Abstract: Traditional single prediction model in the coalmine gas concentration prediction can’t objectively and accurately reflect its change law because there are many factors that affect the change of gas concentration and internal relationship is complex. In order to effectively improve the gas concentration prediction precision, a coalmine gas concentration of sub-state prediction model is proposed. Using the maximum Lyapunov exponent the gas concentration time series are divided into non-chaotic state and chaotic state, then the improved Least Squares Support Vector Machines（LS-SVM） and neural network based on the Radial Basis Function（RBF） are used for modeling and training parameters optimization. It gets the best prediction model and the gas concentration time series prediction simulation experiment is carried out. The results show that the sub-state prediction model improves the prediction accuracy effectively and reduces the prediction error, the method can be more objective and accurate to forecast the gas concentration.

Key words: sub-state prediction, phase space reconstruction, chaotic and non-chaotic, Support Vector Machine（SVM）, neural network

摘要： 由于影响瓦斯浓度变化的因素很多且内部关系复杂，传统的单一预测模型无法客观准确地反映其变化规律，导致预测精度较低。为有效提高瓦斯浓度预测精度，提出一种基于分态的预测模型。应用最大李雅普诺夫指数（Lyapunov指数）对瓦斯浓度时间序列的混沌特性进行识别，将其分为非混沌态和混沌态，接着分别采用改进的最小二乘支持向量机（LS-SVM）和基于径向基函数（Radial Basis Function，RBF）的神经网络进行建模和训练参数的优化，最终得到最佳预测模型并对瓦斯浓度时间序列进行预测。结果表明，分态预测模型有效提高了预测精度，降低了预测误差，用该方法可以更加客观准确地对瓦斯浓度进行预测。

关键词: 分态预测, 相空间重构, 混沌和非混沌, 支持向量机, 神经网络