Comparative Study of Latent Structure Modeling Approaches with Its Application to Prediction Dioxin Emission Concentration

Dioxin (DXN) is a kind of pollutant commonly discharged during municipal solid waste incineration (MSWI). In practical industrial processes, the concentration of DXN emission is measured by using offline analysis, but this method is constrained by long time lag and high cost. This study aims to develop soft measuring model for DXN emission concentration by using easy-to-measure MSWI process variables with the latent structure algorithm. Three latent structure algorithms, namely, linear projection to latent structure (PLS), nonlinear kernel PLS (KPLS), and a new improved general algorithm-based selective ensemble KPLS (IGASENKPLS), are applied to build the DXN estimation model. Results show that the latent structure algorithm can successfully generate DXN models with good prediction performance. Nonlinear KPLS can extract more variations from the dataset than linear PLS, but IGASENKPLS can enhance prediction performance even further. The proposed approach demonstrates the feasibility of using latent structure algorithm to model DXN emission concentration by using collinear, nonlinear, and small-size sampling data.