Power grid real-time topology identification system and method based on neural network VGG model
Through improved VGG model and data preprocessing technology, the application problem of large models on embedded devices is solved, efficient, accurate identification and real-time monitoring of power grid topology are achieved, and the recognition accuracy is improved.
Patent Information
- Application Number
- CN202311834628.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-28
- Publication Date
- 2025-07-08
AI Technical Summary
The existing end-to-end perception network model is too large to be effectively applied on embedded devices, and traditional grid topology recognition methods are affected by noise and data incomplete identification, so the recognition accuracy is insufficient.
Using an improved VGG model, through data preprocessing and feature extraction, combined with hollow convolution and regularization technology, the model depth is reduced and feature extraction capabilities are improved, adapted to embedded devices, and the grid topology is monitored in real time.
It realizes efficient and accurate grid topology recognition on embedded devices, reduces computing complexity, improves recognition accuracy, and can monitor the status of the power system in real time.
Smart Images
Figure CN120278199A_ABST
Abstract
Description
[0001] Technical Field: The present invention relates to the field of electricity, and particularly to a power grid topology recognition technology that uses an improved VGG model to improve the recognition accuracy.
[0002] Background Art: The topology recognition of power systems is crucial for system monitoring and operation and maintenance. Traditional recognition methods are limited by noise, data incompleteness, and computational complexity. The currently available end-to-end perception network (MMNet) captures multi-scale features generated by different-sized filters and integrates two different levels of multi-scale features to improve recognition performance through end-to-end training. However, this method has an overly large model and requires a high-performance computer, and is not suitable for embedded devices. In the present invention, the VGG model is adopted and improved to achieve relatively accurate recognition with a smaller model. The present invention aims to apply this method to power grid topology recognition.
[0003] Summary of the Invention: The technical problem to be solved by the present invention is to overcome the problems existing in the above-mentioned prior art and provide a power grid real-time topology recognition system and method based on the neural network VGG model. To achieve the above invention objective, the present invention adopts the following technical solutions:
[0004] 1: Data acquisition: Deploy power line carrier communication devices, such as smart meters, device control units (DCUs), etc., at key nodes of the power system. These devices can transmit digital signals on the power line, including current, voltage, switch status, and other relevant data.
[0005] 2: Data reception and preprocessing: At the receiving end, the device receives the data signal on the power line. The data preprocessing module demodulates, denoises, and filters the received data for subsequent analysis.
[0006] 3: Data parsing and feature extraction: Parse the preprocessed data and convert the digital signal into the status information of each device in the power grid.
[0007] 4: Topology modeling and analysis: Based on the parsed and feature-extracted data, start topology modeling and analysis. This includes detecting information such as the connection status of power lines, the switch status of devices, and the current load distribution.
[0008] 5: Data analysis and visualization: Analyze the data obtained from the topology analysis to identify any anomalies or potential problems. Visualize the topology data so that operators can intuitively understand the status of the power system.
[0009] 6: Real-time monitoring: Continuously receive and analyze power line carrier data to monitor the status and topology of the power system in real time. The real-time monitoring system can detect any changes or problems in a timely manner and issue an alarm.
[0010] 7. Data preprocessing module, which adopts data two-dimensionalization to convert one-dimensional power grid data into a two-dimensional format to meet the input requirements of the convolutional neural network;
[0011] 8. Modeling and analysis module, which is implemented using an improved VGG model. The original 19-layer convolutional operation is changed to 9 layers, the fully connected layers are removed, the convolutional kernel size of each layer is adjusted, and dilated convolution is added to improve the feature extraction ability.
[0012] Furthermore, the data preprocessing in 2 adopts data two-dimensionalization: converting one-dimensional power grid data into a two-dimensional format to meet the input requirements of the convolutional neural network. This can be achieved by reorganizing the data into a two-dimensional matrix according to the time and sensor dimensions.
[0013] Furthermore, the modeling and analysis in step 4 are implemented using an improved VGG model. After the VGG is simplified, the network depth is insufficient and the feature extraction is not sufficient. Dilated Convolution, also known as atrous convolution or expansive convolution, is used in the last three layers. By introducing a dilation parameter in the convolutional kernel, the receptive field of the convolutional kernel is expanded, thereby increasing the receptive field range and feature extraction ability of the network. Adding regularization technology (Dropout) can accelerate the training speed of the VGG model. By randomly setting the outputs of some neurons to 0 during the training process, the dependence of the model on certain inputs can be reduced, thereby improving the generalization performance. The optimized algorithm can converge faster and extract effective features, shortening the training time. Overfitting is a common problem in neural network models, especially when there are many parameters. By introducing techniques such as Dropout, the overfitting phenomenon is effectively reduced.
[0014] Furthermore, the VGG model runs according to the following process:
[0015] Step1: Input image: Pass the input image to the network model for processing.
[0016] Step2: Convolutional layer: The VGG model constructs a feature extractor with multiple consecutive convolutional layers. Each convolutional layer consists of one or more convolutional kernels, and each convolutional kernel performs a convolutional operation on the input image to generate a set of feature maps. The convolutional operation is performed on the input image by means of a sliding window, and an activation function (such as ReLU) is used to introduce non-linearity.
[0017] Step3: Pooling layer: A pooling layer is added after each convolutional layer to reduce the spatial size of the feature map. Average pooling operation is usually used. By summarizing and reducing the dimension of the feature values in the local area, the main features are retained and the computational amount is reduced. The maximum pooling formula:
[0018]
[0019] Step 4: Output layer: The final predicted density map is output through three convolutional layers, and the last layer uses a 1×1 ordinary convolutional layer to output the result. Description of the drawings:
[0020] Figure 1 It is a comparison diagram between the improved VGG model and the original model.
[0021] Figure 2 It is a comparison diagram between ordinary convolution and dilated convolution.
[0022] Figure 3 It is the structure diagram of the real-time topology recognition system for this power grid. Specific implementation manner:
[0023] The present invention will be further described below with reference to the accompanying drawings. The following content is only used to more clearly illustrate the technical solution of the present invention and cannot be used to limit the protection scope of the present invention.
[0024] 1. Data acquisition: [Data reception and preprocessing module]
[0025] Power line carrier communication devices are deployed at key nodes (such as substations) in the power system. The specific data collected by these devices are:
[0026] Current value (in amperes): It is collected in real time through smart meters.
[0027] Voltage value (in volts): It is obtained through power line carrier communication devices.
[0028] Device switch state: It is obtained through the device control unit (DCU).
[0029] 2. Data preprocessing: [Data parsing and feature extraction module] [Data preprocessing module]
[0030] The received raw data is first subjected to a series of processes:
[0031] Demodulation: Convert the analog signal into a digital signal.
[0032] Denosing: Remove the noise that may be introduced during signal transmission.
[0033] Filtering: Use a digital filter to smooth the data and ensure that high-frequency noise does not affect the analysis result.
[0034] Subsequently, two-dimensional processing of the data is performed, and the one-dimensional time series data and sensor dimensions are reorganized into a two-dimensional matrix.
[0035] 3. Improvement of the neural network model: [Modeling and analysis module]
[0036] We use an improved VGG model for topology recognition. The improvements mainly include:
[0037] Atrous convolution operation: Introduce atrous convolution to improve the network's perception ability.
[0038] Regularization technique (Dropout): Randomly turn off neurons during training to prevent overfitting.
[0039] 4. Specific calculation steps for topology recognition: [Topology Modeling and Analysis Module]
[0040] Input the preprocessed data into the neural network for topology recognition:
[0041] Input layer: Input the current, voltage, and switch state data into the neural network.
[0042] Convolutional layer: Multiple convolutional layers extract the features of the input data.
[0043] Pooling layer: Reduce the spatial size of the feature map and retain the main features.
[0044] Output layer: The last convolutional layer outputs the predicted density map.
[0045] 5. Result analysis: [Real-time Monitoring Module] [Data Analysis and Visualization Module]
[0046] By analyzing the output results of the neural network:
[0047] Node recognition: Determine the main nodes of the power system by detecting the peak values in the density map.
[0048] Analyze the density map of the connection relationship, identify the connection relationship, and establish the topology of the power system.
Claims
1. A power grid real-time topology recognition system and method based on the neural network VGG model, characterized in that Including: (1) A data acquisition module for deploying power line carrier communication devices at key nodes of the power system. The devices can transmit digital signals on the power line, including current, voltage, switch status, and other relevant data; (2) A data receiving and preprocessing module for receiving data signals transmitted on the power line at the receiving end, and then performing processing such as demodulation, denoising, and filtering on the received data to prepare the data for subsequent analysis; (3) A data parsing and feature extraction module for parsing the preprocessed data and converting the digital signals into the status information of each device in the power grid; (4) A topology modeling and analysis module for performing topology modeling and analysis based on the parsed and feature-extracted data, including detecting connection status of power lines, switch status of devices, current load distribution, and other information; (5) A data analysis and visualization module for analyzing the data obtained from topology analysis to identify any abnormal conditions or potential problems, and visualizing the topology data so that operators can intuitively understand the status of the power system; (6) A real-time monitoring module for continuously receiving and analyzing power line carrier data to monitor the status and topology of the power system in real time, promptly detecting any changes or problems, and issuing alerts; (7) A data preprocessing module that uses data two-dimensionalization to convert one-dimensional power grid data into a two-dimensional format to meet the input requirements of the convolutional neural network; (8) A modeling and analysis module implemented using an improved VGG model, which includes convolutional layers, pooling layers, and output layers. The original 19-layer convolutional operation is changed to 9 layers, the fully connected layer is removed, and the convolutional kernel size of each layer is adjusted, and dilated convolution is added to improve the feature extraction ability.
2. The power grid real-time topology recognition system and method based on the neural network VGG model according to claim 1, characterized in that The improved VGG model includes a general dilated convolution operation to expand the receptive field range of the convolutional kernel.
3. The power grid real-time topology identification system and method based on the neural network VGG model according to claim 1, characterized in that, Data preprocessing uses data two-dimensionalization: converting one-dimensional power grid data into a two-dimensional format to meet the input requirements of the convolutional neural network; the improved VGG model includes a regularization technique (Dropout) to reduce overfitting. Dropout is a method to prevent overfitting. By randomly setting the outputs of some neurons to 0 during training, the dependence of the model on certain inputs can be reduced, thereby improving the generalization performance.
4. The power grid real-time topology recognition system and method based on the neural network VGG model according to claim 1, characterized in that Among them, the above-mentioned improved VGG model converges faster and extracts effective features, thereby shortening the training time.
5. The power grid real-time topology recognition system and method based on the neural network VGG model according to claim 1, characterized in that, The improved VGG model operates according to the following process: Step 1. Input: Pass the input data to the network model for processing; Step 2. Convolutional layer: Perform convolutional operations using convolutional kernels to generate a set of features; Step 3. Pooling layer: Add a pooling layer after each convolutional layer, using average pooling to reduce the size of the feature map. The following is the average pooling formula: Step 4. Output layer: Output the final prediction through convolution. The last layer uses a 1×1 convolutional layer to output the result.