Deep reinforcement learning for real-time latching control of wave energy converter in non-predicted irregular wave environments

This paper proposes a data-driven coupled model for the latching control of the Edinburgh Duck wave energy converter. To solve the complex state space problem of irregular wave environments caused by inherent randomness and wide spectrum, the coupled model consists of a deep reinforcement learning (DRL) algorithm based on the Soft Actor-Critic (SAC) method and a numerical wave flume implemented using computational fluid dynamics (CFD) methods. The wave dataset used by the algorithm comprises irregular waves generated through the Pierson-Moskowitz (P-M) spectrum. The wave-making capability of the numerical wave flume is validated, and the latching control agent is trained by coupling the DRL algorithm with multiple parallel numerical wave flume (NWF) environments. In the irregular wave testing set, the non-predictive DRL method driven by the environmental state is compared with both predictive and non-predictive benchmark control methods. Under the testing wave, the energy capture efficiency of the DRL control is 8.85 % higher than that of the predictive benchmark method and 17.3 % higher than that of the non-predictive benchmark method. Additionally, the study of the peak angular velocity of the wave energy converter (WEC) under different control methods demonstrates the load advantage of DRL control over the benchmark methods. The DRL control neural network output delay results confirm the algorithm's real-time performance. The generalization capability of DRL control was further validated under extreme waves and different water depths. This research demonstrates the effectiveness of a discrete latching action reinforcement learning model in the irregular wave environment and proves the practicality of the DRL method in terms of both energy capture efficiency and application.