A convolution–attention network predicts boiler slagging and warns of waterwall overheating from the furnace temperature field
Abstract
Slagging and wall-temperature overheating rank among the costliest threats to safe and economical coal-fired boiler operation, yet plants still diagnose them separately although both grow out of the same non-uniform furnace temperature field. This paper treats them jointly. Starting from the heat-transfer balance that couples gas temperature, near-wall flux, ash deposition and tube metal temperature, we build a network that reads the spatial structure of the field through convolution and reweights it through channel, spatial and temporal attention before branching into a slagging classifier and a wall-temperature regressor. The data come from one 660 MW tangentially fired supercritical unit burning a bituminous blend whose ash softens at 1265 °C, supplemented by CFD fields checked for grid independence and matched against plant measurements, and a confidence-gated graded criterion converts the two outputs into staged overheating warnings. Averaged over five random seeds, the model reaches a coefficient of determination of 0.958 ± 0.004 for peak metal-temperature prediction and, for warning, an F1 of 0.931 ± 0.008 with a precision–recall AUC of 0.912 ± 0.013 at a median lead time of 40 s. Against the strongest baseline the gains in regression error, in ranking quality and in lead time are statistically resolvable, whereas the F1 gain is not, and the comparison is reported as such. Ablation and attention-weight visualisation show that spatial weighting concentrates on the burner-belt deposition band and temporal weighting on the pre-event window. A transfer test across ash fusion characteristics marks the boundary of the method: applied without fine-tuning to a lower-softening ash, F1 falls to 0.842. The framework is therefore offered as a reproducible soft-sensor layer for a single, well-characterised unit rather than as a general-purpose tool, with code and a released data extract supplied under Related files.
Related
- YOLO-RMF: A Real-Time Surface Defect Detection Model for Hot-Rolled Strip Steel with Reparameterized Convolution and Axial Attention
- A parallel convolution attention network for crop genomic prediction
- Asymmetric encoder-decoder U-net with hybrid attention and residual dilated convolution for brain tumor segmentation
- ACMNet: An attention-convolution mixed neural network for semantic segmentation of high-resolution remote sensing images
- Design of monocular vision human arm-robotic arm interaction system based on attention feature fusion and sequential graph convolution
BibTeX
@article{huang2026convolutionattention,
title = {{A convolution–attention network predicts boiler slagging and warns of waterwall overheating from the furnace temperature field}},
author = {Linbin Huang and Lucan He and Long Liu and Guoqing Chen and Jun Xiang},
journal = {Scientific Reports},
year = {2026},
month = sep,
doi = {10.1038/s41598-026-72202-w},
url = {https://www.nature.com/articles/s41598-026-72202-w},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - A convolution–attention network predicts boiler slagging and warns of waterwall overheating from the furnace temperature field
AU - Linbin Huang
AU - Lucan He
AU - Long Liu
AU - Guoqing Chen
AU - Jun Xiang
PY - 2026
DA - 2026/09/30
JO - Scientific Reports
DO - 10.1038/s41598-026-72202-w
UR - https://www.nature.com/articles/s41598-026-72202-w
AB - Slagging and wall-temperature overheating rank among the costliest threats to safe and economical coal-fired boiler operation, yet plants still diagnose them separately although both grow out of the same non-uniform furnace temperature field. This paper treats them jointly. Starting from the heat-transfer balance that couples gas temperature, near-wall flux, ash deposition and tube metal temperature, we build a network that reads the spatial structure of the field through convolution and reweights it through channel, spatial and temporal attention before branching into a slagging classifier and a wall-temperature regressor. The data come from one 660 MW tangentially fired supercritical unit burning a bituminous blend whose ash softens at 1265 °C, supplemented by CFD fields checked for grid independence and matched against plant measurements, and a confidence-gated graded criterion converts the two outputs into staged overheating warnings. Averaged over five random seeds, the model reaches a coefficient of determination of 0.958 ± 0.004 for peak metal-temperature prediction and, for warning, an F1 of 0.931 ± 0.008 with a precision–recall AUC of 0.912 ± 0.013 at a median lead time of 40 s. Against the strongest baseline the gains in regression error, in ranking quality and in lead time are statistically resolvable, whereas the F1 gain is not, and the comparison is reported as such. Ablation and attention-weight visualisation show that spatial weighting concentrates on the burner-belt deposition band and temporal weighting on the pre-event window. A transfer test across ash fusion characteristics marks the boundary of the method: applied without fine-tuning to a lower-softening ash, F1 falls to 0.842. The framework is therefore offered as a reproducible soft-sensor layer for a single, well-characterised unit rather than as a general-purpose tool, with code and a released data extract supplied under Related files.
SN - 2045-2322
ER -