Deep reinforcement learning based optimal control for wave energy converters: practical implementation and experimental validation

Abstract

Ocean wave energy resources are immense and, if harnessed, can serve as a reliable source to support a significant portion of electricity needs. However, the technology to convert wave power into useful electrical power is still cost-intensive. One significant challenge is inefficient Power Take-Off (PTO) control technologies. To address this challenge, the research team has previously developed a Deep Reinforcement Learning (DRL)-based optimal control for Wave Energy Converters (WECs) and has shown promising performance in terms of power production and quality (in contrast to state-of-the-art advanced PTO controls). However, the study remains on paper, and the practical performance of the control is unknown. Therefore, in this study, we present the practical implementation and experimental validation of the DRL-based optimal control through wave tank testing, aiming to provide meaningful insights into (1) the challenges and limitations associated with implementing the DRL control in practice; and (2) the real-world power production performance (in contrast to numerical predictions), which is vital knowledge for the practical viability of DRL control (and, more broadly, other advanced controls). More specifically, the control is demonstrated using the Laboratory Upgrade Point Absorber (LUPA) WEC at Oregon State University and tested under a variety of regular and irregular wave conditions. Important takeaways from the testing results are: (1) The real-time implementation of the DRL control is straightforward with a low computational cost; (2) the robustness of the control under uncertainties and nonlinearities in real-world applications can be improved by introducing significant randomness during training; (3) The power produced in the real world is significantly less than that predicted by the numerical simulations due to various types of losses in the system (e.g., mechanical friction, electromagnetic damping, etc.), and (4) The DRL control is able to achieve good power production compared to practical control.