A hybrid system for short-term wind speed forecasting

Highlights: • Develop a novel similarity-based short-term wind speed forecasting system. • Propose a valid strategy of sample selection. • Improve the similarity of building sample in the case of continuity. • Discuss the relationship between the similarity with the modelling requirement. Wind speed forecasting is important for high-efficiency utilization of wind energy. Correspondingly, numerous researchers have always focused on the development of reliable forecasting models of wind speed, which is often noisy, unstable and irregular. Current approaches could adapt to various wind speed data. However, many of these usually ignore the importance of the selection of the modeling sample, which often results in poor forecasting performance. In this study, a hybrid forecasting system is proposed that contains three modules: data preprocessing, data clustering, and forecasting modules. In this system, the decomposing technique is applied to reduce the influence of noise within the raw data series to obtain a more stable sequence that is conducive to extract traits from the original data. To extract the characteristic of similarity within wind speed data, a kernel-based fuzzy c-means clustering algorithm is used in data clustering module. In the forecasting module, a sample with a highly similar fluctuation pattern is selected as training dataset, and which could reduce the training requirement of model to improve the forecasting accuracy. The experimental results indicate that the developed system outperforms the discussed traditional forecasting models with respect to forecasting accuracy.