Purpose This study develops and validates a hybrid wind energy forecasting model for accurate multi-horizon predictions, addressing key data preprocessing challenges, particularly outlier detection and feature selection, to enhance model robustness. Design/methodology/approach The proposed two-stage hybrid machine learning framework, DBSCAN-RFE-XGBoost, integrates time-series data from historical wind energy generation. In the first stage, an automated density-based spatial clustering (DBSCAN) algorithm is introduced, employing a K-dist plot and Knee Point Detection Algorithm to determine the optimal epsilon (EPS) value and remove outliers. In the second stage, recursive feature elimination (RFE) selects the most relevant features for the XGBoost forecasting model. The model is trained and validated using empirical data collected from a grid-connected wind farm. Findings The proposed knee-point detection-based EPS automation in DBSCAN achieved optimal clustering with an EPS value of 1.44, outperforming manual tuning. For a one-hour horizon, it yielded, MSE = 0.00607, RMSE = 0.07791 and MAE = 0.05261. Feature selection identified seven features for short-term and eight for long-term accuracy. Compared with DBSCAN-RFE-SVM, LSTM, LightGBM and KNN, the proposed DBSCAN-RFE-XGBoost model achieved better results across all horizons using traditional and tail-sensitive metrics, confirming its robustness under extreme prediction deviations. Research limitations/implications Although integrating clustering with feature selection improves long-term forecasting performance, the DBSCAN-RFE-XGBoost model still exhibits limited accuracy over extended horizons, representing a practical limitation. Originality/value The novelty of our method lies in automating the DBSCAN clustering process and integrating it with feature selection through recursive elimination, creating a unified two-staged forecasting model.
Khan et al. (2026) studied this question.