A new optimal disturbance rejection control method is proposed for the system with disturbances via a compound neural network prediction approach in this paper. The disturbances caused by external disturbances and model mismatches can be estimated by a disturbance observer, and the estimation of disturbances is introduced into the neural network predictive model to make the predictive output more accurate. Then based on the new compound neural network predictive model, a controller, which ensures both optimal performance by the receding horizon optimization and strong disturbance rejection ability, is obtained. The proposed scheme is applied to control the temperature of a simplified jacketed stirred tank heater (JSTH). Simulation results demonstrate the effectiveness of the proposed control method. (C) 2014 Elsevier Ltd. All rights reserved.