Active control of point absorber wave energy converters using deep reinforcement learning in open-water and wall-included environments
Abstract
With the rise of artificial intelligence, machine learning is increasingly applied to wave energy problems. This study proposes a performance-driven, adaptive control framework for Wave Energy Converters (WECs) using Deep Reinforcement Learning (DRL), trained with CFD-generated physical data. By integrating OpenFOAMbased hydrodynamic modeling with a Soft Actor-Critic (SAC) agent, the framework dynamically optimizes a cylindrical Point Absorber (PA) WEC through real-time wave-structure interaction. In open water (20-300 s), the DRL policy, from the proposed framework, achieves an average power of 0.63 W, compared with 0.61 W for Model Predictive Control (MPC) and 0.33 W for Resistive Control, while regulating WEC motion within safe operational bounds. The DRL policy demonstrates strong generalization under unseen wave conditions. To evaluate performance under more realistic conditions, a reflective wall was introduced behind the WEC. With low-cost fine-tuning, the resulting WALL-DRL policy maintains stable energy capture. It consistently outperforms baseline strategies under small WEC positional variations, and cumulative energy curves show smoother growth than MPC. These results demonstrate that the control framework achieves adaptive and high-efficiency control in both open-water and wall-included environments. It maintains stable performance under wave variability and minor WEC positional variations. This approach offers a promising solution for intelligent wave energy conversion.