Ship DC micro-grid fault pole selection method
By generating time-frequency grayscale images using the valve-side current and voltage of smoothing reactors in a ship's DC microgrid, and combining this with a convolutional neural network to train a fault polarity selection model, the problem of difficulty in distinguishing fault types in existing technologies is solved, achieving fast and accurate fault polarity selection and protection.
Patent Information
- Application Number
- CN202511101293.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-07
- Publication Date
- 2025-10-21
AI Technical Summary
Existing fault diagnosis methods for ship DC microgrids mainly rely on analytical modeling of signal processing, which makes it difficult to distinguish fault types in ships with short lines and small line impedances. Furthermore, the application of deep learning algorithms in high-voltage AC and DC transmission lines is not yet widespread, resulting in slow fault selection speed and insufficient noise interference resistance.
By utilizing the current and voltage on the valve side of the smoothing reactor in the transmission line of the ship's DC microgrid generator, a time-frequency grayscale image is generated through S-transform. A fault polarity selection model is trained using a convolutional neural network, and fault identification and polarity selection are performed solely based on rectifier-side data, enabling rapid identification of faulty lines.
It achieves fast and accurate fault polarity selection, can disconnect faulty lines in a timely manner, reduce the spread of faults, improves the ability to resist noise interference and transition resistance, and supports the rapid recovery of DC microgrids.
Smart Images
Figure CN120824744A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of ship power systems, and in particular to a fault pole selection method for a ship DC microgrid. Background Art
[0002] Compared to traditional AC grids, marine DC microgrids, as a ship's electric propulsion system, effectively eliminate prime mover speed restrictions, enabling stable grid connection and operation of generator sets with large frequency differences, while reducing prime mover fuel consumption. The use of power electronics in DC microgrids eliminates the need for large-capacity distribution transformers, reducing equipment size and increasing available space on board. Furthermore, DC microgrids effectively avoid the reactive power issues associated with AC systems caused by voltage and current phase differences and significantly reduce cable weight.
[0003] Current methods for fault diagnosis in ship power systems primarily rely on analytical modeling based on signal processing. Unlike traditional HVDC systems, ship medium-voltage DC systems have shorter cable lines and very low line impedance. Therefore, when a fault occurs in a DC system line at either pole, relying solely on the short-circuit fault current setting as a basis for judgment makes it difficult to distinguish the fault type. Deep learning algorithms offer strong signal processing capabilities and robust noise immunity. Currently, fault diagnosis and pole selection methods based on deep learning algorithms are primarily applied to high-voltage AC / DC transmission lines, but their application to shipboard DC microgrids remains largely unexplored. Summary of the Invention
[0004] In response to the defects in the prior art, the purpose of the present invention is to provide a method for selecting poles for faults in a ship DC microgrid. The method uses the current and voltage on the valve side of the smoothing reactor in the transmission line at the generating end of the ship DC microgrid to train a pole selection model. When a line fault occurs, the fault data is used for fault identification. There is no need to collect data from the other ends of the multi-terminal network except the generating end. The pole selection speed is fast and the method has good resistance to noise interference and transition resistance.
[0005] In order to solve the above problems, the technical solution of the present invention is:
[0006] A method for selecting poles in a ship DC microgrid fault, comprising the following steps:
[0007] Simulate the ship DC microgrid model in different working states, different fault locations, and different fault types, and collect the fault voltage and fault current values on the valve side of the smoothing reactor on the rectifier side;
[0008] Use S-transform to perform dimension transformation and time domain analysis on the collected feature data to generate a time-frequency grayscale image;
[0009] The generated grayscale image is used to learn and train the fault selection model through a convolutional neural network;
[0010] When a line fault occurs, the real-time voltage and current values collected on the valve side of the smoothing reactor on the rectifier side are used through the pole selection model to select the pole corresponding to the faulty line.
[0011] According to the pole selection results, the auxiliary relay protection system disconnects the corresponding fault line.
[0012] Preferably, the step of using S-transform to perform dimensional transformation and time domain analysis on the collected feature data to generate a time-frequency grayscale image specifically includes: each fault data contains four sets of feature data, S-transform is performed on each set of features to obtain its corresponding time-frequency graph, and the four feature data are processed in the same way to obtain a set of time-frequency graphs; the image is cropped to remove other targets in the image, and the image is resized so that the image can be processed by a convolutional neural network, and the processed data is divided into a training set and a test set, wherein the training set is 80% of the total data and the test set is 20% of the total data.
[0013] Preferably, in the step of using the generated grayscale image to learn and train the fault selection model through a convolutional neural network, the fault selection model training uses a convolutional neural network:
[0014]
[0015] Where l is the number of network layers, B represents the bias, represents the l-th output of layer r, represents the i-th output of the r-1 layer, Represents convolution; the images of the training set are sequentially subjected to a 3×3 convolution layer and a pooling layer with a pooling radius of 2, a 3×3 convolution layer and a pooling layer with a pooling radius of 2 for feature extraction in the convolutional neural network, and an adaptive decay learning rate is used.
[0016] Preferably, in the step of using the generated grayscale image to learn and train the fault pole selection model through a convolutional neural network, the pole selection and testing process includes: adding a Dropout layer after each layer of convolution, with a rate of 25% and 50% respectively, to prevent the model from overfitting; the fully connected layer uses 256 neurons and a softmax activation function to convert the neuron output into a probability distribution; the test set reflects the fault recognition performance of the model, and the accuracy is selected as the evaluation indicator of the model.
[0017] Preferably, the step of assisting the relay protection system in disconnecting the corresponding fault line according to the pole selection result specifically includes: when a fault occurs in the ship's DC microgrid line, if the pole selection system determines that it is a positive line fault, then disconnecting the positive line in the system; if it is a negative line fault, then disconnecting the negative line; if it is an inter-pole fault, then disconnecting the two-pole lines, so as to avoid further expansion of the fault, and facilitate subsequent fault troubleshooting and timely restoration of DC microgrid operation.
[0018] Compared with the existing technology, the present invention's ship DC microgrid fault pole selection method uses the current and voltage on the valve side of the smoothing reactor in the power transmission line at the generating end of the ship DC microgrid to train the pole selection model. When a line fault occurs, the fault is identified based on the fault data. There is no need to collect data from the other ends of the multi-terminal network except the generating end, and the task of "offline training and online pole selection" can be completed. Only the fault data on the rectifier side is used, the pole selection speed is fast, and the method has good resistance to noise interference and transition resistance. BRIEF DESCRIPTION OF THE DRAWINGS
[0019] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments with reference to the following drawings:
[0020] Figure 1 This is a flowchart of the fault pole selection method for a ship DC microgrid according to the present invention;
[0021] Figure 2 This is a schematic diagram of the process of the fault pole selection method for a ship DC microgrid according to the present invention;
[0022] Figure 3 This is a schematic diagram of the topological structure of the ship DC microgrid;
[0023] Figure 4 Schematic diagram of the convolutional neural network structure;
[0024] Figure 5 Confusion matrix diagram of the fault selection model results in the test set. DETAILED DESCRIPTION
[0025] The present invention will be described in detail below with reference to specific embodiments. The following examples will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those skilled in the art, several changes and improvements can be made without departing from the scope of the present invention. These all fall within the scope of protection of the present invention.
[0026] Specifically, the present invention provides a method for selecting poles in a ship DC microgrid fault. Figure 1 and Figure 2 As shown, the method includes the following steps:
[0027] S1: Simulate the ship DC microgrid model in different working states, different fault locations, and different fault types, and collect the fault voltage and fault current values on the valve side of the smoothing reactor on the rectifier side;
[0028] Specifically, Figure 3 This is a ±15kV four-terminal DC microgrid model for ships. The model simulates different ship power operating conditions, using a power step size of 5% of the ship's full-load power. Different fault locations are set, with fault points set at intervals of 10% of the total transmission line length. The fault resistance ranges from 0 to 100Ω, with a step size of 5Ω. Faults are set on the positive and negative lines of the DC line, as well as between poles. After setting these different faults, the positive and negative fault voltages and currents on the valve side of the flat reactor on the generator side of the rectifier are collected as characteristic data.
[0029] S2: Use S transform to perform dimension transformation and time domain analysis on the collected feature data to generate a time-frequency grayscale image;
[0030] Specifically, the feature data collected in the previous step is first normalized, and each group of processed data contains four feature data. The processed data is subjected to S transformation. First, the feature data undergoes fast Fourier transform to generate a complex array, where the complex number represents the amplitude and phase information. For each frequency, a Gaussian window in the frequency domain is constructed, and the Gaussian window is convolved with the Fourier transformed signal to localize the frequency information to different positions on the time axis. The convolution result is inverse Fourier transformed to obtain a complex matrix, which represents the signal intensity and phase of different frequencies at different time points. The image is cropped to remove other targets in the image and the image is resized so that the image can be processed by the convolutional neural network. The amplitude spectrum of the processed time-frequency graph is taken to generate a two-dimensional grayscale image. The processed grayscale image is divided into a training set and a test set, where the training set is 80% of the total data and the test set is 20% of the total data.
[0031] S3: The generated grayscale image is used to learn and train the fault selection model through a convolutional neural network;
[0032] Specifically, the grayscale images in the training set are used to learn and train the fault selection model through a convolutional neural network. The convolutional neural network structure is as follows: Figure 4As shown, the grayscale image is first passed through a 3×3 convolutional layer followed by a 2×2 pooling layer, and then a 3×3 convolutional layer followed by a 2×2 pooling layer for feature extraction. The ReLU activation function is used to avoid the limitations of linear expression and accelerate training. An adaptive learning rate decay is used; if the validation loss does not improve after three consecutive cycles, the current learning rate is halved. Dropout layers are added after each convolution layer, with values of 25% and 50%, respectively, to effectively prevent model overfitting. The fully connected layer uses 256 neurons and a softmax activation function to convert the neuron output into a probability distribution, outputting three nodes for positive pole faults, negative pole faults, and inter-pole faults. The test data set is used to test the performance and accuracy of the pole selection model. If the model accuracy and speed meet the requirements, offline training is completed and the model is saved for actual operation when a line fault occurs.
[0033] The pole selection model training is done using convolutional neural network
[0034]
[0035] Where l is the number of network layers, B represents the bias, represents the l-th output of layer r, represents the i-th output of the r-1 layer, represents convolution;
[0036] The images of the training set are sequentially subjected to a 3×3 convolution layer and a pooling layer with a pooling radius of 2, and a 3×3 convolution layer and a pooling layer with a pooling radius of 2 for feature extraction in the convolutional neural network, and an adaptive decay learning rate is used.
[0037] S4: When a line fault occurs, the collected real-time voltage and current values on the valve side of the smoothing reactor on the rectifier side are passed through the pole selection model to select the pole corresponding to the faulty line.
[0038] Specifically, the trained pole selection model is applied to a shipboard DC microgrid. When the microgrid's relay protection system detects a transmission line fault, the real-time voltage and current values collected at the positive and negative poles of the valve side of the rectifier-side smoothing reactor at the time of the fault are used by the pole selection model to identify the faulted line as a positive line, a negative line, or an inter-pole fault. This completes the online pole selection task, providing technical support for the subsequent relay protection system to quickly disconnect the fault and protect the DC microgrid system.
[0039] S5: Based on the pole selection result, the auxiliary relay protection system disconnects the corresponding fault line.
[0040] Specifically, when a fault occurs in a ship's DC microgrid line, if the pole selection system determines that the positive line is faulty, the positive line in the system is disconnected; if the fault is in the negative line, the negative line is disconnected; if the fault is between poles, both poles are disconnected. This prevents the fault from further expanding, facilitates subsequent troubleshooting, and promptly restores the DC microgrid operation.
[0041] like Figure 5 As shown in the figure, after training the ship DC microgrid fault pole selection model using the training set, the model's performance was tested using the test set. The figure shows the confusion matrix of the test results. As can be seen from the figure, the recognition rate for positive line faults in the test set was 100%, fully meeting the pole selection requirements. The pole selection accuracy for positive line faults and inter-pole faults was 99.67% and 99.42%, respectively. The misclassification of the negative line as an inter-pole fault occurred in 0.33% of cases. In this case, the ship DC microgrid system disconnected both lines. Although the disconnection expanded the fault area, the timely disconnection prevented the fault from expanding, successfully protecting the DC microgrid system. Therefore, the confusion matrix of the test set shows that the pole selection model has a high pole selection accuracy.
[0042] Table 1 below shows the simulation results of the ±15kV four-terminal DC microgrid transmission line fault pole selection method involved in this embodiment when ground faults occur at different locations of the line through different transition resistances under different loads of the ship DC microgrid.
[0043]
[0044] Table 1
[0045] As can be seen from the table, the proposed method for detecting faults in a ship's DC microgrid accurately identifies the polarity of the fault line under varying ship loads, at varying fault locations, and through varying transition resistances. It demonstrates excellent reliability, high accuracy, and robustness against noise interference and transition resistance.
[0046] The above describes specific embodiments of the present invention. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art may make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. The embodiments of this application and the features in the embodiments may be combined with each other in any manner unless there is a conflict.
Claims
1. A method for selecting poles in a ship DC microgrid fault, characterized in that: The method comprises the following steps: Simulate the ship DC microgrid model in different working states, different fault locations, and different fault types, and collect the fault voltage and fault current values on the valve side of the smoothing reactor on the rectifier side; Use S-transform to perform dimension transformation and time domain analysis on the collected feature data to generate a time-frequency grayscale image; The generated grayscale image is used to learn and train the fault selection model through a convolutional neural network; When a line fault occurs, the real-time voltage and current values collected on the valve side of the smoothing reactor on the rectifier side are used through the pole selection model to select the pole corresponding to the faulty line. According to the pole selection results, the auxiliary relay protection system disconnects the corresponding fault line.
2. The method for selecting poles in a ship DC microgrid fault according to claim 1, characterized in that: The step of using S-transform to perform dimensional transformation and time domain analysis on the collected feature data to generate a time-frequency grayscale image specifically includes: each fault data contains four groups of feature data, S-transform is performed on each group of features to obtain its corresponding time-frequency graph, and the four types of feature data are processed in the same way to obtain a group of time-frequency graphs; the image is cropped to remove other objects in the image, and the image is resized so that the image can be processed by a convolutional neural network, and the processed data is divided into a training set and a test set, wherein the training set is 80% of the total data and the test set is 20% of the total data.
3. The method for selecting poles in a ship DC microgrid fault according to claim 1, characterized in that: In the step of learning and training the fault selection model using a convolutional neural network for the generated grayscale image, the fault selection model training uses a convolutional neural network: Where l is the number of network layers, B represents the bias, represents the l-th output of layer r, represents the i-th output of the r-1 layer, Represents convolution; the images of the training set are sequentially subjected to a 3×3 convolution layer and a pooling layer with a pooling radius of 2, a 3×3 convolution layer and a pooling layer with a pooling radius of 2 for feature extraction in the convolutional neural network, and an adaptive decay learning rate is used.
4. The method for selecting poles in a ship DC microgrid fault according to claim 1, characterized in that: In the step of using the generated grayscale image to learn and train the fault pole selection model through a convolutional neural network, the pole selection and testing process includes: adding a Dropout layer after each convolution layer, with a rate of 25% and 50% respectively, to prevent the model from overfitting; using 256 neurons in the fully connected layer and a softmax activation function to convert the neuron output into a probability distribution; the test set reflects the fault recognition performance of the model, and the accuracy is selected as the evaluation indicator of the model.
5. The method for selecting poles in a ship DC microgrid fault according to claim 1, characterized in that: The step of assisting the relay protection system in disconnecting the corresponding fault line based on the pole selection result specifically includes: when a fault occurs in the ship's DC microgrid line, if the pole selection system determines that the positive line is faulty, the positive line in the system is disconnected; if the fault is a negative line, the negative line is disconnected; if the fault is an inter-pole fault, the two-pole lines are disconnected, thereby preventing the fault from further expanding and facilitating subsequent fault elimination and timely restoration of the DC microgrid operation.