A novel method for long term forecasting of wind speed distribution is proposed based on the concept of training neural network. A phase space reconstruction method is used to track the evolution of the wind speed distribution function parameters in a dynamic system. Then, the neural network training and forecasting features are used to learn nonlinear model between historical data and next observation. Moreover, different estimators have been applied and compared to fit the annual distribution of the wind speed in the studied sites before applying the proposed approach. The proposed method shows a good performance and could be successfully applied in wind energy yield.