Predicting Coal Heating Values Using Proximate Analysis via a Neural Network Approach

Abstract

This article presents lower and higher heating value predictions of low rank coals by means of a multi-output neural network model, which uses the proximate analysis variables, such as moisture, ash, volatile matter, and fixed carbon. The neural network model is based on feed forward configuration using a back-propagation learning algorithm. In order to find the best algorithm giving a high prediction performance, the network is trained with eight different back-propagation algorithms. Then, the optimal neuron number of the neural network model is investigated to improve the estimation performance. From the optimal architecture analysis of this model, the LevenbergMarquardt algorithm is found as the best method and the optimal neuron number at the hidden layer is determined as 20. The prediction results of the developed neural network model show that the errors between actual and predicted values are within 4.5% for lower and higher heating values.