A method for predicting the tar yield of tar-rich coals based on the BP neural network using multiple indicators of coal petrography and coal quality

Objective Tar yield, the most important coal quality parameter for coal utilization through low-temperature pyrolysis, determines the clean utilization of tar-rich coals. However, various constraints result in limited test data on tar yield in the geological exploration stage of coals, substantially restricting the fine-scale assessment and efficient utilization of tar-rich coals.

Methods To achieve more scientific and accurate fine-scale tar-rich coal assessments, this study examined 1073 sets of lithotype and coal quality data obtained previously from a Jurassic coalfield in northern Shaanxi. From these data, 141 sets with 20 lithotype and coal quality parameters regarding macerals, proximate analysis, ultimate analysis, and ash composition analysis were selected. Then, employing the back propagation (BP) neural network algorithm, this study constructed two tar yield prediction models based on 20 lithotype and coal quality indices and four proximate analysis indices each (also referred to as the first and second models, respectively). Finally, it assessed the accuracy and rationality of the results of both prediction models.

Results and Conclusions The results are as follows: (1) The first model exhibited a mean square error (MSE) of 0.30 in the final training and a mean absolute error (MAE) of 0.65 for the prediction results of the test set data. In contrast, the second model yielded a MSE of 1.07 in the final training, with a MAE of 1.35 for the prediction results of the test set data. For the prediction results of the superset data, the first and second models yielded MAEs of 0.84 and 1.34, respectively, suggesting that the first model features higher goodness of fit and generalization performance. (2) The importance of 20 lithotype and coal quality indices in the first prediction model was further quantitatively analyzed using the Shapley additive explanation (SHAP) algorithm. The results reveal that factors including vitrinite, hydrogen and carbon elements, Fe₂O₃, moisture, volatile constituents, exinite, and oxygen content prove to be the positive factors influencing the tar yield, whereas Al₂O₃, inertinite, fixed carbon, ash content, and SiO₂ content serve as negative factors influencing the tar yield. The intrinsic relationships between both the lithotype and coal quality and the tar yield, derived from the first model, align well with the general understanding of the geological factors influencing the tar yield. Therefore, the first prediction model can effectively predict the tar yield of the Jurassic coalfield in northern Shaanxi, providing support for the clean and efficient utilization of tar-rich coals in northern Shaanxi.