矿山深部开采覆岩沉陷预测的改进粒子群优化BP模型

为提高矿山深部开采覆岩沉陷的预测精度，以吉林省某矿山为例，将覆岩抗压强度、工作面推进距离、采厚、煤层倾角、采深作为开采沉陷预测的主要影响因素，以煤层倾角、工作面推进距离、采厚等参数作为预测模型的输入值，将最大沉陷值作为预测模型的输出值，构建了BP神经网络开采沉陷预测模型。针对经典BP神经网络模型在训练时具有学习速度慢、抗干扰能力弱以及易陷入局部最小值等不足，采用一种改进型粒子群优化算法对其进行了优化，构建了改进粒子群优化BP神经网络模型。试验表明：①所提模型的平方相关系数R₂为0.932、平均绝对误差e_ME为0.195、平均相对误差e_MRE为0.082、训练时间t为21.5 s、均方误差e_MSE为0.067 1；②经典BP神经网络模型的平方相关系数R₂为0.802、平均绝对误差e_ME为0.605、平均相对误差e_MRE为0.255、训练时间t为30.9 s、均方误差e_MSE为0.089 1；③PSO-BP神经网络模型的平方相关系数R₂为0.825、平均绝对误差e_ME为0.382、平均相对误差e_MRE为0.216、训练时间t为23.5 s、均方误差e_MSE为0.078 2。可见，所提模型具有较高的训练效率，拟合效果较好且预测精度较高，对于大幅提升矿山开采沉陷预测精度有一定的借鉴价值。

Overlying Strata Subsidence Prediction of Deep Mining Based on Improved Particle Swarm Optimization BP Model

In order to improve the overlying strata subsidence prediction accuracy of deep mining,taking a mine in Jilin Province as the study example, taking the parameters of comprehensive strength of overlaying strata, advancing distance of working face and mining thickness,dip angle of coal seam and mining depth as the main influence factors of mining subsidence prediction,among them,selecting the factors of dip angle of coal seam,advancing distance of working face and mining thickness as the input values of the prediction model,and the output value of the prediction model is maximum subsidence value,so as to establish BP neural network mining subsidence prediction model. According to the deficiencies of slow learning speed,weak a-bility of anti-interference and easily fall into local minimum value of the classical BP neural network model,so,it is optimized by using a improved particle swarm algorithm,therefore,the improved particle swarm optimization BP model is established. The test results show that:①the improved BP model proposed in this paper,which square correlation coefficient ( R2 ) is 0. 932, mean absolute error (eME) is 0. 195,mean relative error (eMRE) is 0. 082,training time (t) is 21. 5 s,mean square error (eMSE) is 0. 0671;②classical BP neural network model,which which square correlation coefficient (R2) is 0. 802,mean ab-solute error (eME) is 0. 605,mean relative error (eMRE) is 0. 255,training time (t) is 30. 9 s,mean square error (eMSE) is 0. 0891;③PSO-BP neural network model,which square correlation coefficient ( R2 ) is 0. 825,mean absolute error ( eME ) is 0. 382,mean relative error (eMRE) is 0. 216,training time (t) is 23. 5 s,mean square error (eMSE) is 0. 0782. The above study results further show that the training efficiency of the newly proposed model is higher than other two models,the fitting effects is also better than others,what's more,the prediction accuracy of the model is superior to the other two models,which indicated that the model proposed in this paper is help for improving the prediction accuracy of the mining subsidence.