基于多技术融合加权平滑的径向基函数神经网络熔融指数预报  

Melt index prediction based on radial basis function neural network with multi-technology integrated weighting and smoothing

作　　者：陈红梅[1] 刘兴高[2] 

机构地区：[1]潍坊学院机电与车辆工程学院,山东潍坊261061 [2]工业控制技术国家重点实验室(浙江大学),杭州310027

出　　处：《计算机应用》2015年第A01期104-109,共6页journal of Computer Applications

基　　金：国家自然科学基金资助项目(U1162130);山东省高等学校青年骨干教师国内访问学者项目

摘　　要：针对在对聚丙烯熔融指数进行预测时优势数据和优势变量不突出影响预测精度、数据平滑度不够影响泛化性能的问题,提出了基于多技术融合加权平滑的径向基函数神经网络预报模型。综合运用了在时间尺度基于空间欧氏距离加权、在变量维度上基于灰色关联和线性回归误差加权两种数据加权方案,基于过程变量差分序列欧氏距离的平滑和局部线性平滑两种数据平滑方案,解决了模型精度和泛化性低的问题。为进一步改进模型性能,采用带误差补偿的非线性自回归滑动平均模型框架和径向基函数神经网络,利用自校正预测控制算法和分段线性变学习率算法,对模型进行优化。结合某厂真实数据对模型进行验证,预报结果在泛化集上为:平均相对误差(MRE)1.32%、均方根误差(RMSE)0.045 9。与其他方法进行了详细的比较分析,结果表明该模型具有良好的预报精度和泛化性能,在大时滞工业过程领域具有一定的应用价值。In view of the problems that not highlighted predominant data and variables affect the prediction accuracy and not enough data smoothness influences the generalization performance in the prediction of PolyPropylene （ PP ） Melt Index （ MI） , this paper proposed a forecasting model of MI based on Radial Basis Function Neural Network （ RBFNN） with weighting and smoothing of multi-technology integration. The proposed model integratedly applied two data weighting schemes： weighting based on space Euclidean distance in the time scale, weighting based on grey correlation and linear autoregression error in the variable dimension, and also applied two data smoothing methods： smoothing based upon the Euclidean distance of process variable differential sequence and partial linear smoothing, to solve the problems of low prediction precision and generalization ability. To further improve the forecasting capability of the model, based on the Nonlinear Autoregressive Moving Average （ NARMA） model framework with error compensation and the RBFNN, the paper used the self-tuning predictive control algorithm and the piecewise-linear alterable learning rate algorithm to optimize this model. The presented model is validated by the real data from a plant and the prediction results on the generalization database are as follows： Mean Relative Error （ MRE） is 1. 32%, Root Mean Square Error （ RMSE） is 0. 045 9. Compared and analysed detailedly with the report in the literature, the results show that the proposed model in the paper has an excellent forecasting accuracy and generalization ability, and has a certain application value in the industrial process with large time delay.

关 键 词：多技术融合 加权 平滑 自校正预测控制 分段线性变学习率 径向基函数神经网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]