CO Emission Prediction Model of MSWI Process Combined Sample Output and Feature Space Semi-Supervised Drift Detection

The municipal solid waste incineration process is susceptible to fluctuations in waste composition, equipment wear and maintenance, seasonal changes, and other variables, leading to concept drift-a time-dependent variability in pollutant emission concentration modeling data. This study introduces a semi-supervised drift detection method that integrates sample output and feature space analysis. Initially, an unsupervised approach based on principal component analysis identifies concept drift samples in the feature space. Subsequently, pseudo-truth values for these concept drift samples are labeled using a semi-supervised method based on time-difference learning in the sample output space. Detection of concept drift is then performed using the Page-Hinkley method. Finally, the LSTM model is updated with both new and historical samples. Experimental results based on real CO process dataset validate the effectiveness of the proposed method.