Data Driven Turbine Control in Tidal Power Generation: Development and comparison of data driven turbine control algorithms in a simulated underwater kite environment

Sammanfattning

Minesto’s underwater kite extracts energy from ocean and tidal flows through across-flow motion and aims to convert as much energy as possible from the surroundingwater-flow. To achieve this, designing a generator control signal that appliesoptimal torque on the turbine shaft is essential. This project aims to developdata-driven control systems for the kites generator that increases the power outputrelative to a baseline. ODE-based dynamical models of the kite and generator arederived to create a full simulation environment. The simulation is used to generatesensor data, as well as test and validate new control strategies. Two differentapproaches are investigated, a supervised predictive controller and a deep reinforcementlearning controller where the control signal is generated by the learned policy.The supervised predictive controller approach involves creating a supervised datasetfrom the simulation and training a recurrent neural network to forecast future inflowwater velocities. The knowledge of future inflow is then incorporated in the designof two new control methods, one that is built on the existing controller and anotherthat is developed as a standalone control method. The deep reinforcement learningcontroller instead utilizes the simulation directly by iteratively stepping through thesimulated environment and learning an optimal controller from the outcomes. Thisis achieved through the use of the state-of-the-art deep reinforcement learning algorithmSoft Actor-Critic. The solution also incorporates pre-training on generateddata to validate a possible simulation-to-reality adaptation. Within the adoptedsimulation setting, the main finding is that predictive turbine control based on forecastedinflow can outperform a reactive baseline controller. The recurrent predictivecontrollers consistently improved generated power across unseen evaluation environments,indicating that short-term prediction and the periodicity of the kite motionare useful for control. The Soft Actor-Critic approach demonstrated learning capabilitybut was more sensitive to reward design, partial observability, and tuning.Because the study relies on a simplified simulation model, the results should be interpretedas proof-of-concept and comparative evidence rather than as direct claimsabout real-system performance.