Forecasting of UV-Vis absorbance time series using artificial neural networks combined with principal component analysis

This work proposes a methodology for the forecasting of online water quality data provided by UV-Vis spectrometry. Therefore, a combination of principal component analysis (PCA) to reduce the dimensionality of a data set and artificial neural networks (ANN) for forecasting purposes was used.
The results obtained were compared with those obtained by using discrete Fourier transform (DFT). The proposed methodology was applied to four absorbance time series data sets composed by a total number of 5705 UV-Vis spectra. Absolute percentage errors obtained by applying the proposed
PCA/ANN methodology vary between 10% and 13% for all four study sites. In general terms, the results obtained were hardly generalizable, as they appeared to be highly dependent on specific dynamics of the water system; however, some trends can be outlined. PCA/ANN methodology gives better results than PCA/DFT forecasting procedure by using a specific spectra range for the following conditions: (i) for Salitre wastewater treatment plant (WWTP) (first hour) and Graz West R05 (first 18 minutes), from last part of UV range to all visible range, (ii) for Gibraltar pumping station (first 6 minutes) for all UV-Vis absorbance spectra and (iii) for San Fernando WWTP (first 24 minutes) for all of
UV range to middle part of visible range.