Ultra-short-term wind-power forecasting must handle both stable conditions and rapid ramp events, yet many architectures do not explicitly distinguish positive from negative ramp evidence. We evaluate whether that distinction improves multi-step predictions under a chronological wind-power forecasting protocol. RAA-Net combines a persistence-anchored baseline with a causal GRU and separate mappings for upward and downward ramps. A calibration-set blending rule substitutes persistence for selected low-volatility cases. We evaluated two turbines from the public SDWPF dataset at 40–160 min horizons, using chronological train, validation, calibration, and test partitions and three-seed neural-model comparisons against eight baselines. RAA-Net records the lowest reported RMSE at every horizon for both turbines. On the primary turbine, it also records the best reported MAE, $R^2$, nRMSE, and nMAE at each horizon. Component ablations quantify the effects of the directional module, anchor, and blending rule. These findings support direction-specific ramp representations for the evaluated turbines and protocol.