Knowledge extraction for fischer - tropsch synthesis using local learning approaches

Abstract

In this study, local empirical models for Fischer-Tropsch Synthesis (FTS) were constructed by using machine learning algorithms on the experimental data published in the literature. CO conversion in FTS was modeled as a function of catalyst design variables, physical properties of the catalyst, and operating conditions by using gradient boosting method (GBM), artificial neural network (ANN) and random forest in R 3.2.2 environment. The FTS database was taken from a previous study. Before modeling FTS database, local and global learning techniques were compared using three different simpler and known databases involving selective CO oxidation reaction, water gas shift reaction and steam reforming of methane reaction. These databases were modeled by using GBM, ANN and random forest with both local and global techniques to see the difference in a known model. Local learning showed better performance in each database on the prediction of unseen data. Then the Fischer-Tropsch synthesis database was analyzed by using local learning with a different number of query data. Optimal parameters were determined for each method by training the model with each combination of parameters the standard deviation and mean were calculated and compared to determine the optimal parameters. The model with the best results was chosen as the final model. The R2 values ranged between 0.86 and 0.87 with GBM, 0.68 and 0.75 with ANN and 0.82 and 0.86 with random forest algorithms. RMSE values ranged between 8.60 and 9.16 with GBM, 12.24 and 13.98 with ANN and 9.23 and 10.32 with random forest algorithms. R2 and RMSE values were better than global learning results. GBM algorithm performed better on FTS database both in terms of R2 and RMSE.