Precise and reliable wind-speed prediction is vital for wind-farm operational planning. However, wind speed series usually have complex features, such as non-linearity and volatility, which makes the wind energy forecasting highly difficult. Aimed at this challenge, this paper proposes a forecasting architecture based on a new hybrid decomposition technique (HDT) and an improved flower-pollination algorithm (FPA)-back propagation (BP) neural network prediction algorithm. The proposed HDT combines the complete ensemble empirical mode decomposition adaptive noise (CEEMDAN) and the empirical wavelet transform (EWT), which is unique, since the EWT is specifically employed to further decompose the high frequency intrinsic mode functions (IMFs) generated by CEEMDAN to reduce prediction complexity. And then an improved BPNN with the flower-pollination algorithm is applied to forecast all of the decomposed IMFs and modes. To investigate the forecasting ability of the proposed model, the wind speed data collected from two different wind farms in Shandong, China were used for multi-step ahead forecasting. The experimental results show that the proposed model performs remarkably better than all of the other considered models in one-step to five-step wind speed forecasting, which indicates that the proposed model is highly suitable for non-stationary multi-step wind speed forecasting. (C) 2018 Elsevier Ltd. All rights reserved.