Fault identification method and device for small current grounding system

Abstract

The present application relates to a kind of small current grounding system fault identification method and device.Small current grounding system fault identification method includes model training method and model utilization method;Model training method is: using simulation system to simulate each type of fault of small current grounding system, obtain the required number of simulation waveform data reflecting each type of fault condition as training case;Using training case to train, adjust and test deep learning network, obtain the deep learning network for fault identification;Model utilization method is: obtaining the real fault waveform of small current grounding system, using the deep learning network for fault identification and obtaining fault identification result.Small current grounding system fault identification device, it includes embedded device, starting element, fault confirmation element.The present application uses single algorithm, can simply, efficiently to small current grounding system for fault identification, and the accuracy of fault identification is higher.

Description

Technical Field

[0001] This invention relates to the field of power system fault identification technology, specifically to a method and apparatus for fault identification in a low-current grounding system. Background Technology

[0002] Low-current grounding systems refer to systems in medium-voltage power distribution networks where the neutral point of the power transformer uses one of the following three grounding methods: ungrounded, grounded through an arc suppression coil, or impedance (resistance or reactance, or a combination of both) grounding. There is also a classification method based on the magnitude of the grounding current, where low-current grounding systems refer to systems where the single-phase grounding arc can be extinguished instantaneously (below 10A). Because the grounding current is too small to accurately determine the fault line solely based on the current threshold or the phase relationship between current and voltage, a dedicated low-current grounding fault location device is essential.

[0003] However, current low-current grounding fault location devices generally have two major shortcomings. First, there are too many algorithms. Multiple algorithms are often integrated into the same device, and each algorithm is given different weights under different circumstances, so there is bound to be some trade-offs in the selection process. Second, the accuracy of fault location needs to be improved. Once the fault location fails, the on-duty personnel need to manually select the fault location again based on their experience, which expands the scope of unnecessary power outages. Summary of the Invention

[0004] The purpose of this invention is to provide a method for fault identification in low-current grounding systems that simplifies the algorithm and improves the accuracy.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A fault identification method for a low-current grounding system includes a model training method and a model utilization method. The model training method involves simulating various faults in the low-current grounding system using a simulation system to obtain a required number of simulated waveform data reflecting various fault conditions as training cases. The training cases are then used to train, adjust, and test a deep learning network to obtain a deep learning network for fault identification. The model utilization method involves determining whether a fault identification activation condition has been triggered. If so, the actual fault waveform of the low-current grounding system is obtained, and the fault identification is performed using the deep learning network for fault identification to obtain the fault identification result.

[0007] In the model utilization method, whether the residual current reaches a preset current threshold is used as the fault identification activation condition.

[0008] In the model utilization method, after fault identification, whether the residual voltage reaches a preset voltage setting value is used as the condition for whether to output the fault identification result.

[0009] In the model utilization method, after fault identification and obtaining the fault identification result, the system pauses the determination of whether the fault identification start condition is triggered within a preset delay period.

[0010] In the model training method described above, multiple simulation software programs are used in conjunction to perform simulation.

[0011] In the model training method, the training cases include training cases for training the deep learning network, adjustment cases for checking and adjusting the deep learning network, and test cases for testing the deep learning network. The ratio of the number of training cases, adjustment cases, and test cases is 7:1:2.

[0012] The simulated waveform data is fault current and voltage waveform data based on fault transients.

[0013] The deep learning network employs a convolutional neural network comprising a convolutional layer, a first fully connected layer, a second fully connected layer, an activation layer, and an output layer.

[0014] Due to the application of the above technical solutions, the present invention has the following advantages compared with the prior art: the fault identification method for low current grounding systems of the present invention adopts a single algorithm, which can easily and efficiently identify faults in low current grounding systems, and the accuracy of fault identification is high.

[0015] This invention also provides a simple and accurate fault identification device for low-current grounding systems, the solution of which is as follows:

[0016] A fault identification device for a low-current grounding system is provided for implementing the aforementioned fault identification method for a low-current grounding system. The fault identification device for a low-current grounding system includes:

[0017] An embedded device is provided, in which the deep learning network for fault identification is deployed. The embedded device is used to acquire real fault waveforms of a low-current grounding system, and uses the deep learning network for fault identification to perform fault identification and obtain fault identification results. The deep learning network for fault identification is obtained using a model training method, which involves simulating various faults in the low-current grounding system using a simulation system to obtain a required number of simulated waveform data reflecting various fault conditions as training cases. The deep learning network is then trained, adjusted, and tested using the training cases to obtain the deep learning network for fault identification.

[0018] A startup element is connected to the embedded device. The startup element is used to determine whether the fault identification startup condition is triggered. If so, the startup element outputs a signal to start the embedded device.

[0019] The fault identification device for the low-current grounding system also includes:

[0020] A fault confirmation element is connected to the embedded device. The fault confirmation element is used to determine whether the residual voltage reaches a preset voltage setting value. If so, the fault confirmation element outputs a signal to cause the embedded device to output the fault identification result.

[0021] Due to the application of the above technical solutions, the present invention has the following advantages compared with the prior art: the fault identification device for low current grounding system of the present invention has a simple structure, is easy to use, and has good effect. Attached Figure Description

[0022] Appendix Figure 1 This is a schematic diagram of the structure of a deep learning network.

[0023] Appendix Figure 2 This is a timing diagram of the operation of a fault identification device for a low-current grounding system. Detailed Implementation

[0024] The present invention will be further described below with reference to the embodiments shown in the accompanying drawings.

[0025] Example 1: A fault identification method for a low-current grounding system, comprising two parts: a model training method and a model utilization method.

[0026] The model training method is as follows: use a simulation system to simulate various faults in a low-current grounding system to obtain the required number of simulation waveform data reflecting various fault conditions as training cases; use the training cases to train, adjust and test the deep learning network to obtain a deep learning network for fault identification.

[0027] Deep learning is a subfield of machine learning, which in turn is a subfield of artificial intelligence. Deep learning is developed based on neural networks. It draws on the experience of neural networks from the 1980s, making improvements in areas such as weight initialization, non-linearity (activation functions), and backpropagation. A key reason for the improved performance of deep learning is the advancement of computer hardware, which allows for the use of more layered networks for gradient descent—hence the name "deep" learning. Another reason for its widespread adoption is the large-scale deployment of sensors (especially cameras), making data collection much easier today than it was decades ago. In short, the two defining characteristics that distinguish deep learning from previous neural networks are: massive amounts of training data and multi-layered networks.

[0028] The application of deep learning in the relay protection industry, especially in low-current ground fault location, is due to the following reasons: 1) Deep learning can help improve accuracy: Compared with other protection algorithms, the accuracy of low-current ground fault location still needs improvement and there is still considerable room for enhancement; 2) It is suitable for automatic rule extraction from massive amounts of data: For transient waveform recognition, deep learning can automatically find suitable parameters from massive amounts of waveforms, eliminating the need for engineers to search for and compare formulas and parameters. As long as there are sufficient computing resources (computer power + computing time), it can definitely find an expression suitable for all use cases to distinguish between faults inside and outside the fault zone; 3) Low-current ground fault location does not have high requirements for speed: The massive calculations of artificial intelligence may require a long processing time on embedded devices. The application of low-current ground fault location requires accuracy, not speed. Artificial intelligence algorithms can provide a judgment result after sufficient calculation; 4) Low-current ground fault location is essentially a form of image recognition: All current low-current ground fault location algorithms in the industry are based on the fault transient current and voltage waveforms for judgment, which is essentially a form of waveform recognition and belongs to the field of image recognition.

[0029] Deep learning can be categorized into supervised learning and unsupervised learning. This solution uses supervised learning, which requires manual annotation of the data. The learning network compares its learning results with the manually annotated results and then adjusts its parameters until the accuracy reaches the set target.

[0030] In the above model training method, multiple simulation software programs are used collaboratively to obtain a large number of training cases. The obtained simulation waveform data are fault current and voltage waveform data based on fault transients. For the requirement of massive amounts of data, simulation is the only feasible method. This is because the number of waveforms from real-world experiments and actual faults is too small to reflect the system's variability and cannot cover all situations, such as fault type, transition impedance, capacitance current magnitude, fault occurrence at different distances, and voltage phase angle at the time of fault occurrence. Furthermore, even some actually recorded waveforms are often unusable as training cases due to unclear system and fault descriptions, unsuitable channels, etc. To ensure that the simulation results are as close to reality as possible, EMTP / ATP simulation software is used as the primary simulation method, supplemented by RelaySimTest as a secondary simulation software, and verified using waveforms from the RTDS laboratory and data from real-world experiments on the Finnish power grid. The distribution network system model used in the simulation is selected or designed as needed. To ensure that the simulation results cover as many situations as possible, the parameters of the simulation network need to be adjusted as much as possible to cover all representative and extreme cases. A co-simulation platform combining MATLAB and EMTP / ATP can be used. Network parameters can be modified in batches using MATLAB, and simulations can be run via the EMTP / ATP simulation interface. The simulated waveforms can then be imported into MATLAB in batches and organized into a CSV table for use by Python's training algorithm. Since the fault location is known during simulation, waveforms can be marked as in-zone / out-of-zone in the CSV table. For the simulation of arc grounding, a voltage-controlled switch in EMTP / ATP is used for simplification. The detection time window is selected during the fault transient state because, in a low-current grounding system, the most valuable information occurs at the very beginning of the fault, during the instantaneous charging and discharging process of the line through the three-phase-to-ground capacitance. Once the fault enters a steady state, the measurable current is quite small, and due to the complex operating conditions of the power system, it may no longer be of value. The next step is to control the generation of the arc using a voltage-controlled switch in EMTP / ATP and extinguish the arc with a chopping current. V-fl is the arcing voltage, set to 0.9 times the rated value of the phase voltage. The arc begins to conduct when certain conditions are met. After the minimum burning time T-de, the arc extinguishes when the absolute value of the current is less than the cutoff current Imar. Conduction resumes when the voltage reaches V-fl. The cutoff current Imar is set to 10 amperes, and the minimum burning time T-de is set to 3 ms after comparison with waveforms recorded on-site.

[0031] The above EMTP / ATP and MATLAB co-simulation generated a dataset of 1.8 million faults for use as training cases, with in-zone and out-of-zone faults each accounting for 50%. The results are stored in a CSV table. The training cases include training cases for training the deep learning network, tuning cases for checking and adjusting the deep learning network, and test cases for testing the deep learning network, with a ratio of 7:1:2.

[0032] 70% of the current and voltage waveforms are used as training cases to input into the training network, which then begins adjusting according to the initialized parameters. After each adjustment, the network determines whether the current is positive or negative based on the existing parameter values, and then adjusts the parameters accordingly based on the actual situation. After each round of parameter adjustment, the accuracy is checked on 10% of the adjusted cases. If the result is unsatisfactory, the network is asked to perform a new round of adjustment until the accuracy on the adjusted cases is satisfactory. At this point, the trained neural network and parameters are used to judge 20% of the test cases. This part of the data is data that the network has never seen before. If the network can correctly judge this part of the data, it is considered that the algorithm has found a sufficiently accurate and general pattern, rather than simply forcibly "memorizing" previous waveforms.

[0033] As attached Figure 1 As shown, the deep learning network employs a convolutional neural network comprising convolutional layers, a first fully connected layer, a second fully connected layer, an activation layer, and an output layer. The convolutional layers extract features such as shape, current / voltage magnitude, and relative timing. The fully connected layers make hypotheses about the extracted features, attempting to extract features that influence the results. The activation layers perform non-linear processing on the numerical calculations of the fully connected layers, preventing excessive weighting of any very large value. The output layer normalizes the results of the activation layers to a vector space for distinguishing between in-region and out-of-region faults. Since the algorithm's primary application is waveform image recognition, the convolutional neural network layer, best suited for image recognition, is chosen. The convolutional layer can include n convolutional kernels: kernel 1, kernel 2, ..., kernel n. Experimental observations and consideration of computational complexity demonstrate that a single convolutional layer is sufficient to perfectly complete this one-dimensional waveform image recognition task. The fully connected layer after the convolutional layer is a standard neural network layer. Similarly, in order to control the amount of computation and parameters, the number of layers and nodes is chosen to be as small as possible. The batch size and learning rate are selected with appropriate values ​​based on the rate of loss descent and the test results on the validation set.

[0034] The following is a detailed example. Figure 1The grounding fault detection neural network training process shown, and the process of using the trained grounding fault detection neural network to calculate the label vector V corresponding to the current signal I.

[0035] For ease of explanation, the following settings are made for this example: the sampling resolution is k = 48, the sampling range each time is N = 1 cycle, the number of convolutional kernels included in the convolutional layer of the grounding fault direction detection neural network is 2, that is, n = 2, the sizes of convolutional kernels 1 and 2 are 5×5, the cross-entropy loss function L = -(ylogy^+(1 - y)log(1 - y^)) is used for handling the loss function, and the handling loss threshold T L is set to 0.01, and the activation function σ(z)=1 / ((1 + e^(-z))) is used in the activation layer.

[0036] For the out-of-zone fault current signal sample, its true label vector is [1, 0], indicating that the probability that this sample is an out-of-zone fault current signal is 1, and the probability that it is an in-zone fault current signal is 0; for the in-zone fault current signal sample, its true label vector should be [0, 1], indicating that the probability that this sample is an out-of-zone fault current signal is 0, and the probability that it is an in-zone fault current signal is 1.

[0037] The training sample M1 is provided to the grounding fault direction detection neural network. It is known that M1 is an out-of-zone fault current signal sample, and its true label vector is [1, 0].

[0038] The grounding fault direction detection neural network outputs the label vector y1 corresponding to M1, for example, [0.4, 0.6]. That is to say, at this time, the grounding fault direction detection neural network believes that the probability that sample M1 is an out-of-zone fault current signal is 40%, and the probability that it is an in-zone fault current signal is 60%. This shows that the classification of sample M1 by the grounding fault direction detection neural network at this time is relatively incorrect.

[0039] Based on the loss function L, the grounding fault direction detection neural network can determine its handling loss L = 0.91.

[0040] The grounding fault direction detection neural network updates its parameters according to the backpropagation algorithm. These parameters include the weight matrices of convolutional kernels 1 and 2 with a size of 5×5 in the convolutional layer, the weight matrix W1 with a size of 20×44 in the first fully connected layer, and the weight matrix W2 with a size of 2×20 in the second fully connected layer.

[0041] After the update is completed, the training sample M2 is provided to the grounding fault direction detection neural network, and then the feature extraction, classification, handling loss calculation, etc. of the training sample M2 are performed. The training ends until the handling loss L < TL (0.01).

[0042] After training, we will obtain a 2×5×5 weight matrix corresponding to the two convolutional kernels of the convolutional layer, a 20×44 weight matrix corresponding to the weight matrix W1 of the first fully connected layer, and a 2×20 weight matrix corresponding to the weight matrix W2 of the second fully connected layer. These parameters, together with the previously set structure, loss function, activation function, etc. of the ground fault directionality detection neural network, constitute the trained ground fault directionality detection neural network.

[0043] After completing the above model training method, the model utilization method can be executed when fault identification is required. The model utilization method is as follows: determine whether the fault identification start condition is triggered; if so, obtain the actual fault waveform of the low-current grounding system, use the deep learning network for fault identification to perform fault identification, and obtain the fault identification result.

[0044] In this model, the residual current is used as a preset current threshold as the fault identification activation condition. After fault identification, the residual voltage is used as the condition for outputting the fault identification result. Furthermore, after fault identification and obtaining the fault identification result, the determination of whether to trigger the fault identification activation condition is paused within a preset delay time (return delay).

[0045] Deploying the aforementioned fault identification mechanism for low-current grounding systems into a device yields a fault identification device for low-current grounding systems, comprising an embedded device, a starting element, and a fault confirmation element. The embedded device houses a deep learning network for fault identification. This embedded device acquires real fault waveforms of the low-current grounding system, uses the deep learning network for fault identification to perform fault identification, and obtains the fault identification result. The deep learning network for fault identification is pre-trained using a model training method. This method involves simulating various faults in the low-current grounding system using a simulation system to obtain a required number of simulated waveform data reflecting various fault conditions as training cases. The deep learning network is then trained, adjusted, and tested using these training cases to obtain the deep learning network for fault identification, thus completing the deployment of the deep learning network. The starting element is connected to the embedded device and determines whether the fault identification start condition is triggered. If so, the starting element outputs a signal to start the embedded device. The fault confirmation element is also connected to the embedded device and checks whether the residual voltage reaches a preset voltage setting value. If so, the fault confirmation element outputs a signal to cause the embedded device to output the fault identification result.

[0046] After the structure and parameters of the deep learning network are adjusted in the host computer Python environment, a parameter network for discrimination is obtained, i.e., a deep learning network for fault identification. Compared with the massive computing resources required for the learning process, the computing resources required for the discrimination network are greatly reduced, and it can be deployed in embedded devices. Because in the simulation environment, all waveforms are fault waveforms, the network only needs to determine whether it is inside or outside the fault zone. However, in actual operation, the device does not know when the fault will occur, so it is still necessary to set a start-up condition. The start-up is based on the sudden change in the sampled value of the residual current. The start-up value takes into account the sampling accuracy of the small current transformer (CT) of the device and the large current transformer (CT) of the zero-sequence bushing in the field, and is set as low as possible to ensure the fastest and most reliable start-up, ensuring that the algorithm can capture the waveform of the fault head. Once the residual current threshold is exceeded, the deep learning network starts working and caches the results. To avoid false alarms, residual voltage is used to confirm the fault. Only for confirmed faults will the device output the cached results. The working timing of the fault identification device for the small current grounding system is shown in the attached figure. Figure 2 As shown. It is important to note that: 1. The voltage used is the residual voltage measured at the protection installation location. A lower setting can detect faults with relatively high transition resistance, but setting it too low may cause false alarms due to disturbances caused by operating switches, switching loads, etc. The default value is recommended. 2. After the starting element detects a possible ground fault, the ground fault location confirmation delay begins timing. After the set time, it checks whether the residual voltage is greater than the confirmation voltage setting. If so, it indicates the direction of the ground fault. The default value is recommended. 3. After a ground fault is detected and confirmed, the return delay begins timing. During this period, new intermittent faults and arcing ground faults will not trigger the location function. This setting is to prevent an intermittent fault from triggering multiple actions and alarms in a short period. The location function will reopen when the residual voltage remains below the ground fault confirmation voltage for the return delay.

[0047] The above scheme uses only one algorithm, allowing deep learning to automatically find the optimal solution applicable globally; and improves the accuracy of line selection by adjusting the structure of the learning network.

[0048] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A fault identification method for a low-current grounding system, characterized in that: The fault identification method for low-current grounding systems includes a model training method and a model utilization method. The model training method involves using a simulation system to simulate various faults in the low-current grounding system and obtaining a required number of simulation waveform data reflecting various fault conditions as training cases. The deep learning network is trained, adjusted, and tested using the training cases to obtain a deep learning network for fault identification. The model is used as follows: it is determined whether the fault identification start condition is triggered. If so, the real fault waveform of the low current grounding system is obtained, and the fault identification is performed using the deep learning network for fault identification to obtain the fault identification result. In the model utilization method, whether the residual current reaches a preset current threshold is used as the fault identification activation condition. After fault identification, whether the residual voltage reaches the preset voltage setting value is used as the condition for whether to output the fault identification result. After fault identification is performed and the fault identification result is obtained, the system pauses the determination of whether the fault identification start condition is triggered within a preset delay period. The condition for the line selection to be reopened is that the duration for which the residual voltage is continuously lower than the preset voltage setting value reaches the duration of the delay period.

2. The fault identification method for low-current grounding systems according to claim 1, characterized in that: In the model training method described above, multiple simulation software programs are used in conjunction to perform simulation.

3. The fault identification method for low-current grounding systems according to claim 1, characterized in that: In the model training method, the training cases include training cases for training the deep learning network, adjustment cases for checking and adjusting the deep learning network, and test cases for testing the deep learning network. The ratio of the number of training cases, adjustment cases, and test cases is 7:1:

2.

4. The fault identification method for low-current grounding systems according to claim 1, characterized in that: The simulated waveform data is fault current and voltage waveform data based on fault transients.

5. The fault identification method for low-current grounding systems according to claim 1, characterized in that: The deep learning network employs a convolutional neural network comprising a convolutional layer, a first fully connected layer, a second fully connected layer, an activation layer, and an output layer.

6. A fault identification device for a low-current grounding system, used to implement the fault identification method for a low-current grounding system as described in any one of claims 1 to 5, characterized in that: The fault identification device for the low-current grounding system includes: An embedded device is provided, in which the deep learning network for fault identification is deployed. The embedded device is used to acquire real fault waveforms of a low-current grounding system, and uses the deep learning network for fault identification to perform fault identification and obtain fault identification results. The deep learning network for fault identification is obtained using a model training method, which involves simulating various faults in the low-current grounding system using a simulation system to obtain a required number of simulated waveform data reflecting various fault conditions as training cases. The deep learning network is then trained, adjusted, and tested using the training cases to obtain the deep learning network for fault identification. A startup element is connected to the embedded device. The startup element is used to determine whether the fault identification startup condition is triggered. If so, the startup element outputs a signal to start the embedded device.

7. The fault identification device for a low-current grounding system according to claim 6, characterized in that: The fault identification device for the low-current grounding system also includes: A fault confirmation element is connected to the embedded device. The fault confirmation element is used to determine whether the residual voltage reaches a preset voltage setting value. If so, the fault confirmation element outputs a signal to cause the embedded device to output the fault identification result.

Images