Chemometric advances in COD analysis: Overcoming turbidity interference with a Hybrid PLS-ANN approach

Abstract

In the current context of environmental protection, the accuracy of water quality analyses is crucial, especially chemical oxygen demand (COD) analyses. COD is a key indicator of organic pollutants in water, and is often compromised by turbidity interference at UV wavelengths. Faced with this turbidity challenge, this innovative study proposes an integrated approach, using in situ UV sensors and advanced Chemometric techniques, exploiting the synergy between PLS (Partial Least Square) and ANN (Artificial Neural Network). The research is divided into three main sections: (i) assessment of turbidity interference, (ii) development of a turbidity prediction model and (iii) elaboration of the COD compensation and prediction model. Turbidity significantly alters the UV-Visible absorbance spectrum of COD. By analyzing mixed solutions of COD and turbidity, we quantified this interference and established a turbidity prediction model using spectral areas between 303.5 and 700 nm. Interval PLS identified the most informative spectral regions for COD concentration, highlighting the 200–250 nm interval. To address turbidity interference, we developed a hybrid model combining PLS and ANN regression, and turbidity measurements are incorporated as an explanatory term in the model. This hybrid approach relies on in situ UV sensors to directly capture UV absorbance data in the field and applies robust chemometric models to accurately distinguish the UV absorbance contributions of organic compounds and turbidity. The method not only compensates for the interfering effect of turbidity, but also allows turbidity information to be used to refine COD predictions. Model performance, evaluated using R², RMSE, and MAE, showed a significant increase in R² to 0.9972, and decreases in RMSE to 0.94 and MAE to 0.64, demonstrating the method's effectiveness in correcting turbidity-induced deviations and improving COD prediction accuracy. The results of the evaluation on real data show high performance metrics, with recovery percentages close to 100 % and low RMSE and MAE values, indicating the model’s robust ability to predict COD in the presence of suspended particles.