In this paper, constrained model predictive control (MPC) based on parallel neural network optimization is proposed to apply to pulse width modulation (PWM) rectifier and improve power quality. An decoupled model of three-phase rectifier in abc coordinates is built. Then, the constrained MPC method is proposed. This method breaks the limits of predictive control with finite set and without constraints. Neural network optimization is used to solve online optimization of MPC and accelerate single step optimization. A approach is proposed to guarantee unity power factor and also provides a regulated output dc-voltage with fast dynamics response against sudden changes in the load. The simulation results demonstrate that the output dc-voltage error and the total harmonics distortion (THD) of the input current are small and the unity power factor rectifier performs very well.