Neural network modeling of sorption isotherms of longan (Dimocarpus longan Lour.)

摘要

A multilayer neural network model was developed to predict the equilibrium Moisture content of longan (Dimocarpus longan Lour.) and the model was trained using a back-propagation algorithm. The predictive power of the model Was found to be high (R-2 = 0.9998) after it was adequately trained. The artificial neural network (ANN) model was better than the well known phenomenological Guggenheim, Anderson. and de Boer (GAB) model previously developed by the authors [Janjai, S., Bala, B.K., Tohsing, K., Mahayothee, B., Heawsungcharern, M., Muhlbauer, W., Muller, J, 2006. Equilibrium moisture content heat of sorption of longan (Dimocarpus longon). Drying Technology 24, 1691-1696]. The ANN model was programmed in C++. The isosteric heat of sorption of longan is predicted by a power law model developed ill this Study, which was found to have better fit than the exponential model previously developed by the authors [Janjai, S., Bala, B.K., Tohsing, K., Mahayothee, B., Heawsungcharern, M., Muhlbauer, W., Muller, J., 2006. Equilibrium moisture content heat of sorption of longan (Dimocarpus longan). Drying Technology 24, 1691-1696]. Also a power law model was developed for entropy of sorption. The net isosteric heats of sorption were compared for longan (D. longan Lour.), litchi (Litchi chinensis Sonn.) and mango (Mangifera indica L. cv. Nam Dok Mai). Longan and litchi have the same pattern of variation in heat of sorption with moisture contents which might be due to similar biological structure of both of these two fruits. However, at a moisture content of above 50% (d.b.) the isosteric heat of sorption of longan is lower than and it is higher at a Moisture content below 50% (d.b.) when compared with that of mango. This set of two equations (isosteric heat and entropy) would be useful in the simulation of storage of dried longan. The artificial neural network model predicts equilibrium moisture contents more accurately and hence better equations for heat of sorption and entropy are developed based on data from the neural network model.