Liquid Density Prediction of Ethanol/Water, Using Artificial Neural Network

Abstract

In this work, our objective was to get a reliable model for predicting liquid density ethanol-water and use it again later in modeling the ethanol production process from biomass. Hence, the unreliability of the Peng-Robinson equation of state to predict this property was shown. The average absolute deviation of this prediction is equal to 14.72 %. To have a reliable model, an artificial neural network (ANN) method was followed. Levenberg–Marquardt algorithm is used to choose the optimized ANN structure that has ten neurons in the hidden layer, three neurons in the input layer, and one neuron in the output layer, with a tangent-sigmoid and linear transfer functions, in the hidden and the output layers, respectively. The model training was done using 348 experimental data points from published experiments, realized at different liquid mole fraction range, pressure (0.10 to 10.00MP), and temperature (298.15 K to 476.2 K). The correlation coefficient between the experimental and liquid phase density was 0.9999 for training, validation, and testing the model. Statistical analysis is employed to evaluate the accuracy of the ANN, showing that the average absolute deviation, root mean square, and the Bias are 0.047 %, 0.003 %, and -0.004 %, respectively. So the ANN model gives a good estimation of liquid density, for mixture ethanol/water, with a relative importance of pressure, composition, and temperature equal to 41%, 34 %, and 25 %, respectively.