煤层属性空间变异的差分进化径向基神经网络插值

Spatial interpolation of coalbed properties by radial basis function neural network based on differential evolution

为提高煤层属性空间变异的插值精度,建立了径向基函数神经网络(RBFNN)预测模型。为提高差分进化算法(DE)的全局寻优能力,提出基于非均匀变异的最优克隆算子,使之融入DE,形成最优克隆差分进化算法(OCDE);并应用OCDE优化RBFNN的参数,构成了差分进化径向基神经网络插值方法。以贵州省织纳煤田为例,应用于煤层属性预测,分别设立插值方法的拟合精度评价指标——标准均方根误差(ENRMS)和预测精度评价指标——平均相对误差百分比(EMRP)。差分进化径向基神经网络方法在84个样本时,煤层厚度属性插值的ENRMS和EMRP值分别为23.31%和11.63%。在样本容量为84、74、64、54、44、34个训练样本集条件下,该方法的ENRMS和EMRP值都小于相应训练样本集的Kriging方法,插值的拟合精度和预测精度都显著好于Kriging方法。

For improving the interpolation precision of spatial variability of coalbed properties,based on optimal clone differential evolution algorithm（OCDE）,a radial basis function neural network（RBFNN） was designed for interpolating coalbed properties using limited data.The OCDE that was applied to find optimal parameters for RBFNN was generated by the integration of differential evolution algorithm（DE） and optimal clone operator（OC） with non-uniform mutation.Normalized root mean square error（ENRMS） and mean relative error percent（EMRP） respectively came into use for evaluating the levels of fitting and prediction precision of RBFNN and kriging for spatial interpolation of coalbed properties.In the case study,the ENRMS and EMRP of RBFNN interpolation for coalbed thickness are 23.31% and 11.63% when the sample size is at 84.For the coalbed thickness and other properties,the ENRMS and EMRP of RBFNN interpolation are fewer than that of Kriging with different sample sizes.