A novel probabilistic modeling framework for wind speed with highlight of extremes under data discrepancy and uncertainty

Wind power is serving as a rapidly growing renewable energy resource for global sustainable development, and thus an important task for optimal utilization of this resource is to accurately describe and estimate the uncertain wind speed. In this regard, this paper proposes a novel probabilistic modeling framework for an accurate description of wind speed with the highlight of extremes, where two sources of discrepancy inside wind speed data are taken into account. One is the discrepancy between the non-tail part and the tail (extremes) of wind speed data, which can separate the whole dataset into homogeneous subsets. Then, a dual-model estimation model relying on the existing probability distributions and machine learning algorithm is proposed to facilitate an accurate description of a feasible region with extremely high wind speed values. Another source of the discrepancy lies in the collected data in various lengths of time series, and thus the estimated model should be continuously changed to adapt to the ever-growing time series. As a case study, this proposed framework is verified in a real-world wind speed dataset, where three scenarios with different time lengths are designed for deep investigation. It is found that: (1) Extremes of wind speed can be well separated in the proposed probabilistic modeling with a theoretical derivation of the boundary. (2) The distribution equation-based method can easily provide a more precise description of the extremes than the support vector regression (SVR)-based model, reaching a smaller Bayesian risk and maximal value of the loss. (3) Probabilistic modeling depends on the length of the time series of data collection. (4) Additional experiments under some degree of uncertainty are set to verify the robustness of the proposed method. The novelties lie in the theoretical boundary identification and automatic description of high wind speed distribution, which can potentially provide data-driven evidence for the development of wind power plants.