Imagen de Portada

Robust Dynamic Modeling of Bed Temperature in Utility Circulating Fluidized Bed Boilers Using a Hybrid CEEMDAN-NMI–iTransformer Framework

Guardado en:
Publicado en:Processes vol. 13, no. 3 (2025), p. 816
Autor principal: Li, Qianyu
Otros Autores: Wang, Guanglong, Li, Xian, Yu, Cong, Bao, Qing, Lai, Wei, Li, Wei, Ma, Huan, Si, Fengqi
Publicado:
MDPI AG
Materias:
Outliers (statistics)
Pollutants
Accuracy
Deep learning
Fluid dynamics
Long short-term memory
Power plants
Dynamic models
Fluidized bed boilers
Combustion stability
Industrial plant emissions
Robustness
Boilers
Efficiency
Adaptive algorithms
Heat transfer
Machine learning
Fluidized beds
Predictions
Carbon
Temperature
Neural networks
Variables
Distributed control systems
Correlation analysis
Algorithms
Flue gas
Coal
Energy levels
Acceso en línea:Citation/Abstract
Full Text + Graphics
Full Text - PDF
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
Resumen:Circulating fluidized bed (CFB) boilers excel in low emissions and high efficiency, with bed temperature serving as a critical indicator of combustion stability, heat transfer efficiency, and pollutant reduction. This study proposes a novel framework for predicting bed temperature in CFB boilers under complex operating conditions. The framework begins by collecting historical operational data from a power plant Distributed Control System (DCS) database. Next, the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) algorithm is employed to decompose the raw signals into distinct modes. By analyzing the trade-offs of combining modes with different energy levels, data denoising and outlier reconstruction are achieved. Key features are then selected using Normalized Mutual Information (NMI), and the inflection point of NMI values is used to determine the number of variables included. Finally, an iTransformer-based model is developed to capture long-term dependencies in bed temperature dynamics. Results show that the CEEMDAN-NMI–iTransformer framework effectively adapts to diverse datasets and performs better in capturing spatiotemporal relationships and delivering superior single-step prediction accuracy, compared to Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM), and Transformer models. For multi-step predictions, the model achieves accurate forecasts within 6 min and maintains an R2 above 0.95 for 24 min predictions, demonstrating robust predictive performance and generalization.
ISSN:2227-9717
DOI:10.3390/pr13030816
Fuente:Materials Science Database