Short-term Wind Speed Forecasting with the Extreme Learning Machine using Tukey’s Biweight Criterion and L1 Regularization

This paper presents a novel short-term wind speed forecasting approach based on the extreme learning machine using Tukey’s biweight criterion and L1 regularization (ELM-TL). In the proposed ELM-TL algorithm, a robust loss function with the adaptive penalty term is designed. Different from the squared loss-based models, ELM-TL is insensitive to outliers in both input variables and responses. To learn the parameters of ELM-TL, the accelerated proximal gradient is employed. Real wind speed data collected from a commercial wind farm in Shandong, China is utilized to evaluate the forecasting performance of the proposed approach. Besides, the proposed approach is compared with three commonly applied methods, support vector machine, kernel ridge regression and k nearest neighbors. Computational results demonstrate that the proposed ELM-TL model outperforms the benchmarking models for short-term wind speed forecasting.