Energy-Efficient Scheduling and Control of Aluminum Melting Processes Using Genetic Algorithms and Reinforcement Learning

Abstract

Manufacturing is a primary driver of global energy consumption, necessitating advanced production strategies to maintain industrial competitiveness. This study explores opportunities for thermal optimization within an industrial aluminum melting process in Thailand. While traditional manual furnace regulation and heuristic scheduling provide operational stability, the inherent complexity of multi-stage processes presents opportunities to further minimize holding and reheat losses through automated synchronization. However, existing studies typically address batch-level thermal control and system-level scheduling separately, and few explicitly target the holding and reheat losses caused by unsynchronized transfer between induction and melt-and-hold stages. To address this gap, we propose an integrated framework that combines deep reinforcement learning (RL) for batch-level power control with a genetic algorithm (GA) for system-level energy-aware scheduling. The RL component uses a deep Q-network (DQN) to learn an expert-guided power-control policy, while the GA incorporates a just-in-time (JIT) Gate, a synchronization mechanism that delays or releases melting batches to align induction-furnace output with downstream melt-and-hold (M&H) demand. Under repeated mixed-start evaluation, the revised RL controller achieved 597.96 ± 18.23 kWh per batch, slightly improving upon the original DQN result (600.79 ± 19.96 kWh) and the fixed expert-profile baseline (601.73 kWh), while preserving 100% target attainment and zero executed safety violations. At the system level, the GA-based scheduler eliminated reheat energy from 2,833.3 kWh to 0 kWh and reduced peak demand by 23.9%. These results indicate that RL provides robust, near-expert batch-level control, while GA-based JIT synchronization delivers the major energy-saving benefit at the production-system level.