%A %T Theoretical Analysis and Verification of Prediction Variances and Confidence Limits for Global Temperature Compensation Modeling Approaches %0 Journal Article %D 2017 %J SPECTROSCOPY AND SPECTRAL ANALYSIS %R 10.3964/j.issn.1000-0593(2017)04-1058-06 %P 1058-1063 %V 37 %N 04 %U {https://www.gpxygpfx.com/CN/abstract/article_9059.shtml} %8 2017-04-01 %X Temperature fluctuations affect the action between hydrogen groups, which causes the changes of absorption intensity and peak position of the near infrared (NIR) spectrum,which results in less accuracy of the NIR analyzer. This article studies the prediction accuracy of a type of global temperature compensation models for NIR spectrometric analysis from the aspects of prediction variance and confidence limit respectively. In the designed experiments with continuous temperature changing, NIR spectra are collected with an equal time interval. Hence, the continuous impact on the principal compoments of the NIR spectra caused by temperature is observed and analyzed, which illustrates the mechanism of temperature effection to the model prediction results. As verified by the experimental,the measurement of viscosity of an industrial polymeric material is carried out combined with different modelling methods. According to the experimental verification, the accuracy of non-temperature compensation model and global temperature compensation model are as follows, respectively: RMSEC=0.243 0, Rc=0.871 6, RMSEP=0.243 2, Rp=0.869 3; RMSEC=0.258 2, Rc=0.870 6, RMSEP=0.265 2, Rp=0.856 0. The maximum prediction confidence intervals of these two types of models are about 1.8 and 0.9 kPa·s respectively. Therefore, it can be observed that the modeling accuracy of global temperature compensation model is slightly worse, but the predition precision is much better compared with non-temperature compensated model. Both results of theoretical analysis and experimental verification illustrate that the global temperature compensation modelling methods offer the more accurate models and better robustness and reliability.