Efficient long-term prediction of unsteady flow using hybrid neural network with non-intrusive reduced-order modeling

Abstract This study proposes a non-intrusive reduced-order model based on the proper orthogonal decomposition (POD) and the hybrid neural network (CNN-long short-term memory (LSTM)-efficient channel attention (ECA)) to address long-duration flow prediction challenges. The proposed framework initially decomposes the flow field via POD to extract spatial modes representing distinct energy content and dynamic characteristics, along with their corresponding temporal coefficients. A hybrid neural network is then trained to predict the evolution of multiple temporal coefficients simultaneously. This approach is validated on two types of flow including flow past a cylinder and tandem cylinders. By integrating ECA, the proposed model effectively resolves error amplification caused by multi-frequency coupling–a phenomenon originating from concurrent multi-time-coefficient training. This approach significantly reduces the need for recurrent time-coefficient retraining. Under multi-step prediction conditions, the proposed model achieves a lower root mean square error with a slower error accumulation rate compared to LSTM and temporal convolutional network, thereby demonstrating superior capability in long-term predictions of unsteady flow fields.