Real-time prediction method of three-dimensional flow field for pumping station units operation under geometrically variable conditions based on reduced-order model and machine learning

Abstract

In large-scale water diversion projects, the rapid and accurate evaluation of pumping station unit performance is crucial to ensure that flow rates meet delivery requirements. Computational fluid dynamics (CFD) is effective in analyzing unit performance but is constrained by its high computational complexity and time consumption. Reduced-order models (ROMs) partially alleviate these issues; however, their application is restricted in scenarios involving geometric variability, such as adjustable blade angles, where re-simulation after mesh adjustments leads to inconsistent node configurations. To address these limitations, this study proposes an efficient method for predicting three-dimensional flow fields under varying geometric conditions. A unified snapshot matrix, constructed using interpolation and CFD data, ensures consistent data representation across different geometries. Machine learning is combined with ROMs to achieve efficient and accurate flow field predictions. Compared to the 600.84 s required by traditional CFD simulations, the proposed method reduces computation time to just 1.67 s while maintaining an accuracy of over 90 %. This approach resolves the computational and geometric challenges of traditional CFD and ROMs, providing an efficient solution for real-time evaluation of pumping station unit performance Moreover, it provides a foundation for developing digital twin systems to enhance decision-making efficiency in pumping station management.