Explicit and interpretable nonlinear soft sensor models for influent surveillance at a full-scale wastewater treatment plant

Abstract

In wastewater treatment plants, the most adopted sensors are those with the properties of low cost and fast response. Soft sensors are alternative solutions to the hardware sensor for online monitoring of hard-to-measure variables, such as chemical oxygen demand (COD) and total phosphorus (TP). The purpose of this study is to obtain a modelling approach which is able to identify the nonlinearity of influent and explain the correlation of inputs-outputs. Thus, the variation of influent characteristics was investigated at the first stage, which provided the basis to build global and local multiple linear regression models. Secondly, a nonlinear modelling tool multivariate adaptive regression splines (MARS) was applied for influent COD and TP prediction. Satisfactory prediction accuracy was obtained in terms of root mean square error (RMSE) and R2. Unlike other machine learning techniques which are "black box" models, MARS provided interpretable models which explained the nonlinearity and correlation of inputs-outputs. The MARS models can be used not only for prediction, but also to provide insight of influent variation.