This study presents a novel model that combines higher-order dynamic mode decomposition (HODMD) with a backpropagation neural network (BPNN) for predicting wind pressure on building roofs, termed HODMD-BPNN. Additionally, this paper also optimized the method for determining the HODMD embedding dimension. The HODMD-BPNN model was developed and validated through wind tunnel experiments using a 1∶20 scale model of a low-rise building. The results confirm that the optimized embedding number selection method reliably determines the HODMD embedding number. Second, the HODMD-BPNN model outperforms conventional POD-BPNN in predicting mean and fluctuating wind pressure coefficients. Spectral analysis reveals superior accuracy in low-frequency bands (0–10 Hz) for nonGaussian regions while maintaining better high-frequency (10–160 Hz) performance despite underestimation tendencies. Temporal extrapolation analysis shows that prediction errors amplify in nonGaussian areas when exceeding training time dimensions. In contrast, Gaussian regions exhibit minimal error growth due to relatively steadier flow characteristics. Overall, the proposed HODMD-BPNN model demonstrates enhanced accuracy, robustness, and potential for practical application in structural wind engineering.
Liu et al. (Thu,) studied this question.