Abstract
Accurate wind power forecasting is essential for grid stability and energy market efficiency, yet traditional prediction methods often neglect the relative ordering and temporal consistency of outputs-factors critical for ranking-based decisions in grid dispatch, bidding, and reserve management. To address these shortcomings, this paper introduces a novel wind power forecasting framework that embeds ranking consistency and temporal smoothness directly into the learning objective. The proposed model employs a composite multi-objective loss that simultaneously minimizes point-wise prediction errors, maximizes rank alignment across forecasted values, and enforces temporal rank regularization to avoid instability in ordered outputs. A deep neural architecture based on attention mechanisms is trained end-to-end using historical wind speed, direction, turbulence, and meteorological covariates as inputs. To validate the model, we construct a high-resolution dataset comprising 12 wind farms over 24 months with synchronized SCADA, meteorological, and geographic information. Multiple wind regimes-including low, ramping, and saturation scenarios-are explicitly labeled to facilitate regime-aware evaluation. Extensive numerical experiments demonstrate that the proposed model outperforms baseline methods such as LSTM, Transformer, and LambdaMART in terms of MAE, RMSE, and normalized discounted cumulative gain (NDCG), particularly under high-fluctuation regimes. Moreover, we introduce a Temporal Rank Stability Index (TRSI) to quantify the consistency of ordinal outputs across time, with our model achieving up to 35% improvement over state-of-the-art alternatives. This study offers three core contributions: (1) a theoretically grounded multi-objective loss for ranking-aware and temporally robust wind forecasting, (2) a novel wind regime-labeled dataset supporting both prediction and ranking evaluation, and (3) a suite of visualization tools and metrics that reveal deeper dynamics in ordinal wind forecasting tasks. The results suggest new directions for learning-to-rank in renewable energy forecasting and demonstrate the practical feasibility of incorporating rank-sensitive intelligence into grid-scale forecasting pipelines.