The dissolved oxygen (DO) concentration is of great importance to wastewater treatment due to its influence on effluent quality and operational costs. However, the DO concentration is difficult to be controlled owing to the timevarying and nonlinear characteristics. Considering these issues, a neural network predictive controller (NNPC) based on Levenberg—Marquardt (LM) algorithm was proposed. Firstly, a simplified DO model was established after reasonable hypotheses and constrains in terms of activated sludge model No.1 (ASM1) proposed by International Water Association (IWA). Then the NNPC was applied to the simplified DO model through system identification with BP neural network and model prediction. Furthermore, the LM algorithm integrated the advantages of the gradient steepest descent and Newton methods was used to improve the general BP neural network, which overcame the drawbacks of falling into local minimum easily and slow convergence speed. The simulation results indicated that the improved neural network had good performance in system identification with error less than 3%. Compared with conventional PID control and model predictive control (MPC), the NNPC achieved smoother and better tracking performance and brought obvious improvement. Finally, two measured disturbances were added and good adaptability and robustness were obtained by NNPC. In this way, this method not only can achieve the standard of effluent water quality, but also can reduce the energy consumption of aeration significantly.