Power distribution network fault identification method based on joint deep learning under typhoon disaster

Abstract

The invention provides a power distribution network fault identification method based on joint deep learning under typhoon disasters. The method comprises the steps that a three-phase voltage image of a fault position is acquired; determining a fault feature image based on the three-phase voltage image; the fault feature image is input into a fault recognition module, a determined fault type is output, and the fault recognition module is obtained based on deep learning and comprises a high-precision recognition model and a rapid recognition model. According to the method provided by the invention, the fault identification module obtained based on deep learning is adopted, and the fault identification model corresponding to the demand scene can be selected according to different demand scenes, so that the fault type can be accurately and rapidly identified.

Description

Technical Field

[0001] This application relates to the field of power grids, and in particular to a method for identifying distribution network faults based on joint deep learning under typhoon disasters. Background Technology

[0002] With the rapid development of power distribution networks, power distribution systems have received increasing attention, especially in terms of fault diagnosis. As the end point of the power system network, the distribution network is directly connected to users and is the core of the power distribution process. Ensuring the safe and stable operation of the distribution network is crucial; any fault will cause significant inconvenience to users and may also result in serious economic losses.

[0003] When faults occur in a distribution network, problems such as poor patterning of fault signals and unclear characteristics often arise, posing a significant challenge to fault classification. Generally, distribution network faults can be classified into 10 types: single-phase-to-ground short circuit in phase A, single-phase-to-ground short circuit in phase B, single-phase-to-ground short circuit in phase C, phase-to-phase short circuit between phases AB, phase-to-phase short circuit between phases AC, phase-to-phase short circuit between phases BC, two-phase-to-ground short circuit between phases AB, two-phase-to-ground short circuit between phases AC, two-phase-to-ground short circuit between phases BC, and three-phase-to-phase short circuit between phases ABC. Incorrectly estimating the fault type can potentially cause secondary damage to the power grid. Therefore, how to quickly identify the fault type in different scenarios is a pressing technical problem that needs to be solved. Summary of the Invention

[0004] To overcome the aforementioned technical deficiencies, this application provides a method for distribution network fault identification based on joint deep learning under typhoon disasters. To achieve the above objective, this application implements the following technical solution:

[0005] This application provides a method for distribution network fault identification based on joint deep learning under typhoon disasters, including:

[0006] Obtain the three-phase voltage image at the fault location;

[0007] Based on the three-phase voltage image, a fault characteristic image is determined;

[0008] The fault feature image is input into the fault identification module, which outputs a determination of the fault type. The fault identification module is based on deep learning and includes a high-precision identification model and a fast identification model.

[0009] Optionally, determining the fault feature image based on the three-phase voltage image includes:

[0010] Wavelet transform is applied to the three-phase voltage image to determine the fault feature image.

[0011] Optionally, the training method for the fault identification module is as follows:

[0012] A fault probability model for distribution networks is constructed based on the parameterization of engineering experience with multiple probability distributions.

[0013] The fault probability model of the distribution network was simulated multiple times to determine the original dataset of the three-phase voltage when the fault occurred;

[0014] The fault identification module is trained based on the original dataset of the three-phase voltages.

[0015] Optionally, the fault identification module trained based on the original dataset of the three-phase voltages includes:

[0016] The original dataset of the three-phase voltage is input into the continuous wavelet algorithm for feature extraction, and the fault feature dataset is output.

[0017] Based on the fault feature dataset, a fault identification module is trained.

[0018] Optionally, the fault identification module trained based on the fault feature map includes:

[0019] The fault feature dataset is divided into a fault feature data training set and a fault feature data validation set according to a predetermined ratio;

[0020] A fault identification module is trained based on the fault feature data training set and the fault feature data validation set.

[0021] Optionally, the fault identification module trained based on the fault feature data training set and the fault feature data validation set includes:

[0022] Vgg-16 neural network model, Resnet-152 neural network model, and Alexnet neural network model were selected as the models to be trained.

[0023] Different number of iterations are set, and the fault feature data training set is input into the Vgg-16 neural network model, Resnet-152 neural network model, and Alexnet neural network model respectively for training. The training results of Vgg-16 neural network, Resnet-152 neural network, and Alexnet neural network are output.

[0024] The fault identification module is trained by iteratively training the Vgg-16 neural network training results, the Resnet-152 neural network training results, the Alexnet neural network training results, and the fault feature data validation set.

[0025] Optionally, the step of iteratively training the Vgg-16 neural network training results, the Resnet-152 neural network training results, the Alexnet neural network training results, and the fault feature data validation set to obtain the fault identification module includes:

[0026] Based on the training results of the Vgg-16 neural network and the fault feature data validation set, the average classification accuracy of the first model and the average training time of the first model are determined.

[0027] Based on the training results of the Resnet-152 neural network and the validation set of the fault feature data, the average value of the second classification accuracy and the average training time of the second model are determined.

[0028] Based on the training results of the Alexnet neural network and the validation set of the fault feature data, the average accuracy of the third classification and the average training time of the third model are determined.

[0029] A high-precision identification model is determined based on the average accuracy of the first classification, the average accuracy of the second classification, and the average accuracy of the third classification.

[0030] Based on the average training time of the first model, the average training time of the second model, and the average training time of the third model, a fast discrimination model is determined.

[0031] Optionally, a high-precision identification model is determined based on the average of the first classification accuracy, the average of the second classification accuracy, and the average of the third classification accuracy; including:

[0032] The average classification accuracy of the first classification, the average classification accuracy of the second classification, and the average classification accuracy of the third classification are compared to determine the maximum value of the average classification accuracy, and the model corresponding to the maximum value is taken as the high-precision identification model.

[0033] Optionally, determining the fast discrimination model based on the average training time of the first model, the average training time of the second model, and the average training time of the third model includes:

[0034] The average training time of the first model, the average training time of the second model, and the average training time of the third model are compared to determine the minimum average training time of the models, and the model corresponding to the minimum value is used as the fast discrimination model.

[0035] This application has the following beneficial effects:

[0036] The method proposed in this application employs a fault identification module based on deep learning, which can select a fault identification model corresponding to different demand scenarios, thereby accurately and quickly identifying the type of fault.

[0037] In addition to the purposes, features, and advantages described above, this application has other purposes, features, and advantages. The application will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0038] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0039] Figure 1 This is a flowchart illustrating a method for identifying power distribution network faults based on joint deep learning under typhoon disasters, provided in an embodiment of this application. Detailed Implementation

[0040] The embodiments of this application are described in detail below with reference to the accompanying drawings, but this application can be implemented in many different ways as defined and covered by the claims.

[0041] Therefore, in order to solve the above problems, such as Figure 1 As shown, this application proposes a distribution network fault identification method based on joint deep learning under typhoon disasters, including:

[0042] Step S101: Obtain the three-phase voltage image at the fault location;

[0043] After a fault is detected, high-voltage isolated acquisition equipment (such as current transformers and insulation probes) is used to record the voltage waveforms before and after the fault through online monitoring or on-site detection. Then, a visualized three-phase voltage image is generated through the instrument's built-in function or post-processing software.

[0044] Step S102: Based on the three-phase voltage image, determine the fault characteristic image;

[0045] Since the three-phase voltage image obtained by the above method has low information density, it is necessary to first perform feature engineering on the three-phase voltage image to obtain new feature maps, so as to enhance the fault features and improve the recognition accuracy of the subsequent recognition module.

[0046] This application selects the continuous wavelet transform algorithm to perform wavelet transformation on the three-phase voltage image, thereby determining the fault feature image. Continuous wavelet transform is a commonly used signal processing tool that can provide both time-domain and frequency-domain information of a signal, offering unique advantages in analyzing non-stationary signals. The core of continuous wavelet transform is to perform a series of convolutions between the signal and wavelet functions at different scales and locations, thereby detecting the local features of the signal. For the input signal (fault feature image) x(t), the basic formula for continuous wavelet transform is:

[0047]

[0048] In the formula, 'a' is a scaling parameter used to control whether the wavelet function is broadened or compressed in time and frequency. When 'a' increases, the wavelet function is stretched along the time axis, improving frequency resolution; conversely, when 'a' decreases, the wavelet function is compressed along the time axis, improving time resolution; 'b' is a translation parameter that can move the wavelet function along the time axis to transform the entire original signal; Ψ * (t) is the complex conjugate of the wavelet function, which plays an important role in calculating the inner product of the transform.

[0049] Based on signal characteristics, different signals can be analyzed using different wavelet functions. This application employs the Morlet wavelet to extract features from the three-phase voltage image at the fault location. The Morlet wavelet possesses excellent time-frequency localization capabilities and high frequency concentration, effectively analyzing frequency components in a signal. It offers flexible scale selection and is suitable for analyzing non-stationary signals. Its mathematical expression is:

[0050]

[0051] In the formula, It is the plane wave part, where ω0 is the center frequency, providing frequency selectivity for wavelets; σ is a Gaussian window function, where σ is the standard deviation of the Gaussian window, used to ensure the locality of the signal in the time domain. Through continuous wavelet transform, we effectively captured the differences in the three-phase voltages at the time of the fault, obtaining a more accurate fault characteristic image.

[0052] Step S103: Input the fault feature image into the fault identification module and output the fault type. The fault identification module is based on deep learning and includes a high-precision identification model and a fast identification model.

[0053] After inputting the aforementioned fault feature images into the fault identification module, the corresponding fault type can be determined. To adapt to different scenarios and requirements for fault type identification, the fault identification module generally employs deep learning, comprising a high-precision identification model and a fast identification model. To ensure the accuracy of fault identification by the high-precision and fast identification models, corresponding image recognition models must be selected for training. Based on the training results, the appropriate high-precision and fast identification models are then selected to form the fault identification module.

[0054] Generally speaking, the more samples in the dataset, the wider the coverage of various fault conditions, and the better the training effect of the model.

[0055] For power distribution systems, their lines are susceptible to typical natural disasters, leading to line failures for various reasons. Field surveys in typhoon-prone areas revealed that power distribution network failures during typhoons and heavy rains are primarily caused by tree obstruction, while some are directly caused by excessive wind force resulting in line breaks. This step, based on the parameterization of engineering experience using multiple probability distributions, constructs a power distribution network failure probability model, as shown in the following equation:

[0056] p fc,k (v(t))=1-(1-p fw,k (v(t)))(1-αp ftr,k (v(t))) (3)

[0057]

[0058] In the formula: p fc,k (v(t)) represents the time-varying fault probability of the conductor; p fw,k (v(t)) represents the probability of a direct wind-induced time-varying fault in conductor k; F fw,k (v(t)) and F fo,k (v(t)) represent the wind force and the maximum vertical force that conductor k can withstand, respectively; p ftr,k (v(t)) represents the time-varying fault probability induced by the fallen tree on conductor k; S k The wind intensity at point k on the conductor (0-1 continuous scale) can be calculated by dividing the local typhoon wind speed by the maximum wind speed in the study area; D H It is the diameter at breast height (cm at a height of 1.35m); k s It is a factor that takes into account the influence of micro-topography and is selected based on land cover information near the guide k; a s b s and c s It is a parameter related to the tree species, h(S) k) is an intermediate parameter, which can be used to establish the relationship between the wind strength at point k on the conductor, tree species, environment and other factors and the probability of time-varying faults induced by fallen trees.

[0059] After obtaining the above distribution network fault probability model, multiple simulations are performed to determine the original dataset when the fault occurs. Then, based on the original dataset of the three-phase voltage, the fault identification module is trained.

[0060] After obtaining the raw dataset of three-phase voltages, the raw dataset is input into a wavelet transform algorithm for feature extraction, outputting a fault feature dataset. This fault identification module is then trained based on this raw fault feature dataset. It should be noted that the process of feature extraction using wavelet transform has been explained in detail above and will not be elaborated upon here.

[0061] After obtaining the original fault feature dataset, it is divided into a fault feature data training set and a fault feature data validation set according to a predetermined ratio (e.g., 8:2). Then, a fault identification module is trained based on the fault feature data training set and the fault feature data validation set.

[0062] This application selects the Vgg-16 neural network model, the Resnet-152 neural network model, and the Alexnet neural network model as the models to be trained, and provides a detailed description of the above models as follows:

[0063] The Vgg-16 neural network model's core principle is to build a deep network by stacking small convolutional kernels to achieve a larger receptive field. Vgg-16 contains a 16-layer network architecture, with 13 convolutional layers for feature extraction, followed by ReLU activation functions, and interspersed with 5 pooling layers for dimensionality reduction. This model has a clear and well-structured architecture and demonstrates excellent performance in image classification and feature extraction tasks.

[0064] The ResNet-152 neural network model, also known as the residual neural network model, innovatively introduces two core concepts: "residual learning" and "jump connections." This solves the gradient vanishing and gradient exploding problems in the training process of deep learning networks, making it possible to train deeper neural networks.

[0065] The core idea of ​​residual neural networks is to change the learning objective from directly approximating the underlying mapping H(x) to learning the residual function:

[0066] H(x)=F(x)+x (5)

[0067] In the formula, F(x) is the residual mapping function (representing the residual), and x is the feature (image) representing the input residual block, i.e. the information unit to be processed.

[0068] In traditional deep neural network architectures, learning performance often fails to improve with increasing depth. This phenomenon mainly arises because some layers may perform poorly during convolution, leading to an abnormally large loss function and negatively impacting the final output. The residual neural network architecture shown in Figure 2 is designed to reduce the weights of the convolutional layer to zero when an increase in the loss function is detected. This design effectively ensures skip connections for gradients, allowing gradients to propagate to earlier layers and avoiding the vanishing gradient problem.

[0069] The AlexNet neural network model: At its core, this model uses five convolutional layers for feature extraction and three fully connected layers for feature integration and final prediction. Although the AlexNet model has limited depth, the use of larger convolutional kernels (such as 5×5) enables it to efficiently extract global features from images. Combined with GPU acceleration during training, AlexNet laid the foundation for developing deeper network architectures.

[0070] In other words, among the three neural network models mentioned above, the corresponding high-precision recognition model and fast recognition model were determined. The determination process is as follows:

[0071] First, different iteration numbers are set. Then, the fault feature data training set is input into the Vgg-16 neural network model, the Resnet-152 neural network model, and the Alexnet neural network model for multiple iterations of training. The training results of the Vgg-16 neural network, the Resnet-152 neural network, and the Alexnet neural network are output. Then, based on the training results of the Vgg-16 neural network, the output results of the Resnet-152 neural network, the output results of the Alexnet neural network, and the fault feature data validation set, the fault identification module is trained.

[0072] After training for the current number of iterations is completed, comparing the training results of the Vgg-16 neural network with the results of the validation set of the fault feature data can determine the number of correctly predicted samples, n. r Then, based on the number of correctly predicted samples and the total number of samples n in the fault feature data training set... va The average classification accuracy α of the model at the current iteration number is determined as follows:

[0073]

[0074] Then, based on the average classification accuracy α of the model at the current iteration number, the average classification accuracy of the first model at different iteration numbers can be calculated. During training, the training time t corresponding to the current iteration number is recorded. Then, based on the training time t corresponding to the current iteration number, the average training time of the first model can be calculated.

[0075] Based on the same calculation process described above, we can determine the average second classification accuracy and average second model training time corresponding to the Resnet-152 neural network model, as well as the average third classification accuracy and average third model training time corresponding to the Alexnet neural network training model.

[0076] The average accuracy of the first classification, the average accuracy of the second classification, and the average accuracy of the third classification are compared to determine the maximum value of the average classification accuracy. The model corresponding to the maximum value is then used as the high-precision identification model.

[0077] The average training time of the first, second, and third models is compared to determine the minimum average training time. The model corresponding to the minimum average training time is selected as the fast identification model. In other words, among the Vgg-16, ResNet-152, and AlexNet neural network models, the one with the highest training accuracy is selected as the high-precision identification model, and the model with the shortest training time, i.e., the fastest at identifying fault types, is selected as the fast identification model.

[0078] Based on the actual training results, the ResNet-152 architecture has a significant advantage in accuracy in distribution network fault classification. It can reach more than 90% after 10 iterations, and can further improve with the number of iterations, making it an ideal choice for high-precision fault diagnosis scenarios. With its shallower architecture, AlexNet performs well in terms of iteration speed, and its advantage becomes more obvious with the number of iterations, making it suitable for scenarios that require rapid diagnosis and classification.

[0079] In summary, the method proposed in this application employs a fault identification module based on deep learning, which can select a fault identification model corresponding to different demand scenarios, thereby accurately and quickly identifying the type of fault.

[0080] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for fault identification in power distribution networks based on joint deep learning under typhoon disasters, characterized in that, include: Obtain the three-phase voltage image at the fault location; Based on the three-phase voltage image, a fault characteristic image is determined; The fault feature image is input into the fault identification module, which outputs a determination of the fault type. The fault identification module is based on deep learning and includes a high-precision identification model and a fast identification model.

2. The method according to claim 1, characterized in that, The process of determining the fault feature image based on the three-phase voltage image includes: Wavelet transform is applied to the three-phase voltage image to determine the fault feature image.

3. The method according to claim 1, characterized in that, The training method for the fault identification module is as follows: A fault probability model for distribution networks is constructed based on the parameterization of engineering experience with multiple probability distributions. The fault probability model of the distribution network was simulated multiple times to determine the original dataset of the three-phase voltage when the fault occurred; The fault identification module is trained based on the original dataset of the three-phase voltages.

4. The method according to claim 3, characterized in that, The fault identification module, trained based on the original dataset of the three-phase voltages, includes: The original dataset of the three-phase voltage is input into the continuous wavelet algorithm for feature extraction, and the fault feature dataset is output. Based on the fault feature dataset, a fault identification module is trained.

5. The method according to claim 4, characterized in that, The fault identification module trained based on the fault feature map includes: The fault feature dataset is divided into a fault feature data training set and a fault feature data validation set according to a predetermined ratio; A fault identification module is trained based on the fault feature data training set and the fault feature data validation set.

6. The method according to claim 5, characterized in that, The fault identification module, trained based on the fault feature data training set and the fault feature data validation set, includes: Vgg-16 neural network model, Resnet-152 neural network model, and Alexnet neural network model were selected as the models to be trained. Different number of iterations are set, and the fault feature data training set is input into the Vgg-16 neural network model, Resnet-152 neural network model, and Alexnet neural network model respectively for training. The training results of Vgg-16 neural network, Resnet-152 neural network, and Alexnet neural network are output. The fault identification module is trained by iteratively training the Vgg-16 neural network training results, the Resnet-152 neural network training results, the Alexnet neural network training results, and the fault feature data validation set.

7. The method according to claim 6, characterized in that, The fault identification module is trained by iteratively training the Vgg-16 neural network training results, the Resnet-152 neural network training results, the Alexnet neural network training results, and the fault feature data validation set, including: Based on the training results of the Vgg-16 neural network and the fault feature data validation set, the average classification accuracy of the first model and the average training time of the first model are determined. Based on the training results of the Resnet-152 neural network and the validation set of the fault feature data, the average value of the second classification accuracy and the average training time of the second model are determined. Based on the training results of the Alexnet neural network and the validation set of the fault feature data, the average accuracy of the third classification and the average training time of the third model are determined. A high-precision identification model is determined based on the average accuracy of the first classification, the average accuracy of the second classification, and the average accuracy of the third classification. Based on the average training time of the first model, the average training time of the second model, and the average training time of the third model, a fast discrimination model is determined.

8. The method according to claim 7, characterized in that, Based on the average of the first classification accuracy, the average of the second classification accuracy, and the average of the third classification accuracy, a high-precision identification model is determined; including: The average classification accuracy of the first classification, the average classification accuracy of the second classification, and the average classification accuracy of the third classification are compared to determine the maximum value of the average classification accuracy, and the model corresponding to the maximum value is taken as the high-precision identification model.

9. The method according to claim 7, characterized in that, The step of determining a fast discrimination model based on the average training time of the first model, the average training time of the second model, and the average training time of the third model includes: The average training time of the first model, the average training time of the second model, and the average training time of the third model are compared to determine the minimum average training time of the models, and the model corresponding to the minimum value is used as the fast discrimination model.

Images