Short-term Wind Speed Prediction System Based on Vision Mamba

Accurate short-term wind speed prediction requires models that represent regional spatial heterogeneity and temporal evolution without future information in preprocessing. This study presents VMamba-WSP, a Vision Mamba-based system that maps 24 hourly 13 × 13 scalar 10 m wind-speed fields to the next 12 h. Each hourly field is embedded by two convolutions and processed by an explicit four-route SS2D-style directional scan. Spatially pooled hourly tokens are then modeled with rotary temporal encoding, custom Mamba-style state-space blocks, and temporal attention pooling. The model has 1.997 million parameters and uses a configured empirical composite objective with a physics-inspired spatial-gradient mismatch regularizer, which does not enforce atmospheric conservation equations. The principal evaluation uses training-only gridwise normalization, no temporal smoothing or clipping, 2020 to 2022 for training, 2023 for validation, and a chronological 2024 year-held-out evaluation. Across seeds 42, 43, and 44, VMamba-WSP obtains mean absolute error 0.9348 ± 0.0084 m/s and root mean square error 1.2771 ± 0.0116 m/s. TRNet-lite remains more accurate, with mean absolute error 0.8149 ± 0.0019 m/s. Complete 2025 evaluation preserves this ranking. Removing the four-route scan increases 2024 mean absolute error from 0.9444 to 0.9588 m/s, with a seven-day moving-block bootstrap full-minus-ablation difference of −0.0143 m/s. VMamba-WSP is also slower than all evaluated controls. The results support directional spatial scanning as a useful adaptation, but not claims of accuracy dominance, recovered atmospheric dynamics, or edge deployment readiness.