Performance of adaptive optics (AO) system is limited by the delay errors caused by the servo system and photoelectron noise at the wavefront sensor. A multi-model univariate prediction model is proposed, which is based on the two-layer back propagation neural network with Levenberg-Marquardt learning algorithm. Using the multi-core processors, a novel predictive controller with parallel processing capabilities is designed that is able to predict the control voltage in the closed-loop AO system and eliminate the delay errors. Through numerical simulation, the prediction performance and parallel efficiency are studied. The control voltages of the AO system and the Strehl ratio are calculated and compared for the multi-model univariate prediction algorithm and proportional integral (PI) control algorithm. The results show that the residual error caused by servo delay in the system and Strehl ratio are improved effectively by using the predictive controller than by using the PI control algorithm. The prediction time is reduced by using multi-model univariate prediction algorithm.