Progressive horizon extrapolation with wavelet-based directional modeling for regional wind vector field prediction

Regional wind vector field prediction is an important spatiotemporal prediction task for safe and efficient maritime activities. Recurrent neural network (RNN)-based methods model temporal evolution through step-by-step autoregressive prediction but suffer from error accumulation. In contrast, recurrent-free frameworks avoid this issue by predicting all horizons simultaneously; however, they may overlook the progressive evolution of wind during forecasting. Moreover, wind vector fields contain multi-scale spatial structures associated with different frequency components, while such frequency-dependent spatial information remains insufficiently exploited by existing methods. To address these limitations, this paper proposes PHWDNet, a novel framework with the Progressive Horizon Extrapolation (PHE) strategy and the Wavelet-Based Directional Modeling (WDM) block for regional wind vector field prediction. Specifically, PHE bridges strictly autoregressive and fully parallel prediction by extrapolating future states through short horizon segments, where each subsequent segment is predicted from historical observations and previously predicted segments. This enables progressive modeling of wind evolution while mitigating error accumulation. In addition, WDM enables multi-scale directional spatial modeling in a frequency-adaptive manner through wavelet decomposition and Frequency-Adaptive Directional Convolutions (FDC). Experiments on two real-world wind vector field datasets demonstrate that PHWDNet outperforms advanced baselines.