Answers are grounded in this paper and cite their sources.
Accurate wind power forecasting is essential for improving grid stability, optimizing energy management, and facilitating the large-scale integration of renewable energy. However, forecasting based on real-world SCADA measurements remains challenging because of the stochastic and nonlinear nature of wind generation. This study presents a comprehensive and reproducible benchmarking framework for short-term wind power forecasting by comparing fifteen forecasting approaches spanning statistical baselines, machine learning, and deep learning paradigms. The evaluated models include Persistence, Linear Regression, ARIMA, SARIMA, XGBoost, LightGBM, CatBoost, ExtraTrees, ANN, DNN, RNN, LSTM, GRU, CNN, and Informer. The comparison is conducted using a 24-month industrial SCADA dataset (2024–2025) collected at 10-min intervals from an operational wind farm in northern Morocco. All models are evaluated under a common chronological protocol with the same forecasting target, horizon, test period, and evaluation criteria, while model-specific input representations and optimization procedures are adopted according to the requirements of each forecasting paradigm. All models perform one-step-ahead forecasting at the native 10-min SCADA resolution. The resulting power predictions are subsequently aggregated into daily, weekly, monthly, and yearly accumulated energy quantities for multi-scale evaluation, rather than being generated as independent direct forecasts at these horizons. Performance is then assessed using multiple statistical metrics, a MeanRankScore framework, and formal statistical significance tests (Diebold–Mariano and Wilcoxon signed-rank tests). The results show that the Persistence model provides the highest forecasting accuracy under the investigated conditions, while XGBoost is the best-performing learning-based model, consistently outperforming the remaining machine learning and deep learning approaches. In contrast, complex recurrent and Transformer-based architectures, including GRU and Informer, exhibit substantially larger prediction errors, particularly during low-power operating periods. These findings demonstrate that increasing model complexity does not necessarily improve forecasting accuracy for structured industrial SCADA datasets and highlight the value of rigorous benchmarking for selecting reliable forecasting models in practical wind energy applications.
No takes yet. Share an insight, caveat, or question.
Bousla et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: