Digital Twin Framework and Federated Learning for Multi-plant Knowledge Sharing in Decision Support for Electric Steelmaking and Beyond

The processes of electric steelmaking are complex and difficult to control to achieve sustainable production. To strive towards competitiveness and green transformation, steelmakers apply the Electric Arc Furnace (EAF) to circulate scrap into new products. This saves resources compared to iron-ore-based production, enabling both circular economy and green energy sources. However, efficient EAF operation faces difficulties in state monitoring and control decisions. Fortunately, the control can be facilitated with optimization based on Artificial Intelligence (AI) and Digital Twins (DT). Still, DT accuracy can suffer from input data fluctuation or coverage limitations in development and validation datasets. The fluctuation stems from the environment-related variation, especially the scattering and not exactly known chemical composition of secondary raw materials. For decision support, this work suggests a DT framework with Federated Learning (FL) for multi-plant schemes, focusing on electric steelmaking. The framework can deliver both historical data and message-oriented online data to the DTs. It builds upon a container orchestration system (Kubernetes) for software lifecycle management and resource scaling. Importantly, the framework implements FL to exploit network-wide knowledge. That is, the DTs share knowledge with a centralized server that aggregates a global model distributed to the participants, broadening data diversity. Still, all data remain local, which preserves privacy. The framework applies FL for two types of process DTs, EAF and the subsequent Ladle Furnace (LF). FL can optimize EAF parameters although EAF is not AI but a dynamic model. Conversely, the LF model is composed of a set of neural networks. The results from a prototype system with actual data prove the concept. Firstly, the DTs accurately estimate process variables online, such as the chemical composition and temperature. Secondly, FL experiments indicate potential for model parameter optimization and enhanced performance. Besides, the framework concept is applicable for even more DTs and across industries.