Estimation of heavy and light rare earth elements of coal by intelligent methods

Since last two decades, several investigations in various countries have been started to discover new rare earth element (REE) resources. It was reported that coal can be considered as a possible source of them. REE of coal occur in low concentrations, and their detection is a complicated process; therefore, their predictions based on conventional coal properties (proximate, ultimate and major elements (ME)) may have several advantages. However, few studies have been conducted in this area. This study examined relationships between coal properties and REE (HREE and LREE) for a wide range of coal samples (708 samples). Variable importance measure (VIM) by Mutual information (MI) as a new feature selection method was applied to consider the heterogeneous structure of coal and assess the individual relation between coal parameters and REE to select the compact subsets as input variables for modeling and improve the performance of prediction. VIM by MI showed that Si- Carbon, and Al-Hydrogen are the best subsets for the prediction of HREE and LREE concentrations, respectively. A boosted neural network (BNN) model as a new predictive tool was used for REE prediction. BNN can significantly reduce generalization of error. Results of BNN models showed that the HREE and LREE concentrations can satisfactory estimate (R2: 0.83 and 0.89, respectively). Results of this investigation were approved that MI-BNN can be used as a potential tool for prediction of other complex problems in energy and fuel areas.