CWT-CNN-based high-voltage DC power transmission system fault identification method

Abstract

The invention provides a CWT-CNN (Continuous Wavelet Transform-Convolutional Neural Network)-based high-voltage direct-current power transmission system fault identification method, which comprises the following steps: collecting a trigger angle signal of a single end of a high-voltage direct-current power transmission system, and carrying out normalization processing on the trigger angle signal; performing time-frequency analysis on the normalized trigger angle signal by using continuous wavelet transform, extracting local disturbance characteristics and generating a two-dimensional grayscale image; dividing the grayscale image into a training set and a test set, training a convolutional neural network model by using the training set, extracting deep features, and verifying and storing optimal model parameters by using the test set; performing fault identification on the network output by using a Softmax classifier to obtain a probability distribution result corresponding to a fault type; the performance of the model is evaluated according to the recognition result, and a basis is provided for subsequent protection strategy and control system application. The method has the advantages of being simple in model structure, high in feature extraction capacity, good in fault recognition effect, high in anti-noise capacity and the like.

Description

Technical Field

[0001] This invention relates to the field of power system fault diagnosis technology, and more specifically, to a fault identification method for high-voltage direct current transmission systems using continuous wavelet transform-convolutional neural network (CWT-CNN). Background Technology

[0002] High-voltage direct current (HVDC) transmission is widely used in modern power systems due to its advantages such as long distance, large capacity, and grid connection flexibility. However, the complex operating environment of DC systems leads to a variety of fault types, and traditional protection strategies (such as single-ended sampling protection and dual-ended communication protection) often have problems in accurately distinguishing between external faults and high-resistance line faults.

[0003] In recent years, artificial intelligence has been widely studied in power system fault diagnosis, with deep learning methods possessing powerful feature extraction capabilities that can effectively improve identification accuracy. However, in practical engineering applications, problems such as inaccurate identification of high-resistance grounding faults, severe external disturbances, and model dependence on complex physical features are frequently encountered. Especially in scenarios where information from the opposite side is unavailable or communication delays exist, single-ended voltage and current signals are often insufficient to accurately determine the fault type and polarity. Furthermore, while some neural network-based methods possess strong classification capabilities, their input feature selection lacks engineering specificity, model training is highly dependent on fault data, and their generalization ability is insufficient, making direct deployment in real-world systems difficult. Therefore, there is an urgent need for an effective method based on more stable and representative single-ended signals that can achieve fault type identification even under conditions of incomplete communication. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a fault identification method for high-voltage direct current transmission systems based on CWT-CNN. This invention can achieve high-precision fault identification using only single-end firing angle information. It has advantages such as simple model structure, strong feature extraction capability, good identification effect for high-resistance faults, and strong anti-noise capability. It is applicable to the construction of fault detection and protection criteria for high-voltage direct current transmission systems under various complex operating conditions.

[0005] To solve the above problems, the technical solution of the present invention is as follows:

[0006] A fault identification method for high-voltage direct current transmission systems based on CWT-CNN includes the following steps:

[0007] The firing angle signal of a single end of a high-voltage direct current transmission system is collected and normalized.

[0008] The normalized firing angle signal is analyzed in time and frequency using continuous wavelet transform to extract local disturbance features and generate a two-dimensional grayscale image.

[0009] The grayscale images are divided into training and test sets. The training set is used to train the convolutional neural network model and extract deep features. At the same time, the test set is used to verify and save the optimal model parameters.

[0010] The Softmax classifier is used to identify faults in the network output, and the probability distribution results corresponding to the fault types are obtained.

[0011] The performance of the model is evaluated based on the identification results, providing a basis for subsequent protection strategies and control system applications.

[0012] Preferably, in the step of acquiring the firing angle signal of a single end of the high-voltage direct current transmission system and performing normalization processing, the normalization processing adopts the minimum-maximum normalization method, the expression of which is:

[0013] x'=(xx min ) / (x max -x min )

[0014] Where x is the original trigger angle value, x min With x max These are the minimum and maximum values ​​of the signal, respectively.

[0015] Preferably, the step of performing time-frequency analysis on the normalized trigger angle signal using continuous wavelet transform, extracting local perturbation features, and generating a two-dimensional grayscale image specifically includes: performing continuous wavelet transform on the normalized trigger angle signal. Continuous wavelet transform is used to convert the time-domain information of the signal into time-frequency domain features and extract multi-scale features of the signal. The continuous wavelet transform uses the Mexican Hat mother wavelet function. Through continuous wavelet transform, the signal is decomposed into different frequency components, and corresponding two-dimensional grayscale images are generated as input data for the CNN model.

[0016] Preferably, in the step of dividing the grayscale image into a training set and a test set, using the training set to train the convolutional neural network model, extracting deep features, and using the test set to verify and save the optimal model parameters, the convolutional neural network model includes two sets of 3×3 convolutional layers, a ReLU activation function and a 2×2 pooling layer, followed by a fully connected layer containing 256 neurons and a Softmax classifier. During network training, 25% and 50% Dropout layers are set respectively to prevent overfitting.

[0017] Preferably, the step of using a Softmax classifier to identify faults in the network output and obtain the probability distribution results corresponding to the fault types specifically includes: using the trained CNN model, using a Softmax classifier to classify the fault types in the network output results. The Softmax classifier converts the output layer of the convolutional neural network into a probability distribution, and the model output is mapped to four categories: positive electrode fault, negative electrode fault, inter-electrode fault, and out-of-area fault. A corresponding probability value is generated for each category as the fault type identification result output.

[0018] Preferably, the step of evaluating the model's performance based on the identification results to provide a basis for subsequent protection strategies and control system applications specifically includes: evaluating the model's performance based on the identification results, including the model's accuracy, feature extraction capability, and noise resistance index; and based on these evaluation indexes and combined with the requirements of the actual system, providing data support for the protection system and helping to decide whether to implement subsequent protection measures.

[0019] Compared with existing technologies, this invention presents a fault identification method for high-voltage direct current (HVDC) transmission systems based on CWT-CNN. It utilizes control disturbance information reflected by the converter firing angle signal as a feature source, combining continuous wavelet transform and convolutional neural networks to extract fault features. This enables the identification of faults within and outside the transmission zone, as well as fault types. The method boasts advantages such as simple model structure, strong feature extraction capability, good fault identification performance, and strong noise resistance. It is applicable to fault detection and protection criterion construction in HVDC transmission systems under various complex operating conditions. Furthermore, this method has low sampling rate requirements, is easy to implement in hardware, and is suitable for large-scale power systems. Attached Figure Description

[0020] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0021] Figure 1 This is a flowchart of the fault identification method for high-voltage direct current transmission systems based on CWT-CNN of the present invention;

[0022] Figure 2 This is a detailed flowchart of the fault identification method for high-voltage direct current transmission systems based on CWT-CNN of the present invention;

[0023] Figure 3 This is a schematic diagram comparing the simulation results of different schemes for the system when different faults occur in a high-voltage direct current transmission line. Detailed Implementation

[0024] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0025] Specifically, this invention provides a fault identification method for high-voltage direct current transmission systems based on CWT-CNN, such as... Figure 1 As shown, the method includes the following steps:

[0026] S1: Collect the firing angle signal of a single end of the high-voltage direct current transmission system and normalize it;

[0027] Specifically, the firing angle signal at one end of the high-voltage direct current (HVDC) transmission system is acquired. This signal, generated by the converter's control system, reflects the converter's operating status. For each converter in the system, the firing angle signal provides effective information regarding current control and voltage regulation. The acquired signal is then normalized to eliminate amplitude differences in input signals between different systems or over different time periods, ensuring the stability and accuracy of subsequent processing.

[0028] The normalization process employs the minimum-maximum normalization method, the expression of which is:

[0029] x'=(xx min ) / (x max -x min )

[0030] Where x is the original trigger angle value, x min With x max These are the minimum and maximum values ​​of the signal, respectively.

[0031] S2: Use continuous wavelet transform to perform time-frequency analysis on the normalized firing angle signal, extract local disturbance features and generate a two-dimensional grayscale image;

[0032] Specifically, a continuous wavelet transform (CWT) is performed on the normalized firing angle signal. The CWT effectively converts the time-domain information of the signal into time-frequency domain features and can extract multi-scale features. In this invention, the CWT uses the Mexican Hat mother wavelet function, which has good time-frequency localization characteristics and can effectively capture local disturbances in the signal. Through the CWT, the signal is decomposed into different frequency components, and corresponding two-dimensional grayscale images are generated as input data for the CNN model.

[0033] S3: Divide the grayscale image into a training set and a test set. Use the training set to train the convolutional neural network model and extract deep features. At the same time, use the test set to verify and save the optimal model parameters.

[0034] Specifically, such as Figure 2 As shown, the generated 2D grayscale images are divided into training and test sets. A convolutional neural network (CNN) model is constructed and trained on the training set. The CNN model includes multiple convolutional and pooling layers. Convolutional operations extract high-level features of the signal, while pooling layers reduce the image dimensionality, thus reducing computational complexity. During training, the CNN extracts grayscale image features through convolutional and pooling layers and maps them to fault type classification labels through fully connected layers. After training, the model is validated using the test set to ensure its generalization ability and select the optimal model parameters.

[0035] In this embodiment, the convolutional neural network model includes two sets of 3×3 convolutional layers, a ReLU activation function, and a 2×2 pooling layer. After pooling, a fully connected layer containing 256 neurons and a Softmax classifier are added. During network training, 25% and 50% Dropout layers are set respectively to prevent overfitting.

[0036] S4: Use the Softmax classifier to identify faults in the network output and obtain the probability distribution results corresponding to the fault types;

[0037] Specifically, the trained CNN model is used to classify fault types using a Softmax classifier. The Softmax classifier transforms the output layer of the convolutional neural network into a probability distribution, mapping the model's output to four categories (positive electrode fault, negative electrode fault, inter-electrode fault, and out-of-zone fault), and generating a corresponding probability value for each category as the fault type identification result. This method effectively distinguishes different types of faults and has high identification accuracy, especially in complex fault scenarios.

[0038] S5: Evaluate the model's performance based on the identification results to provide a basis for subsequent protection strategies and control system applications.

[0039] Specifically, the model's performance is evaluated based on the identification results. The evaluation results include metrics such as model accuracy, feature extraction capability, and noise resistance. Based on these evaluation metrics and the requirements of the actual system, data support can be provided to the protection system, helping to decide whether to execute subsequent protection measures, such as restart commands or other fault recovery operations. By analyzing the output results of fault types, the nature of the fault can be determined more accurately, leading to more efficient and reliable protection decisions.

[0040] like Figure 3 As shown, simulation analysis of the protection device of this embodiment under different types of faults in high-voltage DC transmission lines shows that the fault recognition model based on CWT-CNN developed using the training set on the test set has improved fault recognition rate and model convergence speed compared with the CNN-based model and the TW-CNN model.

[0041] Table 1 below shows the simulation results of the fault identification method for high-voltage direct current transmission systems using control system characteristics and CWT-CNN in this embodiment when there is a high-resistance fault at the end of the line.

[0042] To verify the performance of the model in identifying and distinguishing between faults within and outside the fault zone, the proposed scheme was compared with the fault identification schemes proposed in Scheme A and Scheme B. Scheme A: Utilizes the voltage waveform characteristics at both ends; Scheme B: Compares the voltage and current setpoints of the smoothing reactor on the rectifier side with the setpoints on the reactor valve side.

[0043]

[0044] Table 1

[0045] As can be seen from Table 1, the fault identification method for high voltage direct current transmission systems based on CWT-CNN proposed in this invention achieves high accuracy in identifying high resistance faults at the end of the line.

[0046] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A fault identification method for high-voltage direct current transmission systems based on CWT-CNN, characterized in that, The method includes the following steps: The firing angle signal of a single end of a high-voltage direct current transmission system is collected and normalized. The normalized firing angle signal is analyzed in time and frequency using continuous wavelet transform to extract local disturbance features and generate a two-dimensional grayscale image. The grayscale images are divided into training and test sets. The training set is used to train the convolutional neural network model and extract deep features. At the same time, the test set is used to verify and save the optimal model parameters. The Softmax classifier is used to identify faults in the network output, and the probability distribution results corresponding to the fault types are obtained. The performance of the model is evaluated based on the identification results, providing a basis for subsequent protection strategies and control system applications.

2. The fault identification method for high-voltage direct current transmission systems based on CWT-CNN according to claim 1, characterized in that, In the step of acquiring the firing angle signal of a single end of the high-voltage direct current transmission system and normalizing it, the normalization process adopts the minimum-maximum normalization method, the expression of which is: x'=(x-x min ) / (x max -x min ) Where x is the original trigger angle value, x min With x max These are the minimum and maximum values ​​of the signal, respectively.

3. The fault identification method for high-voltage direct current transmission systems based on CWT-CNN according to claim 1, characterized in that, The step of performing time-frequency analysis on the normalized trigger angle signal using continuous wavelet transform, extracting local perturbation features, and generating a two-dimensional grayscale image specifically includes: performing continuous wavelet transform on the normalized trigger angle signal. Continuous wavelet transform is used to convert the time-domain information of the signal into time-frequency domain features and extract multi-scale features of the signal. The continuous wavelet transform uses the Mexican Hat mother wavelet function. Through continuous wavelet transform, the signal is decomposed into different frequency components, and corresponding two-dimensional grayscale images are generated as input data for the CNN model.

4. The fault identification method for high-voltage direct current transmission systems based on CWT-CNN according to claim 1, characterized in that, In the steps of dividing grayscale images into training and testing sets, using the training set to train the convolutional neural network model, extracting deep features, and using the testing set to verify and save the optimal model parameters, the convolutional neural network model includes two sets of 3×3 convolutional layers, ReLU activation function and 2×2 pooling layer, followed by a fully connected layer containing 256 neurons and a Softmax classifier. During network training, 25% and 50% Dropout layers are set respectively to prevent overfitting.

5. The fault identification method for high-voltage direct current transmission systems based on CWT-CNN according to claim 1, characterized in that, The step of using a Softmax classifier to identify faults in the network output and obtain the probability distribution results corresponding to the fault types specifically includes: using the trained CNN model, the Softmax classifier is used to classify the fault types in the network output results. The Softmax classifier converts the output layer of the convolutional neural network into a probability distribution. The output of the model is mapped to four categories: positive electrode fault, negative electrode fault, inter-electrode fault, and out-of-area fault. A corresponding probability value is generated for each category as the output of the fault type identification result.

6. The fault identification method for high-voltage direct current transmission systems based on CWT-CNN according to claim 1, characterized in that, The step of evaluating the model's performance based on the identification results to provide a basis for subsequent protection strategies and control system applications specifically includes: evaluating the model's performance based on the identification results, including the model's accuracy, feature extraction capability, and noise resistance index; and based on these evaluation indexes and the requirements of the actual system, providing data support for the protection system and helping to decide whether to implement subsequent protection measures.

Images