One parametric approach for short-term JPDF forecast of wind generation

The time domain correlation information of wind generation is important for wind power utilization. This paper proposes a parametric approach for short-term multi-period joint probability density function (JPDF) forecast of wind generation, which contains such correlation information. The approach makes a spot forecast of wind generation by using Support Vector Machine (SVM), and the probability distribution of SVM forecast errors are estimated using Sparse Bayesian Learning (SBL), which assumes the forecast errors follow Gaussian distribution. Then, the SVM forecast results are corrected by the expectation of the forecast errors. The correlation coefficient matrix of wind generation forecast errors during forecast periods is estimated from historical SVM forecast errors. By combining the variance information obtained by SBL and the correlation coefficient matrix, the covariance matrix of the forecast errors within multiple successive forecast periods is formed. Thereby, the JPDF of wind generation is obtained. Data from an actual wind farm are used for the study. The spot and distribution forecast accuracy of the proposed approach is assessed by quantitative indices. The study results illustrate the effectiveness of the proposed approach. Furthermore, the justification of the Gaussian distribution assumption of SBL is also explained in this paper.