A Hybrid Probabilistic Wind Power Prediction Based on An Improved Decomposition Technique and Kernel Density Estimation

The high uncertainty of non-stationary wind power time series is a challenging issue in optimal operation and planning of power systems. An efficient way to show wind power uncertainty is to use high-quality prediction intervals (PIs). This paper proposes a hybrid probabilistic wind power prediction model based on improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) technique, extreme learning machine (ELM) and kernel density estimation (KDE). First, using ICEEMDAN, the original wind power time series is decomposed to components with different frequency ranges. Then, sample entropy (SampEn) technique is employed to group components to three main time series trend, cycle, and noise with diverse complexity levels. The first two components are deterministically predicted while the noise component is probabilistically predicted using the combination of KDE technique and direct plug-in as a well-known bandwidth selection technique. The lower and upper bounds of final PI are found using the summation of lower and upper bounds of noise component with trend and cycle predicted points. The efficacy of the proposed prediction model is depicted by generating reliable and sharp PIs for real wind power datasets in Canada and comparing with other conventional PI construction approaches.