| High voltage transmission system is one of the most important structures in power systems.However,due to the precipitous route of overhead transmission lines,the lines are often attacked by natural disasters such as lightning strikes,resulting in line failures becoming the most important threat factor to high-voltage transmission lines in addition to operational failures.Relevant statistics show that in recent years,tripping accidents caused by lightning strikes account for more than 48% of transmission line failures except operational failures,It can be seen that line faults caused by lightning have seriously threatened the safe operation of power systems.Therefore,accurately identifying the types of lightning overvoltage on transmission lines can effectively reduce the time and economic losses caused by faults,and plays a significant role in ensuring the safety and smooth operation of power supply systems.In order to explore accurate and efficient lightning overvoltage identification methods for transmission lines,this paper uses PSCAD/EMTDC software to simulate and build a lightning strike model for overhead transmission lines.These models include a dual exponential function model,a segmented wave impedance model,a zinc oxide arrester model,and a circuit breaker model.By changing parameters such as the type of lightning current model,initial phase angle,and lightning current amplitude,three fault waveforms of induced lightning stroke,counter lightning stroke,and bypass lightning stroke are simulated;The short circuit fault waveform is simulated by changing the short circuit transition resistance,short circuit type,and fault location.Then,the above four waveforms are recognized and classified through image recognition technology.The specific identification method is to use a deeper convolutional neural network-VGG-16 network model,and optimize and improve it based on the characteristics of lightning overvoltage,the main improvements are as follows: reducing the number of convolutional groups;Crop the obtained lightning waveform image to 224 × Size of224,changing the later convolution kernel to 2 × 2 convolution kernel in convolutional groups;Removes the full connectivity layer from traditional network structures.The experimental results show that the improved VGG-16 network through the method in Chapter 4 has the highest recognition rate,and the robustness and nonlinear ability of the model have been improved.This further demonstrates the effectiveness of the method in this paper.Compared to the traditional VGG-16 model,it can achieve better results,and both the accuracy and recall rates have been improved by more than 1%,It can be used to identify lightning overvoltage on transmission lines.In order to further improve the efficiency and accuracy of network classification and identification,Chapter 5 proposes an improved residual network(ResNet)method for lightning overvoltage identification of transmission lines.Improvements to the traditional ResNet-50 network include the following three aspects: First,replace the original 7 × Replace 7 large convolution cores with 3 3 × 3 small convolution cores to improve the feature extraction ability of the network;Secondly,replace the Re LU activation function of the traditional ResNet-50 network with the Re LU-Softplus activation function,which improves the network convergence rate while retaining data less than 0;Then,adjust the ResNet-50 network structure to strengthen the role of batch standardization.Finally,combined with migration learning and improved residual network,a lightning overvoltage identification model for transmission lines is constructed.The experimental results show that the improved residual network converges at step 70,with a recognition accuracy of 97.25%,and a recognition accuracy sum that is superior to other models. |
