巨厚多分层露天煤矿煤层厚度动态预测模型

Dynamic prediction model for coal seam thickness in open-pit coal mines with ultra-thick multilayer coal seams

  • 摘要: 针对巨厚多分层露天煤矿煤层厚度空间变化复杂、静态预测难以及时利用随采揭露信息导致预测精度较低的问题,构建了融合极端梯度提升树(XGBoost)与普通克里金(OK)的煤层厚度动态预测模型。以平面坐标、底板标高、倾角、煤层层位及邻域连续性等特征为输入,采用XGBoost拟合煤层厚度的全局确定性趋势;采用球状模型对空间局部残差拟合半变异函数,并通过OK进行空间补偿,形成回归克里金(RK)预测模型;进一步引入基于滑动时窗的随采数据同化机制,随工作面推进更新预测模型,实现煤层厚度动态预测。基于大南湖二矿403个钻孔、覆盖21层可采煤层的1 864条建模样本,采用按钻孔分组的10折交叉验证评价模型性能,并通过模拟随采揭露实验检验动态更新效果。结果表明:该模型的合并决定系数(R2)为0.894,层内R2为0.684,煤量相对偏差为−0.045%;与XGBoost模型相比,总体预测精度相近,层内预测能力提高,煤量估计偏差减小;在模拟随采更新场景中,累计同化近80个新增见煤点后,工作面前方煤层厚度预测的均方根误差较静态预测降低39%。该模型能够利用新增揭露信息修正煤层厚度预测,可为露天煤矿剥采方案优化与煤量核算提供数据支持。

     

    Abstract: To address the complex spatial variation in coal seam thickness in open-pit coal mines with ultra-thick multilayer coal seams and the low prediction accuracy caused by the inability of static prediction methods to promptly incorporate information revealed during mining, a dynamic coal seam thickness prediction model integrating Extreme Gradient Boosting (XGBoost) and Ordinary Kriging (OK) was proposed. Planar coordinates, floor elevation, dip angle, coal seam horizon, and neighborhood continuity were used as input features, and XGBoost was used to fit the global deterministic trend in coal seam thickness. A spherical model was used to fit the semivariogram of local spatial residuals, and OK was used for spatial compensation to form a Regression Kriging (RK) prediction model. A sliding-time-window-based mechanism for assimilating data acquired during mining was then introduced to update the model as the working face advanced, enabling dynamic prediction of coal seam thickness. Model performance was evaluated using 10-fold cross-validation grouped by borehole, based on 1 864 modeling samples covering 21 mineable coal seams from 403 boreholes at Dananhu No.2 Mine. The effectiveness of dynamic updating was tested by simulating the progressive exposure of coal seam data during mining. The results showed that the model achieved a pooled coefficient of determination (R2) of 0.894, a within-seam R2 of 0.684, and a relative coal tonnage bias of −0.045%. Compared with XGBoost model, the model achieved similar overall prediction accuracy, improved within-seam predictive performance, and reduced coal tonnage estimation bias. In the simulated updating scenario, after nearly 80 newly exposed coal observation points had been cumulatively assimilated, the root mean square error of coal seam thickness predictions ahead of the working face was 39% lower than that of the static prediction model. The model can use newly exposed information to correct coal seam thickness predictions and provide data support for optimizing stripping and mining plans and accounting for coal tonnage in open-pit coal mines.

     

/

返回文章
返回