Optimizing Control of Waste Incineration Plants Using Reinforcement Learning and Digital Twins

More than 60 waste incineration plants (WIPs) are active across Germany. WIPs have various levels of automation, but they still rely on manual operations by human operators. Human operators tend to rely on a few operational levers and infrequent interventions to manage the complex combustion process. Consequently, the combustion process is managed rather inefficiently, and steam outputs and emission levels are not optimal. This article investigates how reinforcement learning (RL) can help enhance process automation and, thus, optimize the combustion process, e.g., by making more frequent and diverse interventions. An RL agent is trained via trial and error with a reward function that includes the optimization criteria. Since the actual equipment, i.e., the real WIP, cannot be used as the training environment, a digital twin is built using original plant data and a neural network. The RL agent is then trained in this offline environment with the deep Q-network algorithm. First, our work demonstrates that a digital twin of a WIP can be built in a data-driven way. Second, we show that the RL agent outperforms the human operator. Thus, the application of RL might benefit the plant operator in financial terms and the environment in terms of reduced emission levels.