Analysis and prediction of thermal stress distribution on the membrane wall in the arch-fired boiler based on machine learning technology. (1st February 2022)

• Record Type: Journal Article
• Authors: Wen, Du; Pan, Yuqing; Chen, Xiaole; Aziz, Muhammad; Zhou, Qulan; Li, Na
• Is Part Of: Thermal science and engineering progress. Volume 28(2022)
• Publication Date: 2022-02-01
• Subjects: Arch-fired boiler -- Membrane wall -- Thermal stress prediction -- Neural networks -- Anomaly detection; Heat engineering -- Periodicals; Heat engineering; Thermodynamics; Periodicals; 621.402
• DOI: 10.1016/j.tsep.2021.101137
• Languages: English
• ISSNs: 2451-9049

Abstract

Highlights: Machine learning technology was first applied to the thermal stress prediction on the membrane wall in an arch-fired boiler. The neural network models were trained and validated by the experimental data coming from a lab-scale arch-fired boiler. High-temperature distribution is the primary impact factor of high-stress generation under the full load condition, while boiler configuration dominates it under the half load condition. The classification neural network model is the optimal method because of the high accuracy and generality. It is suitable for real application. Abstract: The arch-fired boiler easily suffers from membrane wall deformation and rupture caused by high stress. Thus, it is necessary to predict thermal stress distribution on the membrane wall for the sake of early warning. In this study, a lab-scale arch-fired boiler was constructed to achieve thermal stress distribution on the membrane wall, and the reason for high-stress formation was found. Besides, machine learning technology was first applied to predict thermal stress based on experimental data. The results show that both high-temperature distribution and boiler configuration play an important role on high-stress distribution. As for the prediction, the regression neural network model presents an admissible result (R 2 = 0.724), but has a relatively poor performance on the high-stress prediction due to the skewed data distribution. Specifically, the relatively high prediction errors present on the bending and connection areas. Future work should consider the effect of boiler configuration and different heat loads. In contrast, the binary classification neural network model has a more accurate result with a high F1 score (0.893). It is the best choice for practical application considering the generality and accuracy. This work is meant to be valuable in optimizing the prediction model and applying it in practice.