Speaker
Description
The electric arc furnace (EAF) generates crude metal from virgin materials and recycled scrap.
Meeting the right concentrations of alloying elements in the crude metal (e.g., Cu, Ni) is crucial.
Scrap with higher uncertainty in the concentrations of such alloys is usually cheaper than virgin materials but might threaten product quality.
Using the cheapest input materials while complying with crude metal quality requirements are important goals of an EAF plant operator.
To support operators, we developed the Metallics Optimiser (MO) Application that calculates the cheapest commodity mix by integrating machine learning techniques into physical process models.
The physical models capture mass and energy balances of the EAF and consider weight and composition of the hot heel, slag weight and composition, dust losses and off-gas composition.
The machine learning models are estimating the concentrations of alloys in the scrap:
Based on historical charge mixes and chemical analysis of the crude steel, the MO backward calculates the most likely alloy composition of all used input materials over time.
This can be used in future heats to predict the chemical composition of different charge mixes.
The hot heel of the previous heat is, apart from the scrap, an incoming material stream for the EAF.
Thus, the hot heel properties have an influencing factor on the steel quality and must be considered in the prediction.
However, the exact weight of the hot heel cannot be determined for each heat, introducing an instability in the estimation.
Early results show that the MO is able to find scrap mixes that result in lower production costs than current systems while maintaining requested quality requirements and even decreasing process uncertainty.
Using a dynamic scrap element characterization allows to recycle more scrap, leading to higher savings compared to static element characterizations and static charge mix optimizations.