Optimizations of NARX lag space selection for a Multi-Layer Perceptron (MLP)-based model of a down-flowing steam distillation system using Particle Swarm Optimization (PSO)

This paper presents an application of the Particle Swarm Optimization (PSO) algorithm to perform lag space selection for a Multi-Layer Perceptron (MLP)-based Non-Linear Autoregressive Model with Exogenous Inputs (NARX) model of a down-flowing steam distillation system. PSO represents each candidate lag space as integer values. These candidate lag spaces were then used to construct the training data, which was then used to train the MLP. Two datasets have been used in this experiment. These datasets were separated into training and validation data using the interlacing technique. The PSO-based optimization results were then compared with the ARX structure selection function in MATLAB. The results suggest that the proposed method managed to improve the OSA model fit compared to the MATLAB ARX structure selection method.