A Review of the Accuracy and Computational Cost of AI Regression Models for Short-Term Wind Power Forecasting

Short-term wind power forecasting models must meet both accuracy requirements and computational limits. Published comparisons use different datasets, forecast horizons, and timing boundaries. In this focused narrative review, twelve publications on artificial intelligence (AI) regression models and their evaluation are examined. Support vector regression, random forests, XGBoost, and long short-term memory networks are the main model families discussed. Prediction errors and computational costs are compared within studies, and changes in input information are examined. Under the reported experimental conditions, lower prediction errors do not always require higher computational costs. Runtime rankings can also differ between training and inference. Meteorological inputs and data enrichment affect accuracy and the performance gaps between models. However, cost reporting remains incomplete. The evidence cannot establish whether these accuracy gains require longer computation times or more resources. A conditional screening procedure is therefore suggested. The forecasting task and input availability should be checked first. Parameter search, training, and inference costs should then be compared separately. Results should also be tested across several time windows. These recommendations can support model screening under operational constraints. Their applicability still requires validation at the target wind farm.