In this paper, a modified EMD-FNN model (empirical mode decomposition (EMD) based feed-forward neural network (FNN) ensemble learning paradigm) is proposed for wind speed forecasting. The nonlinear and non-stationary original wind speed series is first decomposed into a finite and often small number of intrinsic mode functions (IMFs) and one residual series using EMD technique for a deep insight into the data structure. Then these sub-series except the high frequency are forecasted respectively by FNN whose input variables are selected by using partial autocorrelation function (PACF). Finally, the prediction results of the modeled IMFs and residual series are summed to formulate an ensemble forecast for the original wind speed series. Further more, the developed model shows the best accuracy comparing with basic FNN and unmodified EMD-based FNN through multi-step forecasting the mean monthly and daily wind speed in Zhangye of China. ► Multi-step forecasting for nonlinear and non-stationary wind speed series. ► EMD technique is used for a deep insight into the data structure. ► Use the robust FNN model for forecasting the sub-series obtained from EMD. ► PACF is used to select the input variables for FNNs. ► Get rid of the high frequency to improve forecasting precision.