An Improved Spectral Clustering Algorithm for Large-Scale Wind Farm Power Prediction

Abstract

Aiming at reaching the balance between calculation efficiency and power prediction accuracy of wind farms, two improved spectral clustering (SC) algorithms and their application framework are proposed. For classical k-way Ng-Jordan-Weiss SC, the clustering sample space is composed of k eigenvectors, which may lose part of structural information and may not reach accurate clustering results. To improve the accuracy and stability, we proposed to cluster with feature expansion and the Cuckoo Search (CS) algorithm. We extended the clustering eigenspace from k eigenvectors to 2k to improve the clustering accuracy. To avoid following into local optimum while extending the eigenspace, the CS algorithm was introduced to search for better initial points instead of the random choice method. To apply the proposed algorithm for wind power prediction, wind turbines with similar wind regime were designated to the same group using the proposed SC algorithm. The power prediction model was established for each wind turbine group, and the output power of the entire wind farm was obtained by superposition. Experimental results indicated that the clustering accuracy is improved and the results of multiple clustering hold steady, which meets the requirement of accurate and timely prediction of wind farm power.