矿区开采沉陷预计的改进BP神经网络模型

Mining Subsidence Prediction Based on Improved BP Neural Network Model

为精确预计锦界矿某工作面开采沉陷,首先结合该工作面的地质资料、采掘工作平面图及孔柱状图,采用FLAC3D软件建立了该工作面开采沉陷仿真模型,得到工作面推进100、300、500、700 m时的开采沉陷数据;其次基于该类数据对BP神经网络预计模型进行训练和验证,建立沉陷数据与工作面推进距离的非线性关联;然后用粒子群优化算法(Particle swarm optimization,PSO)对BP神经网络模型的结构参数和连接权值阈值进行优化,并引入遗传算法(Genetic algorithm,GA)中的自适应变异因子以一定概率初始化部分变量,以解决PSO算法易陷入局部最优解的问题,避免BP神经网络模型易陷入局部最小值、训练收敛速率低以及PSO算法易早熟收敛等问题。分别采用BP神经网络模型、PSO-BP神经网络模型以及所提模型进行试验对比,并引入偏差平方和(Sum of squares for total,SST)对各模型的预计精度进行评价,研究表明:在工作面分别推进100,300,500 m的情况下,BP神经网络模型的SST值分别为0.056,0.062,0.066,PSO-BP神经网络模型的SST值分别为0.049,0.054,0.048,所提模型的SST值分别为0.028,0.026,0.031,明显小于前两者,表明该模型有助于提高矿区开采沉陷预计精度,有一定的实用价值。

In order to predict the mining subsidence of a mining working face of Jinjie Mine,firstly, based on the geological data, mining working plans and bore log charts,the mining subsidence simulation model of the mining working face is established based on FLAC3D software, the mining subsidence data of the mining working face advancing distance of 100,300, 500,700 m are obtained;then, the BP neural network prediction model is conducted training and validation based on the above mining subsidence data, the nonlinear correlation relationship between mining subsidence data and mining working face advan- cing distance is established;thirdly, the structural parameters and the thresholds of connection weights are optimized by using particle swarm optimization （PSO） algorithm, part of the variables of BP neural network prediction model is initialized by a- dopting the adaptive mutation factor of genetic algorithm （GA） so as to solve the problems of local optimal solutions of PSO al- gorithm, besides that, it can also contribute to solve the problems of existence of local minimum values and low rate of training convergence and the premature convergence of PSO algorithm. The experiment of BP neural network model, PSO-BP neural network model and the improved BP neural model proposed in this paper is conducted,the sum of squares for total （SST） is used to evaluate the prediction precise of the above three models, the results show that：under the conditions of the mining working face advancing distance of 100,300,500 m, the SST value of BP neural network model is 0. 056,0. 062,0. 066 respectively,the SST value of PSO-BP neural network model is 0. 049,0. 054,0. 048 respectively, the SST value of the model proposed in this paper is 0. 028,0. 026,0. 031 respectively,which are lower than the others significantly,and it is further show thatthe improved BP neural network proposed in this paper is good to improve the mining subsidence prediction precise, and it also has certain practical value.