Abstract: Flow around a circular cylinder, a classical topic in fluid–structure interactions, has broad engineering applications, particularly in structural load analyses. To address the high computational cost and inefficiency of traditional fluid dynamics methods, a reduced-order modeling (ROM) offers an effective alternative by simplifying flow structures while retaining key physical features. This study proposes two novel ROM methods: the AE-FCNN ROM, combining an autoencoder (AE) with fully connected neural networks (FCNN), and the AE-FCNN-POD-RBFNN (AFPR) ROM, which further incorporates proper orthogonal decomposition (POD) and radial basis function neural networks (RBFNN). The validation on unsteady cylinder flow simulations shows that these two methods can efficiently reconstruct flow fields. Notably, the AFPR ROM improves reconstruction accuracy by 14% over AE-FCNN ROM with only 2% additional computational time, and achieves 84.75% higher efficiency than the traditional method. Research results demonstrate that the proposed hybrid ROM has good engineering adaptability in balancing accuracy and efficiency, providing a new idea for the rapid force analysis of Marine structures under the action of wave and current.