Multi-fidelity enhanced few-shot time series prediction model for structural dynamics analysis

These days, deep learning (DL) techniques are regarded as an effective substitution of refined finite element models to conduct structural dynamics analysis . Nevertheless, the efficacy of DL methods overwhelmingly depends on the quality and quantity of the data. Since high-fidelity (HF) data are accurate enough but usually time-consuming and costly, while low-fidelity (LF) data are low-cost and efficient but inaccurate, the DL models cannot efficiently and accurately achieve prediction of structural dynamic response solely using either HF or LF data. Under this circumstance, a multi-fidelity enhanced few-shot time series prediction approach is proposed that can significantly promote the efficiency of structural dynamics analysis. Moreover, the K-shape clustering method is utilized to select representative training sample to lessen the need for HF data, further improving efficiency of structural dynamics analysis. For validating the accuracy and efficiency of this method, a case study about the prediction for seismic response on a reinforced concrete frame is conducted, wherein fiber element models are utilized to generate accurate but limited HF seismic response data, while multi-degree of freedom models are utilized to develop a larger LF response database. Eventually, diverse vital factors impacting the performance of the proposed method are also investigated. The results demonstrate that the proposed method can efficiently implement structural dynamics analysis without compromising precision.