Devolatilization predicting model based on coal heterogeneous chemical structure from micro-Raman spectroscopy with neural network

Abstract

This study proposes a novel devolatilization model based on a convolutional neural network (CNN), employing quantified coal chemical structure features as input. Initially, a reliable method was developed to quantify the average/heterogeneous properties of coal structures using high-resolution micro-Raman spectroscopy. The evolution of chemical structure with increasing coal rank was investigated using ¹³C Nuclear Magnetic Resonance (NMR) and micro-Raman spectroscopy. A strong positive correlation was observed between the parameters derived from these two techniques, highlighting their complementarity and enhancing the capability of micro-Raman for analyzing heterogeneous chemical structures. As coal rank increased, the distribution features of different structural types exhibited significant changes, particularly in the heterogeneity of fluorescence-rich and aromatic ring structures, which initially increased and subsequently decreased. Furthermore, the potential of heterogeneous chemical characteristics for predicting coal devolatilization was explored. While the general distribution model (GDM) demonstrated substantial potential in predicting devolatilization, its precision was found to be insufficient. To address this limitation, a CNN was introduced to improve prediction accuracy. The results revealed that compared to the direct use of chemical distribution parameters as input (CNN model), the GDM-CNN model, which incorporates the results from GDM, achieved the highest and most balanced precision. The absolute prediction error for raw and blended samples was consistently below 23.5 °C. This work introduces a methodology for establishing a high-precision devolatilization model by combining quantitative chemical structure analysis with neural networks. This approach can be extended to other characterization techniques and solid fuel samples, demonstrating broad applicability.