High-fidelity temporal prediction model for vehicle wake based on deep learning and bidirectional information fusion

Abstract

Vehicle wake has complex, nonlinear, and multi-timescale characteristics, consuming expensive time and costs in simulations or experiments. A deep learning-based flow field time series prediction model is a relatively good solution. However, current models have evident limitations. One difficulty is avoiding error accumulation, and these models’ performance in complex nonlinear problems, such as vehicle wake, is unknown. Therefore, this study proposes a high-fidelity temporal prediction model for vehicle wake based on deep learning and bidirectional information fusion, and conducts research for Ahmed body wake prediction. Temporal prediction over long time periods is divided into multiple short-time-period prediction processes by establishing a mapping between low-time-resolution and high-time-resolution flow fields. Bidirectional information fusion uses the flow field at adjacent time steps to obtain the final flow field by mapping along the forward and reverse temporal paths. Results indicate that the proposed model exhibits high accuracy for complex vehicle wake. The model accurately predicts velocity gradients and flow details within the flow field, with average peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and mean absolute percentage error (MAPE) of 27.33 dB, 0.974, and 5.10% respectively. The BIFTSR accurately captures the vortex shedding characteristics at the wake of the Ahmed body. The proposed model exhibits good robustness and effectively suppresses noise of varying intensities. When noise intensity increases in 0.1 intervals, the reduction in PSNR and SSIM is suppressed within 0.8 dB and 0.008, respectively. The BIFTSR demonstrates good generalization for unseen data when inflow velocity or geometry changes. Lastly, ablation study validates the contribution and importance of bidirectional information fusion.