Machine Learning-Based Time-Series Forecasting for Off-Gas Characterization and Decarbonization in Electric Arc Furnace Steelmaking

Electric arc furnace (EAF) steel production substantially reduces CO2 emissions from steelmaking but is not fully carbon neutral. To facilitate full decarbonization of this batch process, machine learning-based time-series forecasting models are constructed to predict the EAF off-gas composition (CO2, CO and H2O) on a second-by-second basis. These models will be used to drive the design and testing of carbon capture technologies that can handle dynamically changing flue gases. The resultant models successfully forecasted the behavior of the off-gas composition, including the stochastic and erratic behavior of the variables, for both a scrap-fed EAF and a DRI-fed EAF. Generation of numerous predictions for the same future heat creates a distribution of the most likely operating conditions at each point in time. This ensures that carbon capture equipment for downstream decarbonization can be appropriately designed to handle the full breadth of the operational variability of the system.