A universal framework for high-fidelity flow field prediction in scroll machinery based on POD and time series forecasting: A case study of co-rotating scroll hydrogen recirculation pump

Solely relying on computational fluid dynamics (CFD) is insufficient to support the development of scroll machinery digital twins, primarily due to its high computational cost and poor real-time performance. With the rapid advancement of data-driven technology, reduced order models (ROMs) have emerged as a promising pathway to facilitate it. This paper proposes a universal framework for high-fidelity flow field prediction in scroll machinery, utilizing Proper Orthogonal Decomposition (POD) for dimensionality reduction combined with time series forecasting. To address the time-varying grid topology of scroll machinery, an innovative data extraction method based on fixed background grid is proposed. Furthermore, modal coefficients are grouped via wavelet transform and K-means clustering analysis for improved prediction model training efficiency, with the grouped counterparts predicted using the Patch Time Series Transformer (PatchTST). The results indicate that the POD of the flow field in scroll machinery is firstly realized through background grid interpolation. Compared with individual prediction, group prediction reduces the number of prediction models required for training to 1/20. With 99.99% energy truncation, the method achieves a mean absolute percentage error of approximately 0.67% for flow field prediction, while requiring only 0.16 s to reconstruct the flow field at each rotation angle. The proposed framework substantially lowers computational costs while satisfying the requirements for real-time, high-precision performance characterization in digital twin applications. This novel methodology can be extended to the flow field prediction and digital twin construction for other types of positive displacement fluid machinery.