Real-time optimization of a chemical plant with continuous flow reactors via reinforcement learning

Reinforcement learning (RL) has many new applications in recent years, and its results often exceed human performance, especially in environments where the action space is discrete. However, it is challenging to use RL in the chemical industry, where variables are often continuous and various constraints are complex. This study applies RL with continuous actions to maximize the productivity of a continuous process. The RL agent provides optimal setpoints of flow rates and temperatures while the concentrations of raw materials are changing. Two environments with one and six actions were established after the sensitivity analysis. In the one-action environment, the agents SAC, PPO and A2C showed similar performances, but A2C needed fewer timesteps for training. SAC outperforms PPO and A2C in the environment with six actions. This paper shows the successful RL applications in a continuous process and the high applicability of SAC in both low-dimension and high-dimension environments.