A method to compensate the hysteresis nonlinearity of piezoelectric steering mirror (tip/tilt mirror, TTM) based on neural network models is presented, which may improve the steering mirror control accuracy in adaptive optics system. The hysteresis of TTM is modeled by using back propagation (BP) neural network via the response data of TTM. We simulate the process of hysteresis response and compute the voltage applied to TTM in real time, so as to achieve the forward compensation control of TTM. In order to meet the real time intention, we use the function form of BP neural network instead of time-consuming simulation operation, which increases the computing speed and insures the operation precision. The results show that by using this compensation approach the hysteresis nonlinearity is reduced by about 70%. Moreover, the global linearity and the controlling accuracy are obviously improved.