CO Emission Modeling Based on Fixed-window Drift Detection for the Municipal Solid Waste Incineration Process

Carbon monoxide (CO) from municipal solid waste incineration (MSWI) is a pivotal process parameter that can characterize the stability of the combustion process. Addressing at the dynamic time-varying characteristics of the MSWI process, a modeling method for CO emission modeling method based on fixed-window drift detection is proposed. First, based on the historical dataset, the k-means algorithm was used to obtain the typical sample pool (TSP). An offline prediction model based on the long short-term memory (LSTM) neural network and a drift index calculation model based on kernel principal component analysis (KPCA) were constructed. Subsequently, for each online collected sample, online prediction was made based on the historical LSTM neural network model once the preset fixed window was not filled. When the preset fixed-window was filled, the historical KPCA model was used for drift detection. The Hotelling’s T² and squared prediction error (SPE) were used to determine whether drift occured. If no drift occured, the prediction process was returned to the new window period. Otherwise, historical data and drift data were integrated to obtain new TSP, LSTM model, and KPCA model. The rationality and effectiveness of the proposed method were verified by the simulation of actual data in the industrial field.