CO Emission Prediction for MSWI Process Based on Dual-Space Nested Dual-Window Drift Detection

Abstract

The concentration of carbon monoxide (CO) emissions is intricately linked to the operational status and combustion efficiency of municipal solid waste incineration (MSWI) processes, which are characterized by complex, dynamic, and time-varying behaviors. In order to tackle the challenge of predicting CO emissions, this article introduces a novel method based on nested dual-window drift detection (NDWDD). Initially, a typical sample pool (TSP) is generated using the $k$ -means algorithm. An offline prediction model combining long short-term memory (LSTM) with a feature space drift detection model based on robust principal component analysis (RPCA) is then developed. The control limit for error space prediction accuracy is set using the fast Hoeffding drift detection method (FHDDM). The NDWDD employs a unique combination of external feature space drift detection and nonparametric drift detection within the internal error space, using a nested mechanism to enhance detection efficiency and reduce the influence of inherent noise factors in industrial processes. Finally, the dual-space drift sample collection facilitates updates to the TSP, historical prediction models, RPCA model, and FHDDM control limits. Experimental results from a Beijing MSWI power plant demonstrate that the proposed method can predict CO emissions both robustly and effectively.