Heterogeneous ensemble prediction model of CO emission concentration in municipal solid waste incineration process using virtual data and real data hybrid-driven

Abstract

Carbon monoxide (CO) is a hazardous gas discharged during municipal solid waste incineration (MSWI). Its emission concentration serves as a vital indicator for assessing the stability of the MSWI process. Therefore, accurate prediction of CO emissions is crucial. While existing research predominantly relies on historical real data-driven models, it often overlooks the effective utilization of the combustion mechanism. This article introduced a novel approach: a heterogeneous ensemble prediction model that integrates virtual and real data. Firstly, virtual mechanism data was obtained through a multi-condition mechanism model constructed using coupled numerical simulation software of FLIC and Aspen Plus. Secondly, based on this virtual mechanism data, a linear regression decision tree (LRDT) algorithm was employed to establish the mechanism mapping model. Simultaneously, a real historical data-driven model based on a long short-term memory (LSTM) neural network algorithm was developed. In the offline training verification phase, the heterogeneous models were combined using an inequality-constrained random weighted neural network (CIRWNN) after aligning virtual and real samples representing operating conditions based on the k-nearest neighbor (KNN) approach. Subsequently, in the online testing verification stage, CO online prediction was achieved by ensemble the LRDT-based mechanism mapping model and.the LSTM-based historical data-driven model. The proposed method's effectiveness and rationality were validated through an industrial case study of MSWI process in Beijing.