Interest in renewable and clean energy sources is becoming significant due to both the global energy dependency and detrimental environmental effects of utilizing fossil fuels. Therefore, increased attention has been paid to wind energy, one of the most promising sources of green energy in the world. Wind speed forecasting is of increasing importance because wind speeds affect power grid operation scheduling, wind power generation and wind farm planning. Many studies have been conducted to improve wind speed prediction performance. However, less work has been performed to preprocess the outliers existing in the raw wind speed data to achieve accurate forecasting. In this paper, Support Vector Regression (SVR), a learning machine technique for detecting outliers, has been successfully combined with seasonal index adjustment (SIA) and Elman recurrent neural network (ERNN) methods to construct the hybrid models named PMERNN and PAERNN. Then, this paper presents a medium-term wind speed forecasting performance analysis for three different sites in the Xinjiang region of China, utilizing daily wind speed data collected over a period of eight years. The experimental results suggest that the hybrid models forecast the daily wind velocities with a higher degree of accuracy over the prediction horizon compared to the other models. (C) 2014 Elsevier Ltd. All rights reserved.