Comparison of neural networks. An estimation model in yield of monoglycerides from biodiesel by-product

Abstract

Biodiesel is generally manufactured by transesterification, obtaining glycerol as a by-product. The transesterification of methyl stearate selectively produced monoglycerides, for glycerol valuation. Mixed oxides containing lithium catalysed the reaction. The purpose of this work was to develop and compare mathematical models obtained through artificial neural networks (ANN), capable for characterising the relationship between the mole percent conversion of methyl stearate and the yield of the products mono-, di- and triglycerides. The lowest mean squared error (MSE), the highest correlation coefficient (R), similarity in the evolution of validation and simulation errors and absence of data overlearning were considered to select the best model. Three ANNs with backpropagation structures were compared. They evidenced high correspondence between the estimated product yield values and the interpolated experimental ones. The ANN containing 35 neurons with sigmoid transfer function in the hidden layer and a linear neuron in the output one was the simplest. Consequently, the 5, 15 and 60 neurons were also explored in the hidden layer. The ANN structured with an intermediate number of neurons (35) achieved the most adequate MSE, considering mono- and diglyceride products (0.011193, 0.000489). The development of these models contributes to the dynamic estimation of the process.