A differential evolution based approach for short-term predictive maintenance optimization of an offshore wind turbine

Offshore wind energy is growing rapidly but faces high maintenance costs. To implement predictive maintenance (PdM) in this field, this paper proposes a novel optimization approach for short-term PdM of a single offshore wind turbine (OWT). It combines an adaptive differential evolution algorithm with optional archive (JADE) and Gaussian process regression (GPR) models through a reinforcement learning (RL) framework. Temperatures from four OWTs are processed by an autoencoder to create a database. Using this database, a case study involving an OWT with five components is conducted. First, GPR models generate probabilistic remaining useful life (RUL) predictions for each component. An RL agent then simulates future degradation trajectories based on these predictions. JADE is applied across different scenarios to optimize maintenance by interacting with the virtual environment, considering external factors such as electricity and component prices, wind speeds, and maintenance duration. Results show that the maintenance plan based on pessimistic RUL forecasts exhibits the highest robustness across scenarios. This approach delivers highly precise and robust PdM schedules under prognostic uncertainty. It bridges a critical gap in integrating failure prognostics with maintenance optimization for OWTs.