A spliced XGBoost–truncated GPD framework for probabilistic prediction of wind speed within the 3–25 m/s operational range in Bangladesh
Аннотация
Accurate wind speed prediction is essential for renewable energy planning. While machine learning methods often provide strong predictive accuracy for ordinary observations, they frequently underperform in the upper tail of the distribution, particularly for extreme wind events. This study proposes an applied spliced framework integrating Extreme Gradient Boosting (XGBoost) with Peaks-Over-Threshold extreme value modeling to predict the conditional distribution of daily wind speed given that it falls within the retained operational range of 3–25 m/s, rather than the complete daily wind speed distribution. For each station day, an XGBoost regressor estimates the centre of the non-extreme regime, a Platt-scaled XGBoost classifier estimates the conditional probability of exceeding the threshold, and a truncated generalized Pareto distribution (TGPD) models daily exceedance magnitudes between threshold and upper truncation bound. Out of fold residual bootstrap sampling converts the XGBoost bulk prediction into a predictive distribution, and Monte Carlo simulation combines the bulk and tail components on the same daily probability scale. The approach is applied to daily observations from 35 Bangladesh Meteorological Department (BMD) stations during 2000–2020. Relative to an XGBoost only residual bootstrap benchmark, the spliced model reduced the reported continuous ranked probability score (CRPS) from 1.0195 to 0.9926 (2.64%) and reduced pinball loss by 6.36–12.95% across the 0.90–0.99 quantiles. Results show that standalone XGBoost systematically underpredicts high wind speed, whereas the proposed spliced model improves CRPS and reduces quantile losses at upper quantiles. The method provides a practical bridge between machine learning and extreme value theory, yielding improved probabilistic performance relative to the matched XGBoost benchmark and upper-quantile prediction loss while retaining predictive performance in the bulk region within the operational range.