ABSTRACT Calorific value provides a strong measure of useful energy during coal utilization. Previously, different AI techniques have been used for the prediction of calorific value; however, one model is not valid for all geographic locations. In this research, Lower Calorific Value (LCV) of the Thar coal region in Pakistan is predicted from proximate analysis of 693 drill holes extending to 9,000 sq. km. Researchers have applied different techniques to produce the best model for prediction of calorific value; however, Gradient Boosting Trees (GBT) has not been used for this purpose. A comparison of GBT, Back-propagation Neural Networks (BPNN), and Multiple Linear Regression (MLR) is presented to predict the calorific value from a total of 8,039 samples with 1 m support interval. The samples were split randomly into 70:15:15 for training, testing, and validation of GBT, BPNN, and MLR models, reporting correlations of 0.90, 0.89, and 0.80, respectively. The features’ importance was reported by the intuitive and best-performing GBT model in decreasing order of importance as: Volatile Matter, Fixed Carbon, Moisture, and Ash with corresponding feature importance values of 0.50, 0.30, 0.12, and 0.08.

中文翻译：

---

预测塔尔褐煤矿床的热值：反向传播神经网络 (BPNN)、梯度提升树 (GBT) 和多元线性回归 (MLR) 之间的比较

摘要 热值提供了煤炭利用过程中有用能量的有力衡量标准。此前，不同的人工智能技术已被用于预测热值；然而，一种模型并非对所有地理位置都有效。在这项研究中，通过对延伸至 9,000 平方公里的 693 个钻孔的近似分析，预测了巴基斯坦塔尔煤区的低热值 (LCV)。研究人员应用了不同的技术来生成预测热值的最佳模型；然而，梯度提升树 (GBT) 尚未用于此目的。提出了 GBT、反向传播神经网络 (BPNN) 和多元线性回归 (MLR) 的比较，以预测来自总共 8,039 个样本的热值，支持间隔为 1 m。样本被随机分成 70:15:15 进行训练、测试、和 GBT、BPNN 和 MLR 模型的验证，报告的相关性分别为 0.90、0.89 和 0.80。特征的重要性由直观且性能最佳的 GBT 模型按重要性递减顺序报告为：挥发性物质、固定碳、水分和灰分，相应的特征重要性值为 0.50、0.30、0.12 和 0.08。