Best-performing models on the test set.

Accurate wind energy forecasting is critical for integrating wind power into electrical grids due to its inherent variability and uncertainty. This study introduces a systematic framework that integrates large-scale oceanic climate indices and time-lagged features with advanced machine-learning models to enhance short-term wind power prediction. We evaluate four experimental configurations: (A) a baseline using only wind speed; (B) wind plus contemporaneous indices; (C) the addition of 1–12 month lags for both wind and index variables; and (D) MRMR-based feature selection applied to the full lagged set. A comprehensive benchmark using 25 state-of-the-art models is conducted on monthly data from the Pawan Danavi wind farm in Sri Lanka (2015–2019). Results reveal that raw indices alone can degrade forecast accuracy, while incorporating lagged features significantly reduces RMSE and enhances . MRMR pruning of the 156 lagged predictors distills the set to three key variables: current wind speed, a nine-month lag of the Atlantic Meridional Mode, and a six-month lagged wind speed. This yields a minimum RMSE of and . The proposed approach delivers robust, computationally efficient forecasts, supporting more reliable grid operations and informing future integration of climate teleconnections in renewable energy forecasting.