基于遗传-支持向量机的分布式光纤监测矿压时序预测

Time Series Prediction of Distributed Optical Fiber Monitoring Rock Pressure Based on GA-SVR

为了有效地掌握岩层内部变形,准确预测开采过程中的矿压显现规律,采用分布式光纤监测覆岩内部变形并结合支持向量机计算方法,将光纤频移变化度作为主要特征参数,构建混沌矿压数据相空间,采用遗传算法(genetic algorithm,GA)对支持向量机回归(support vector machine regression,SVR)超参数寻优。开展相似材料模型试验,模拟工作面开采,并引入光纤频移变化度概念,建立GA-SVR时序预测模型。预测结果与传统回归模型和BP神经网络模型(BP neural network,BPNN)进行比较。结果表明,BPNN容易发生过拟合,传统SVR模型依赖超参数选取,GA-SVR模型在超参数选取上更科学,不容易发生过拟合,预测精度高于上述两种算法,为矿压时序预测定量化提供科学依据。

The research aimed to effectively grasp the mechanism of internal deformation of rock strata and make accurate prediction of rock pressure law in excavation process.Distributed optical fiber was used to monitor the internal deformation of rock mass.Combined with the support vector machine,the optical fiber frequency shift variation was used as the main characteristic parameter to construct the phase space of chaotic rock pressure data.The genetic algorithm(GA)was used for support vector machine regression(SVR)for super-parametric optimization.By model tests of similar materials,mining working faces working face was simulated,and the concept of frequency shift variation of optical fibers was introduced to establish a GA-SVR Time Series Prediction Model.The prediction results were compared with traditional regression model and BP neural network model(BPNN).The results show that BPNN is prone to over-fitting,and the traditional SVR model depends on the selection of super-parameters.GA-SVR model is more scientific in super-parameter selection and not prone to overfitting.The prediction accuracy is higher than the above two algorithms.The quantification of forecasts provides a scientific basis.