Missing values challenge the probabilistic wind power forecasting at both parameter estimation and operational forecasting stages. In this paper, we illustrate that we are allowed to estimate forecasting functions for each missing patterns conveniently, and propose an adaptive quantile regression model whose parameters can adapt to missing patterns. For that, we particularly design a feature extraction block within the quantile regression model, where parameters are set as a function of missingness pattern and only account for observed values. To avoidthe quantile-crossing phenomena, we design a multi-task model to ensure the monotonicity of quantiles, where higher quantiles are derived by the addition between lower quantiles and non-negative increments modeled by neural networks. The proposed approach is distribution-free and applicable to both missing-at-random and missing-not-at-random cases. Case studies demonstrate that the proposed approach achieves the state-of-the-art in terms of the continuous ranked probability score.

翻译：缺失值在参数估计和运行预测两个阶段给概率风电功率预测带来挑战。本文表明，可以针对每种缺失模式便捷地估计预测函数，并提出一种参数能自适应缺失模式的分数位回归模型。为此，我们特别在分位数回归模型内设计了一个特征提取模块，其参数被设为缺失模式的函数且仅考虑观测值。为避免分位数交叉现象，我们设计了一个多任务模型来确保分位数的单调性，其中较高分位数由较低分位数与神经网络建模的非负增量相加得到。所提方法无分布假设，适用于随机缺失和非随机缺失两种情况。案例研究表明，该方法在连续排序概率评分方面达到了当前最优水平。