Retrieval-Augmented Non-Parametric Probabilistic Wind Power Forecasting under Regime-Dependent Uncertainty

Abstract Reliable ultra-short-term probabilistic wind-power forecasting is important for renewable-energy integration, especially during ramps and extreme fluctuations. Existing deep probabilistic models typically encode historical operating regimes only implicitly, while conventional non-parametric estimators rely on low-dimensional similarity measures and coarse sample weighting. This paper proposes a retrieval-augmented non-parametric framework that reconstructs the conditional distribution of wind power from physically relevant historical analogs. A physically informed BiLSTM–Attention network extracts regime features from meteorological and historical power inputs; candidate historical samples are then retrieved and weighted by combining deep-feature similarity, wind-speed proximity, and temporal continuity, and the weighted neighbors are passed to a diffusion-smoothed kernel density estimator that produces both point forecasts and prediction intervals. Experiments on two heterogeneous wind farms in Zhejiang and Inner Mongolia show that, on Dataset A at the 60-min horizon, the method lowers MAE and RMSE by 21.2% and 22.7% relative to a Transformer baseline and reduces the 80% Winkler score by 24.4% relative to CKDE. A retrieved-neighbor analysis further makes each forecast traceable to physically meaningful historical samples.