Optimization based comparative study of machine learning methods for the prediction of bio-oil produced from microalgae via pyrolysis

Abstract

Production of bio-oil from the pyrolysis microalgae is an effective and alternative fuel resources. However, to examine the correlation between pyrolysis conditions, ultimate, and proximate analysis with bio-oil production is an intricate and a challenging task for the experimental technique. Therefore, an efficient and well-organized model must be created to reliably predict the effect of input parameters on the bio-oil yield. A novel particle swarm-based and genetic algorithm-based selection of features and hyperparameters optimization is used in this study and based on these optimization techniques five different machine learning models were developed and compared. It was found that Gaussian Process Regression model performed better and the values of R2 (Coefficient of determination) = 0.997 and RMSE (Root mean Square Error) = 0.0185 while using PSO based features selection and R2=.994, RMSE = 0.0120 for GA performed better and is highly recommended. The result of SVM was the worst one R2 = 0.43, RMSE = 7.01 for PSO and R2 = 0.55, RMSE = 5.80 for GA. The values of R2 for DT, ANN, Ensembled tree were 0.91, 0.92, 0.83 for PSO based study and for GA based algorithms the values of R2 were 0.62, 0.93, and 0.94 respectively. The significance of independent input factors on dependent output responses was thoroughly examined using partial dependence plots and Shapley method. Moreover, an easily usable software (GUI) was developed by applying GPR model to predict yield of bio-oil. The difference between the yield predicted by GUI and experimental study was found to be 0.568, 1.48, .06, 0.42. This study offers new intuitions in the pyrolysis of microalgae and to enhance production of bio-oil.