type may be selected from the group consisting of ABC, Ag, BC, CA-g.

[0130] The zone classifier can receive as input the rate of change of voltage for three phases and the rate of change of current for three phases and output a zone.

[0131] The zones may be selected from two zones or more than two zones. The method may include updating the phase classifier, the zone classifier, and / or the protection function during field operation of the IED.

[0132] The method may include retraining the machine learning model during field operation of the IED.

[0133] Retraining the machine learning model can include retraining a random forest.

[0134] The method can be used to generate phase classifiers, zone classifiers, protection functions, or other decision logic for transmission lines of power systems having IBRs, particularly for distance protection.

[0135] The phase classifier and / or zone classifier may be determined in a self-configuring manner, providing autonomous generation of decision logic.

[0136] The transmission line protection may include high impedance fault protection. Generating the phase classifier and / or zone classifier can be performed using a data set that includes data for a number of distinct source-line impedances.

[0137] A computing system according to one embodiment has an interface or memory operative to receive or store training data indicative of a faulty line voltage and / or current, the computing system has at least one integrated circuit operative to determine a fault onset time from the faulty line voltage and / or current, determine a rate of change value indicative of a rate of change of voltage and / or indicative of a rate of change of current during at least one sample time that is dependent on the fault onset time, and use the rate of change value to generate a decision logic using machine learning.

[0138] The decision logic may be at least one of a distance protection function, a phase classifier for fault classification, and / or a zone classifier for faulty zone identification.

[0139] The computing system may be an intelligent electronic device for distance protection IED.

[0140] The computing system may have an interface for deploying phase classifiers, zone classifiers, or decision logic derived therefrom (such as protection function decision logic) in intelligent electronic devices IEDs for distance protection.

[0141] The computing system is operable to generate intelligent electronic device (IED) decision logic based on the phase classifier and the zone classifier and to deploy the IED decision logic for execution.

[0142] The IED decision logic may be a decision tree. The IED decision logic may be an interpretable IED decision logic.

[0143] The IED may be a distance relay. The computing system may be operable such that generating the phase classifier and generating the zone classifier comprises an ensemble machine learning method.

[0144] The computing system can be operable to derive features for training the machine learning model from training data including voltages and currents.

[0145] The computing system can be operable such that deriving the features can include determining a rate of change of voltages of multiple phases of the transmission line and a rate of change of currents of multiple phases of the transmission line at one or more times that are dependent on the fault onset time.

[0146] The computing system can be operable such that generating the phase classifier and generating the zone classifier can include random forest training to generate a random forest.

[0147] The computing system may be operable such that features of nodes in an input layer of a random forest decision tree may be selected from rate of change values.

[0148] The computing system may be operable such that features of a node in an input layer of the random forest decision tree may be selected from at least m×n rate of change values ​​indicative of a rate of change of voltage and m×n rate of change values ​​indicative of a rate of change of current, where m≧2 indicates a number of phases and n≧1 indicates a number of sample times.

[0149] The computing system may be operable such that the features of a node in an input layer of the random forest decision tree may be selected from at least m×n rate of change values ​​indicative of a rate of change of voltage and m×n rate of change values ​​indicative of a rate of change of current, where m≧2 indicates a number of phases and n≧2 indicates a number of sample times that is dependent on the fault onset time.

[0150] The computing system may be operable such that the features of a node in an input layer of the random forest decision tree may be selected from at least m×n rate of change values ​​indicative of a rate of change of voltage and m×n rate of change values ​​indicative of a rate of change of current, where m=3 indicates a number of phases and n≧2 indicates a number of sample times that is dependent on the fault onset time.

[0151] The features of the nodes in the input layer of the random forest decision tree can be selected from sample times taken from up to a quarter cycle after the fault onset time.

[0152] The features of the nodes in the input layer of the random forest decision tree can be selected from sample times that are not obtained from the time one-quarter cycle after the fault onset time.

[0153] The computing system may be operable such that random forest training is performed with a data set of L>1 fault lines, each including voltages and currents for three phases.

[0154] The computing system may be operable to train a first machine learning model to generate a phase classifier and to train a second machine learning model to generate a zone classifier.

[0155] The computing system can be operable such that the first machine learning model comprises a first random forest having a first plurality of decision trees, and the second machine learning model comprises a second random forest having a second plurality of decision trees different from the first plurality of decision trees.

[0156] The computing system may operate such that the first machine learning model may be a first random forest classifier and the second machine learning model may be a second random forest classifier.

[0157] The computing system may be operative such that determining the failure onset time includes determining a sliding-window moving average and a sliding-window standard deviation.

[0158] The computing system may be operable such that determining the failure onset time includes comparing an absolute value of the deviation from a sliding window moving average to a threshold value.

[0159] The computing system may be operated such that the threshold value is dependent on the sliding window standard deviation.

[0160] The computing system may be operable such that determining the fault onset time includes performing, for each phase, a comparison of the absolute value of the deviation of the voltage from a sliding-window moving average of the voltage to a phase-specific voltage threshold that depends on a sliding-window standard deviation of the voltage for that phase, and a comparison of the absolute value of the deviation of the current from a sliding-window moving average of the current to a phase-specific current threshold that depends on a sliding-window standard deviation of the current for that phase, respectively. For any phase, and for either the voltage or the current, the earliest time at which the threshold is reached may be determined to be the fault onset time.

[0161] The computing system includes determining a rate of change value for each phase, determining a filtered voltage, and filtering the filtered voltage. Ringed The method may be operative to include determining a current and calculating a rate of change value from the filtered voltage and the filtered current.

[0162] The computing system is operable such that the phase classifier receives as input the rate of change of voltage for the three phases and the rate of change of current for the three phases and outputs a fault type.

[0163] The computing system may be operable such that the fault type is selected from the group consisting of: ABC, Ag, BC, CA-g.

[0164] The computing system is operable such that the zone classifier receives as input the rate of change of voltage for three phases and the rate of change of current for three phases and outputs a zone.

[0165] The computing system may be operated such that the zone is selected from two zones or more than two zones.

[0166] The computing system may operate to update the phase classifier, the zone classifier, and / or the protection functions during field operation of the IED.

[0167] The computing system may be operable to retrain the machine learning model during field operation of the IED.

[0168] The computing system may be operable such that retraining the machine learning model includes retraining a random forest.

[0169] An intelligent electronic device IED according to one embodiment comprises an interface for receiving a power system characteristic, such as a voltage and / or a current for at least one phase of a power transmission line. The IED may be operable to determine a rate of change of the power system characteristic (e.g., a rate of change of the voltage and / or a rate of change of the current) and input the determined rate of change as an input to a decision logic.

[0170] The power system may be or may include a transmission line. The decision logic may be at least one of a distance protection function, a phase classifier for fault classification, and a zone classifier for faulty zone identification.

[0171] The decision logic may be operable to receive as inputs the rates of change of voltage for three phases and the rates of change of current for the three phases of the power transmission line.

[0172] The IED may be operable to determine a filtered voltage from the voltage, determine a filtered current from the current, calculate a rate of change of the voltage from the filtered voltage, and calculate a rate of change of the current from the filtered current.

[0173] The phase classifier may include or be based on a first random forest.

[0174] The zone classifier may include or be based on a second random forest.

[0175] The phase classifier may comprise a first random forest including a first plurality of decision trees; or Based on that and the zone classifier may include or be based on a second random forest including a second plurality of trees.

[0176] The number of decision trees in the second plurality may exceed the number of decision trees in the first plurality. The first number of variables available for splitting at each tree node in the first random forest is greater than the second number of variables available for splitting at each tree node in the second random forest.

[0177] The IED may be operable to update the phase classifier, zone classifier, and / or protection functions during field operation of the IED.

[0178] The IED may be operable to retrain the machine learning model during field operation of the IED.

[0179] The IED can be operated such that retraining the machine learning model includes retraining a random forest.

[0180] The power system comprises a power transmission line and an intelligent electronic device IED according to an embodiment operative to perform a distance protection function for the power transmission line.

[0181] The power system may further include one or more inverter-based resources.

[0182] The inverter-based resource may be a renewable energy resource. The power system may further comprise a computing system operative to receive training data indicative of a voltage and / or current of the faulted line, determine a fault onset time from the voltage and / or current of the faulted line, determine a rate of change value indicative of a rate of change of voltage and / or indicative of a rate of change of current during at least one sample time that is dependent on the fault onset time, use the rate of change value to generate decision logic that is at least one of a distance protection function, a phase classifier for fault classification, and / or a zone classifier for faulted zone identification using machine learning, and deploy the decision logic to an intelligent electronic device (IED).

[0183] According to an embodiment, which may be used in combination with or independently of the above-described embodiments, a computer-implemented method for implementing power system protection comprises: generating a phase classifier; generating a zone classifier; Includes.

[0184] The method may include generating the phase classifier and the zone classifier by performing random forest training.

[0185] Generating the phase classifier and generating the zone classifier may include training a machine learning (ML) model.

[0186] Generating the phase classifier and generating the zone classifier can include determining features from the labeled dataset to train a decision logic model. The decision logic model can include an ML model. The decision logic model can include a first and a second ML model.

[0187] Determining features for training may include determining a fault onset time from voltages and / or currents in the labeled dataset and determining a rate of change value indicative of a rate of change of the voltage and / or the current with respect to at least one sample time, where the at least one sample time is dependent on the fault onset time.

[0188] The method comprises: A phase classifier; A zone classifier; a protection function generated based on a phase classifier and / or a zone classifier; in at least one intelligent electronic device (IED).

[0189] The method includes: A phase classifier; A zone classifier; Protection functions and The method may include performing at least one of the steps in a field deployment of the IED.

[0190] According to an embodiment, which may be used in combination with or independently of the above-mentioned embodiments, there is provided the use of at least one integrated circuit for generating features indicative of a rate of change of at least one electrical characteristic of a power system, the features being input to an input layer of at least one machine learning (ML) or artificial intelligence (AI) model, the ML or AI model having an output layer providing a signal indicative of a fault and / or a type of fault and / or a zone of fault.

[0191] The use may further include deploying the ML or AI model in an intelligent electronic device (IED) to generate the signal.

[0192] The use may further include executing, by an intelligent electronic device (IED), the ML or AI model to generate the signal.

[0193] The use may further include executing, by an intelligent electronic device (IED), the ML or AI model for power system protection and / or monitoring.

[0194] The ML or AI model may be the protection logic. The ML or AI model may be or may implement a distance protection function.

[0195] The use may include using ML or AI models for transmission line protection.

[0196] The ML or AI model may be or comprise at least one of a protection function, a phase classifier for fault classification, and a zone classifier for faulty zone identification.

[0197] The signal may be or may include at least one of a circuit breaker control signal, a switch control signal, an alarm, a warning, status information, an output for output via a human machine interface (HMI).

[0198] The use may include determining a fault onset time from the at least one electrical characteristic.

[0199] Determining the fault onset time may include determining a sliding-window moving average of the at least one electrical characteristic and comparing an absolute value of a deviation of the at least one electrical characteristic from the sliding-window moving average to at least one threshold value.

[0200] Determining the fault onset time can include determining a sliding-window standard deviation of the at least one electrical characteristic.

[0201] At least one threshold may be dependent on a sliding window standard deviation. The fault onset time can be identified as the earlier of the time when the absolute value of the deviation of the voltage from a sliding-window moving average of the voltage reaches or exceeds a phase-specific voltage threshold and the time when the absolute value of the deviation of the current from a sliding-window moving average of the current reaches or exceeds a phase-specific current threshold.

[0202] Generating the features indicative of the rate of change may include, for each phase, determining a filtered voltage, determining a filtered current, and calculating a rate of change value from the filtered voltage and the filtered current.

[0203] The filtered voltage may be determined by averaging N>1 sample values ​​of the voltage, and the filtered current is determined by averaging N>1 sample values ​​of the current.

[0204] Generating features indicative of the rate of change may include determining a voltage difference between the filtered voltage sample value at time k and the filtered voltage sample value at time kN, and a current difference between the filtered current sample value at time k and the filtered current sample value at time kN.

[0205] The rate of change value of the voltage may be determined as the change in the voltage difference between a sample time and a previous sample time.

[0206] The rate of change value of the current may be determined as the change in the current difference between a sample time and a previous sample time.

[0207] The use may include deploying an ML or AI model to implement a phase classifier and / or a zone classifier for execution by an IED for distance protection.

[0208] The use may include using the features to generate one or both of a phase classifier and a zone classifier.

[0209] The use may include using ensemble machine learning (ML) methods to generate a phase classifier and to generate a zone classifier.

[0210] Generating the phase classifier and generating the zone classifier can include random forest training using the features to generate a random forest.

[0211] The features of the nodes in the input layer of the random forest decision tree can be selected from the rate of change values.

[0212] The features of the nodes in the input layer of the random forest decision tree can be selected from at least m × n rate of change values ​​indicating the rate of change of voltage and m × n rate of change values ​​indicating the rate of change of current, where m ≧ 2 indicates the number of phases and n ≧ 1 indicates the number of sample times.

[0213] The features of the nodes in the input layer of the random forest decision tree can be selected from at least m × n rate of change values ​​indicating the rate of change of voltage and m × n rate of change values ​​indicating the rate of change of current, where m ≥ 2 indicates the number of phases and n ≥ 2 indicates the number of sample times depending on the fault onset time.

[0214] The features of the nodes in the input layer of the random forest decision tree can be selected from at least m × n rate of change values ​​indicating the rate of change of voltage and m × n rate of change values ​​indicating the rate of change of current, where m = 3 indicates the number of phases and n ≥ 2 indicates the number of sample times depending on the fault onset time.

[0215] The features of the nodes in the input layer of the random forest decision tree can be selected from sample times taken from up to a quarter cycle after the fault onset time.

[0216] The features of the nodes in the input layer of the random forest decision tree can be selected from sample times that are not obtained from the time one-quarter cycle after the fault onset time.

[0217] Using the features indicative of the rate of change to generate one or both of the phase classifier and the zone classifier can include training a first machine learning model with the features to generate the phase classifier.

[0218] Using the features indicative of the rate of change to generate one or both of the phase classifier and the zone classifier can include training a second machine learning model with the features to generate the zone classifier.

[0219] The phase classifier can receive as input the rate of change of voltage for three phases and the rate of change of current for three phases and output a fault type.

[0220] The zone classifier can receive as input the rate of change of voltage for three phases and the rate of change of current for three phases and output a zone.

[0221] The use may include generating phase classifiers and / or zone classifiers in a self-configuring manner.

[0222] The ML or AI model may be the decision logic for transmission line protection. The transmission line protection may include high impedance fault protection.

[0223] An ML or AI model can be trained to generate a phase classifier and / or a zone classifier using a dataset that includes data for several distinct source-line impedances.

[0224] Various effects and advantages are achieved by the devices, systems, and methods according to the embodiments.

[0225] The methods, devices, and systems enable phase classification, zone classification, and / or protection functions to be generated using a data-driven approach. The methods, devices, and systems enable phase classification, zone classification, and / or protection functions to be determined in a self-configuring manner. The methods, devices, and systems provide fast and safe fault and zone classification of power systems having IBR.

[0226] The methods, devices, and systems provide trained AI / ML models that are system independent and therefore do not require system data every time.

[0227] The method, device, and system provide fast and safe operation of transmission line relays in systems with high penetration of inverter-based resources (renewable energy).

[0228] The methods, devices, and systems provide interpretable AI / ML models. By way of example, a trained AI / ML model may be, or may be convertible into, a decision tree that is interpretable by a human expert.

[0229] The methods, devices, and systems provide techniques for generating practical deployable code for easily importing protection logic into an IED.

[0230] Devices, methods, and systems according to embodiments may be used in conjunction with, but are not limited to, IBR systems. Devices, methods, and systems may be used to improve distance protection functions performed by distance relays.

[0231] BRIEF DESCRIPTION OF THE DRAWINGS The subject matter of the present invention will now be explained in more detail with reference to preferred exemplary embodiments illustrated in the accompanying drawings. [Brief description of the drawings]

[0232] [Figure 1] FIG. 2 is a schematic partial view of a system with a dual circuit line and a protection device. [Diagram 2] FIG. 2 is a block diagram of a protection device. [Diagram 3] 1 is a flow chart of a method. [Figure 4] FIG. 1 is a schematic diagram of the first and second random forests. [Diagram 5] 1 is a flow chart of a method. [Figure 6] 1 is a flow chart of a method. [Figure 7] 1 is a flow chart of a method. [Figure 8] A diagram of voltages and currents in a phase of a faulty transmission line. [Figure 9] A diagram of voltages and currents in a phase of a faulty transmission line. [Figure 10] FIG. 11. Accuracy of phase classifiers in 10-fold cross-validation. [Figure 11] FIG. 13 is a plot of average producer accuracy of the phase classifier. [Figure 12] FIG. 13 is a plot of average user accuracy of the phase classifier. [Figure 13] FIG. 11. Accuracy of zonal classifiers in 10-fold cross-validation. [Figure 14] FIG. 13 is a plot of average producer accuracy of the zone classifier. [Figure 15]FIG. 13 is a plot of average user accuracy of the zone classifier. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0233] Detailed Description of the Preferred Embodiments Exemplary embodiments of the invention are described with reference to the drawings, in which like or similar reference numerals indicate like or similar elements. Although described in the context of certain embodiments, the methods and devices described in detail below can be used in a wide variety of systems.

[0234] Features of the embodiments can be combined with each other unless otherwise stated. According to embodiments of the present invention, methods, devices, and systems are provided that operate to perform protection functions, particularly distance protection functions, for power systems. The methods, devices, and systems can be used for distance protection of transmission lines in power systems having inverter-based resources such as renewable energy.

[0235] According to an embodiment of the present invention, devices, systems, and methods are provided that operate to generate and / or use phase and zone classifiers for distance protection. The devices, systems, and methods can determine features from voltages and currents of three phases of a transmission line, the features being indicative of the rate of change of voltage and the rate of change of current. The features can be used to train one or several machine learning models to generate classifiers (training) and / or to use decision logic based on the generated classifiers (live operation of an intelligent electronic device (IED)).

[0236] In a particular embodiment, a random forest may be generated using features derived from training data. The features may be indicative of the rates of change of voltage and current of the three phases of a faulty transmission line (the training data encompasses voltage and current before and after the occurrence of a fault). The random forest may be trained to generate a first random forest of first decision trees implementing a phase classifier indicative of the type of fault and a second random forest of second decision trees implementing a zone classifier indicative of the faulty zone. Other machine learning (ML) / artificial intelligence (AI) models may be trained using features indicative of the rates of change of voltage and current of the three phases of a faulty transmission line.

[0237] As used herein, the term "phase classifier" refers to a classifier that outputs a fault type classification. For a three-phase transmission line, the phase classifier can classify into a class selected from the group consisting of types A-C, Ag, BC, CA-g.

[0238] As used herein, the term "zone classifier" refers to a classifier that outputs a zone. For purposes of illustration, but not limitation, the zone classifier may classify into a class selected from the group consisting of two zones (Zone 1, Zone 2).

[0239] In methods, devices, and systems according to embodiments, a second random forest having a second plurality of decision trees may be generated to operate as a zone classifier that receives as input the rate of change values ​​of the voltage in the three phases and the rate of change values ​​of the current in the three phases, performs zone classification, and outputs information regarding the zone type. In response to receiving as input the rate of change values ​​of the voltage in the three phases and the rate of change values ​​of the current in the three phases, the zone classifier may operate to classify the fault into one of a plurality of distinct zones (Zone 1, Zone 2, etc.).

[0240] 1 is a schematic, partial diagram of a power system 10. Power system 10 has a source 12 that may include an inverter-based resource, such as a renewable energy resource.

[0241] The power system has a transmission line 11. A source 12 may be coupled to the transmission line 11 by a transformer 13. The power system may include a dual circuit transmission line 11, 11'.

[0242] The protection device 20 may be implemented as a protection relay, in particular a distance relay. The protection device 20 may receive the transmission line current from a current transformer 14. The protection device 20 may receive a transmission line voltage measurement, such as a phase-to-ground voltage, at the terminals of the transmission line (e.g., a local bus) on which the protection device 20 is located. The voltage may be received from a transformer 17.

[0243] The protection device 20 is operable to perform distance protection functions. The protection device 20 may perform decision logic, such as a distance protection function that uses as inputs the rates of change of current and voltage of several phases of the transmission line 11. The decision logic may include protection functions that output phase classification, zone classification, or a decision to perform tripping or not.

[0244] To protect the dual-circuit transmission line 11, 11', a further protection device 20' may be provided that receives current and voltage measurements from the further current transformer 14' and the transformer 17 (or a dedicated further transformer). The operation, structure and functionality of the further protection device 20' may correspond to the operation, structure and functionality of the protection device 20. Thus, although reference is made hereinafter to the protection device 20, it will be understood that the techniques disclosed herein may be applied to generate decision-making logic for each of several protection devices (i.e. protection devices located at the local and remote ends of the dual-circuit transmission line) and / or during operation of the protection devices.

[0245] The decision logic executed by the protection device 20 may be determined by the computing system 15. The computing system 15 may be, but need not be, distinct and separate from the protection device 20.

[0246] The computing system 15 is operable to receive training data 16 from a data repository and process the training data 16 to generate phase and zone classifiers. The phase and zone classifiers or decision logic dependent thereon may be deployed in a protection device 20 for use in performing transmission line protection.

[0247] The training data 16 may include scenarios of various fault types, fault locations, source-line impedance ratios at terminals A and B, fault resistances, and initiation angles.

[0248] The computing system 15 can be operative to provide artificial intelligence (AI) / machine learning (ML) based phase identification (for fault classification) and zone identification for transmission line systems. The computing system 15 can be operative to perform fault onset time detection in the training data 16 using analytical techniques. The computing system 15 can be operative to determine a sliding window moving average and a sliding window standard deviation of the current and / or voltage, and compare the absolute value of the deviation from the sliding window moving average to a threshold that depends on the sliding window standard deviation, thereby determining the fault onset time when generating the phase and zone classifiers. The computing system 15 can be operative to determine the rate of change of the current and the rate of change of the voltage for the three phases of the transmission line represented by the training data 16, and use these data to generate the phase and zone classifiers.

[0249] The training data 16 may be or include labeled training data that includes information on fault types and / or fault locations in addition to currents and voltages of several phases of the transmission line. Information on fault types and / or fault locations can be used when performing ML training. Alternatively, the training data 16 may be or include unlabeled training data. Expert knowledge can be used when processing the training data 16 to generate the phase and zone classifiers. Various ML techniques, such as Random Forest (RF), have the ability to perform missing data substitution. Correlations between features are determined from a dataset and used to calculate, often iteratively, substitute values ​​for the missing data. Thus, even if only a portion of the training data includes fault types and fault locations in addition to currents and voltages, this may be sufficient to perform automatic classifier generation using ML-based techniques.

[0250] The computing system 15 can be operative to determine a fault onset time from the voltages and / or currents of the phases from training data associated with the transmission line. The computing system 15 can be operative to determine values ​​indicative of the rate of change of the voltage and / or the rate of change of the current. These values ​​are referred to as "rate of change values" for brevity and typically include both the rate of change of the voltage for several (particularly three) phases of the transmission line and the rate of change of the current for several (particularly three) phases. The rate of change values ​​are taken at least one time that depends on the determined fault onset time. The computing system 15 can be operative to use the rate of change values ​​to generate a phase classifier for fault classification and / or a zone classifier for faulty zone identification. The computing system 15 can generate the phase classifier and the zone classifier using machine learning in a data-driven and self-configuring manner.

[0251] The rate of change values ​​used to generate the decision logic can be selected from sample times taken from up to a quarter cycle after the fault initiation time.The rate of change values ​​used to generate the decision logic can be selected from sample times taken from up to a quarter cycle after the fault initiation time.

[0252] Based on the phase and zone classifiers determined by the computing system 15, a decision logic for the protection functions (particularly the distance protection function) can be determined. This may be done automatically by the computing system 15. Alternatively, the computing system 15 can assist an expert in setting the decision logic by providing an interpretable classifier or decision logic for use in the protection device 20.

[0253] The phase classifier, zone classifier, or protection functions (particularly distance protection functions) generated therefrom may be loaded into protection device 20 for execution. In field operation of protection device 20, protection device 20 may receive currents and voltages for phases of transmission line 11, process the currents and voltages to determine features (particularly features indicative of rates of change of current and rates of change of voltage), and use the determined features as inputs to the phase classifier, zone classifier, or protection function.

[0254] The phase classifier, zone classifier, or protection function may be or be based on a trained ML model generated by the computing system 15.

[0255] 2 is a block diagram of the protection device 20. The protection device 20 may be a distance relay for transmission line protection in a power system with high IBR penetration.

[0256] The protection device 20 controls the voltage and current for one or several phases of the transmission line. receiveThe protection device 20 has an interface 25 for receiving and transmitting data from the protection device 20. The protection device 20 has one or several integrated circuits (ICs) 21 for performing processing functions. The one or several ICs 21 may include one or several of a processor, a microprocessor, a controller, a microcontroller, a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), or any combination thereof.

[0257] One or several ICs 21 may implement a phase classifier 22. The phase classifier may operate to classify a fault into one of the types ABC, Ag, BC, CA-g in response to features that depend on the currents and voltages of the phases of the transmission line 11. The features that may be determined by IC 21 from the currents and voltages may depend on the rate of change values ​​of the voltages in the three phases and the rate of change values ​​of the currents in the three phases.

[0258] One or several of the ICs 21 may implement a zone classifier 23. The zone classifier may be operable to classify the transmission line 11 into one of a number of distinct zones (zone 1, zone 2, etc.) in response to features that depend on the currents and voltages of the phases. The features that may be determined by the ICs 21 from the currents and voltages may depend on rate of change values ​​of the voltages in the three phases and rate of change values ​​of the currents in the three phases.

[0259] One or several ICs 21 may implement a distance element 24 to perform a distance protection function. The distance element may operate to determine a corrective response action to be performed in response to a characteristic quantity that depends on the currents and voltages of the phases of the transmission line 11. The characteristic quantity that may be determined by IC 21 from the currents and voltages may depend on rate of change values ​​of the voltages in the three phases and rate of change values ​​of the currents in the three phases.

[0260] The features used as inputs to the phase classifier 22, the zone classifier 23, or the distance element 24 can be determined from the voltages and currents in the same or similar manner as described below in connection with classifier training.

[0261] The phase classifier 22, the zone classifier 23, or the distance element 24 may be or be based on a trained ML model, such as provided by the computing system 15.

[0262] 3 is a flow chart of a method 30. The method 30 may be performed by the computing system 15. The method 30 may be performed to process the training data 16, thereby generating phase and zone classifiers.

[0263] At step 31, feature determination is performed. The feature determination may include process currents and voltages included in the training data 16. The feature determination may include determining the rate of change of voltages and currents for several phases at or near the fault initiation time for each line represented by data in the training data 16. It will be appreciated that feature determination may be performed for each of several separate scenarios included in the training data 16. The training data 16 may encompass various fault types, fault locations, source-line impedance ratios at terminals A and B, fault resistances, and initiation angles.

[0264] In step 32, machine learning (ML) model training is performed. The ML model training may be or may include a random forest technique. The ML model training may be based on the features determined in step 31. The ML model training may utilize features indicative of the rate of change of current and the rate of change of voltage determined from the scenarios in the training data. If a random forest (RF) technique is used, step 32 may include generating a first random forest of first decision trees implementing a phase classifier indicative of a type of fault and generating a second random forest of second decision trees implementing a zone classifier indicative of a faulty zone. The number of first decision trees may be different from the number of second decision trees.

[0265] Model validation may optionally be performed after training, in step 32. If validation fails, the method may loop back to step 31.

[0266] In step 33, the trained ML model, or decision logic based on the trained ML model (such as a classifier or distance protection function), can be deployed to a distance relay 20 or other protection device 20 to perform a transmission line protection function.

[0267] As explained above, ML training may involve RF techniques, as exemplarily shown with reference to FIG.

[0268] 4 is a schematic diagram showing a first random forest 40 and a second random forest 50. The first random forest 40 may be generated to implement a topological classifier. The second random forest 50 may be generated to implement a zonal classifier.

[0269] The first random forest 40 includes a first plurality of first decision trees 41, 42, 43. A leaf node 48, 49 of each first decision tree may indicate a disorder type. The disorder type may be selected from the group consisting of ABC, Ag, BC, CA-g.

[0270] With the random forest technique, a first random forest 40 includes trees with different features at their root nodes, which in turn have different features at the child nodes of the root node, and so on. By way of example, a first decision tree 41 may have a root node 45 associated with a feature indicative of a rate of change of a voltage VA in a phase of the transmission line 11. Another first decision tree 42 may have a root node 46 associated with a feature indicative of a rate of change of a voltage VB in another phase of the transmission line 11. Yet another first decision tree 43 may have a root node 47 associated with a feature indicative of a rate of change of a current IC in one of the phases of the transmission line 11. Similarly, nodes further down in the decision trees 41-43 may vary from one decision tree to another within the random forest (although this need not be the case for all trees).

[0271] The second random forest 50 includes a second plurality of second decision trees 51-54. A leaf node 59 of each second decision tree may indicate a zone. The zone may be selected from a group of several distinct zones, such as zone 1 and zone 2.

[0272] With the random forest technique, the second random forest 50 includes trees with different features at their root nodes, and similarly with different features at the child nodes of the root node, and so on. By way of example, the second decision tree 51 may have a root node 55 associated with a feature indicative of a rate of change of the voltage VA in a phase of the transmission line 11. The other second decision tree 52 may have a root node 56 associated with a feature indicative of a rate of change of the voltage VB in another phase of the transmission line 11. A further second decision tree 53 may have a root node 57 associated with a feature indicative of a rate of change of the current IC in one of the phases of the transmission line 11. A further second decision tree 54 may have a root node 58 associated with a feature indicative of a rate of change of the current IA in another one of the phases of the transmission line 11. Similarly, the nodes further down the decision trees 51-54 may vary from one decision tree to another within the random forest (although this need not be the case for all trees).

[0273] The number of variables available for splitting at each tree node (i.e., the Mtry hyperparameter) may be different between the first random forest and the second forest. For illustrative purposes, the number of variables available for splitting at each tree node may be greater in the first random forest than in the second random forest.

[0274] 5 is a flow chart of a method 60. Method 60 may be performed by computing system 15. Method 60 may be performed to process training data 16, thereby generating phase and zone classifiers. Method 60 may be performed to perform ML model training in step 32 and deployment in step 33 of method 30.

[0275] In step 61, a first random forest is generated from the training data. Generating the first random forest includes: using features that indicate the rate of change of current and the rate of change of voltage for various scenarios in the training data; using information about fault types available for the scenarios in the training data; may include.

[0276] If information about fault types is unavailable for some of the scenarios, they may still be used in generating the first random forest while applying missing data replacement techniques that iteratively provide improved estimates of the missing data (here, missing fault types). Each first decision tree in the first random forest may include non-leaf nodes associated with the rates of change of current and voltage, respectively, for several phases of the transmission line. Each first decision tree in the first random forest may output a fault type (such as a fault type selected from the group consisting of ABC, Ag, BC, CA-g) for any set of inputs including the rates of change of current and voltage for several phases of the transmission line. The first random forest may implement a first ML model that provides a phase classifier for fault classification.

[0277] At step 62, a second random forest is generated from the training data. Generating the second random forest includes: using features that indicate the rate of change of current and the rate of change of voltage for various scenarios in the training data; using information about fault locations available for the scenarios in the training data; may include.

[0278] If information about fault locations is unavailable for some of the scenarios, they may still be used in generating the second random forest while applying missing data replacement techniques that iteratively provide improved estimates of the missing data (here, missing fault locations). Each second decision tree in the second random forest may include non-leaf nodes associated with the rates of change of current and voltage, respectively, for several phases of the transmission line. Each second decision tree in the second random forest may output a fault zone (e.g., a fault zone selected from the group consisting of zone 1, zone 2, etc.) for any set of inputs including the rates of change of current and voltage, respectively, for several phases of the transmission line. The second random forest may implement a second ML model that provides a zone classifier for faulty zone classification.

[0279] In step 63, an interpretable decision logic (e.g., an interpretable code for distance protection) may be generated and deployed to the protection device. The decision logic may depend on a phase classifier and / or a zone classifier.

[0280] 6 is a flow chart of a method 70 for determining features that may be used in generating phase and zone classifiers and / or when using decision logic within protection device 20. Method 70 may be performed by computing system 15 and / or protection device 20.

[0281] In step 71, fault onset times are determined from currents and voltages for several (eg, three) phases of the transmission line.

[0282] In step 72, the current and voltage are filtered to determine an average current and voltage (note that the filtering and averaging occurs over time, not over different phases). The filtering may be performed using a moving average filter.

[0283] In step 73, the rate of change of the average current and the rate of change of the average voltage are calculated. The rate of change of the average current and the rate of change of the average voltage may be calculated for at least one or several sample times following the determined fault onset time. For each phase, the rate of change of the average voltage may be calculated based on a comparison comparing a value of the average voltage for a first sample time to a value of the average voltage for a second sample time, the time difference between the first sample time and the second sample time being dependent on the number of samples per cycle used for filtering in step 72. For each phase, the rate of change of the average current may be calculated based on a comparison comparing a value of the average current for a first sample time to a value of the average current for a second sample time, the time difference between the first sample time and the second sample time being dependent on the number of samples per cycle used for filtering in step 72.

[0284] The rate of change versus sample time following the fault onset time is used to determine features that are input to the phase classifier 22, zone classifier 23, and / or distance element 24 for ML model training by the computing system 15 prior to field operation of the decision logic, or by the protective device 20 during field operation.

[0285] 7 is a flow chart of a method 80 for determining fault onset times. The method 80 may be used when processing the training data 16 to generate phase and zone classifiers to determine fault onset times for scenarios in the training data 16. The method 80 may be used in the protection device 20 to generate features for input to the phase classifier 22, the zone classifier 23, and / or the distance element 24.

[0286] In step 81, a moving average of the voltage is determined and a moving average of the current is determined. This may be done for every phase of the transmission line for which fault scenarios are provided in the training data 16. A sliding window technique may be used to determine the moving averages.

[0287] In step 82, a moving standard deviation of the voltage is determined and a moving standard deviation of the current is determined. This may be done for every phase of the transmission line for which fault scenarios are provided in the training data 16. A sliding window technique may be used to determine the moving standard deviation.

[0288] At step 83, a threshold comparison may be performed to determine a fault initiation time. The absolute value of the difference between the voltage and the average voltage may be compared to a first threshold. The first threshold may be a constant multiple of the moving standard deviation of the voltage. The absolute value of the difference between the current and the average current may be compared to a second threshold. The second threshold may be a constant multiple of the moving standard deviation of the current. The earlier of the times at which the first or second threshold comparison indicates that the absolute value of the difference exceeds the associated threshold may be considered the fault initiation point or fault initiation time.

[0289] Feature determination Detailed embodiments of the feature determination are provided below. The feature determination may be performed to determine features that may be used in generating phase and zone classifiers and / or when using decision logic within the protection device 20. The feature determination may be performed by the computing system 15 and / or the protection device 20.

[0290] The feature determination can operate on three-phase voltage and current signals. The steps performed in preparing the features are described below.

[0291] Step 1: Determine the outage start time The line voltage and current are monitored continuously (i.e., continuously). The line voltage and current may be monitored continuously. For training data 16, data is available for each fault scenario in the training data. The line voltage and current are processed to determine the fault onset time. This can be done using A-D below.

[0292] · One sample difference (i.e. the difference between successive samples) of the three-phase voltage and current signals is calculated (e.g. consecutively, for various sample times).

[0293]

number

[0294]

number

[0295] The moving average and standard deviation of a sliding window of size w are calculated (e.g., continuously, for each sample time)

[0296]

number

[0297]

number

[0298] where w is the sliding window size. The sliding window size w can be adapted to improve the accuracy of the trained ML model. sd represents the standard deviation. mean represents the mean function to calculate the average value.

[0299] The fault initiation time for each phase voltage and current is identified when

[0300]

number

[0301]

number

[0302] Therefore, the threshold comparison of equations (5) and (6) is performed on the deviation from the moving average. The threshold (the right-hand side of equations (5) and (6)) may depend on the moving standard deviation. c is a constant that may be the same in equations (5) and (6) or may be different in equations (5) and (6).

[0303] The smallest of the voltage and current fault starting points of each phase is regarded as the fault starting point. 8 and 9 show the voltage 91 and current 92 for a phase of a transmission line. At time or sample 93, the current 92 meets the threshold criterion (6), which indicates the fault onset time for this phase.

[0304] Step 2: Moving average filter The voltage and current of each phase are processed through a moving average filter to remove unwanted ripple. This can be done according to the following:

[0305]

number

[0306]

number

[0307] where N is the number of samples per cycle and k is the sample number. Step 3: Determine rate of change Determine the rate of change of voltage (ΔV) and current (ΔI) for each phase. This can be done using the following relationships:

[0308]

number

[0309]

number

[0310] The difference between the right-hand sides of equations (9) and (10) depends on the number of samples N per cycle used in the immediately preceding filtering.

[0311] Step 4: Sample values Rate of change of voltage (ΔV avg ) and the rate of change of current (ΔI avg ) may be obtained for each phase from the instance of the fault starting point detected from step 1.

[0312] The rate of change sample values ​​used to generate the decision logic can be selected from sample times taken from up to a quarter cycle after the fault initiation time.The rate of change sample values ​​used to generate the decision logic can be selected from sample times taken from up to a quarter cycle after the fault initiation time.

[0313] Step 5: Calculate the variation The voltage and current variations of each phase using the following relationships:

[0314]

number

[0315]

number

[0316] Random forest model training The features determined as described above can be used for random forest (RF) model training, which may involve developing two RF classification models.

[0317] Topological Classification Models - for Fault Classification

[0318]

number

[0319] Output: [Fault type (ABC,Ag,BC,CA-g)] Zone Classification Model - for fault zone identification

[0320]

number

[0321] Output: [Zone(Zone1,Zone2)] Random forest model training may be performed based on the training data using the features determined as described above and at least the fault type and faulty zone information contained in the training data.

[0322] Working Example Training Dataset A total of 9072 scenarios of line faults are simulated using Power System CAD (PSCAD) / EMTDC software to train and validate the RF model. These scenarios are simulated to cover the range of possible combinations of line faults with respect to different loads, fault locations, SIRs, fault resistances and opening angles. The range of these different scenarios is summarized in Table 1.

[0323] [Table 1]

[0324] RF Model Training A total of 27216 samples covering 9072 fault types are used for training. A first RF model and a second RF model are trained. Cross-validation techniques are used to identify the best parameters (e.g. number of trees) for both RF classification models. The identified parameters are shown in Table 2 along with the inputs and objectives considered.

[0325] [Table 2]

[0326] Mtry is the number of variables that can be split at each tree node. Model Validation To validate the performance of the fitted RF classifier model, k-fold cross-validation is used along with majority voting technique. The set of observations is randomly divided into k groups or folds of approximately equal size. The first fold is treated as the validation set, and the XGBoost model is fitted using the remaining k-1 folds. The misclassification error is then calculated for the held-out folds. To calculate the misclassification error, the output is assigned to a class that is calculated as the majority vote of the predicted observation classes. By repeating this procedure k times, k estimates of the test error are calculated. Each time, a different set of observations is treated as the validation set.

[0327] Topological Classification Model: The faulty topological classification random forest model (i.e., the topological classifier) ​​is validated with 10-fold cross-validation. The overall classification accuracy obtained for each fold is shown in Figure 10. The average producer and user accuracies for each fault type class are provided in Figures 11 and 12, respectively.

[0328] Zonal Classification Model: The impaired zonal classification random forest model (i.e., zonal classifier) ​​is validated with 10-fold cross-validation. The overall classification accuracy obtained for each fold is shown in Figure 13. The average producer and user accuracies for each zonal prediction class are provided in Figures 14 and 15, respectively.

[0329] As shown in Figures 10-12, fault classification with an average accuracy of 99.59% can be achieved within 3-5 milliseconds after a disturbance. As shown in Figures 13-15, fault zone identification with an average accuracy of 99.3% can be achieved within 3-5 milliseconds after a disturbance.

[0330] The technology of the embodiments provides an ML / AI based technique that can be interpreted and used to generate protection logic as shown in the following figure. The generated protection logic can be loaded into the protection IED.

[0331] Protection devices, methods, and systems according to embodiments provide various advantages: Fault onset time identification can be performed accurately; Smart data window selection (e.g., by determining the rate of change of voltage and current based on the difference in sample times that depends on the number of samples per cycle used in the preceding filtering step) contributes to improved accuracy of faulty phase and zone identification;

[0332] The protection device, method, and system provide a system independent technique that is easy to generate protection logic. The protection logic is interpretable and easy to implement in IEDs. Accurate and fast fault and zone classification can be performed. The protection device, method, and system are suitable for, but not limited to, use with IBR systems.

[0333] Devices, methods and systems according to embodiments provide accurate and robust results. Devices, methods, and systems according to embodiments may be used to provide distance protection to transmission networks of an electrical grid that includes renewable energy sources or other inverter-based resources.

[0334] Although the present invention has been described in detail in the drawings and the foregoing description, such description should be considered as illustrative or exemplary and non-limiting. Variations of the disclosed embodiments can be understood and achieved by those skilled in the art practicing the claimed invention from a study of the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other elements or steps, and the indefinite articles "a" or "an" do not exclude a plurality. The mere fact that certain elements or steps are recited in separate claims does not indicate that combinations of these elements or steps cannot be advantageously used, and in particular, any further meaningful combination of claims should be considered as disclosed in addition to the actual claim dependencies.

Claims

1. 1. A computer-implemented method, the method comprising: and calculating features for use in generating decision logic operative to generate an operative signal for use in the power system, the calculating the features comprising: determining a fault onset time from at least one electrical characteristic of the sampled data for a given type of power system; determining a rate of change value indicative of a rate of change of the at least one electrical characteristic with respect to at least one sample time, the at least one sample time being dependent on the fault onset time; and using the features to generate the decision logic; and deploying the decision logic in an intelligent electronic device (IED) to generate the signal; A computer-implemented method comprising:

2. 2. The computer-implemented method of claim 1, wherein using the features to generate the decision logic comprises performing machine learning (ML) model training with the computed features to generate one or both of a phase classifier and a zone classifier.

3. The computer-implemented method of claim 2 , wherein the ML model has an input layer that receives the rate of change values.

4. The decision logic comprises: Protection functions and a topological classifier for fault classification; a zone classifier for faulty zone identification; 2. The computer implemented method of claim 1, comprising at least one of:

5. The signal is A circuit breaker control signal; A switch control signal; An alarm and A warning and Status information, an output for output via a human machine interface (HMI); 2. The computer implemented method of claim 1, wherein the at least one of:

6. The failure start time is the time at which the absolute value of the deviation of the voltage from a sliding window moving average of the voltage reaches or exceeds a phase-specific voltage threshold; the time at which the absolute value of the deviation of the current from a sliding window moving average of the current reaches or exceeds a phase-specific current threshold; 2. The computer implemented method of claim 1, wherein the first and second inputs are identified as the earlier of:

7. 2. The computer-implemented method of claim 1, wherein determining the fault onset time comprises determining a sliding window moving average and comparing an absolute value of a deviation of the at least one electrical characteristic from the sliding window moving average to at least one threshold value.

8. The computer-implemented method further comprises determining a sliding-window standard deviation of the at least one electrical characteristic, the at least one threshold value being dependent on the sliding-window standard deviation; and / or The failure start time is the time at which the absolute value of the deviation of the voltage from a sliding window moving average of the voltage reaches or exceeds a phase-specific voltage threshold; the time at which the absolute value of the deviation of the current from a sliding window moving average of the current reaches or exceeds a phase-specific current threshold; 8. The computer implemented method of claim 7, wherein the first and second inputs are identified as the earlier of:

9. Determining the rate of change value includes, for each phase: determining a filtered voltage; determining a filtered current; calculating the rate of change value from the filtered voltage and the filtered current; 2. The computer implemented method of claim 1, comprising:

10. the filtered voltage is determined by averaging N>1 sample values ​​of the voltage, and the filtered current is determined by averaging N>1 sample values ​​of the current; Optionally, calculating the rate of change value comprises: a voltage difference between a sample value of the filtered voltage at time k and a sample value of the filtered voltage at time k-N; a current difference between a sample value of the filtered current at time k and a sample value of the filtered current at time k-N; determining Further optionally, the rate of change value of the voltage is determined as a change in the voltage difference between a sample time and a previous sample time; 10. The computer-implemented method of claim 9, wherein the rate of change value of the current is determined as a change in the current difference between the sample time and the previous sample time.

11. The computer-implemented method of claim 1 , wherein deploying the decision logic includes deploying a phase classifier and / or a zone classifier for execution by an IED for distance protection.

12. generating the decision logic includes using the computed features to generate one or both of a phase classifier and a zone classifier.

10. The computer implemented method of claim 1.

13. Using the computed features to generate the decision logic includes using an ensemble machine learning (ML) method to generate a phase classifier and to generate a zone classifier; and / or and / or generating the phase classifier and generating the zone classifier comprises random forest training using the computed features to generate a random forest; Using the rate of change values ​​to generate one or both of the phase classifier and the zone classifier, training a first machine learning model using the computed features to generate the topological classifier; training a second machine learning model using the computed features to generate the zone classifier; and Including, 13. The computer implemented method of claim 12.

14. the phase classifier receives as input the rate of change of voltage for the three phases and the rate of change of current for the three phases and outputs a fault type; and / or 13. The computer implemented method of claim 2 or claim 12, wherein the zone classifier receives as inputs rates of change of voltage for three phases and rates of change of current for the three phases and outputs a zone.

15. 13. A computer-implemented method according to claim 2 or claim 12, wherein the phase classifier and / or the zone classifier are generated in a self-configuring manner.

16. 13. The computer implemented method of claim 2 or claim 12, wherein the phase classifier and / or the zone classifier are continuously updated during field operation.

17. the decision logic is decision logic for transmission line protection, optionally the transmission line protection includes high impedance fault protection; and / or generating the phase classifier and / or zone classifier is performed using a data set including data for a number of distinct source-line impedances; 3. The computer implemented method of claim 2.

18. 1. An intelligent electronic device (IED), comprising: an interface for receiving a voltage and / or a current for at least one phase of a power transmission line; The IED is operative to determine a rate of change of the voltage and / or a rate of change of the current, and to input the determined rate of change of the voltage and / or the determined rate of change of the current as input to decision logic based on a machine learning (ML) model trained using features based on the determined rate of change of the voltage and / or the determined rate of change of the current.

19. A power transmission line; 20. An intelligent electronic device (IED) as recited in claim 18 operative to perform a distance protection function for the power transmission line; A power system comprising:

Images