矿井瓦斯体积分数的Stacking集成预测研究
Study on Stacking integrated prediction of mine gas volume fraction
摘要
本文提出基于 Stacking 集成学习的瓦斯体积分数预测模型(MLR-MLP),有效融合多元线性回归与多层感知机优势以解决非线性特性.模型首层利用异构基模型分别提取瓦斯波动的线性趋势与复杂非线性特征,第二层元模型通过自适应学习最优融合参数,实现预测结果的智能加权融合.以某矿埋深 600 m 的 3025 工作面为对象,经 5 折交叉验证,模型测试集 EMR、EMA、ERMS分别为 0.220 3、0.047 0、0.057 1,误差较单一模型显著降低.实验显示,该模型在瓦斯突变点预测误差降低 15%以上,显著提升了精度与泛化能力.研究证实 Stacking 策略能有效协同模型的可解释性与表达能力,为矿井安全监测提供新范式.
This paper proposes a methane volume fraction prediction model(MLR-MLP)based on Stacking ensemble learning to address the nonlinear and time-varying characteristics of gas fluctuations.The first layer utilizes heterogeneous base models to capture linear trends and complex nonlinear features respectively,while the meta-model achieves intelligent weighted fusion of results through an adaptive weight fusion mechanism.Taking the 3025 working face at a depth of 600 m as the study subject,the model demonstrated superior performance after 5-fold cross-validation,with EMR,EMA,and ERMS reaching 0.220 3,0.047 0,and 0.057 1,respectively.Experimental results show that the model reduces prediction errors at gas mutation points by over 15%,effectively enhancing accuracy and generalization capability.This study confirms that the Stacking strategy synergistically leverages model interpretability and expressive power,providing a new paradigm for intelligent safety monitoring in coal mines.
年，卷(期)：
2026, 45(10)
分类号：
TD712
机标分类号：
TD712
TP391
F224
在线出版日期：
2026年10月29日
