<p>Traditional composite analysis in cyclone studies, which compares variables at fixed grid points relative to the cyclone center, is limited by spatial misalignments caused by cyclone rotation. This misalignment blurs composite results and hinders accurate structural analysis. To address this issue, we propose the Cyclone Alignment Network (CAN) method. CAN aligns variables into a unified coordinate system by learning an affine transformation matrix, improving classification and composite results. Specifically designed for cyclones, CAN utilizes a Transformer structure with Rotary Position Embedding (RoPE) to effectively capture relative positional information, unlike typical Convolutional Neural Networks (CNNs). Its classification network, informed by cyclone development equations, concentrates coordinate transformation within the affine matrix. Our evaluation using a cyclone dataset shows that CAN-based composites outperform traditional methods, yielding more significant results and more coherent variable coupling. CAN reveals several key common features: (1) cyclone rapid intensification in spring is dominated by cold air activity; (2) topography significantly impacts intensification; and (3) downstream ridge structures potentially influence intensification by causing anomalous subsidence, which leads to low-level dynamic uplift and limited baroclinic energy release. CAN effectively analyzes cyclone circulation and structure.</p>