Midterm colloquium George Vitanov
12 October 2022 15:00 till 15:30 - Location: lecture room D (james watt), 3me - By: DCSC | Add to my calendar
"Transferring Domain Knowledge to a Data-Driven Controller"
Tata Steel in IJmuiden has devised and developed a technology to make steel in the future with 20% less CO2 emissions: HIsarna, an alternative to the blast furnace process. HIsarna sidesteps the need of a pellet plant for ore or coal and the raw materials can be directly injected into the furnace in powder form simplifying the hot metal production process while simultaneously reducing carbon emissions. Currently the technology is at Pilot Plant level, and the next goal is to understand the process through data and synthesize an automated controller that can keep the basicity of the hot iron inside the plant at a desired setpoint.
The challenge in building the controller is to distinguish between the effects of an intended control action and other factors of influence. As the result of our literature study we propose a stochastic chemical model based approach for modelling the state of HIsarna. Our stochastic system model and control objective satisfy all criteria of a Markov Decision Process, and thus Dynamic Programming was selected as the most appropriate controller synthesis method. After controller synthesis the robustness of the controller is examined, and comparison is made with the performance of a human operator. The achieved controller is deemed suitable for application in a trial of an advisor system to the human operators of HIsarna as it performs better than the human operators on a probabilistic basis with a high confidence threshold.
Prof.dr.ir. T. Keviczky and Dr. P. Mohajerin Esfahani