Reinforcement learning for the adaptive speed regulation on a metallurgical pickling line

Data Fest Online 2020 Manufacturing, Energy, and Logistics track Speakers: Boris Voskresenskii, Chief Digital Officer, Severstal Digital Kseniia Kingsep, Head of Data Science, Severstal Digital Anna Bogomolova, Team Lead, Seversal Digital RL controls the speed of the continuous pickling line (NTA-3) at Cherepovets Steel Mill. Previously, the control speed was set once for each roll manually. A mathematical model controls the speed of the unit, taking into account in real time about a hundred parameters including the length, width and thickness of the roll, the steel grade, the temperature of the metal and many others. In the end of March Severstal improved this model with a RL module. Technological processes of the pickling line heavily depend on the parameters of steel strips going through the line. Our RL based agent uses steel strip parameters synthesized by generative adversarial network (GAN). Today, a mathematical model and the RL based agent work cooperatively in the real-time producing more than 5% additional steel on this unit. Register and get access to the tracks: Join the community:
Back to Top