Dynamic soft sensors for wall temperature in a down-fired boiler using split-transform-merge strategy in Lightweight CNN modeling

Intense water-wall temperature fluctuations during deep peak-shaving operations pose a significant risk of overheating-induced tube bursts, critically undermining unit safety and load-following flexibility. To address the challenge, a data-driven framework for dynamic water-wall temperature prediction was proposed, integrating feature selection with a lightweight convolutional neural network (CNN). First, the maximum information coefficient (MIC) quantifies variable correlations to identify key features and eliminate redundancy. A Lego-CNN was then developed, which innovatively employs a low-dimensional Lego filter to replace conventional kernels, incorporating a split-transform-merge strategy. This design reduces parameter redundancy while preserving high feature extraction capability. A hybrid local loss function, combining prediction loss and similarity matching loss, further enhances accuracy despite model simplification. Validated with operational data from a 600 MW supercritical down-fired boiler, the proposed Lego-CNN outperforms baseline CNN, increasing R 2 by 2.74% and reducing MAE and RMSE by 59.22% and 59.37%, respectively. It also demonstrates superior robustness under variable and transient loads. These results confirm its effectiveness and suitability for reliable water-wall temperature prediction in deep peak-shaving scenarios.