基于SSA-XGBoost模型的地表下沉系数预测研究

Study on Prediction of Surface Subsidence Coefficient Based on SSA-XGBoost Model

为解决当前地表下沉系数预测模型精度有限、模型不统一、计算繁琐且不利于泛化等问题,对现有35组样本数据中地表下沉系数影响因素进行分析,建立了基于麻雀搜索算法(SSA)优化极限梯度提升树(XGBoost)的预测模型,通过SSA-XGBoost模型迭代学习地表下沉系数与煤层采高、煤层倾角、覆岩岩性、深采比、基载比和基采比之间的非线性映射关系,得到了基于SSA-XGBoost模型预计的下沉系数预测值,并利用拟合优度、预测均方根误差和平均绝对百分比误差对下沉系数预测值进行精度分析。结果表明:基于SSA-XGBoost模型建立的地表下沉系数预测组合模型的拟合优度为0.9516,预测均方根误差仅为0.0206,平均绝对百分比误差仅为2.47%;SSA-XGBoost模型相对于XGBoost模型、BP神经网络模型、随机森林算法模型拟合优度分别提升了15.79%、111%和62.81%,预测均方根误差分别降低了43.25%、76.61%和73.72%,平均绝对百分比误差分别降低了46.99%、73.52%和75.99%;基于SSA-XGBoost的下沉系数预测结果拟合效果较好且模型精度较高,研究结果可为地表下沉系数的预测提供参考。

In order to solve the problems of limited accuracy,inconsistent model,complicated calculation and poor generalization of the current prediction model of surface subsidence coefficient,a prediction model based on Extreme Gradient Boosting Tree(XGBoost)optimized by Sparrow Search Algorithm(SSA)was established by analyzing the influencing factors of surface subsidence coefficient in 35 existing sample data sets.The SSA-XGBoost model was used to iteratively learn the nonlinear mapping relationship between surface subsidence coefficient with coal seam mining height,coal seam dip angle,overburden lithology,ratio of buried depth and mining height,ratio of bedrock thickness and overburden thickness,as well as ratio of bedrock thickness and mining height.The predicted surface subsidence coefficient based on the SSA-XGBoost model was obtained,and its accuracy was verified by using goodness of fit,root mean square error,and average absolute percentage error.The results show that the goodness of fit of SSA-XGBoost model is 0.9516,the root mean square error of prediction is only 0.0206,and the average absolute percentage error is only 2.47%.Compared with other models(XGBoost model,BP neural network model and random forest algorithm model),the goodness of fit of SSA-XGBoost model is improved by 15.79%,111%and 62.81%respectively,the root mean square error of prediction is reduced by 43.25%,76.61%and 73.72%respectively,and the average absolute percentage error is reduced by 46.99%,73.52%and 75.99%respectively.It is found that the prediction results of subsidence coefficient based on SSA-XGBoost have a better fitting effect and a higher accuracy.