A Convolution Neural Network Method for Power System Oscillation Type Identification

The identification and suppression of oscillation are very important for the stability and security of modern power system. The commonly used oscillation identification methods assume the types of oscillation according to experience. For example, FFT and Prony that are used to identify low frequency oscillation, regard the measured signal as a stationary and time-invariant signal of low frequency oscillation. These are not fit for the modern power system with high proportion of renewable energy, where the types of oscillation are diverse and the system status is fast time-varying. In this paper, a convolution neural network method is proposed for identification of oscillation type in the modern power system. Firstly, obtain the training data of neural network by modeling multi-order mixed exponentially damped sinusoid with random parameters and strong time-varying property. Then, transform the problem of frequency parameter estimation into a classification problem by establishing convolution neural network. Finally, consider average loss index based on cross entropy error function as the accuracy index and testing networks trained with different signal noise ratios and time-varying characteristics to acquire the optimal neural network. Simulations show that the proposed approach can identify oscillation types quickly and accurately.