Digital Twin-Driven TD3-CPER Control for Continuous Casting

Abstract

Continuous casting is a key process in intelligent steel manufacturing, and mold level control directly affects slab quality and production stability. However, strong nonlinearity, multivariable coupling, time delays, and operating disturbances make real-time control optimization challenging for conventional control strategies. To address this problem, this paper proposes a digital twin-driven reinforcement learning control method based on Twin Delayed Deep Deterministic Policy Gradient with Composite Prioritized Experience Replay, termed TD3-CPER. A data-driven digital twin environment is constructed using real production data to support closed-loop policy training and evaluation. In TD3-CPER, the replay mechanism integrates temporal-difference error, reward feedback, replay frequency, and K-Medoids clustering to improve sample utilization and training stability. Simulation results show that TD3-CPER increases the number of Grade I slabs by 56.2% compared with manual control and achieves a 14.7% higher average cumulative reward than the strongest baseline reinforcement learning method. These results suggest that TD3-CPER is a feasible approach for quality-oriented control optimization in continuous casting.