Primary Frequency Response of Floating Offshore Wind Turbines Via Deep Reinforcement Learning and Domain Randomization

This paper proposes a deep reinforcement learning scheme for the primary frequency response of floating offshore wind turbines (FOWTs). Considering the fact that current FOWT simulators cannot meet the demand for simulating numerous FOWTs, we first develop LoFT, an open-source framework for the low-order modelling and fast simulation of large amounts of FOWTs. Then, with LoFT, a deep reinforcement learning scheme is established. The scheme is a three-phase procedure, which includes training, validation and fine-tuning. With the aim of guaranteeing control performance under various operational scenarios, domain randomization is introduced to the training process. By this means, the proposed scheme demonstrates resilience to challenging environments and model mismatch. Finally, simulation results on IEA 15MW semi-submersible FOWTs verify the effectiveness of the proposed scheme.