A large-scale CO2 Absorption/Desorption pilot plant has been used to investigate the feasibility of applying a Nonlinear Model Predictive Control technique - Receding Horizon OptimarContror (RHC) - and to investigate the performance of such algorithms when applied to a real system where mismatch between plant and model is always present. The performance, stability and robustness of the nonlinear algorithm are compared to those of a standard linear technique, Dynamic Matrix Control (DMC) applied to the same system. The critical nature of the various components of the RHC (model, estimator of unmeasured states and parameters, optimisation algorithm) has also been assessed. Moreover, the differences in performance of the algorithms in both simulations and experimental studies has been explored. Results showed that both the linear and nonlinear optimal control algorithms performed extremely well in simulation studies and were able to achieve excellent control while minimising the overall cost. Implementation of the algorithms on the real plant, however, showed that good performance was critically dependent upon reducing the plant/model mismatch. This was especially true in the case of the nonlinear RHC algorithm. A problem which had to be resolved in on-line experimental implementation is the trade-off between improved model accuracy and computation time for both modelling and optimisation.