论文部分内容阅读
为实现对瓦斯浓度(体积分数)的准确预测,基于海量煤矿瓦斯监测数据和多元分布滞后模型(MDL)建立了多变量瓦斯浓度时间序列预测模型.基于惩罚最小二乘法和自回归的思想,提出了新的变量选择和定阶方法——Adjust Group最小绝对值压缩与选择(LASSO)方法.该方法以岭估计及局部二次近似迭代算法实现了预测模型的构建,通过有效选取具有解释性的自变量子集,提高模型的解释性,采用广义交叉检验准则(GCV)确定惩罚参数,并通过分组惩罚来实现变量筛选与滞后变量的定阶.结果表明:Adjust Group LASSO方法预测得到的残差平方和为0.433 0,具有较高的精度,能够较好的预测工作面瓦斯浓度的动态变化,与LASSO、最小角回归算法(LARS)以及其他瓦斯预测常用方法相比,大大提高了预测的准确性.
In order to achieve accurate prediction of gas concentration (volume fraction), a multivariate gas concentration time series prediction model based on mass coal mine gas monitoring data and multivariate distribution lag model (MDL) was established.Based on the idea of penalty least squares and autoregressive A new method of variable selection and ranking - Adjust Group Minimum Absolute Compression and Selection (LASSO) method is proposed in this paper.The construction of the prediction model is realized by ridge estimation and local quadratic approximation iterative algorithm. By selecting validly explanatory The subset of independent variables improved the interpretability of the model, and adopted the generalized cross-checking criterion (GCV) to determine the penalty parameters, and the grouping penalty was used to achieve the rank-order of variable selection and lagged variables.The results showed that the residuals predicted by Adjust Group LASSO The square sum is 0.433 0, which has high precision and can predict the dynamic change of gas concentration in the working face. Compared with LASSO, LARS and other commonly used methods of gas prediction, the prediction accuracy is greatly improved Sex.