Prediction of lactic acid concentration using artificial neural network

Abstract

This study investigates the application of artificial neural network in model development for lactic acid production. The current measurement of lactic acid concentrations is conducted offline, resulting in time delay in obtaining the results, not to mention that current analysis method is expensive and in need of specially trained personnel. In view of this, two model of artificial neural network; multilayer perceptron (MLP) and radial basis function (RBF) network, have been employed. For the development of MLP model, normalization method, the size of input layer, size of hidden layer and activation function have been varied. Effects of input combinations on the MLP performance have also been investigated. For RBF model development, effects of the tolerance (MSE), radius (s) value, the number of input variables and input combinations on the RBF performance have been investigated. The results show that the optimum structure of MLP has four input variables (biomass concentration, glucose concentration, temperature and reaction time) and a transfer function of log sigmoid in the hidden layer and linear in the output layer. This model is capable of producing the error index (EI) test of 7.26% and R-value test of 0.9909 with seven nodes in the hidden layer. Also, the RBF model was able to obtain EI test of 6.48% and R-value of 0.9926 with a model of three input variables (biomass concentration, glucose concentration and reaction time) and a radius (s) equal to 1.5. The optimum structure of the RBF model was 3-7-1. Both models exhibit comparable and good generalization ability. However, the RBF model out-performed the MLP model with regard to its generalization ability and reproducibility but overall both models have displayed satisfying ability in estimation of lactic acid concentration for the identified process.