Predicting the entire pyrolysis process of representative charring material infiltrated with kerosene using an improved artificial neural network

It is crucial to predict the entire pyrolysis process of charring materials infiltrated with flammable liquids to provide guidance for reducing the fire hazards and implementing numerical simulations of fire accidents caused by the leakage of flammable liquids. In the present study, pyrolysis data obtained from the thermogravimetric experiments were preprocessed using linear interpolation and data normalization methods. The processed data set was then used to train the artificial neural network (ANN) framework to predict the pyrolysis behaviors of typical charring material (Mongolian Scots pine) infiltrated with typical flammable liquid (kerosene) under three different conditions: (1) pyrolysis at a constant mass fraction of kerosene with multiple heating rates; (2) pyrolysis at multiple mass fractions of kerosene with a constant heating rate; and (3) pyrolysis at multiple mass fractions of kerosene and multiple heating rates. The ANN, with the double hidden layer structure (6−3), was capable to well predict the entire pyrolysis behaviors of Mongolian Scots pine under all the three scenarios. Besides, five data transformation function (multiplication, exponential function with base e, logarithmic function with base e, square root, and exponentiation function) were applied to the present three input variables (temperature, mass fraction of kerosene, and heating rate) to generate new input variables in order to extract more characteristics among the pyrolysis data. Therein, the ANN utilizing the data transformation functions of multiplication and exponential function with base e exhibited the best predictive ability. Among the three input variables in the ANN (temperature, mass fraction of kerosene, and heating rate), temperature was identified as the most sensitive one to the prediction performance, followed by mass fraction of kerosene, with heating rate being the least sensitive. [Display omitted] •Data preprocessing can significantly improve the prediction accuracy of ANN.•ANN with double hidden layer can predict pyrolysis curves in all scenarios.•The data transformation function can change the effect of ANN prediction.•Temperature has the highest sensitivity among pyrolysis parameters.