Abstract Probabilistic forecasting of distribution tails (i.e., quantiles below 0.05 and above 0.95) is challenging for non-parametric approaches since data for extreme events are scarce. A poor forecast of extreme quantiles can have a high impact in various power system decision-aid problems. An alternative approach more robust to data sparsity is extreme value theory (EVT), which uses parametric functions for modelling distribution’s tails. In this work, we apply conditional EVT estimators to historical data by directly combining gradient boosting trees with a truncated generalized Pareto distribution. The parametric function parameters are conditioned by covariates such as wind speed or direction from a numerical weather predictions grid. The results for a wind power plant located in Galicia, Spain, show that the proposed method outperforms state-of-the-art methods in terms of quantile score.

中文翻译：

---

预测风能的条件极端分位​​数

摘要 分布尾部（即低于 0.05 和高于 0.95 的分位数）的概率预测对于非参数方法具有挑战性，因为极端事件的数据很少。对极端分位数的错误预测可能会对各种电力系统决策辅助问题产生重大影响。另一种对数据稀疏性更稳健的替代方法是极值理论 (EVT)，它使用参数函数对分布的尾部进行建模。在这项工作中，我们通过直接将梯度提升树与截断的广义帕累托分布相结合，将条件 EVT 估计器应用于历史数据。参数函数参数由来自数值天气预报网格的风速或风向等协变量调节。位于西班牙加利西亚的风力发电厂的结果，