A method for detecting low voltage ac series arc faults
By collecting line current data, calculating the difference between adjacent currents, and combining it with a 1DCNN model, the problems of low accuracy and large computational load in low-voltage AC series arc fault detection are solved, and efficient arc fault detection under multi-load conditions is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- 国网山东省电力公司日照供电公司
- Filing Date
- 2022-09-22
- Publication Date
- 2026-07-07
AI Technical Summary
Existing technologies are difficult to efficiently identify low-voltage AC series arc faults, especially in multi-load lines where accuracy is low, computation is large, and hardware requirements are high.
By collecting line current data and calculating the difference between adjacent wave currents, the occurrence and start period of faults are determined using dual thresholds. Fault identification is then performed using a one-dimensional convolutional neural network (1DCNN), reducing computational load and hardware requirements.
It achieves high-accuracy arc fault detection under different load conditions, reduces computational load and hardware requirements, and has strong adaptability and resistance to load disturbances.
Smart Images
Figure CN115598475B_ABST
Abstract
Description
Technical Field
[0001] A method for detecting low-voltage AC series arc faults, belonging to the field of power line fault detection. Background Technology
[0002] Arc faults in low-voltage AC power lines, caused by conductor breakage, loose terminals, or insulation aging and cracking, are prone to generating high temperatures at the fault point, potentially leading to electrical fires. Series arc faults are difficult to identify due to their small fault current. Furthermore, the electrical characteristics of series arc faults are easily affected by line load characteristics, exhibiting randomness, making detection a current challenge.
[0003] Existing methods for identifying series arc faults primarily rely on recognizing line voltage and current fault characteristics. Due to the uncertainty of the arc's location, methods based on fault voltage characteristics require multiple data acquisition points, making field application difficult. Current research on series arc identification methods often utilizes current distortion characteristics such as fault current similarity, "shoulder-to-shoulder zero-rest," frequency domain variations, and arc randomness. However, frequency domain and "zero-rest" characteristics are still inevitably affected by non-faulty branch currents, nonlinear load characteristics, and arc randomness, reducing reliability in multi-load lines.
[0004] During a series arc fault, the fault current exhibits randomness, a characteristic that can distinguish arc faults from normal load operation. However, the amplitude of the arc's randomness fluctuation within the current difference is significant. Detection methods based solely on a single-cycle randomness index have low computational complexity and are relatively easy to implement, but their accuracy is unsatisfactory. Existing technologies also include some solutions for AC series arc fault detection, such as using convolutional neural networks (CNNs). However, these CNN-based fault detection solutions involve extremely high computational demands and require sophisticated hardware. Summary of the Invention
[0005] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a method for detecting low-voltage AC series arc faults. This method utilizes the characteristic that the adjacent wave current difference exhibits two significant fluctuations before and after the arc fault to initially select the fault occurrence and fault initiation period, and uses a neural network to discriminate the period after the fault initiation period. This method reduces the computational load and hardware requirements while ensuring detection accuracy.
[0006] The technical solution adopted by this invention to solve its technical problem is: a method for detecting low-voltage AC series arc faults, characterized by comprising the following steps:
[0007] Step a: Collect current data in the line;
[0008] Step b: Calculate the difference between adjacent wave currents;
[0009] Step c: Determine whether the decision-making start condition is met. If the decision-making start condition is met, proceed to step d. If the decision-making start condition is not met, return to step a.
[0010] Step d: Determine the starting period for waveform recognition;
[0011] Step e: Determine the series arc fault state for each cycle using the trained neural network model;
[0012] Step f: Determine whether a series arc fault is detected using the trained neural network model. If a series arc fault is detected, proceed to step g; otherwise, return to step a.
[0013] Step g: Record and alarm the fault cycle of the identified series arc fault.
[0014] Preferably, in step c, the absolute value change of the difference between adjacent current sampling points during a single cycle is used as the index k, and a threshold k is set simultaneously. set.1 When k <k set.1 At this time, a series arc fault may occur in the line, indicating that the judgment start condition is met.
[0015] Preferably, the formula for calculating the index k is:
[0016]
[0017] Where: Δi a x Δi represents the current difference at sampling point a in period x. b x+1 This represents the current difference at the b-th sampling point in the (x+1)-th cycle, where n is the number of sampling points per cycle.
[0018] Preferably, in step d, the method for determining the waveform recognition start period is: setting a threshold k set.2 k set.2 >k set.1 First, the initial waveform is selected to identify the starting period: k is determined. <k set.1 The period is the initial period. If k > k occurs within N waiting periods after the initial period... set.2 If k>k is not found within N waiting periods after the initial period, then this period is defined as the identification start period. set.2 When the start-up conditions are met, the third cycle after which the start-up period is set as the starting cycle.
[0019] Preferably, the k set.1 Set to 0.7–0.9, where k set.2Set it to 1.4~1.6.
[0020] Preferably, the k set.1 The value of k is 0.8. set.2 It is 1.5.
[0021] Preferably, the waiting period N is set to 8.
[0022] Preferably, when performing step e, the neural network model detection is terminated when there are no faults detected for five consecutive cycles.
[0023] Preferably, before performing step e, normalized waveform data of adjacent current difference in a single cycle is used as training data to obtain an arc random waveform recognition model. The arc random waveform recognition model uses a network with two convolutional layers plus one pooling layer to extract data features, and a three-layer fully connected network structure to classify the extracted fault features.
[0024] Compared with the prior art, the beneficial effects of this invention are:
[0025] In this method for detecting low-voltage AC series arc faults, the method utilizes the characteristic that the adjacent current difference fluctuates twice before and after the arc fault to initially select the fault occurrence and fault initiation period. This reduces the computational load and hardware requirements while ensuring detection accuracy.
[0026] In this method for detecting low-voltage AC series arc faults, a fault identification algorithm is proposed based on the numerical and waveform variation characteristics of the adjacent current difference, and an experimental platform is built to verify it. The two significant fluctuations in the adjacent current difference before and after the arc fault can be used for initial selection of the fault occurrence and fault initiation periods. The arc randomness characteristics of the normalized waveform of the adjacent current difference in series arc faults still exhibit a relatively uniform time-domain distribution under different loads and air gap spacing conditions. Based on this, this method selects a 1DCNN to construct the recognizer to achieve the determination of series arc faults after the initial selection.
[0027] This method for detecting low-voltage AC series arc faults not only boasts high accuracy for single loads but also demonstrates strong identification capabilities for branch faults in untrained multi-load circuits, and exhibits immunity to load switching disturbances. Compared to other series arc fault identification methods, this method is less affected by load characteristics, achieves higher accuracy, and demonstrates good adaptability to different load types and combinations.
[0028] Convolutional Neural Networks (CNNs) are a type of deep learning intelligent algorithm that can effectively improve the accuracy of complex classification problems and are widely used in image, speech, and text recognition. One-dimensional Convolutional Neural Networks (1DCNNs) exhibit good invariance to waveform translation, scaling, and distortion, and are more suitable for time-series data. This network can be used to identify load types, improving the accuracy of single-class load classification and enhancing the ability to classify mixed multi-class loads that have not been trained.
[0029] The model detection can only be terminated after five consecutive cycles of fault-free detection, in order to reduce the impact of a small number of false detection cycles. Attached Figure Description
[0030] Figure 1 This is a flowchart of a method for detecting low-voltage AC series arc faults. Detailed Implementation
[0031] Figure 1 This is the preferred embodiment of the present invention, which is described below in conjunction with the accompanying drawings. Figure 1 The present invention will be further described below.
[0032] like Figure 1 As shown, a method for detecting low-voltage AC series arc faults includes the following steps:
[0033] Step 1001: Collect current data in the line;
[0034] Current data in the line is collected by current transformers installed in the line.
[0035] Step 1002: Calculate the difference between adjacent wave currents;
[0036] When the power supply and load are stable and fault-free in each cycle, the current fluctuation between adjacent cycles is negligible. The adjacent current difference when the load is operating normally contains only a noise component. When a series arc fault exists, the adjacent current difference is composed of the current variation between cycles caused by noise and the randomness of the arc. When a series arc fault occurs suddenly, the line parameters change significantly before and after the fault, resulting in a non-negligible current fluctuation between cycles. The adjacent current difference is composed of the current fluctuation caused by the fault and noise. By identifying the state changes of the adjacent current difference, the series arc can be identified.
[0037] Step 1003: Does the start condition for determination meet?
[0038] Determine whether the start-up conditions are met. If the start-up conditions are met, proceed to step 1004. If the start-up conditions are not met, return to step 1001.
[0039] In this method for detecting low-voltage AC series arc faults, the starting condition is determined by the absolute value change of the difference between adjacent current sampling points during a single cycle, as index k.
[0040]
[0041] Where: Δi a x Δi represents the current difference at sampling point a in period x. b x+1 This represents the current difference at the b-th sampling point in the (x+1)-th cycle, where n is the number of sampling points per cycle.
[0042] Set two thresholds: k set.1 k set.2 . (The last part is incomplete and likely refers to a different context.) set.1 In comparison, when k <k set.1 At this time, a series arc fault may occur in the line, indicating that the judgment start condition is met and the fault identification is initiated.
[0043] Step 1004: Determine the waveform recognition start period;
[0044] The method for determining the starting period of waveform recognition is as follows: First, initially select the starting period of waveform recognition: determine if k < k set.1 The period is the initial period. If k > k occurs within N waiting periods after the initial period... set.2 If k > k is not found within N waiting periods after the initial period, then this period is designated as the identification start period. set.2 When the start-up conditions are met, the third cycle after which the start-up period is set as the starting cycle.
[0045] When the load remains constant and no faults occur, k will not exceed the limit. This part can eliminate a large number of cycles, reducing the amount of calculation. Detecting sudden faults in k set.1 Set to 0.7–0.9, preferably 0.8; k set.2 The value is set to 1.4 to 1.6, preferably 1.5. To reduce the interference of the k value exceeding the limit during load switching on and off, which may affect fault identification, the waiting period N is set to 8 to accommodate situations where some long-running high-power loads require multiple cycles to complete the startup.
[0046] Step 1005: Perform waveform processing on the current waveform;
[0047] Further processing is performed on the subsequent cycle data of the initial fault cycle. Mean filtering is applied to the single-cycle adjacent-wave absolute difference current to reduce sampling glitches, and the data is normalized and added to the dimension for input into the model for detection.
[0048] Step 1006: Fault identification is performed using a neural network;
[0049] Before using a neural network model to identify faults, it is necessary to define the structure of the arc random waveform identification model and train it to obtain the model.
[0050] The model training uses normalized waveform data of adjacent-wave current difference in a single cycle. Data is categorized and labeled according to the fault and normal states of different types of loads before training. The trained 1DCNN model identifies the randomness characteristics of the arc in the normalized absolute difference of adjacent waves, determines the series arc fault state in each cycle, and eliminates false starts caused by load switching.
[0051] The structure of the random waveform recognition model for electric arc is set as a nine-layer structure as shown in Table 1: data feature extraction is performed by a network with two convolutional layers plus one pooling layer, and the extracted fault features are classified by a three-layer fully connected network structure.
[0052] Table 1 1DCNN Layers and Layer Parameters
[0053] level name Hierarchical settings Output data Convolutional layer 1 32*8，4，4 32*501 Convolutional layer 2 2*8，4，4 64*126 Max pooling layer 1 2，2 64*63 Convolutional layer 3 2*4，2，2 128*32 Convolutional layer 4 2*4，2，2 256*17 Max pooling layer 2 2，2 256*8 Fully connected layer 1 2056*600 600 Fully connected layer 2 600*100 100 Fully connected layer 3 100*6 6 loss function CrossEntropyLoss /
[0054] Table 1 shows the convolutional layer parameters as kernel number, kernel length, stride, and boundary padding; the pooling layer parameters are kernel size and stride. Based on the characteristics of waveform fault features, numerous local features, and long time spans of the feature intervals in the absolute difference between adjacent waveforms, the bottom convolutional layers are configured with the most parameters, decreasing sequentially in subsequent layers. Max pooling is used to extract prominent waveform features. The features after convolution and pooling are 256-dimensional data sets of length 8. These data sets are then processed through a three-layer fully connected network to obtain the probability of their classification, with the label of the highest probability being used as the classification result.
[0055] This method for detecting low-voltage AC series arc faults can be developed using Python, employing a 1DCNN model to identify the randomness of series arcs. The 1DCNN model is trained using the PyTorch deep learning framework with a learning rate of 0.0005 and an adaptive moment estimation gradient descent algorithm (Adam) as the iterator. Once training is complete, this method for detecting low-voltage AC series arc faults can begin identifying such faults.
[0056] Step 1007: Has a series arc fault been detected?
[0057] The trained 1DCNN model is used to determine whether a series arc fault is detected. If a series arc fault is detected, step 1008 is executed. If no series arc fault is detected, the process returns to step 1001. The model detection is terminated only after five consecutive cycles of no fault detection, in order to reduce the impact of a small number of false detections.
[0058] Step 1008: Record the series arc fault;
[0059] The fault cycle of the series arc fault identified by the 1DCNN model is recorded and an alarm is triggered.
[0060] The following example will be used to verify the detection method for low-voltage AC series arc faults:
[0061] 1. First, an experimental platform was built according to the national standard GB 14287.4-2014 to simulate a series arc fault in a low-voltage line. The platform consists of a 220V AC power supply, an arc fault generator, loads, and a waveform recorder, with a data sampling rate of 100kbps. The arc generator has an 8mm graphite rod as the stationary contact and a 10mm copper rod as the moving contact. After the arc is stably controlled, current data is collected. This paper uses several common household loads for experiments. Based on the waveform fault characteristics, the samples are labeled according to the method of classifying different types of loads under normal conditions into one category and loads with the same number of internal combustion arc cycles into another category. The information and labels of each load are shown in Table 2.
[0062] Table 2: Load Types, Parameters, and Label Classification
[0063] Load type Load Name Power (W) Normal label Fault Label Resistive (low power) Incandescent lamp 120 4 1 Resistive (high power) electric heater 800 4 1 Resistance (eddy current) induction cooker 2100 3 1 Resistive inductance (motor) Vacuum cleaner 1300 3 1 Switching power supply computer 230 2 1 Rectifier electric heater 400 4 0
[0064] 2. The testing process is shown in Table 4:
[0065] Table 3 Single-load recognition accuracy
[0066] Load and its type Label accuracy % Fault identification accuracy % Incandescent lamp (resistive) 99.66 100 Electric heater (resistive) 99.66 99.83 Induction cooker (resistive) 100 100 Vacuum cleaner (resistive) 100 100 Computer (Nonlinear) 100 100 Electric heater (rectified half-wave) 99.31 100
[0067] As shown in Table 3, the model achieved an accuracy of over 99% on the test set for the training load. The final overall accuracy stabilized at 99.87%.
[0068] The model achieves a fault identification rate of over 98% and a label recognition accuracy of over 90% for a certain type of untrained load combination. The model exhibits good adaptability to parameter variations within a certain type of load, as shown in Table 4.
[0069] Table 4 shows the recognition accuracy for a certain type of untrained load combination.
[0070] Post-table status Load Combination Accuracy A% Normal (resistance) Incandescent bulb + electric heater 100 Normal (resistance) Induction cooker + vacuum cleaner 99.31 Normal (non-linear) Computer A + Computer B 98.61 Fault (obstruction) Incandescent bulb + electric heater 99.65 Fault (resistance / inductance) Induction cooker + vacuum cleaner 100 Fault (nonlinear) Computer A + Computer B 100
[0071] 3. Testing of multiple parallel branches under untrained conditions
[0072] Table 5 shows the test results for determining faults in untrained multi-load line branches based on trunk line data:
[0073] Table 5 shows the recognition accuracy for untrained multi-class load combinations.
[0074]
[0075] When no series arc fault occurs and the branch load power is not significantly different, the model's accuracy in label identification is above 99%, and its accuracy in identifying normal conditions is 100%.
[0076] When a series arc fault occurs in a branch, this low-voltage AC series arc fault detection method detects that the main circuit current k value exceeds the limit, thus completing the initial fault selection. Except for resistive fault branches, the identification accuracy rate for each combination where the total load power of the normal branch is approximately 4 times that of the faulty branch is above 90%.
[0077] The accuracy of load startup process identification is shown in Table 6:
[0078] Table 6. Accuracy of load startup process identification
[0079] Normal operating load Startup load Algorithm startup % Accuracy A% Incandescent lamp computer 100 98.13 Incandescent lamp Rectifier electric heater 100 96.81 Incandescent lamp Incandescent lamp 100 96.98 electric heater Vacuum cleaner 100 93.89 electric heater induction cooker 100 94.72 electric heater electric heater 100 95.69
[0080] As shown in Table 6, when a new load is applied to the line, this low-voltage AC series arc fault detection method detects changes in downstream line parameters and initially selects the fault cycle by monitoring threshold exceedances. Using a 1DCNN model, this low-voltage AC series arc fault detection method correctly identifies over 93% of the cycles.
[0081] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.
Claims
1. A method for detecting low-voltage AC series arc faults, characterized in that: Includes the following steps: Step a: Collect current data in the line; Step b: Calculate the difference between adjacent wave currents; Step c: Determine whether the decision-making start condition is met. If the decision-making start condition is met, proceed to step d. If the decision-making start condition is not met, return to step a. Step d: Determine the starting period for waveform recognition; Step e: Determine the series arc fault state for each cycle using the trained neural network model; Step f: Determine whether a series arc fault is detected using the trained neural network model. If a series arc fault is detected, proceed to step g; otherwise, return to step a. Step g: Record and alarm the fault cycle of the identified series arc fault; In step c, the absolute value of the change in the difference between adjacent current sampling points during a single cycle is used as the indicator. k At the same time, a threshold k is set. set.1 ,when k < k set.1 At this time, a series arc fault may occur in the line, indicating that the judgment start condition is met; In step d, the method for determining the waveform recognition start period is as follows: set a threshold k. set.2 k set.2 >k set.1 First, the initial waveform identification starts with the period: This will determine... k < k set.1 The period of time is the starting period. If it is after the starting period... N Within a waiting period k > k set.2 If the period is after the initial period, then that period is defined as the identification start period. N No occurrence within the waiting period k > k set.2 When the conditions for starting the identification are met, the third cycle after which the identification start cycle is set as the identification start cycle.
2. The method for detecting low-voltage AC series arc faults according to claim 1, characterized in that: The aforementioned indicators k The calculation formula is: in: Indicates the first x Cycle number a Current difference at sampling points Indicates the first x+ 1st cycle b Current difference at sampling points n This represents the number of sampling points per cycle.
3. The method for detecting low-voltage AC series arc faults according to claim 1, characterized in that: The k set.1 Set to 0.7~0.9, where k set.2 Set it to 1.4~1.
6.
4. The method for detecting low-voltage AC series arc faults according to claim 3, characterized in that: The k set.1 The value of k is 0.
8. set.2 It is 1.
5.
5. The method for detecting low-voltage AC series arc faults according to claim 1, characterized in that: The waiting period N Set it to 8.
6. The method for detecting low-voltage AC series arc faults according to claim 1, characterized in that: When performing step e, the neural network model detection is terminated when there are no faults detected for five consecutive cycles.
7. The method for detecting low-voltage AC series arc faults according to claim 1, characterized in that: Before performing step e, normalized waveform data of adjacent current difference in a single cycle is used as training data to obtain an arc random waveform recognition model. The arc random waveform recognition model uses a network with two convolutional layers plus one pooling layer to extract data features, and a three-layer fully connected network structure to classify the extracted fault features.