This paper addresses the issue of developing feasible advanced control strategies for the operation of industrial fed-batch multi-stage sugar crystallization processes. The operation of such processes poses very challenging problems mainly those inherent to its batch nature and also those due to the difficulties in measuring key process variables. Inadequate control policies lead to out-of-spec batches, with consequent losses resulting from the need of product recycling. In order to address these problems, a modification of the general Nonlinear Model Predictive Control (NMPC) is proposed in this paper, where the NMPC is executed only when the tracking error is outside a pre-specified bound alpha. Once the error converges towards the alpha-strip, the NMPC is switched off and the control action is kept constant. In order to further reduce the complexity of the control system, the proposed modification, termed Error Tolerant MPC (ETMPC), is provided with a Recurrent Neural Network (RNN) predictive model. The ETMPC + RNN control scheme was extensively tested on a crystallizer dynamic simulator, tuned with data from two industrial units, and compared with the classical NMPC and PI strategy. The results demonstrate that both NMPC and ETMPC controllers lead to improved end point process specifications, when compared with the PI controller. The explicit introduction of the error tolerance in the optimization relaxes the computational burden and can complement several other suggestions in the literature for feasible industrial real time control. (C) 2010 The Institution of Chemical Engineers. Published by Elsevier B.V. All rights reserved.