基于改进CNN-LSTM的电力系统宽频振荡辨识

Identification of Power System Wide-band Oscillation Based on Improved CNN-LSTM

宽频振荡的强非线性和强时变性会导致参数准确辨识困难,提出了改进CNN-LSTM神经网络的参数辨识方法。首先,采用卷积神经网络(CNN)提取输入的宽频振荡信号的振荡特征,并通过Softmax分类器输出振荡模态数量。然后,根据模态数量对长短期记忆网络(LSTM)辨识的模态个数进行定阶;同时,通过对CNN输出矩阵进行1×1卷积运算替代LSTM中的矩阵乘法,实现LSTM模型对高维输入的可行性。最后,以卷积运算结果作为LSTM的输入,辨识振荡频率和衰减因子。实测数据分析结果证明改进的CNN-LSTM对宽频振荡的频率和衰减因子都具有较高的辨识精度,在处理宽频振荡频率漂移现象和衰减因子突变等问题上有突出的优势。

It is difficult to accurately identify the parameters due to strong nonlinearity and strong time-varying nature. Therefore,the paper proposes a parameter identification method based on improved CNN-LSTM neural network. Firstly,convolutional neural network(CNN)is used to extract the oscillation characteristics from the input wide-band oscillation signal,and the number of oscillation modes is output through Softmax classifier. Then the order of the number of oscillation modes identified by long short-term memory(LSTM) network is determined according to the number of the modes;At the same time,1 x1-convolution instead of matrix multiplication in LSTM is used for CNN output matrix,making LSTM model suitable for high-dimensional input. Finally,the convolution result is used as the input of LSTM model to identify the oscillation frequency and attenuation factor. The analysis results of the measured data verify that the improved CNN-LSTM has high identification accuracy for the frequency and attenuation factor of the wide-band oscillation,as well as the outstanding advantages in dealing with the wide-band oscillation frequency drift and the sudden change of the attenuation factor.