<p>Accurate prediction of ship fuel consumption under complex sea conditions is of great significance for promoting green shipping. However, achieving consistently high accuracy is challenging due to the dynamic nature of maritime environments, where static prediction models often fail to adapt effectively. This study proposed a multi-agent prediction method for predicting ship fuel consumption, using reinforcement learning to dynamically select the optimal prediction model to adapt to changes in sea state. The proposed method leverages Proximal Policy Optimization (PPO) to guide the selection process, enabling the system to dynamically adapt to environmental variations. Experiments based on real-world seagoing ship data demonstrate that the PPO-based strategy consistently achieves high predictive accuracy, maintaining <i>R</i>² above 0.89 across all sea states and achieving substantial error reductions compared with conventional single-model approaches. The results highlight the method’s engineering significance in achieving adaptive and energy-efficient ship operation under uncertain sea conditions.</p>