EFJS']JSP-IBDR: Energy-efficient flexible job shop scheduling under incentive-based demand response via graph reinforcement learning with dual attention mechanism

As energy demand grows and the penetration of intermittent renewable energy, optimizing energy consumption patterns in industrial production processes has become an increasing challenge for sustainable manufacturing. This study investigates an energy-efficient flexible job shop scheduling problem with incentive-based demand response (EFJSP-IBDR), aiming to enhance energy consumption efficiency and cost-effectiveness. Given the time-sensitive and complex state features of EFJSP-IBDR, we propose an end-to-end graph reinforcement learning (GRL) approach. Initially, we decompose EFJSP-IBDR into a set of preference-based subproblems to facilitate multi-objective optimization. On this basis, a Markov decision process (MDP) with heterogeneous graph states is employed to model these subproblems. Subsequently, we construct graph neural network schedulers that integrate dual attention mechanisms to extract the features necessary for decision-making in each subproblem. Additionally, a similarity-based parameter transfer strategy is designed to accelerate the training process of all schedulers. Empirical results demonstrate that our approach can reduce peak demand by 47.22% and average load by 51.01% during the manufacturing cycle. The effectiveness of the proposed method suggests its potential to improve load-side energy consumption flexibility and grid resilience.