Design and Tank Testing of Reinforcement Learning Control for Wave Energy Converters

This paper introduces a model-free control strategy utilizing reinforcement learning (RL) to improve the electrical power generation of a point absorber wave energy converter (WEC). While model-based methods may suffer from control performance degradation due to modeling errors, such as inherent Coulomb-type friction, RL-based approaches are well-suited for the WEC environment, where system dynamics are complex or unknown. The strength lies in their ability to learn from interactions with the environment, bypassing the necessity for precise models. To enhance the control performance in electrical power generation, a control-oriented loss model is established, and a force penalty term is introduced into the reward function to avoid the WEC system operating in high-loss, low-efficiency regions. To further eliminate the reliance on wave information and improve applicability, an analysis is conducted to examine the contribution of each state feature to the training outcomes and a loss-considering and wave information-independent RL-based control scheme is developed. The RL-based controller is further validated on a point absorber WEC prototype in the wave tank experiment, demonstrating effective implementation and commendable performance in both regular and irregular waves.