A deep learning-based multi-feature fusion single-phase ground fault type identification method

Abstract

The application discloses a multi-feature fusion single-phase ground fault type identification method based on deep learning, and mainly comprises the following steps: 1) obtaining fault recording data collected by a field fault recording device; 2) pre-processing the fault recording data, and performing time-frequency decomposition on the pre-processed fault recording data based on HHT to obtain corresponding time-frequency information features; 3) constructing the time-frequency information features of the fault recording data obtained in the step 2 into a primary data set; 4) building and training a basic learner Resnet18 to extract complex nonlinear features in the data set; 5) building and training a basic learner LSTM to extract time sequence correlation features in the data set; 6) splicing and fusing the complex nonlinear features and the time sequence correlation features obtained by learning and extracting in the steps 4 and 5 to construct a secondary data set, and identifying specific single-phase ground fault types in combination with a decision tree model. The application has good accuracy and robustness, and is good in universality. The application is suitable for identifying various single-phase ground fault types including intermittent arc ground faults and high-resistance ground faults, and reliable basis can be provided for subsequent development of targeted fault treatment measures according to the identification results.

Description

Technical Field

[0001] The present invention relates to the field of smart distribution networks, and in particular to a method for identifying single-phase grounding fault types based on multi-feature fusion based on deep learning. Technical Background

[0002] Safety, reliability, quality, and economy are fundamental requirements for power system operation. However, single-phase grounding faults frequently occur in my country's medium and low voltage distribution networks, impacting the reliable operation of these systems. Because single-phase grounding faults are subtle, the characteristics that differentiate between different fault types are low, making accurate fault identification difficult.

[0003] Existing research results have achieved certain results in single-phase grounding fault detection, but most of them only select some characteristics of the distribution network, that is, the unique attributes of a certain type of fault for analysis, resulting in insufficient description of fault information. They can only identify a specific fault type and do not comprehensively classify single-phase grounding fault types. The algorithm is not versatile enough, which is not conducive to dispatchers to formulate targeted fault handling measures.

[0004] Deep learning is increasingly used in engineering. It is very good at automatically learning complex and useful features from high-dimensional data sets. Compared with many excellent manual feature extractors that have appeared in the past, such as scale-invariant feature transform, Gabor filter and histogram of oriented gradients, deep learning models can learn features of different properties and levels by building different structures and adjusting the number of hidden layers. They can directly implement end-to-end task training or extract abstract features for learning downstream tasks.

[0005] In summary, the present invention considers comprehensive identification of seven types of single-phase ground faults, including high-resistance ground faults and intermittent arc ground faults. This requires the extraction of more comprehensive fault characteristics. The recorded data collected on-site from the distribution network serves as an "electrocardiogram" of the network's real-time status and operation, providing the most direct and accurate basis for single-phase ground fault type identification. The key to identifying specific single-phase ground fault types lies in leveraging deep learning techniques to automatically learn complex and useful features from high-dimensional recorded data. Summary of the Invention

[0006] The purpose of the present invention is to solve the problems existing in the prior art.

[0007] The technical solution adopted to achieve the purpose of the present invention is as follows: a multi-feature fusion single-phase grounding fault type identification method based on deep learning mainly includes the following steps:

[0008] 1) Obtaining fault recording data collected by the on-site fault recording device;

[0009] 2) Preprocessing the fault recording data and performing time-frequency decomposition based on the Hilbert-Huang transform on the preprocessed fault recording data;

[0010] 3) constructing the time-frequency information features of the fault recording data obtained by processing in 2) into a primary data set;

[0011] 4) Use the primary dataset to train the basic learner Resnet18 model to extract the complex nonlinear features in the primary dataset;

[0012] 5) Use the primary dataset to train the basic learner LSTM model and extract the temporal correlation features in the primary dataset;

[0013] 6) Concatenate the complex nonlinear features and temporal correlation features learned and extracted in 4) and 5) to construct a secondary dataset;

[0014] 7) The decision tree model is trained using the secondary data set, and the trained decision tree model is used as a secondary learner to identify the specific type of single-phase grounding fault.

[0015] Furthermore, the step 2) specifically includes:

[0016] 2.1) Intercept the fault recording data in the fault state stage;

[0017] 2.2) Divide all electrical quantity features in the fault recording data into key feature parts and non-key feature parts;

[0018] 2.3) Perform Hilbert-Huang transform on the key feature parts of the fault recording data to extract the IMF component, instantaneous amplitude, instantaneous frequency and Hilbert spectrum.

[0019] Furthermore, the step 3) specifically includes:

[0020] 3.1) concatenating the Hilbert-Huang transform result of the key feature portion in step 2.3) with the original non-key feature portion in step 2.2);

[0021] 3.2) Normalize the spliced ​​features and convert them into grayscale images;

[0022] 3.3) Construct a primary data set, where each data sample corresponds to a single-phase ground fault type label 0, 1, 2, ..., m;

[0023] 3.4) Divide the constructed primary dataset into a training set and a test set. The training set is used to train the two basic learners, and the test set is used to test the learning effects of the two basic learners.

[0024] Furthermore, the step 4) specifically includes:

[0025] 4.1) Building the Resnet18 model as a basic learner. Based on the original Resnet18 model structure, we made some modifications based on the characteristics of the fault recording data. These modifications included: changing the convolution kernel of the first conv layer to 3×3, retaining the fully connected layers and softmax layers during model training; and removing the fully connected layers and softmax layers after model training to extract the complex nonlinear features of the primary dataset.

[0026] 4.2) Training the base learner ResNet-18 model. Use the constructed primary dataset to train the base learner ResNet-18 model. When the model's validation accuracy reaches the expected level, save the model's parameter settings. 4.3) Use the trained base learner ResNet-18 model to extract complex nonlinear features from the primary dataset. The extracted feature dimensions are the output dimensions of the last convolutional layer of the ResNet-18 model.

[0027] Furthermore, the step 5) specifically includes:

[0028] 5.1) Build the basic learner LSTM model. The LSTM model structure is derived from the fault recording data and experimental results combined with hyperparameter adjustment.

[0029] 5.2) Basic Learner LSTM Model Training. Use the constructed primary dataset to train the basic learner LSTM model. When the model's verification accuracy reaches the expected effect, save the model parameter settings.

[0030] 5.3) Use the trained LSTM model to extract temporal correlation features from the primary dataset. The output layer of the LSTM model is removed during feature extraction, so the extracted feature dimension is the number of nodes set in the last hidden layer.

[0031] Furthermore, the step 6) specifically includes:

[0032] 6.1) Constructing a secondary dataset. Using the Resnet18 and LSTM models, we extract the complex nonlinear features and time-series correlation features contained in the recorded data. These two sets of features are then concatenated with the true type labels of single-phase grounding faults to form a secondary dataset. The label values ​​range from 0 to m-1, representing m different types of single-phase grounding faults.

[0033] Furthermore, the step 7) specifically includes:

[0034] 7.1) Decision Tree Model Training: The CART model is selected as the decision tree model. It is trained using the secondary dataset constructed in 6.1) and the grid search algorithm is used to find the optimal parameter settings for the CART model.

[0035] 7.2) Based on the trained secondary learner, identify the specific single-phase grounding fault results.

[0036] The technical benefits of this invention are undeniable. In the context of the power Internet of Things, this invention proposes a method for identifying single-phase ground fault types using multi-feature fusion based on deep learning. This method demonstrates excellent accuracy, robustness, and versatility. It is suitable for identifying a variety of single-phase ground fault types, including intermittent arc ground faults and high-resistance ground faults. The identification results provide a reliable basis for the subsequent development of targeted fault handling measures. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is the overall process architecture diagram of the single-phase grounding fault type identification method;

[0038] Figure 2 This is the structural framework diagram of the Resnet18 model of the basic learner;

[0039] Figure 3 LSTM model structure framework diagram for basic learner

[0040] Figure 4 This is the framework diagram of the single-phase ground fault identification model

[0041] Figure 5 Comparison of the prediction and classification effects of the single-phase grounding fault type identification method proposed by the present invention and the Resnet18 or LSTM model alone DETAILED DESCRIPTION

[0042] The present invention will be further described below with reference to the following examples, but it should not be understood that the scope of the present invention is limited to the following examples. Without departing from the above technical ideas of the present invention, various substitutions and modifications can be made according to common technical knowledge and customary means in the art, and all should be included in the scope of protection of the present invention.

[0043] Example

[0044] 1) Obtaining fault recording data collected by the on-site fault recording device;

[0045] The recording data used in the present invention comes from a real distribution network test site in China. By changing the neutral point grounding operation mode, the type of grounding medium, the size of the grounding resistance, etc., different types of single-phase grounding fault recording data are obtained. Among them, the neutral point grounding mode covers mainstream forms such as no grounding, grounding through arc suppression coils, and grounding through small resistances; the grounding medium includes common fault types such as intermittent arc grounding, stable arc grounding, soil grounding, and resistance grounding; the grounding resistance value selects typical values ​​such as 250Ω, 1000Ω, 2000Ω, and 5000Ω, which include a total of 7 types of single-phase grounding faults. The experiment generated 420, 600, and 240 fault recording data respectively under the three grounding operation modes; the sampling frequency of the recording device is 10kHz, the sampling period includes 12014 sampling points, and each segment of recording data contains 291 electrical quantities.

[0046] 2) Preprocessing the fault recording data and performing time-frequency decomposition based on the Hilbert-Huang transform on the preprocessed fault recording data to obtain the corresponding time-frequency information features;

[0047] Furthermore, step 2) specifically includes the following:

[0048] 2.1) Time Period Capture: Each segment of single-phase ground fault recording data is not in a fault state throughout the entire sampling period. It also includes the normal state before the fault occurs and the transition state after the fault occurs. To clarify the observation object and reduce the amount of data to be analyzed, the recording data in the fault state is captured. Through observation of the recording data, the captured time period includes 600 sampling points.

[0049] 2.2) Key Feature Classification: Many of the 291 electrical quantities in the recorded data remain constant over time or exhibit periodic variations, making them less useful for single-phase ground fault type identification. Based on the experience of distribution network operation and maintenance engineers, all electrical quantity features in the recorded data are divided into two parts: key feature part and non-key feature part.

[0050] Among them, the key features include: bus zero sequence current 3I0, bus zero sequence voltage 3U0, I true training 1 three-phase current 3I0 Ⅰ ,ⅡTrue training 1 three-phase current Ⅲ True training 1 three-phase current Ⅳ Three-phase current The remaining electrical quantity characteristics are non-critical characteristics;

[0051] 2.3) Performing Hilbert-Huang transform on the key features of the distribution network fault recording data to extract the intrinsic mode function, instantaneous amplitude, instantaneous frequency, and Hilbert spectrum;

[0052] Furthermore, the step 3) specifically includes:

[0053] 3.1) concatenating the Hilbert-Huang transform result of the key feature portion in step 2.3) with the original non-key feature portion in step 2.2);

[0054] 3.2) Normalize the spliced ​​features and convert them to grayscale. 3.3) Construct a primary dataset. Seven fault categories, including intermittent arc grounding fault, grounding fault via 250Ω resistor, grounding fault via 1000Ω resistor, grounding fault via 2000Ω resistor, grounding fault via 5000Ω resistor, grounding fault via soil, and stable arc grounding fault, are represented by 1, 2, ..., 7. The fault recording dataset processed in step 2) and the fault labels together constitute the primary dataset.

[0055] 3.4) The constructed primary dataset is divided into a training set and a test set with a ratio of 8:2. The training set is used to train the two basic learners, and the test set is used to test the learning performance of the two basic learners.

[0056] 4) Use the primary dataset to train the basic learner Resnet18 model to extract the complex nonlinear features in the primary dataset;

[0057] Furthermore, the step 4) specifically includes:

[0058] 4.1) Building the Resnet18 model for the basic learner. Based on the original Resnet18 structure, this paper makes some modifications based on the characteristics of fault recording data, including the following: The convolution kernel of the first conv layer is modified to 3×3, and the fully connected layer and softmax layer are retained during the model training phase. After model training, the fully connected layer and softmax layer are removed when extracting the complex nonlinear features of the dataset. The Resnet18 model parameters for feature extraction are shown in the following table.

[0059] Table 1

[0060]

[0061]

[0062] 4.2) Training the Resnet18 base learner model. Use the training set constructed in 3) to train the Resnet18 base learner model, using the cross-entropy loss function. Use the test set constructed in 3) to test the training results of the Resnet18 base learner model. When the accuracy of the Resnet18 base learner model reaches the expected level, save the model parameter settings.

[0063] 4.3) Use the trained Resnet18 model as the base learner to extract complex nonlinear features from the primary dataset. The extracted feature dimension is the output dimension of the last convolutional layer of the Resnet18 model, which is 2048.

[0064] 5) Use the primary dataset to train the basic learner LSTM model and extract the temporal correlation features in the primary dataset;

[0065] Furthermore, the step 5) specifically includes:

[0066] 5.1) Build the basic learner LSTM model. The LSTM model structure was developed based on fault recording data and experimental results from adjusting hyperparameters. Setting the number of LSTM layers to 2 and the number of hidden layer nodes to 100 resulted in a model with both good accuracy and fast convergence.

[0067] 5.2) Training the Base Learner LSTM Model. Use the training set constructed in 3.2) to train the Base Learner LSTM model, using the cross-entropy loss function. Use the test set constructed in 3.2) to test the training effectiveness of the Base Learner LSTM model. When the Base Learner LSTM model achieves the expected accuracy, save the model parameter settings.

[0068] 5.3) Use the trained LSTM model to extract temporal correlation features from the primary dataset. The output layer of the LSTM model is removed during feature extraction, so the extracted feature dimension is the number of nodes in the hidden layer, which is 100.

[0069] 6) Concatenate the complex nonlinear features and temporal correlation features learned and extracted in steps 4) and 5) to construct a secondary dataset;

[0070] Furthermore, step 6) specifically includes: 6.1) Constructing a secondary dataset. The feature dimensions extracted by the base learner ResNet18 model and the base learner LSTM model are (1, 2048) and (1, 100), respectively. These two sets of features are concatenated and combined with the true type label of the single-phase grounding fault (with a value of 0-6 representing seven different types of single-phase grounding faults) to form a secondary dataset, as shown in the following table.

[0071] Table 2

[0072]

[0073] 7) Using the secondary data set to train the decision tree model, the trained decision tree model is used as a secondary learner to identify the specific single-phase grounding fault type;

[0074] Furthermore, the step 7) specifically includes:

[0075] 7.1) Decision Tree Model Training. The CART model was selected for training using the secondary dataset constructed in 6). A grid search algorithm was used to find the optimal parameter settings for the CART model.

[0076] 7.2) Based on the trained secondary learner, i.e., the decision tree model in 7.1), the specific single-phase grounding fault result is identified.

Claims

1. A multi-feature fusion single-phase grounding fault type identification method based on deep learning, characterized by: The main steps include: 1) Obtaining fault recording data collected by the on-site fault recording device; 2) performing preprocessing of the fault recording data and performing time-frequency decomposition of the preprocessed fault recording data based on the Hilbert-Huang transform; 3) constructing the time-frequency information features of the fault recording data obtained in step 2) into a primary data set; 4) Use the primary dataset to train the basic learner Resnet18 model to extract the complex nonlinear features in the primary dataset; 5) Use the primary dataset to train the basic learner LSTM model and extract the temporal correlation features in the primary dataset; 6) Concatenate the complex nonlinear features and temporal correlation features learned and extracted in steps 4) and 5) to construct a secondary dataset; 7) Decision tree model training. The specific decision tree model selected is the CART model. It is trained using the secondary data set constructed in step 6). The grid search algorithm is used to optimize the hyperparameters of the CART model to obtain the optimal parameter settings, construct a secondary learner, and identify the specific single-phase grounding fault type.

2. According to the method for identifying single-phase grounding fault types based on multi-feature fusion based on deep learning in claim 1, the main steps of step 2) are as follows: 2.1) Intercept the fault recording data in the fault state stage; 2.2) Divide all electrical quantity features contained in the fault recording data into key feature parts and non-key feature parts; 2.3) Perform Hilbert-Huang transform on the key feature parts of the fault recording data to extract the IMF component, instantaneous amplitude, instantaneous frequency and Hilbert spectrum.

3. According to the method for identifying single-phase grounding fault types based on multi-feature fusion based on deep learning in claim 2, the main steps of step 3) are as follows: 3.1) concatenating the Hilbert-Huang transform result of the key feature portion in step 2.3) with the original non-key feature portion in step 2.2); 3.2) Normalize the spliced ​​features and convert them into grayscale images; 3.3) Construct a primary data set, where each data sample corresponds to a single-phase ground fault type label 0, 1, 2, ..., m; 3.4) The constructed primary dataset is divided into a training set and a test set. The training set is used to train the two basic learners, and the test set is used to test the learning effects of the two basic learners.

4. According to the method for identifying single-phase grounding fault types based on multi-feature fusion based on deep learning in claim 1, the main steps of step 4) are as follows: 4.1) Building the Resnet18 model as a basic learner. Based on the original Resnet18 architecture, we made some modifications based on the characteristics of the fault recording data. These modifications included: changing the convolution kernel of the first conv layer to 3×3, retaining the fully connected and softmax layers during model training, and removing them after model training to extract the complex nonlinear features of the primary dataset. 4.2) Training the basic learner Resnet18 model: Use the constructed primary dataset to train the basic learner Resnet18 model. When the verification accuracy of the model reaches the expected effect, save the model parameter settings; 4.3) The trained base learner Resnet18 model is used to extract complex nonlinear features of the primary dataset, and the extracted feature dimension is the output dimension of the last convolutional layer of the base learner Resnet18 model.

5. According to the method for identifying single-phase grounding fault types based on multi-feature fusion based on deep learning in claim 1, the main steps of step 5) are as follows: 5.1) Build the basic learner LSTM model. The structure of the basic learner LSTM model is obtained by comparing the fault recording data and the experimental results combined with the adjustment of hyperparameters; 5.2) Training the basic learner LSTM model: Use the constructed primary dataset to train the basic learner LSTM model. When the verification accuracy of the model reaches the expected effect, save the model parameter settings; 5.3) The trained base learner LSTM model is used to extract temporal correlation features of the primary dataset. When extracting features, the output layer of the base learner LSTM model is removed, so the extracted feature dimension is the number of nodes set in the last hidden layer.

6. According to the method for identifying single-phase grounding fault types based on multi-feature fusion based on deep learning in claim 1, the main steps of step 6) are as follows: 6.1) Construct a secondary dataset. Based on the basic learner Resnet18 model and the basic learner LSTM model, extract the complex nonlinear features and time series correlation features contained in the recorded data. These two sets of features are spliced ​​together and then combined with the true type label of the single-phase grounding fault to form a secondary dataset. The label value is 0-m-1, representing m different types of single-phase grounding fault types.

7. According to the method for identifying single-phase grounding fault types based on multi-feature fusion based on deep learning in claim 6, the main steps of step 7) are as follows: 7.1) Decision tree model training. Specifically, the decision tree model selected is a CART model. The secondary dataset in step 6.1) is used to train the CART model. A grid search algorithm is used to find the optimal parameter settings for the CART model and construct a secondary learner. 7.2) Based on the trained secondary learner, identify the specific single-phase grounding fault results.

Images