Clustering and bayesian inference as forecasting tools for UV-Vis absorbance time series

This paper proposes a method designed to forecast online water quality data obtained using UV-Vis spectrometry. The method combines clustering (k-means)—meant to reduce a dataset’s dimensionality—and Bayesian inference (Markov Chains)—meant to aid in forecasting. In addition, Locally Weighted Scatterplot Smoothing (LOWESS) was employed to smooth the forecasted absorbance spectra values. With regards to performance assessment, three absorbance time series datasets consisting of 5705 UV-Vis spectra were used; likewise, Absolute Percentage Errors (APE) were calculated by applying the proposed k-means clustering and Markov Chains (collectively referred to as KmMC). Across the three study sites, APE results vary from 0% to 36%. Generally speaking, these results cannot be generalized, for they are highly dependent on specific water system dynamics. However, that does not mean the data lack trends: KmMC produced APE values under 20% for a variety of circumstances in all three spectra absorbance time series. In sum, the proposed method is best suited for forecasting determinants within the UV spectrum (e.g. NO2-Nitrites and NO3-Nitrates).