DeepThermal: Combustion Optimization for Thermal Power Generating Units Using Offline Reinforcement Learning

• DOI: 10.1609/aaai.v36i4.20393
• Publication Date: 2022-07-04T11:06:18Z
• Authors: Xianyuan Zhan; Haoran Xu; Yue Zhang; Xiangyu Zhu; Honglei Yin; Yu Zheng
• Keywords: Offline learning; Online and offline

Optimizing the combustion efficiency of a thermal power generating unit (TPGU) is highly challenging and critical task in energy industry. We develop new data-driven AI system, namely DeepThermal, to optimize control strategy for TPGUs. At its core, model-based offline reinforcement learning (RL) framework, called MORE, which leverages historical operational data TGPU solve complex constrained Markov decision process problem via purely training. In we first learn simulator from dataset. The RL agent MORE then trained by combining real as well carefully filtered processed simulation through novel restrictive exploration scheme. DeepThermal has been successfully deployed four large coal-fired plants China. Real-world experiments show that effectively improves also report superior performance comparing with state-of-the-art algorithms on standard benchmarks.