A new control scheme is presented for feedforward control of unknown disturbances in the model-predictive control (MPC) scheme. In this control scheme, a neural network is connected in parallel with the MPC controller and trained on-line by minimizing the MPC controller output corresponding to the unmodeled effect. It is applied to distillation column control and nonlinear reactor control to illustrate its effectiveness. The result shows that the neural feedforward controller can cope well with strong interactions, time delays, nonlinearities, and process/model mismatch. The controller also offers such advantages as fault tolerance, generalization capability by interpolation, and learning capability by random input patterns.