Multiscale dense-phase structures in a 700-MW-class cold-state CFB boiler model using sequence-aware optical feature fusion--颗粒学报PARTICUOLOGY

a State Key Laboratory of Complex Nonferrous Metal Resources Clean Utilization, School of Metallurgical and Energy Engineering, Kunming University of Science and Technology, Kunming, 650093, China
b Department of Electrical and Electronic Engineering, School of Engineering, Faculty of Science and Engineering, The University of Manchester, Manchester, M13 9PL, UK
c Guangdong Provincial Key Laboratory of Interdisciplinary Research and Application for Data Science, Faculty of Science and Technology, Beijing Normal-Hong Kong Baptist University, Zhuhai, Guangdong, 519087, China
d Department of Statistics and Data Science, College of AI, Cyber and Computing, The University of Texas at San Antonio, San Antonio, TX, 78249-0634, USA

10.1016/j.partic.2026.07.011

Volume 117, October 2026, Pages 287-302

Received 21 February 2026, Revised 8 July 2026, Accepted 16 July 2026, Available online 29 July 2026, Version of Record 7 August 2026.

E-mail: kaiyang@kust.edu.cn; qingtai.xiao@kust.edu.cn

Highlights

• Fluidized bed image features are mapped to multi-scale calibrated metrics.

• Proposing a leak-free two-stage stack with PSO-tuned short-sequence modelling.

• Reveals slope compression in length scales from wall channels and core flattening.

• Explains high accuracy of bulk solids and interface via monotonic image links.

• Establishes lighting and position diagnostics, improving robustness across flow groups.

Abstract

Dense-phase hydrodynamics in large circulating fluidized bed (CFB) boilers strongly influences performance, yet its spatiotemporal variability is complicated to quantify with conventional diagnostics. An optical data driven workflow is developed on a 700 MW class cold state CFB boiler model to map time resolved recordings to calibrated hydrodynamic quantities that remain valid across gas flow settings and lateral positions. The proposed method combines background modeling, morphological refinement, image descriptor extraction, correlation aware two-level stacking, and short window long short-term memory modeling to estimate dense phase hydrodynamic quantities. Isotonic calibration, conformal prediction, and multitask ridge regression are further used to obtain calibrated estimates and 90% prediction intervals. Under leave-one-group-out validation by gas flow setting, the workflow achieves a macro-averaged R-squared of 0.69 and reduces normalized error by 10-15% relative to the strongest single-model baseline. Overall solids fraction and normalized interface height reach R-squared values of 0.86 and 0.85 with near nominal coverage. Predictions for anisotropic correlation-length surrogates are moderate, and monotone calibration mitigates upper range compression. The results show that the proposed workflow provides physically consistent, uncertainty quantified quantities that are suited to trend monitoring, regime discrimination, and measurement planning in large CFB systems.

Graphical abstract

Keywords

Circulating fluidized bed; Dense-phase hydrodynamics; Stacking ensemble algorithm; Data-driven modeling; Uncertainty quantification

文章导读