煤层瓦斯含量是矿井瓦斯灾害防治及煤层气勘探开发的基础参数，为提高其预测精度及科学性，对典型矿井煤层瓦斯含量的35组实测数据进行了零-均值规范化处理，通过全子集回归和随机森林2种特征选择方法对11类影响煤层瓦斯含量的参数进行不同规律组合，得到17种瓦斯含量特征参数组合。运用高斯过程回归、最小二乘支持向量机、梯度提升回归树和极限回归机等4种经典有监督机器学习算法，分别对17种特征参数组合进行预测，得到68种瓦斯含量预测模型。根据各机器学习算法平均判定系数≥0.800，对68种瓦斯含量预测模型进行初步筛选。综合归一化均方误差≤001以及希尔不等系数≤001，得到21种基于特征选择和机器学习融合的最优预测模型，并取平均值得到了最终预测序列。结果表明：最终预测序列的归一化均方误差为0007，希尔不等系数为0005，判定系数为0993，平均绝对误差为0170 m3/t，平均相对误差为075%，各精度评估指标均符合要求，所构建的多参数组合多算法融合的预测模型具有广泛的普适性且精度较高。

Coalbed gas content is an essential parameter for mine gas disaster prevention and CBM exploration and development. In order to improve its prediction accuracy and scientificity of gas content, 35 sets of measured data of coal seam gas content in typical coal mines have been standardized by zero-mean values. Through the complete subset regression method and the random forest feature selection method, the 11 types of parameters that affect the coal seam gas content were selected and combined in different rules and 17 combinations of gas content feature parameters were obtained. Four classic supervised machine learning algorithms, including Gaussian process regression, least squares support vector machine, gradient boosting regression tree, and limit regression machine, were used to predict 17 feature parameter combinations and 68 gas content prediction models were obtained. According to the average judgment coefficient of each machine learning algorithm ≥0.800, 68 kinds of gas content prediction models were preliminarily screened. combined with normalized mean square error≤0.01 and Hill unequal coefficient≤0.01, and 21 optimal prediction models based on the fusion of feature selection and machine learning were obtained. The final prediction sequence was obtained by averaging. The results show that the normalized mean square error of the final prediction sequence is 0.007, the Hill unequal coefficient is 0.005, the determination coefficient is 0.993, the average absolute error is 0.170 m3/t, and the average absolute error is 0.75%. The accuracy evaluation indicators are all In line with the requirements, and the constructed prediction model of multi-method fusion under multi-parameter combination has a wide range of universality and high accuracy.

0 引言

1 瓦斯含量预测数据处理

   1.1 瓦斯含量预测数据实例获取

   1.2 数据标准化处理

2 煤层瓦斯含量预测模型建立流程及初选

   2.1 预测模型建立流程

   2.2 预测模型初选

3 煤层瓦斯含量预测模型优选及验证

   3.1 煤层瓦斯含量预测模型优选

   3.2 最优融合预测模型验证

4 结论