Abstract: [Objectives]　Zhundong coal is characterized by large reserves and favorable combustion properties, but its high alkali metal content tends to lead to fouling on boiler heating surfaces. Traditional mechanism-based and data-driven methods face challenges such as difficulty in mechanism simplification and insufficient labeled data in fouling monitoring modeling. Therefore, there is an urgent need for accurate and reliable models for fouling quantification characterization to guide sootblowing decisions and optimize diagnostics. To address these issues, a modeling method for heating surface fouling monitoring based on machine learning is proposed. [Methods]　Taking the boiler of a 1 000 MW power plant as the research object, a high-precision dynamic simulation model incorporating the control system is established. On this basis, new characteristic parameters are established, and a fouling quantification characterization modeling method based on autoencoder (AE) and long short-term memory (LSTM) neural network is proposed. [Results]　The fouling monitoring model for heating surface based on new characteristic parameters such as heat transfer deviation and outlet steam temperature deviation shows false alarm rates of 0.6% under stable load conditions and 1.2% under variable load conditions, effectively mitigating the influence of load fluctuations on thermal parameters. The fouling monitoring model based on AE demonstrated high accuracy and robustness. [Conclusions]　The proposed model can accurately monitor the fouling trends of heating surfaces during sootblowing cycles, providing new insights for boiler heating surface fouling monitoring.

Key words: coal-fired combustion, fouling monitoring, autoencoder (AE), long short-term memory neural network (LSTM), Zhundong coal, anomaly detection, characteristic parameters, boiler heating surface