In this paper we develop neural network-based models to estimate the top and bottom product composition in a pilot plant distillation column. We study and compare the performance of several recurrent neural network architectures, namely the multiplayer Neural Network in Parallel-configuraiton (MLNP), the Jordan Sequential Neural network (JSNN), the Elman Recurrent Neural Network (ELNN), the Diagonal Recurrent Neural Network DRNN and the State Predictor Neural Network (SPNN). The models obtained can produce multi-step-ahead predictions, and therefore can be considered an alternative for on-line composition analyzers. We find that the JSNN-based model gives worse results with respect to the other models, when used in the identification of the distillation process under consideration. This may be due to structural limitations of networks of the JSNN type.
展开▼