Accurate short-term wind power forecasting is important for improving the economic and stable operation of power system. Since it is easy to be affected by subjective factor and fall into local optimum in parameters selecting compared with traditional Neural Network, a novel combination forecasting approach based on Improved Crow Search Algorithm (ICSA) to optimize the parameters of Echo State Network (ESN) neural network is proposed to overcome above inadequacies. The Lévy flight is introduced to increase the searching efficiency at initial stages, and during the later stage of iteration, the Gauss function is added aiming at making an appropriate adjustments for the whole trajectory points after evolution, which can guarantee the ability of global optimization and successive approximation; it chooses optimal the weight values of the hidden layer to enhance the efficiency of neural network training by ICSA algorithm. Finally, effectiveness of the proposed forecasting model is tested on two groups of experimental data, the results show that proposed algorithm can effectively cope with the variability and intermittency of wind power time series, having higher modeling precision and faster convergence speed. This work is supported by National Natural Science Foundation of China (No. 51677072).