Classification of Broadband Oscillations Using Wavelet Convolution and Multi‐Channel Attention Network

Accurate identification of broadband oscillation types is a prerequisite for implementing appropriate control strategies. The strongly nonlinear, nonstationary and multi‐modal characteristics of broadband oscillation signals impose higher demands on identification methods. Practical applications face challenges such as coupling effects between different oscillation modes, noise interference in oscillation signals that degrade model identification and classification accuracy and the curse of dimensionality in multi‐dimensional monitoring data that increases algorithmic complexity. To address these challenges, this paper deeply integrates continuous wavelet transform with a multi‐channel attention network to propose a broadband oscillation classification method based on wavelet transform convolutional neural network‐multi‐channel attention network (WTCNN‐MCANet). First, the WTCNN module performs continuous complex wavelet transform on raw signals and employs an adaptive scale selection algorithm to generate time–frequency feature maps. Subsequently, a multi‐scale convolutional attention layer with parallel convolutional paths further extracts time–frequency features from different frequency bands, whereas a channel attention mechanism dynamically weights the contribution of each frequency band. This approach achieves an optimal balance among feature extraction efficiency, classification accuracy and model generalisation capability, providing a novel methodology for rapid and accurate identification of broadband oscillation types in new‐generation power systems.