FDMNet: Frequency-Aware Dual-Branch Modeling for Wind-Vector Prediction

Near-surface wind prediction requires models to represent broad spatial patterns and localized variations. We introduce FDMNet, a frequency-aware dual-branch network for multi-step wind-vector prediction. A variable-wise encoder and multi-head variable attention combine meteorological fields. Latent channels are partitioned between attentive state-space and depthwise-convolution branches, whose outputs are fused using squeeze-and-excitation weights. This partition is motivated by spatial-scale differences without imposing an explicit spectral cutoff. We evaluate the corrected implementation using WeatherBench 2 data from the fifth-generation reanalysis (ERA5) of the European Centre for Medium-Range Weather Forecasts, with 1979–2015 for training, 2016 for validation, and 2017–2018 for testing. The model uses twelve six-hourly input fields to predict the next twelve fields, covering a 72 h horizon. The pooled area-weighted vector root mean square error is 2.7125 m s−1 for FDMNet, 2.7710 for the Temporal Attention Unit (TAU) baseline, and 5.6567 for Persistence. FDMNet reduces the pooled error relative to TAU by 2.11%, with lower errors at ten of twelve leads but higher errors at 12 and 18 h. The two models were trained once each, using different learning rates. Archived regional runs contain an input-grouping error and are not used to validate the corrected model. Repeated training, component ablations, and regional evaluation are needed to assess the mechanism and generality of the observed improvement.