Accurate prediction of wind speed is of great significance for safe operation and efficient power generation of wind farms. Aiming at the inherent defects of the single decomposition strategy used in existing literatures in wind speed prediction and the unstable effect of the optimized prediction model, a hybrid prediction model combining two-stage decomposition and iJaya-ELM is proposed. First, ICEEMDAN decomposition is performed on the original wind speed sequence, and 12 components are obtained, and reconstructed into high frequency terms, middle frequency terms and low frequency terms based on the permutation entropy. Then, the high frequency term is filtered by singular spectrum decomposition to remove the sequence noise. An improved Jaya algorithm, iJaya, is proposed to obtain the optimal connection weights and thresholds of ELM. Finally, the predictive results of each component are linearly integrated to obtain the final results. The model is validated by wind speed data of wind farm in Gansu province of China, and its robustness and universality are tested by wind speed data of Xinjiang region. The experimental results show that the iJaya algorithm is of strong optimization accuracy and stability, and the two-stage decomposition can deeply excavate the characteristics of wind speed series. The hybrid model can effectively improve the wind speed prediction accuracy, and the average absolute error and mean square error are 0. 067 9 and 0. 134 5, respectively.