Abstract
To achieve large-scale wave energy generation, multiple individual devices are integrated into arrays, potentially forming offshore wave farms. To enhance the efficiency of wave energy converter (WEC) arrays, the implementation of advanced control strategies is crucial. However, the complex coupling effects within WEC arrays pose challenges for precise modeling and increase computational burden of model-based methods. Hence, this paper proposes a model-free multi-agent reinforcement learning (MARL)-based method to address these challenges and improve conversion efficiency for WEC arrays. Three MARL control schemes are designed, namely: (i) Centralized training with centralized execution (CTCE), achieving high efficiency but with significant communication demands; (ii) Decentralized training with decentralized execution (DTDE), reducing communication requirements but with lower efficiency; and (iii) Centralized training with decentralized execution (CTDE), offering a balance between efficiency and communication needs. The performance of the MARL-based methods is evaluated on the Simulink platform, benchmarked against optimal centralized model predictive control (C-MPC) and reactive control (RC). Additionally, the impact of delays due to high communication demands and robustness under varying wave conditions are analyzed. Results indicate that the CTDE controller achieved 94% of the optimal efficiency of C-MPC under regular waves and 87% under irregular waves, while reducing computational costs by 99%.