Abstract：The S-transform overcomes the defect that the window width of the short-time Fourier transform is fixed, but the S-transform window function changes with the trend of frequency, and the generalized S-transform realizes the control of the Gaussian window by changing the variance of the Gaussian function, but it is also difficult to select the parameters of the window function. According to the correlation between the time maximum amplitude curve and frequency maximum amplitude curve of generalized S-transform time-frequency matrix and the amplitude and spectrum envelope of power quality signal, an adaptive selection method of Gaussian window function parameters by optimizing generalized S-transform is proposed, which fully preserves the amplitude and frequency characteristics of power quality disturbance. Artificial feature extraction method is difficult to effectively extract deep disturbance features from power quality signals, while deep learning method can automatically extract abstract features from data, but only using a single input for power quality signals leads to inadequate feature extraction and low recognition accuracy for complex power quality disturbances. A hybrid input neural network architecture is proposed. Firstly, convolution neural network is used to automatically extract features from the time-frequency matrix obtained from the optimized generalized S-transform, and then one-dimensional convolutional neural network is used to automatically extract features from the original one-dimensional time sequence signal, and the long short term memory network module is added to enhance the feature extraction effect of the time sequence signal. Finally, the features extracted from the two inputs are combined and the disturbance type is identified by the full connection layer. Through the training and verification of the simulation data set containing 26 kinds of power quality disturbance signals, the results show that the disturbance recognition accuracy of the hybrid input neural network model proposed in this paper is 99.77%, which is higher than that of the traditional single input neural network.
2025, Vol. 46 
