Nonlinear Model Identification Using Recurrent Neural Networks: Application to Acetone Cracking

This paper presents the formulation of a nonlinear model identification method based on recurrent neural network (RNN) Nonlinear AutoRegressive with eXternal input (NARX) model derived from dynamic feedforward neural network (DFNN) by adding feedback connection between output and input layers. The proposed identification method identifies the neural network (NN) model of an input-output system. The identified NN model is then validated using the following three different validation algorithms: (1) one-step ahead cross-validation of the training and test data predicted by the trained network; (2) Akaike’s final prediction error (AFPE) estimate of the average generalization error; and (3) 5-step prediction simulations. The algorithm has been applied to a non-isothermal non-adiabatic distributed chemical reactor (DCR) used for the cracking of acetone into ketone and methane. The neural network training and validation data are obtained from the open-loop simulation of a validated first principles DCR plant model. Simulation results show that the RNN models the reactor to a high degree of accuracy.