Nonlinear dynamic partial least squares modeling of a full-scale biological wastewater treatment plant

Partial least squares (PLS) has been extensively used in process monitoring and modeling to deal with many, noisy, and collinear variables. However, the conventional linear PLS approach may be not effective due to the fundamental inability of linear regression techniques to account for nonlinearity and dynamics in most chemical and biological processes. A hybrid approach, by combining a nonlinear PLS approach with a dynamic modeling method, is potentially very efficient for obtaining more accurate prediction of nonlinear process dynamics. In this study, neural network PLS (NNPLS) were combined with finite impulse response (FIR) and auto-regressive with exogenous (ARX) inputs to model a full-scale biological wastewater treatment plant. It is shown that NNPLS with ARX inputs is capable of modeling the dynamics of the nonlinear wastewater treatment plant and much improved prediction performance is achieved over the conventional linear PLS model.