Artificial neural network implementation for dissolved organic carbon quantification using fluorescence intensity as a predictor in wastewater treatment plants.

Although spectroscopic methods provide a fast and cost-effective means of monitoring dissolved organic carbon (DOC) in natural and engineered water systems, the prediction accuracy of these methods is limited by the complex relationship between optical properties and DOC concentration. In this study, we developed DOC prediction models using multiple linear/log-linear regression and feedforward artificial neural network (ANN) and investigated the effectiveness of spectroscopic properties, such as fluorescence intensity and UV absorption at 254 nm (UV), as predictors. Optimum predictors were identified based on correlation analysis to construct models using single and multiple predictors. We compared the peak-picking and parallel factor analysis (PARAFAC) methods for selecting appropriate fluorescence wavelengths. Both methods had similar prediction capability (p-values >0.05), suggesting PARAFAC was not necessary for choosing fluorescence predictors. Fluorescence peak T was identified as a more accurate predictor than UV. Combining UV and multiple fluorescence peak intensities as predictors further improved the prediction capability of the models. The ANN models outperformed the linear/log-linear regression models with multiple predictors, achieving higher prediction accuracy (peak-picking: R = 0.8978, RMSE = 0.3105 mg/L; PARAFAC: R = 0.9079, RMSE = 0.2989 mg/L). These findings suggest the potential to develop a real-time DOC concentration sensor based on optical properties using an ANN for signal processing.