Optimized fault detection in a power grid

Abstract

The invention relates to the detection of a fault (in particular an arcing fault) in a power grid, on the basis of measurement data. The measurement data are evaluated using a first analysis method in order to detect the fault. In addition, according to the invention the measurement data are evaluated using a second analysis method based on machine learning only if the first analysis method detects the presence of the fault. A fault is only output if the second analysis method also detects the presence of the fault. The first analysis method is designed for less complex analysis than the second method and is configured to make reliable statements regarding the absence of the fault.

Description

[0001] 202415783

[0002] 1

[0003] Description

[0004] Optimized fault detection in a power grid

[0005] The invention relates to a method, a device, a computer program product and a computer program for detecting a fault in a power network (especially arc faults) based on measurement data.

[0006] An arc fault detection device (AFDD) is a type of arc fault circuit interrupter (AFDD) used to detect arc faults. This is necessary because arc faults can cause fires, which pose a significant risk to property and personal safety.

[0007] There are two types of arc faults: parallel and series. Parallel arc faults occur between a phase conductor and earth or protective earth, between two phase conductors, or between a phase conductor and neutral conductor. Such arc faults can often be detected by residual current devices (RCDs). In contrast, series arc faults, which occur within a single conductor, cannot be detected by RCDs.

[0008] This is where AFDD technology comes into play. It is based on monitoring and evaluating current and voltage waveforms, as well as high-frequency noise emissions often referred to as RSSI (Received Signal Strength Indication), as measured by sensors. By analyzing the temporal profiles of these signals, an AFDD can detect the occurrence of an arc fault. Various sensors and approaches have been described in the literature, for example, in US publication 5,729,145 A, which also explains how a suitable sensor is used on a phase line in conjunction with the AFDD.

[0009] Currently, rule-based approaches are frequently used for arc flash detection. However, machine learning (ML)-based detection approaches are proving promising. In practical applications of ML-based detection methods, training data is generated or recorded and used to train an ML algorithm. The trained ML model is then evaluated on a processor or microcontroller in a device to analyze the recorded measurement signals for the presence of arc flashes.

[0010] 2

[0011] An example of such a procedure is described in DE 102021 2045 30 A1.

[0012] The invention aims to optimize ML-based arc flash detection.

[0013] The problem is solved by a method according to claim 1 for detecting a fault (in particular an arc fault) in a power network on the basis of measurement data and a device designed for carrying out a method according to the invention according to claim 11. Advantageous further developments are specified in the dependent claims.

[0014] A method for detecting a fault in a power grid based on measurement data is proposed. The fault is preferably a series arc fault, which is detected, for example, by an AFDD. In principle, however, the method can also be used for detecting other faults. The method is particularly useful where fault assessment can be complex and time-consuming. This is less likely to be the case with other faults protected by circuit breakers, such as residual current (or parallel arc faults), short circuits, or overcurrents (although the method is applicable in principle and this application is explicitly included in the scope of protection). In contrast, the method can also be used to analyze more complex quality requirements for the power supply and to detect problems (instabilities, non-compliance with quality parameters, etc.), which are particularly relevant in the context of...can play a role in network control.

[0015] In the inventive method, detection is carried out in two stages, with the second stage being performed using a machine learning-based analysis method. Such methods are very powerful but also require considerable resources. This is where the invention comes in. A less resource-intensive analysis method is placed before the machine learning-based analysis method. This less resource-intensive method is designed to make reliable statements regarding the absence of the error (i.e., it operates by process of elimination). If the less resource-intensive method rules out an error, no analysis is performed using the more complex machine learning-based analysis method. In other words, the process is as follows:

[0016] • Evaluating the measurement data using an initial analysis procedure to determine an error 202415783

[0017] 3

[0018] • Evaluate the measurement data using a second, machine learning-based analysis method, only if the first analysis method does not determine the absence of the error.

[0019] • Determining the fault if the second analysis method indicates the presence of the fault (e.g., the presence of a serial arc).

[0020] The statement that the first analysis method is designed to make reliable statements regarding the absence of the error should not be interpreted as meaning that there is absolutely no error. While this is desirable and ideally the case, it is not always achievable with a complex technical system. Therefore, the statement should be understood to mean that the first analysis method is at least biased, tending, or pre-set to detect an error in uncertain situations or cases of doubt. For example, the first analysis method might be a (less powerful or less resource-intensive) method that is also based on machine learning and has been trained to avoid "false negatives," i.e., the erroneous exclusion of an error. Depending on the error and the analysis system, the statement that the first analysis method is designed to make reliable statements regarding the absence of an error can be interpreted in more specific ways.to assume the absence of the error, but also to guarantee absolute certainty.

[0021] A response to the detection of a fault may be planned, e.g., the tripping of a circuit breaker in the case of a series arc, or a network control action (switching on / off of supply lines / consumers, etc.) in the case of a detected network instability. The term "evaluating the measurement data" should not be understood to mean that exactly the same measurement data is evaluated for the first and second analysis methods. It is conceivable, for example, that only a subset of the measurement data used for the second analysis method is evaluated for the first method.

[0022] The statement that the first analysis method is designed for a less complex analysis compared to the second method refers to a lower average requirement for a resource, e.g. computing time or storage.

[0023] For the first analysis method, there are several alternatives, including a suitably trained neural network, a simple model such as a small decision tree, a support vector machine, or a low-complexity rule-based system. It can be 202415783

[0024] 4. It could also be a hybrid system, e.g. a formula-based system with parameter learning.

[0025] Depending on the initial analysis system, it may be able to make a statement about the presence of a fault with relatively high certainty in a subset of cases where a statement is made. For example, consider a formula-based system that uses threshold values ​​for parameters (e.g., current values, voltage values, RSSI (Received Signal Strength Indication) measurements, and their derivatives). This system can specify not only threshold values ​​for a potential fault but also stricter threshold values ​​that indicate a fault with greater reliability. In this case, these stricter threshold values ​​can function as a kind of reliability measure, and a fault can be assumed immediately (i.e., without performing the second analysis procedure) if this reliability measure meets a criterion (e.g., stricter threshold values ​​are exceeded by suitable parameters).Accordingly, further training includes the following:

[0026] - Establishing a reliability measure for the detection of the error by the first analysis procedure,

[0027] - Evaluating the measurement data (possibly a subset of the total measurement data used for the second analysis method) using the first analysis method to determine the error,

[0028] - Determining the reliability measure for the detection of the error by the first analysis method and

[0029] - Determining the error without evaluation using the second, machine learning-based analysis method, if the reliability measure meets a criterion.

[0030] According to one embodiment of the inventive method, the first analysis method is based on machine learning, and parameters (features) for analysis by machine learning are determined from the measurement data (either directly measured or calculated from the measurement data). In this embodiment, the first analysis method operates on the basis of fewer parameters than the second (preferably, the first analysis method uses a subset of the features from the second).

[0031] According to one embodiment of the inventive method, the first analysis method is based on machine learning, and a reduced training dataset was used for training the second analysis method. This dataset was reduced by removing the examples that had already been easily and correctly classified during the training of the first analysis method. 202415783

[0032] 5

[0033] According to one embodiment of the inventive method, measurement data are continuously acquired, assigned to (mostly consecutive) time intervals, and the method for detecting a fault (e.g., a serial arc fault) is carried out for each time interval based on the measurement data assigned to that interval. The method can be performed periodically, especially when the measurement data comes from an AC power network. In this case, for the purpose of dividing the measurement data into time intervals, the data can be assigned to half-waves of the current or voltage, or multiples thereof.

[0034] The invention also relates to a device configured to carry out a method according to the invention. This device can be a protective switch (e.g., an AFDD or a circuit breaker with an AFD function) that trips upon the detection of a fault. However, it is also conceivable, for example, that another low-voltage analysis device could operate accordingly (e.g., a control device such as a PMD (power monitoring / management device)).

[0035] The invention also relates to a computer program which performs a method according to the invention when it runs on a processor, as well as a computer program product with such a computer program.

[0036] The invention enables a method for training and evaluating an ML module (i.e., a module designed for machine learning) which can advantageously perform arc detection on the processor or microcontroller with reduced energy consumption.

[0037] This approach can advantageously achieve lower energy consumption by performing fewer calculations on input data that are relatively easy to identify as non-critical. In such cases, for example, the complex evaluation of a neural network can be omitted, and the processing unit designated for evaluation can be switched to a sleep or idle mode with lower energy consumption for the corresponding cycle.

[0038] Energy consumption during the evaluation of machine learning models can be reduced, for example, by using simpler algorithms with lower complexity (such as smaller neural networks) for recognition. However, this generally reduces the recognition rate, resulting in poorer performance.

[0039] 6

[0040] Accuracy is achieved. According to the invention, this can be avoided in a resource-saving manner by performing a more detailed analysis (only) in cases of doubt.

[0041] The invention will now be explained in more detail using an exemplary embodiment and the accompanying figures. These figures show...

[0042] Fig. 1: a schematic representation of a circuit breaker with arc fault detection,

[0043] Fig. 2: a schematic representation of the circuit breaker from Fig. 1 with further details,

[0044] Fig. 3: a procedure for generating training data for training a neural network for arc fault detection,

[0045] Fig. 4: a schematic representation of the architecture of a neural network trained for arc fault detection using training data obtained as shown in Fig. 3,

[0046] Fig. 5: Another procedure for generating training data for training a neural network for arc fault detection,

[0047] Fig. 6: another principal representation of the architecture of a neural network trained for arc fault detection with training data obtained as shown in Fig. 5,

[0048] Fig. 7: an ML-based arc fault detection system for an AC power grid, and

[0049] Fig. 8: an ML-based arc fault detection system according to the invention for an AC power network.

[0050] Fig. 1 shows a schematic representation of a circuit breaker. This circuit breaker comprised the functions of a typical miniature circuit breaker (MCB). Terminals 1 and 3 of a neutral conductor N and terminals 2 and 4 of a phase conductor L are shown. The circuit breaker is typically connected between a power source or supply and a load via these terminals. Both conductors include switching contacts 5, which are opened for protection purposes. A known mechanical mechanism 6 is implemented for separating the switching contacts, as shown in 202415783.

[0051] 7 the contacts 5 can be opened and closed again (e.g. by manually operating a button provided for this purpose on the housing of the switch).

[0052] The typical short-circuit and overcurrent protection functions of MCBs are implemented. In the event of a short circuit, an electromagnet 7 is energized, which then triggers the mechanism 6. Overcurrent protection is achieved by means of a bimetallic strip 8, which is connected to the current path by a flexible conductor 9. The system consisting of the bimetallic strip 8 and the flexible conductor 9 is calibrated so that in the event of an overcurrent, the bimetallic strip deforms, and this deformation triggers the mechanism 6. The switch also features arc-guiding elements 11, through which any arc that may occur during short-circuit tripping is directed into an arc-quenching chamber 10.

[0053] The circuit breaker shown in Figure 1 is additionally equipped with electronics 12. These enable various further functions, which are explained in more detail with reference to Figure 2.

[0054] Further information on a switch providing protective functions in accordance with the switch shown in Fig. 1 is contained in EP 3 358 594 A1.

[0055] For the following explanations, two points are therefore particularly relevant: a) The necessary hardware is available to perform the calculations required for arc fault detection. The software for this can be installed as firmware during the manufacture of the switch. b) Means are available to trigger the opening of a circuit monitored by the arc fault switch.

[0056] The means for opening the monitored circuit (typically isolating contacts and a mechanism acting on them) do not necessarily have to be provided in the same switching unit as the arc fault detection. DE 10 2012 204 964 A1 and DE 10 2014 226 221 disclose arc fault circuit interrupters (AFCIs) that do not include their own tripping means or isolating contacts, but are each coupled to another switch that receives a tripping signal upon arc fault detection by the AFCI. This switch is, for example, a miniature circuit breaker (MCB). Another type of switch with which AFCIs can be combined is the residual current device (RCD), i.e., a protective device for ensuring protection against a fault current or differential current in an electrical installation. 202415783

[0057] 8

[0058] The functions of an arc fault circuit interrupter (AFDD), a miniature circuit breaker (MCB), and a residual current device (RCD) can also be combined in one switch, cf. DE 102010 021 068 A1 (AFDD + MCB), DE 10 2016 223264 A1 (AFDD + RCD) and DE 10 2016218 960 A1 (AFDD + RCBO, where an RCBO (Residual current operated Circuit-Breaker with Overcurrent protection) combines the functions of an RCD and an MCB).

[0059] For the purposes of this application, the term arc fault detection device (AFDD) or fire protection switch is to be understood as encompassing both switches with and without their own isolating contacts, as well as switches with additional functions (e.g., AFDD + MCB, AFDD + RCD, AFDD + RCDO).

[0060] In the embodiment shown in Figures 1 and 2, the functions of an MCB and an AFDD are implemented in one device (AFDD + MCB). Figure 2 shows details of the electronics 12. These will be described below, with elements already described with reference to Figure 1 being discussed in more detail only to the extent necessary for explaining the AFDD function. The neutral conductor, which passes through the switch, is equipped with a sensor arrangement 16, from which a voltage signal is detected. Additionally, a current signal is detected via a shunt resistor 15 on the L conductor. Both signals are further processed by elements 21 and 22 via an ASIC 20, whereby a high-frequency signal (e.g., harmonics > 500 Hz) is extracted from the voltage signal. The signals processed by the ASIC 20 form the input for a block 30, which implements the actual AFDD function.A microcontroller 31 is provided for evaluating the input signals. This microcontroller includes a digital signal processor (DSP) 312, arc fault analysis software 313 (AFA), over voltage analysis software 314 (OVA), and a microcontroller chip 311. This chip performs the calculations for analyzing the input signals for the presence of an arc fault. A self-test component 38, an input element 37, a user interface (HMI) 35, an LED display 26, and release elements 33 and 34 are also connected via G P0 (General Purpose Input / Output) 32. In the event of an arc fault, a coil 33 is energized, which triggers a release element 34. The release element 34 then initiates the mechanism 6 for disconnecting the contacts 5.

[0061] A communication block 40 is connected to the AFDD block, meaning that information can be exchanged between the two blocks (arrow 45). The communication block 40 is formed with a microchip 41, by means of which a signal adapter 43 and a 202415783

[0062] 9

[0063] Antenna 42 enables wireless signal transmission and reception. A high-frequency crystal 47 and a GPIO 44 are also provided. The communication block 40 exchanges data with a consumption measurement block 50, meaning consumption data can be recorded and transmitted externally via the communication block 40.

[0064] The microcontroller 31 transmits the information acquired by sensors 15 and 16 for processing. According to the invention, this processing is carried out by using this information as input for a neural network that has been trained to assess information regarding the occurrence of an arc fault.

[0065] The neural network and its training are described in more detail below using Figures 3 to 6.

[0066] As shown in Figure 3, the starting point (step 60) is the use of measurement data, such as current values, voltage values, and RSSI values ​​(RSSI (Received Signal Strength Indication) values ​​are a measure of the relative quality of a signal received by a device). This measurement data forms the input for the analysis regarding the presence of an arc fault (step 70). A series of criteria (or features F) are used. The entirety of these criteria forms a vector (in block 80, this vector would have entries F1 to F4). For each individual vector (four vectors or rows are shown in block 80), there is also an assignment to an assessment of whether an arc fault is present or not. That is, each vector has a corresponding binary entry that includes a classification into two categories (arc fault, no arc fault). The vectors are then combined into a matrix.

[0067] For evaluating measurement data regarding the presence of an arc flash, numerous criteria or features F are described in the literature, which refer to the measured quantities or quantities derived from them (e.g., change, derivative, combination criteria, etc.). An example is US 5,729,145 A, cited in the introduction to the description. Further information on assessing the presence of an arc flash can be found in DE 102013 224 867 A1, US 6,459,273 BI, DE 102016 209443 A1, 102016209444 A1, and DE 102016209445 A1, as well as in the German application with application number 102020 201 887.1 and in the documents cited by these publications. Such procedures are applied in step 70. 202415783

[0068] 10

[0069] Figure 4 illustrates how a neural network is trained using the matrix input 81 created according to Figure 3. The neural network is schematically represented by three levels 82, 84, and 86, as well as feature embeddings shown between the levels. The training is focused on predicting arc faults 87. The figure employs a feed-forward neural network with a multi-layer perceptron architecture.

[0070] In deep learning, weight sharing can be used. This reduces the number of model parameters, which is particularly advantageous when using sequential models. Figures 5 and 6 show the preparation of the input data and the architecture of the neural network for sequential models, such as convolutional neural networks (CNNs) or long short-term memory networks (NSTMs). In Figure 5, raw data (current, voltage, and spectral current) is used again and employed for feature engineering to evaluate the occurrence of an electric arc (step 170). A sequence of matrices is then created from this data (step 181). This sequence of matrices is then used for a CNN network, as shown in Figure 6.

[0071] This process involves using a 1D CNN layer and feature embedding (Block 182), from which a 1D CNN layer or layers are derived (Step 183). Using Max Pooling, i.e., a strategy for downsampling in convolutional neural networks (CNNs), a hidden layer or layers are obtained. Based on this, statements can then be made regarding the presence of an arc fault (Step 186). The learning process can also involve transfer learning, where only specific weights are adjusted. This approach is useful, for example, when a network already trained for a particular application needs to be adapted for a different application (e.g., when different standards need to be met).

[0072] The neural network, trained as described above, is then loaded into the microcontroller 311 of the switch shown in Figure 2. Within the switch, the neural network then evaluates, based on the input measured by the sensor elements 15 and 16, whether an arc fault is present. In the event of an arc fault, the mechanism 6 is triggered to open the contacts via the coil 33 and the release unit 34. 202415783

[0073] 11

[0074] Fig. 7 shows a typical process of a detection system for arc fault detection. In the first step, temporal data from current, voltage, and noise information (RSSI) are divided into blocks. Such blocks can be determined, for example, by zero crossings and, for 50 Hz signals, are in the range of approximately 10 ms. In the second step, features or indicators are calculated for each block, condensing specific information such as minimum, mean, and maximum values ​​in the raw data. Here, for example, a total of 10–20 features are generated from 125 samples per sensor per block. In the third step, the condensed information is fed into a machine learning model, which, with a lower frequency or higher cycle time (advantageously, for example, approximately 10 ms) compared to the rate at which measurement data is sampled, determines whether an arc fault is present in the respective cycle. In the example of Fig.In cycles 0-10 there is no arc (section “no Arc”), while in cycles 11-13 there is an arc (section “Arc”).

[0075] In contrast, the processing steps of the invention disclosure by hierarchical combination of ML models are shown in Fig. 8 and can be described as follows:

[0076] 1. Segmentation of the raw data into blocks of half-waves

[0077] All signals are segmented into blocks (for example, based on zero crossings).

[0078] 2. Calculation of partial features for the detection of simple cases

[0079] To detect simple cases, a small number of features are calculated (for example, 'Maximum value of stream', 'Maximum value of RSSI').

[0080] 3. Detection of simple cases using ML model 1

[0081] A first machine learning model is evaluated to make decisions between the cases 1. "No Arc" / 2. "Possibly Arc". This model is preferably a simple model such as a small decision tree, a support vector machine, a small neural network, or a low-complexity rule-based system. In the case of the decision 'no Arc', no further computational steps are performed and the decision 'no Arc' is output. In the other case, in more complex situations, the model decides 'possibly Arc' and further processing steps are performed.

[0082] 4. Only in difficult cases: Calculation of remaining features

[0083] If the first ML model 1 outputs the decision "possibly Arc", further features for recognizing more complex data patterns are calculated. These features are numbered 202415783.

[0084] 12. More complex patterns in the data, such as specific patterns in the raw data, specific frequency components, form factors, or similar features, may be included or based on them.

[0085] 5. Only in difficult cases: Detection of complex cases using ML model 2. Another ML model 2 is evaluated with a larger number of potentially more complex features and produces a final decision as to whether an arc was present in the cycle under investigation.

[0086] Figure 8 illustrates a possible application of the proposed method. Simple cases are identified using a reduced set of features, leading to a "no arc" decision with minimal computational effort. The ML model 2 is then evaluated only for the remaining cases.

[0087] The training of ML models 1 and 2 is described in more detail below:

[0088] Training ML Model 1 :

[0089] An initial machine learning model is trained based on the entire dataset. The training dataset contains possible features and their associated labels (arc / no arc), which are typically defined by experts. The cost function for training the network includes the following deviations from conventional training:

[0090] 1. The training aims to minimize the number of false assessments that no arc is present ("false negatives") and maximize the number of correct assessments that no arc is present ("true negatives"). The number of "true positives" and "false positives" is less important. Such training can be achieved, for example, by appropriately and significantly different weightings of the cost terms "false positives" and "false negatives".

[0091] 2. The training aims to use the fewest possible features.

[0092] 3. The training aims for the lowest possible complexity of the ML model 1.

[0093] Training ML Model 2:

[0094] A second machine learning model is trained based on the entire or a reduced dataset. The reduced training dataset can be further narrowed by removing examples that were already easily and correctly classified by machine learning model 1.

[0095] 1. The training of ML model 2 aims for the most optimal decision possible. 202415783

[0096] 13

[0097] 2. The training can use the features of ML model 1 as well as other features, if applicable, as input or input variables.

[0098] 3. The training can also result in a ML model with higher complexity compared to ML model 1.

[0099] Predicting the presence of arcs can, on average, be performed with lower energy consumption by the computational system, since in practice many simple cases can be detected with a very basic machine learning model. For this purpose, a small set of easily computable signal features, such as mean or maximum values, and a learned decision rule may suffice.

[0100] The entire forecast can still be performed with optimal accuracy, as complex cases can be classified using a hierarchically downstream and more powerful ML model.

[0101] The invention is not limited to the example above. Other embodiments and modifications are immediately apparent to those skilled in the art. In particular, the first stage of the inspection can also be based on other, non-ML-based methods. The procedure is also applicable to other, non-arc-related defects.

Claims

202415783 14 Patent claims 1. Method for detecting a fault in a power grid based on measurement data, comprising - Evaluating the measurement data using an initial analysis procedure to determine the error, - Evaluating the measurement data using a second, machine learning-based analysis method, only if the first analysis method detects the presence of the error, and - Determining the fault if the second analysis procedure determines the presence of the fault, whereby - the first analysis method is designed for a less complex analysis compared to the second method, and - the first analysis method is designed to make reliable statements regarding the absence of the error.

2. Method for detecting a fault in a power network according to claim 1, comprising - Establishing a reliability measure for the detection of the error by the first analysis procedure, - Evaluating the measurement data using the first analysis method to determine the error, - Determining the reliability measure for the detection of the error by the first analysis procedure, and - Determining the error without evaluation using the second, machine learning-based analysis method, if the reliability measure meets a criterion.

3. Method according to claim 1 or 2, characterized in that the fault is caused by the presence of an arc flash.

4. Method according to one of the preceding claims characterized in that - the first analysis method based on machine learning, and - was trained to reduce the number of errors where an arc fault is not detected.

5. Method according to claim 4, characterized in that 202415783 15 - for the analysis methods, parameters for analysis using machine learning are determined from the measurement data, and the first analysis method operates on the basis of fewer parameters than the second.

6. Method according to claim 4 or 5, characterized in that training data used for training the first machine learning-based analysis method were used for training the second machine learning-based analysis method.

7. Method according to one of the preceding claims, characterized in that - the first analysis method is formula-based, and - at least one parameter used in a formula was chosen to reduce the number of faults where an arc fault is not detected.

8. Method according to one of the preceding claims, characterized in that the first analysis method operates by means of a decision tree, a support vector machine, a neural network or a rule-based system.

9. Method according to one of the preceding claims, characterized in that - continuous measurement data is collected, - the measurement data are assigned to time intervals, and - for the individual time intervals, the method according to claims 1-8 is carried out to detect a fault based on measurement data assigned to the respective time interval.

10. Method according to claim 9, characterized in that - it concerns measurement data from an alternating current network, and - for the purpose of dividing the measurement data into time intervals, half-waves of the current or voltage or multiples thereof are assigned. 202415783 16 11. Device which is configured to carry out a method according to any one of claims 1 to 10.

12. Device according to claim 11, characterized in that - that the device is a circuit breaker, and - the circuit breaker is designed to trip when a fault is detected by the procedure.

13. Computer program product, comprising a computer program which performs a method according to any one of claims 1 to 10 when running on a processor.

14. Computer program which performs a method according to any one of claims 1 to 10 when running on a processor.

Images