A comparative approach of inverse modelling applied to an irradiative batch dryer employing several artificial neural networks

Abstract

The present work has focused on a comparison between commonly employed artificial neural networks (ANNs) in engineering applications to identify the most efficient ANN for the inverse modelling of an irradiating furnace/dryer in terms of accuracy and computing time. To this end, several ANNs were designed, trained and employed to estimate the heat emitted during the irradiative batch drying process with the aid of NeuroSolution®.As part of the study, different ANNs were designed and trained to play the role of the inverse heat transfer model. The reasons for exploiting these ANNs were derived from various studies in the literature, in which ANNs were employed for engineering modelling purposes. The results showed that the multiple layer perceptron (MLP) with the Levenberg-Marquadt (LM) in the back propagation (BP) was the best ANN among the methods evaluated to solve the inverse heat estimation problems used in irradiative batch drying processes. An important advantage of the ANN method in comparison with the classical inverse heat transfer modelling approaches is that a detailed knowledge of geometrical and thermal properties of the system (such as wall conductivity, emissivity, etc.) is not required. Such properties are difficult to measure and may undergo significant changes during the temperature transient mode.In this study, genetic algorithms (GAs) have been employed to determine the key parameters of the employed ANNs. These parameters are normally found heuristically or by a trial and error brute force process. The results demonstrate that the aforementioned parameters may be estimated much more accurately and faster by the GA method. The performance of the networks has been improved as well and the number of required hidden layers has been discovered using a non trial-error method, which eliminates time-consuming repeating procedures and produces more accurate results. © 2014 Elsevier Ltd.