采空区煤自燃预测的随机森林方法

煤自然发火温度的准确预测是矿井煤自燃防控的关键。为了准确可靠地预测采空区煤自燃温度,在大佛寺煤矿40106综放工作面开展了长期的采空区温度和气体观测实验,提出了一种基于随机森林(RF)方法的采空区煤自燃预测模型,并将预测结果与支持向量机(SVM)和BP神经网络(BPNN)方法对比。采用粒子群优化算法(PSO)对RF和SVM超参数进行优化,建立了参数优化的PSO-RF和PSO-SVM预测模型。结果表明,RF,PSO-RF,SVM和PSO-SVM模型均具有较强的泛化性和鲁棒性; RF在建模过程中拥有宽广的参数适应范围,当树的数量(ntree)超过100后,其训练误差趋于稳定,ntree的改变对预测性能没有实质的影响;虽然PSO算法可以找到RF最优超参数,但默认参数的RF模型就能获得满意的预测性能; SVM预测结果则对超参数十分敏感,PSO优化可以显著提高其预测精度,其预测性能依赖于超参数的最优选择; BPNN模型在训练阶段拥有极佳的预测结果,但易出现"过拟合",导致泛化性弱,测试阶段误差较大。通过在其他矿井煤自燃预测中应用,验证了RF方法的稳定性和普适性,且无需复杂参数设置和优化就能获得良好的预测性能,可进一步应用于其他能源燃料领域。

Random forest method for predicting coal spontaneous combustion in gob

The accurate prediction of coal temperature plays a vital role in preventing and controlling the coal spontaneous combustion in coal mines.To predict the temperature of coal spontaneous combustion in a gob accurately and reliably,a long term observation test of temperature and gases was implemented in the gob of 40106 fully mechanized top coal caving face at Dafosi coal mine.A prediction model of coal spontaneous combustion in the gob based on random forest (RF) method was proposed,and the prediction results were compared with the support vector machine (SVM) and BP neural network (BPNN) methods.The particle swarm optimization (PSO) algorithm was employed to optimize the hyper parameters of RF and SVM for establishing the PSO RF and PSO SVM prediction models with optimized parameters.The results indicate that RF,PSO RF,SVM,and PSO SVM models all had strong generalization and robustness.RF possessed a wide range of parameters adaptation in the modeling process.When the number of trees (ntree) exceeded 100,the training errors tended to be stable,and the change of  n tree had no substantial impact on the prediction performance.Although the PSO algorithm could find the optimal hyper parameters of RF,the RF model with the default parameters could obtain a satisfactory prediction performance.The prediction results of SVM were very sensitive to its hyper parameters,PSO optimization could significantly improve its prediction accuracy,and its prediction performance depended on the optimal choice of hyper parameters.The BPNN model exhibited excellent prediction results in the training stage,but it was prone to “over fitting”,resulting in weak generalization and large errors in the testing stage.Through the application of coal spontaneous combustion prediction in other mines,the stability and universality of the RF method were verified,and good prediction performance could be obtained without complicated parameter settings and optimization,it could be further applied to other energy and fuel fields.