ABSTRACT This study proposes a two‐stage methodological framework that simultaneously expedites model training and enhances predictive fidelity in wind power forecasting. In the first stage, input dimensionality is reduced through an initial correlation coefficient screening, followed by an iterative correlation stability analysis that retains only those instances exhibiting robust and persistent associations with the target variable. In the second stage, the hyperparameters of nonlinear autoregressive models with exogenous inputs are optimally calibrated via a grey wolf optimiser employing integer encoding. Empirical assessments conducted on extensive wind speed and wind power time series reveal that the streamlined feature set, coupled with optimised hyperparameters, reduces training time substantially while yielding superior forecasting accuracy relative to conventional baselines. Moreover, the refined wind speed estimates propagate to wind‐power predictions, delivering additional performance gains. The findings corroborate the efficacy of integrating rigorous dimensionality reduction with metaheuristic hyperparameter tuning for data‐driven wind power forecasting.
Junhuathon et al. (Thu,) studied this question.