Multi-objective optimization of boiler NO emissions and platen superheater overheating based on ensemble learning and deep reinforcement learning
Yuchen Fan, Xiao Liu, Yongqing Zhou, Chang Wei, Xin Liu, Chi Kong Li, Xinying Li, Heyang Wang
Abstract
Air-staged combustion is widely used in coal-fired boilers to reduce NO x emissions. However, it causes the high-temperature flame to move upwards and may lead to tube overheating of the platen superheater (PLSH) located in the upper furnace. This has significantly affected boiler safety and limited the implementation of deep air-staged combustion for greater NO x reduction. To solve this problem, a multi-objective optimization model is proposed to realize the coordinated optimization of NO x emissions and PLSH tube temperatures. Firstly, the ensemble learning method, which integrates the strengths of multiple models (DNN, XGBoost, TabNet, and ELM), is used to construct the prediction models for boiler NO x and PLSH temperatures, which improves the prediction accuracy by 10 % compared to individual models. Secondly, the deep reinforcement learning algorithm (TD3) is utilized to identify the optimal boiler parameters by interacting with the prediction models to realize the coordinated optimization of NO x emissions and PLSH temperatures. Application of this multi-objective optimization model to a 600 MW boiler shows that it can discriminate the main boiler problems at different boiler loads (higher NO x at high loads and high PLSH temperature at all loads) and present different optimization strategies accordingly, which can achieve an average reduction of 44.3 mg/Nm 3 in NO x emissions and an average increase of 14.2 % in PLSH temperature compliance rate without compromising boiler efficiency. This study provides a solution to improve the capability of boilers to apply deeper air-staged combustion to achieve greater NO x reduction while maintaining operational safety and economy.