A data-driven reduced-order model for fast prediction of resonant acoustic flow under vertical vibration based on secondary decomposition

Accurate dimensionality reduction models are crucial for constructing real-time computational digital twin systems for process equipment. To deepen the understanding of acoustic resonance flow mechanisms and optimize equipment design, a high-precision prediction method for acoustic resonance flow fields is required. This study introduces a non-intrusive reduced-order modeling (ROM) approach to learn fluid motion patterns and forecast flow field evolution. The original dataset is derived from an experimentally validated computational fluid dynamics model. The flow field snapshots are decomposed into spatial modes and temporal coefficients using proper orthogonal decomposition. High-complexity temporal coefficients, identified through sample entropy analysis, undergo secondary decomposition using methods such as variational mode decomposition. A specialized BiLSTM-attention network is then employed to learn and predict each component's behavior. By integrating secondary decomposition techniques commonly used in time series prediction with ROM technology, this paper presents a non-intrusive flow field prediction method. Compared to the control method without secondary decomposition, this approach significantly enhances prediction accuracy, with the coefficient of determination for individual components improving by up to 27%. This advancement is instrumental for the development of high-fidelity digital twins in acoustic resonance devices, providing a robust basis for real-time computational process equipment.