CN114755529A

CN114755529A - Multi-feature fusion single-phase earth fault type identification method based on deep learning

Info

Publication number: CN114755529A
Application number: CN202210355956.8A
Authority: CN
Inventors: 范敏; 夏嘉璐; 刘宇彤; 孟鑫余; 彭屿雯; 冯楚瑞
Original assignee: Chongqing University
Current assignee: Chongqing University
Priority date: 2022-04-06
Filing date: 2022-04-06
Publication date: 2022-07-15

Abstract

The invention discloses a multi-feature fusion single-phase earth fault type identification method based on deep learning, which mainly comprises the following steps: 1) acquiring fault recording data acquired by a field fault recording device; 2) preprocessing fault recording data, and performing HHT-based time-frequency decomposition on the preprocessed fault recording data to obtain corresponding time-frequency information characteristics; 3) constructing the time-frequency information characteristics of the fault recording data obtained by the processing in the step 2 into a primary data set; 4) building and training a basic learner Resnet18, and extracting complex nonlinear features in a data set; 5) building and training a basic learner LSTM, and extracting time sequence correlation characteristics in a data set; 6) and (4) splicing and fusing the complex nonlinear features and the time sequence correlation features obtained by learning and extracting in the steps (4) and (5), constructing a secondary data set, and identifying a specific single-phase earth fault type by combining a decision tree model. The invention has good accuracy and robustness, and good universality. The method is suitable for identifying various single-phase earth fault types including intermittent arc earth faults and high-resistance earth faults, and the identification result can provide reliable basis for subsequent establishment of targeted fault treatment measures.

Description

Multi-feature fusion single-phase earth fault type identification method based on deep learning

Technical Field

The invention relates to the field of intelligent power distribution networks, in particular to a multi-feature fusion single-phase earth fault type identification method based on deep learning.

Technical Field

Safety, reliability, high quality and economy are basic requirements for the operation of a power system. However, single-phase earth faults frequently occur in low-voltage distribution network systems in China, and certain impact is caused on the reliable operation of power systems. Because the fault is weak when the single-phase earth fault occurs, the feature discrimination between different fault types is low, and the fault type is difficult to be accurately identified.

The existing research results have a certain effect in single-phase earth fault detection, but most of the existing research results only select partial characteristics of a power distribution network, namely the specific attributes of certain faults to analyze, so that the fault information description is insufficient, only a specific fault type can be identified, the single-phase earth fault types are not comprehensively divided, the algorithm universality is insufficient, and the targeted fault treatment measures made by a dispatcher are not facilitated.

Deep learning is becoming more widely used in the engineering field, and is very adept at automatically learning complex and useful features from high dimensional data sets, compared to many excellent manual feature extractors that have appeared in the past, such as: the Scale Invariant Feature Transform (SIFT), the Gabor filter, the Histogram of Oriented Gradients (HOG) and the like, and the deep learning model can learn to obtain features of different properties and different levels by building different structures and adjusting the number of hidden layers, so that end-to-end task training can be directly realized or abstract features can be extracted for learning of downstream tasks.

In summary, the present invention considers comprehensive identification of seven types of single-phase ground faults including high resistance ground fault and intermittent arc ground fault, for which more comprehensive fault characteristics need to be extracted. The recorded wave data acquired on the site of the power distribution network is an electrocardiogram of the real-time state and the running condition of the power distribution network, and can provide the most direct and accurate basis for identifying the type of the single-phase earth fault. How to automatically learn complex and useful features from high-dimensional recording data by using deep learning technology is the key to identify specific single-phase earth fault types.

Disclosure of Invention

The present invention is directed to solving the problems of the prior art.

The technical scheme adopted for achieving the purpose of the invention is that the method for identifying the type of the multi-feature fusion single-phase earth fault based on deep learning mainly comprises the following steps:

(1) acquiring fault recording data acquired by a field fault recording device;

(2) preprocessing fault recording data, and performing HHT-based time-frequency decomposition on the preprocessed fault recording data to obtain corresponding time-frequency information characteristics;

(3) constructing the time-frequency information characteristics of the fault recording data obtained by the processing in the step (2) into a primary data set;

(4) Building and training a basic learner Resnet18, and extracting complex nonlinear characteristics in a data set;

(5) building and training a basic learner LSTM, and extracting time sequence correlation characteristics in a data set;

(6) and (5) splicing and fusing the complex nonlinear features and the time sequence correlation features obtained by learning and extracting in the steps (4) and (5), constructing a secondary data set, and identifying a specific single-phase earth fault type by combining a decision tree model.

Further, the step (2) specifically includes:

(2.1) intercepting fault recording data in a fault state stage;

(2.2) dividing all electrical quantity characteristics in the fault recording data into a key characteristic part and a non-key characteristic part;

(2.3) performing Hilbert-Huang transform (HHT) on key characteristic parts of fault recording data, and extracting IMF components, instantaneous amplitudes, instantaneous frequencies and Hilbert spectrums;

(2.4) splicing the result of the key characteristic part in the step 2.3 after HHT transformation with the original non-key characteristic part in the step 2.2;

and (2.5) normalizing the spliced characteristics and converting the gray-scale map.

Further, the step (3) specifically includes:

(3.1) constructing a primary data set, wherein each data sample corresponds to a single-phase ground fault type label 0, 1, 2, …, m;

And (3.2) dividing the constructed primary data set into a training set and a test set. The training set is used for training the two basic learners, and the testing set is used for testing the learning effects of the two basic learners;

further, the step (4) specifically includes:

(4.1) building a Resnet18 model of a basic learner. On the basis of the original structure of Resnet18, the invention combines the characteristics of fault recording data to make some changes, including the following: modifying the convolution kernel of the first layer conv to be 3 multiplied by 3, and reserving a full connection layer and a softmax layer in the model training stage; after the model training is finished, when the complex nonlinear features of the data set are extracted, the full connection layer and the softmax layer are removed, and Resnet18 model parameters when the features are extracted are shown in the following table.

TABLE 1

(4.2) training a basic learner Resnet18 model. And (4) training a basic learner Resnet18 by using the training set constructed in the step (3), wherein the loss function used in model training is a cross-entropy loss function. And detects the training effect of the basic learner Resnet18 based on the test set constructed in (3). When the accuracy of the basic learner Resnet18 model reaches the expected effect, the parameter settings of the model are saved.

And (4.3) extracting complex nonlinear characteristics of the data set by using a basic learner Resnet18 model after training. The extracted feature dimension is the output dimension of the last convolutional layer of the Resnet18 model, which is 2048.

Further, the step (5) specifically includes:

and (5.1) building a basic learner LSTM model. And comparing the model structure of the LSTM according to the fault recording data and the experimental result of the combined adjustment over-parameter. The number of LSTM layers is set to be 2, and the number of nodes of the hidden layer is set to be 100, so that the model has better precision and higher convergence rate.

And (5.2) training the LSTM model of the basic learner. And (4) training a basic learner LSTM by using the training set constructed in the step (3), wherein the loss function used in model training is a cross entropy loss function. And detecting the training effect of the basic learner LSTM based on the test set constructed in (3). And when the accuracy of the LSTM model of the basic learner reaches the expected effect, saving the parameter setting of the model.

And (5.3) extracting time sequence correlation characteristics of the data set by using the LSTM model of the basic learner after training is finished. When the characteristics are extracted, the output layer of the LSTM model is removed, so that the characteristic dimensionality obtained by extraction is 100 of the number of nodes arranged in the last hidden layer.

Further, the step (6) specifically includes:

(6.1) constructing a secondary data set. The extracted feature dimensions of the ResNet18 and LSTM models are (1, 2048) and (1, 100), respectively. The two sets of characteristics are spliced and form a secondary data set with true type labels (0-6 respectively represent 7 different types of single-phase earth fault types) of single-phase earth faults.

And (6.2) training a decision tree model. The decision tree model specifically chosen is the CART model, which is trained using the secondary dataset constructed in (6.1). And the optimal parameter setting of the CART model is found by utilizing a grid search algorithm.

And (6.3) identifying a specific single-phase earth fault result based on the decision tree model in the secondary learner of which the training is finished, namely (6.2).

The technical effect of the invention is undoubted. The invention provides a multi-feature fusion single-phase earth fault type identification method based on deep learning in the context of an electric power internet of things. The method has good accuracy and robustness, and good universality. The method is suitable for identifying various single-phase earth fault types including intermittent arc earth faults and high-resistance earth faults, and the identification result can provide reliable basis for subsequent establishment of targeted fault treatment measures.

Drawings

FIG. 1 is a general flow diagram of a single-phase ground fault type identification method;

FIG. 2 is a structural framework diagram of the basic learner Resnet 18;

FIG. 3 is a diagram of the basic learner LSTM structural framework

FIG. 4 is a single-phase earth fault identification model frame diagram

FIG. 5 is a comparison graph of the predicted classification effect of the single-phase earth fault type identification method proposed by the present invention and the Resnet18 or LSTM model alone

Detailed Description

The present invention will be further described with reference to the following examples, but it should be understood that the scope of the subject matter described above is not limited to the following examples. Various substitutions and modifications can be made without departing from the technical idea of the invention and the technical idea of the invention, and the technical idea of the invention is included in the scope of the invention.

Examples

(1) Acquiring fault recording data acquired by a field fault recording device;

the recording data used by the invention comes from a certain real power distribution network test field in China. Different types of single-phase earth fault recording data are obtained by changing the grounding operation mode, the grounding medium type, the grounding resistance and the like of the neutral point. The neutral point grounding mode covers the mainstream modes of ungrounded mode, grounded mode through an arc suppression coil, grounded mode through a small resistor and the like; the grounding medium comprises common fault types such as intermittent arc grounding, stable arc grounding, earth grounding, resistance grounding and the like; the grounding resistance values are selected from typical values of 250 omega, 1000 omega, 2000 omega, 5000 omega and the like, and the grounding resistance values totally comprise 7 single-phase grounding fault types. The test respectively generates 420, 600 and 240 pieces of fault recording data under three grounding operation modes; the sampling frequency of the wave recording device is 10kHz, the sampling period comprises 12014 sampling points, and each section of wave recording data comprises 291 electric quantities.

further, the step (2) specifically comprises the following steps:

and (2.1) intercepting in a time period, wherein each section of single-phase earth fault recording data is not in a fault state in the whole sampling stage, and further comprises a normal state before fault and a transition state which is recovered to be normal after fault. In order to make the observed object more clear and reduce the analyzed data volume, the recording data in the fault state is intercepted. Through observation of the recorded wave data, the intercepted time period contains 600 sampling points.

(2.2) key feature division, wherein for 291 electrical quantities contained in the wave recording data, as many features of the 291 electrical quantities keep a constant value in the whole time period or show periodic regular changes, the significance for identifying the type of the single-phase earth fault is small. The method is obtained through experience of power distribution network operation and maintenance engineering personnel, and all electrical quantity characteristics contained in the wave recording data are divided into two parts: critical feature portions and non-critical feature portions.

Wherein the key features comprise: bus zero sequence current 3I ₀Bus zero-sequence voltage 3U₀I, 1 three-phase current 3I for vacuum culture₀ ^ⅠII, three-phase current of vacuum culture 1

III true culture 1 three-phase current

IV true culture 1 three-phase current

The remaining electrical quantity characteristics are non-critical characteristics.

And (2.3) performing Hilbert-yellow transformation processing on key characteristic parts of fault recording data of the power distribution network, and extracting an intrinsic mode function, an instantaneous amplitude, an instantaneous frequency and a Hilbert spectrum.

Further, the step (3) specifically includes:

(3.1) constructing a primary dataset. Seven fault categories of intermittent arc ground fault, 250 Ω resistance ground fault, 1000 Ω resistance ground fault, 2000 Ω resistance ground fault, 5000 Ω resistance ground fault, earth ground fault, stable arc ground fault, etc. are indicated by numbers 1, 2, …, 7. And (3) the fault recording data set processed in the step (2) and the fault label form a primary data set together.

(3.2) the constructed primary dataset is processed according to the following steps of 8: 2 into training and test sets. The training set is used for training the two basic learners, and the testing set is used for testing the learning effects of the two basic learners;

Further, the step (4) specifically includes:

TABLE 1

Further, the step (5) specifically includes:

And (5.3) extracting time sequence correlation characteristics of the data set by using the basic learner LSTM model after training is finished. When the features are extracted, the output layer of the LSTM model is removed, so the feature dimensionality obtained by extraction is the node number of the hidden layer and is 100.

Further, the step (6) specifically includes:

(6.1) constructing a secondary data set. The extracted feature dimensions of the ResNet18 and LSTM models are (1, 2048) and (1, 100), respectively. These two sets of signatures are concatenated and combined with the true type tags of single-phase ground faults (0-6 for 7 different types of single-phase ground fault types), to form a secondary data set, as shown in the following table.

TABLE 2

And (6.2) training a decision tree model. The decision tree model specifically chosen was the CART model, which was trained using the secondary dataset constructed in (6.1). And the optimal parameter setting of the CART model is found by utilizing a grid search algorithm.

(6.3) identifying a specific single-phase earth fault result based on the decision tree model in the secondary learner at the end of training, namely (6.2).

Claims

1. A multi-feature fusion single-phase earth fault type identification method based on deep learning is characterized by mainly comprising the following steps:

1) acquiring fault recording data acquired by a field fault recording device;

2) preprocessing fault recording data, and performing Hilbert-Huang transform (HHT) -based time-frequency decomposition on the preprocessed fault recording data to obtain corresponding time-frequency information characteristics;

3) constructing the time-frequency information characteristics of the fault recording data obtained by the processing in the step 2 into a primary data set;

4) building and training a basic learner Resnet18, and extracting complex nonlinear features in a data set;

5) building and training a basic learner LSTM, and extracting time sequence correlation characteristics in a data set;

6) and (4) splicing and fusing the complex nonlinear features and the time sequence correlation features obtained by learning and extracting in the steps (4) and (5), constructing a secondary data set, and identifying a specific single-phase earth fault type by combining a decision tree model.

2. The method for identifying the type of the multi-feature fusion single-phase earth fault based on the deep learning as claimed in claim 1, wherein fault recording data acquired by the power distribution network is preprocessed, and corresponding time-frequency information features are obtained through HHT time-frequency decomposition, and the method mainly comprises the following steps:

1) and intercepting fault recording data in a fault state stage.

2) All electrical quantity characteristics contained in the fault recording data are divided into a key characteristic part and a non-key characteristic part.

3) And performing HHT (Hilbert-Huang transform) on key characteristic parts of fault recording data, and extracting IMF (intrinsic mode function) components, instantaneous amplitude values, instantaneous frequencies and Hilbert spectrums.

4) And (4) splicing the result of HHT conversion of the key characteristic part in the step (3) with the original non-key characteristic part in the step (2).

5) And normalizing the spliced features, and converting the gray level image.

3. The method for identifying the type of the multi-feature fusion single-phase earth fault based on the deep learning as claimed in claim 1, wherein a basic learner Resnet18 is constructed and trained for extracting complex nonlinear features in a data set. The method mainly comprises the following steps:

1) and constructing a Resnet18 model of the basic learner. On the basis of the original structure of Resnet18, the invention combines the characteristics of fault recording data to make some changes, including the following: modifying the convolution kernel of the conv of the first layer into 3 multiplied by 3, and reserving a full connection layer and a softmax layer in the model training stage; and after the model training is finished, removing the full connection layer and the softmax layer when extracting the complex nonlinear features of the data set.

2) The basic learner Resnet18 model is trained. The constructed primary dataset is used to train the underlying learner Resnet 18. And when the verification precision of the model reaches the expected effect, saving the parameter setting of the model.

3) And extracting complex nonlinear features of the data set by using a Resnet18 model of a basic learner after training. The extracted feature dimension is the output dimension of the last convolutional layer of the Resnet18 model.

4. The method for identifying the type of the multi-feature fusion single-phase earth fault based on the deep learning of claim 1 is characterized in that a basic learner LSTM is constructed and trained for extracting time sequence correlation features in a data set. The method mainly comprises the following steps:

1) and (4) building an LSTM model of a basic learner. And the model structure of the LSTM is obtained by comparing the fault recording data with the experimental result of the combined adjustment of the hyper-parameters.

2) And training an LSTM model of a basic learner. The constructed primary dataset is used to train the base learner LSTM. And when the verification precision of the model reaches the expected effect, saving the parameter setting of the model.

3) And extracting time sequence correlation characteristics of the data set by using the basic learner LSTM model after training is finished. And removing the output layer of the LSTM model when extracting the characteristics, so that the characteristic dimensionality obtained by extraction is the number of nodes arranged in the last hidden layer.

5. The method for identifying the type of the single-phase earth fault based on the deep learning and the multi-feature fusion as claimed in claims 3 and 4, wherein the complex nonlinear features and the time sequence correlation features contained in the wave recording data are extracted based on a Resnet18 model and an LSTM model of a basic learner, and a secondary data set is constructed by combining a fault type label. And identifying a specific single-phase earth fault type by using the decision tree model. The method mainly comprises the following steps:

1) a secondary data set is constructed. Extracting complex nonlinear features and time sequence correlation features contained in the recording data based on a basic learner Resnet18 model and a basic learner LSTM model, splicing the two groups of features, and forming a secondary data set with real type labels (0-m-1 respectively represents m different types of single-phase earth fault types) of the single-phase earth fault.

2) And (5) training a decision tree model. The specifically selected decision tree model is the CART model, which is trained using the secondary dataset constructed in step 1. And finding the optimal parameter setting of the CART model by utilizing a grid search algorithm.

3) Based on the secondary learner at the end of training, a specific single-phase ground fault result is identified.