Abstract: Focusing on the application of deep reinforcement learning algorithms in chemical industrial processes, the deep Q-network (DQN) algorithm was used to simulate and optimize the operating temperature of the Williams Otto reaction, achieving adaptive adjustment of the reaction temperature and significantly increasing the yield of high-value products and economic benefits. The proximal policy optimization (PPO) algorithm was applied to the optimization of operating parameters in the steam cracking process for ethylene production. Based on the convolutional neural network architecture (D-VGG), a model of the ethylene cracking process was established. The reinforcement learning agent interacted with the ethylene cracking model environment for learning and optimized the operating parameters of the cracking unit, significantly enhancing the yield of ethylene and propylene. The research results not only verified the effectiveness and practicality of deep reinforcement learning algorithms in chemical industrial processes but also provided new ideas and methods for the real-time optimization control of other complex industrial systems.

Key words: reinforcement learning, Markov decision process, chemical processes, real-time control