Local calibration for a UV/Vis spectrometer: PLS vs. SVM. A case study in a WWTP

In this work, the use of Support Vector Machine (SVM) methods to determine functional relationships between absorbance and concentrations of pollutants has been explored. These methods can be applied to establish local calibration functions for UV/Vis spectrometers used for on-line monitoring purposes. Results are obtained through the analysis of a data set from the outlet of pre-treatment steps at the Fontaines-sur-Saône waste water treatment plant in Lyon, France. Ninety-three dry-weather samples collected in 2011 with an American Sigma 900 automatic sampler were at our disposal. For each sample, TSS, COD and filtered COD concentrations, as well as fingerprints measured with a Scan Spectro::lyser UV-Vis probe (with a measuring window of 2 mm) were obtained under laboratory controlled conditions. In order to avoid outliers or bias in the results, 100 random combinations of calibration / verification data sub- sets were used to determine local calibration functions of the probe. For each combination, SVM and more traditional PLS (partial least squares) models were determined. For some combinations, calibration results obtained using PLS models are better than those obtained with SVM models. However, predictability seems to be improved by using SVM models compared to PLS ones for some combinations, which reveals non-linear relations between absorbance and lab- concentrations data.