Combustion Optimization Study of Pulverized Coal Boiler Based on Proximal Policy Optimization Algorithm

Xuecheng Wu,Hongnan Zhang,Huafeng Chen,Shifeng Wang,Lingling Gong

DOI: https://doi.org/10.1016/j.applthermaleng.2024.123857

IF: 6.4

2024-01-01

Applied Thermal Engineering

Abstract:In most industrial sectors, large coal-fired boilers are a source of carbon and pollutant emissions, so it is important to carry out combustion adjustment and optimize energy-saving operation of coal-fired boilers. Traditional combustion adjustment relies on human intervention, but manual adjustment is difficult to achieve synergistic optimization of NOx and thermal efficiency at the same time, so there is a large adjustment space for boiler combustion optimization. Artificial intelligence technology can explore the potential of combustion optimization from boiler operation data. Currently, the boiler combustion optimization method based on supervised learning modeling and optimization algorithms has good optimization effect and high application value. At present, there are problems such as the combination of dynamic model and optimization algorithm is difficult and the optimization time is long, etc. This paper adopts feature classification and multi-model coupling to build a static-dynamic composite prediction model of boiler performance indicators, dynamic prediction model of boiler thermal efficiency and nitrogen oxides (NOx) is established by using long short-term memory (LSTM) and one-dimensional convolutional neural network (1D_CNN). The model is categorized into static and dynamic models based on the input features, and the dynamic model is coupled with BP neural network to establish a static-dynamic composite prediction model and further couples the proximal policy optimization (PPO) reinforcement learning algorithm to establish a boiler in-place optimization strategy. Through the experimental validation of 5619 test cases, the strategy successfully achieves 63.5 % co-optimization of NOx and thermal efficiency, with thermal efficiency increase ranging from 0-0.61 % and NOx reduction ranging from 0-65 mg/m3. Meanwhile, comparing the optimization effect of the PPO algorithm with that of the genetic algorithm (GA) shows that the PPO strategy has a more significant effect on NOx reduction while keeping the thermal efficiency optimized. Moreover, the online decision-making speed of the PPO strategy is much higher than that of the GA, with an average time consumption of only 0.015 s, while the GA requires about 3 min for a single optimization, which indicates that the combustion optimization strategy of the PPO algorithm coupled with the composite prediction model has a significant advantage in realizing high-efficiency and accurate optimization.