Prediction of Acid Values of Vegetable Oils Having High Free Fatty Acids Using Artificial Neural Networks

Non-edible vegetable oils are found to be a good option for diesel fuel, as these oils are renewable energy or green energy and less polluting to the environment. In developing countries like India, especially southern parts of the country (i.e., in Kerala), more land is cultivated with rubber trees. Rubber seed oil, a non-edible oil, can be used as an alternative to diesel fuel. It cannot be used directly in the diesel engine as it results in injector choking and carbon deposits. Rubber seed oil usually has high free fatty acid, and has to be reduced by the chemical modification (two-step pretreatment) in order to increase the yield of rubber seed methyl ester (biodiesel). A two-step pretreatment method was developed to reduce acid value from 48 to 1.72 mgKOH/g, and then it was subjected to transesterification to modify chemically. Due to the complexity of the process, cost, and time for experimentation, an ANN model was designed to find acid value of the non-edible oils. The acid values of rubber seed oil were found experimentally as per ASTM D 974 and the results were used to train the network and test with values set aside, which were not trained. ANN was optimized by changing activation functions or transfer functions and the number of neurons in the hidden layer. The network was trained by 84 data points and tested with data not used for training. Back propagation algorithm with a single hidden layer, Logsigmoid (activation function), Trainlm (training rule), 13 neurons in a hidden layer, and goal error of 10-5 were obtained by optimization of the network. To improve accuracy of the network, regression analysis was done and the correlation coefficient (R2) obtained was 1. This optimized network was used to predict the acid values and it was compared with experimentally found values. The predicted values were comparable with actual values and were found to be approximately equal. The absolute errors were calculated for the values 0.54, 0.19, and 0.39 and relative errors were 8.6, 3.9, and 7.4%, respectively.