Real-Time Shipboard Power Management Based on Monte-Carlo Tree Search

The increasing concern in reducing greenhouse gas emissions and improving the fuel efficiency of marine transportation leads to a higher demand for intelligent power management systems (PMS). Unlike offline PMS with prior knowledge on the load profiles, real-time PMS is more challenging because of unknown load profiles. Two typical real-time PMS methods are equivalent consumption minimization strategy (ECMS) and model prediction control (MPC). However, ECMS can only make instantaneous decisions without the ability to handle large load demand changes, and MPC normally relies on pre-defined input references to guide the optimization. To alleviate these problems, in this paper, a Monte-Carlo Tree Search (MCTS) based method is proposed with a reward function guided by worst-case to minimize the fuel consumption. Meanwhile, a Siamese learning-based model is integrated with MCTS to improve its performance further. To ensure the sustainability of the power for unknown profiles, a time-dependent SoC-shore constraint is introduced, which intends to fully recharge the battery when leaving each of the shore connection stations. This constraint has not been considered in the previous real-time PMS studies. The experimental results demonstrate that the proposed method outperforms other methods on both fuel consumption minimization and the SoC-shore constraint fulfillment.