Power control is required for large grid integration variable speed wind turbine at above rated wind speed. A nonlinear control strategy of Elman network with input and output context layers is proposed in the paper, which can effectively overcomes disadvantages of tradition control that accuracy of pitch angle is limited by specified model and work time. Novel modified Elman network structure merges selffeedback information of input and output nodes to approximate power coefficient inverse model and acts as controller. It can precisely supply an optimal pitch angle set point according to limited power coefficient and tip speed ratio during large wind variation region and maintain power output at rated power. Process equations of wind turbine, network structure and controller algorithm are described in the paper. Simulation results show that proposed control strategy is effective.