A Method for Assessing Transient Overvoltage of AC/DC System Based on Deep Learning

摘要

In order to predict the transient over-voltage of AC/DC system during large disturbances effectively, this paper puts forward a method for assessing transient overvoltage based on deep learning, such as convolutional neural networks (CNN). Firstly, a non-linear network structure with multiple layers of hidden layers is built, based on the basic principle of CNN input feature construction. According to the topology of the power grid nodes, the voltage, phase angle and power of each node collected by the wide area measurement device are used as input layers. The relationship and the occurrence of the fault to the time sequence of the resection are spliced to obtain a matrix characterizing the state of the grid. Then, the hyperparameters of CNN are optimized and adjusted, and the gradient descent method is used for supervised training. The weight matrix between the input layer and the convolutional layer is optimized layer by layer to realize the automatic extraction of key eigenvalues, and the deep structure of CNN is used to construct the temporary structure. A mapping model between the overvoltage and the input data to quickly and accurately estimate the transient overvoltage of the AC/DC system. Finally, the modified Nordic32 AC/DC hybrid system is analyzed to verify the effectiveness and accuracy of the proposed method.