Commonly used wind power forecasts methods choose only one representative wind turbine to forecast the output power of the entire wind farm; however, this approach may reduce the forecasting accuracy. If each wind turbine in a wind farm is forecasted individually, this considerably increases the computational cost, especially for a large wind farm. In this work, a compromise approach is developed where the turbines in the wind farm are clustered and a forecast made for each cluster. Three clustering methods are evaluated: K-means; a self-organizing map (SOM); and spectral clustering (SC). At first, wind turbines in a wind farm are clustered into several groups by identifying similar characteristics of wind speed and output power. Sihouette coefficient and Hopkins statistics indices are adopted to determine the optimal cluster number which is an important parameter in cluster analysis. Next, forecasting models of the selected representative wind turbines for each cluster based on correlation analysis are established separately. A comparative study of the forecast effect is carried to determine the most effective clustering method. Results show that the short-term wind power forecasting on the basis of SOM and SC clustering are effective to forecast the output power of the entire wind farm with better accuracy, respectively, 1.67% and 1.43% than the forecasts using a single wind speed or power to represent the wind farm. Both Hopkins statistics and Sihouette coefficient are effective in choosing the optimal number of clusters. In addition, SOM with its higher forecast accuracy and SC with more efficient calculation when applied into wind power forecasts can provide guidance for the operating and dispatching of wind power. The emphasis of the paper is on the clustering methods and its effect applied in wind power forecasts but not the forecasting algorithms.
展开▼
