Development of Soft Sensors Based on Analytical and Spectral Data on a Real Small Size Wastewater Treatment Plant

Abstract

Advances in measurement and automation have recently enabled the deployment of monitoring systems with high frequency data acquisition in Wastewater Treatment Plants (WWTP). In this context, this work aims to design soft sensors to predict hard-to-measure wastewater quality variables required for mechanistic modeling of biological treatment in a real municipal WWTP. Input data collected at the WWTP using a flowmeter, online spectrophotometric and electrochemical probes, sampling campaigns and off-line analyses, are here used for the development of soft sensors through multivariate methods, i.e., Principal Component Analysis (PCA) and Partial Least Squares (PLS) regression. Exploratory data analysis is performed to detect outliers, patterns and correlations. Soft sensor PLS models are optimized using leave-one-out cross validation and the root mean squared error (RMSE) for the prediction of an independent dataset is computed. Normalized RMSE values for organic nitrogen prediction result in 19.5 % and 18.1 % for sensors using analytical and spectral data, respectively. The possibility of using a single wavelength spectrophotometric probe is here evaluated aiming to reduce the online monitoring investment costs.