Prediction of activation energy for combustion and pyrolysis by means of machine learning

Abstract

Thermogravimetric analysis (TGA) is a widely used technique to determine the activation energy (Ea), which is an important parameter for thermochemical processes. Thus, researchers have recently developed computational methods to minimize the experimental effort. While there are some studies on estimating kinetic parameters such as Ea using artificial neural networks (ANN), these are insufficient for generalization because they involve only one or two operational parameters. Therefore, in this study, Ea estimation was performed by creating a realistic ANN model as a machine learning approach, including operating parameters that were not previously considered. TGA experiments were performed with biomass, coal, and blends under different operating conditions for the training and test data sets of the model. In order for the ANN to give a satisfactory result, the experimental results were enriched with the data from the literature. The dataset was analyzed using statistical tools like correlation map, feature importance etc. Then a feedforward neural network was developed using Levenberg-Marquardt optimization algorithm. As a result, by using an appropriate number of input variables and a sufficient amount of data, it was possible to conduct a reliable TGA simulation to calculate the Ea values, with R2 values greater than 0.96 and mean absolute percentage error values<20%. Furthermore, the results of the statistical analysis applied to the input parameters of the ANN model were found to be consistent with the scientific background. The initial and final temperatures of decomposition are the most significant parameters for the determination of Ea.