Neural networks for the identification of a three component distillation column

The identification of a three-component distillation column was performed using a multilayered neural network trained with the backpropagation algorithm. To find an appropriate network size, several adjustment tests were carried out during the experimentation. These tests included changing the number of hidden layers and number of the nodes in the hidden layer. Validation of the resulting neural model was made by comparison of network and process responses to inputs different from those used during training. The network adequately identified the system. Also, it was observed that the network is able to approximate the nonlinearities of the process with greater accuracy than an ARX model whose parameters were estimated using the classical least squares method