摘要
采动覆岩裂隙带高度决定了卸压瓦斯抽采钻孔终孔或巷道层位布置参数,为进一步提高其预测精度,采集了不同矿区的361组数据,分析了采动裂隙带高度与采高、煤层倾角、工作面斜长、采深、硬岩岩性比例系数之间的关系;采用深度信念网络(DBN)、长短期记忆网络(LSTM)、Elman神经网络(ENN)等3种机器学习算法对采动裂隙带高度进行五折交叉验证,基于判定系数、均方根误差、平均绝对误差、平均绝对百分比误差等常用评价指标,筛选出LSTM为初步预测模型;采用遗传算法(GA)和麻雀搜索算法(SSA),对采动裂隙带高度LSTM预测模型进行优化,得到LSTM、GA-LSTM、SSA-LSTM 3种模型的预测结果.结果表明:SSA-LSTM预测模型较LSTM、GA-LSTM预测模型预测结果更优,其判定系数、均方根误差、平均绝对误差、平均百分比误差分别为0.991、0.329、0.148、0.017,各精度评估指标均符合判定要求,所构建的采动裂隙带高度预测模型精度较高且具有一定普适性.
Abstract
The height of the fracture zone in mined overburden rock determines the gas extraction parameters of the final hole or roadway of the gas extraction borehole.In order to further improve the prediction accuracy,by collecting 361 sets of data from different mining areas,the relationship between the height of the mining-induced fissure zone and parameters like mining height,coal seam inclination angle,working face oblique length,mining depth,and hard rock lithology proportion coefficient was analyzed.Three machine learning algorithms,including deep belief network(DBN),long short-term memory network(LSTM),and Elman neural network(ENN),were used to perform five-fold cross-validation on the height of mining-induced fissure.LSTM was selected as the preliminary prediction model based on the common evaluation indicators such as coefficient of determination,root mean square error,mean absolute error,mean absolute percentage error,etc.The LSTM prediction model of mining-induced fissure was optimized by genetic algorithm(GA)and sparrow search algorithm(SSA),and the prediction results of LSTM,GA-LSTM and SSA-LSTM were obtained.The results show the SSA-LSTM prediction model has the better prediction results than the LSTM and GA-LSTM prediction models.Its coefficient of determination,root mean square error,average absolute error,and average percentage error are 0.991,0.329,0.148,and 0.017 respectively.All accuracy evaluation indicators meet the prediction accuracy judgment requirements.The constructed mining fracture zone height prediction model has resonable high accuracy and adaptability certain universality.
基金项目
陕西省教育创新团队建设科研计划项目(21JP075)
国家自然科学基金重点项目(51734007)
NETL
NSTL
万方数据