State-Aware Dynamic Retrieval and Persistence-Anchored Expert Fusion for Very Short-Term Wind Power Forecasting Without Numerical Weather Prediction

Abstract

Very short-term wind power forecasting without future numerical weather prediction (NWP) must exploit historical observations while controlling operating-state heterogeneity, erroneous analog retrieval, and unstable expert corrections. This study proposes state-aware dynamic retrieval and persistence-anchored expert fusion (SADR-PAEF). A continuous operating-state representation supports training-only historical retrieval, while persistence, analog-trajectory, and neural residual experts are coordinated by a state-aware gate. Persistence-biased initialization, entropy-based contraction based only on gate-weight concentration, and validation-selected anchoring constrain excessive deviations from persistence; gating entropy is treated as a heuristic stabilization factor rather than a calibrated reliability measure. State-Adaptive Online Conformal Prediction (SAOCP) provides state-conditional online interval calibration. Across five runs on the 2024 main-site dataset, SADR-PAEF achieved an MSE of 44.0526 ± 0.2307 MW2, 3.22% lower than DLinear, with significant four-horizon-aggregated squared-error improvements over all seven baselines. Independent retraining on the 2019 site yielded consistent performance under another site distribution, which is interpreted as site-level replicability rather than cross-site transfer or zero-shot generalization. At 90% nominal coverage, SAOCP achieved the lowest overall Winkler score and the lowest Winkler score at all four horizons, although Rolling CP was marginally closer to nominal coverage at 15 min. Persistence retained a slight MAE advantage. The proposed framework therefore primarily reduces squared-error and large-deviation risk while providing adaptive interval calibration under NWP-unavailable conditions.