Non-parametric models for joint probabilistic distributions of wind speed and direction data

Abstract

Abstract Two non-parametric models, namely the non-parametric kernel density (NP-KD) and non-parametric JW (NP-JW) models, are proposed for joint probabilistic modeling of wind speed and direction distributions. In the NP-KD model, a novel bivariate kernel density function, which could consider the characteristics of both wind direction (angular) and speed (linear) data, is firstly constructed and the optimal bandwidth is selected globally through two cross-validation (CV) methods. In the NP-JW model, the univariate Gaussian and von Mises kernel density functions are, respectively, utilized to fit the wind speed and direction data. The estimated wind speed and direction distributions are used to form the joint distribution according to the JW model. Several classical parametric models, including the AG, Weibull, Rayleigh, JW-TNW and JW-FMN models, are also introduced in order for comparisons with the proposed non-parametric models. By conducting various tests on the real hourly wind speed and direction data, the goodness of fit of both parametric and non-parametric models is compared and evaluated in detail. It is shown that the non-parametric models (NP-KD, NP-JW) generally outperform the parametric models (AG,Weibull, Rayleigh,JW-TNW,JW-FMN) and have more robust performance in fitting the joint speed and direction distributions. Among the two non-parametric models, the NP-KD model has better performance in fitting joint distribution, while the NP-JW model has higher accuracy in fitting the marginal speed (or direction) distributions.