Probabilistic and deterministic wind speed forecasting based on non-parametric approaches and wind characteristics information

•Site dependency and distribution heterogeneity of wind speed data is considered.•Generalized Gaussian process models with positive domain likelihoods are proposed.•A statistical model is selected to extract the linear feature from the wind speed.•A pseudo-likelihood with cross validation is applied to improve models’ robustness.•Empirical results statistically verify the performance of our proposed models. As the proportion of the total installed wind capacity continues to increase, precise probabilistic and deterministic wind speed forecasting becoming increasingly significant for wind turbine stable function and operation management. It is worth noting that except for the intermittent nature of wind speed itself, the site dependence of wind energy and the heterogeneous nature of the distribution at different locations can also affect the results of wind speed forecasting. Considering this has often been overlooked in previous studies, this study attempts to introduce characteristic information of wind from wind farms into wind speed probabilistic forecasting. For the sake of acquiring a more rational descriptive statistical interpretation of wind speed, a variety of probability density functions for actual wind speed data are evaluated in this study. And then, Generalized Gaussian Process models are constructed via observation likelihood functions which are determined by obtained probability density functions. Beyond that, to enhance the robustness against model misspecification, a pseudo-likelihood for the Leave-One-Out cross-validation methodology is proposed for approximating the posterior in the Bayesian inference stage. Wind speed forecasts on 6-hour-ahead horizons from two wind farms in China illustrated that the appropriate mining of wind prior statistical characteristic information can contribute to improving the precision of probabilistic wind speed forecasting.