Predictive Ability of Four Statistical Models for Determining the Influence of Coal Thermophysical Properties During the Initial Phase of Coal Spontaneous Combustion
Kai-Qi Zhong,Yang Xiao,Xia Zhao,Lan Yin,Chi-Min Shu,Yuan Tian
DOI: https://doi.org/10.1016/j.fuel.2021.120348
IF: 7.4
2021-01-01
Fuel
Abstract:Studying the thermophysical properties of coal is crucial for mitigating coal spontaneous combustion (CSC). We tested four artificial intelligence statistical models, namely multiple linear regression (MLR), bagging, random forest (RF), and back propagation neural networks (BPNNs), to determine the most accurate prediction model of the thermophysical properties of coal simples. Six impact factors, containing moisture, ash, volatility, fixed carbon, oxygen concentration, and temperature, were obtained using grey correlation analysis. A total of 250 sets of 13 types of coal related data were randomly divided to create 225 training sets and 25 verification sets. The results revealed that MLR is not suitable for the prediction of the thermophysical properties of coal. Although favourable training results could be achieved using BPNNs, its prediction results were inferior. Drilling and forecasting predictions of the bagging and RF models were excellent and their goodness of fit (R2) exceeded 0.91. Moreover, the error (RMSE, MAE, and MAPE) of RF model was lower than the bagging, which exhibited optimal performance. Therefore, the RF model is the most suitable for predicting coal thermophysical properties.