An Adaptive MARL Large Model for Dispatch Strategy Generation in Logistics-Energy Spatiotemporal Coordination of Container Seaports

Logistics-energy coordination significantly enhances energy efficiency in electrified seaports. However, daily changes in environment data necessitate the re-implementation of optimization procedures, causing huge computational burdens. This paper proposes an adaptive multi-agent reinforcement learning (MARL) large model for logistics-energy spatiotemporal coordination of container seaports. The well-trained large model can directly generate optimal policy for each operating day from environment data without re-solving. To achieve this, a comprehensive logistics-energy coordination model is first established considering the spatial and temporal constraints of all-electric ships (AESs), quay cranes (QCs), auto guided vehicles (AGVs), and the seaport power distribution network (SPDN). The model is formulated as a Markov Decision Process (MDP). Then a MARL large model is developed, involving a hypernetwork mapping environment data to optimal policy, and special structures for both hypernetwork and agent policy networks to adapt to any number of daily arrival AESs. Additionally, a cascading action modification layer is designed to ensure correct action outputs within complex spatiotemporal constraints. A tailored training method with two acceleration strategies are developed for the large model. Case studies illustrate that the large model after training can automatically generate optimal policies with little to no fine-tuning, outperforming existing methods that require extensive solution time.