Optimizing The Output Energy Of A Vertical Axis Wind Turbine Using Deep Deterministic Policy Gradient And Proximal Policy Optimization

Official URL: https://risc01.sabanciuniv.edu/record=b3205720

Designing a controller to maximize the output energy of a vertical axis wind turbine can be a complicated task considering that the model equations of the system are required and if available can be highly non-linear. Degradation of generator components due to environmental factors also means that there will be a reduction in performance over a long period of time and the designed controller will need re-tuning. In this thesis we apply powerful deep reinforcement learning techniques, like deep deterministic policy gradient and proximal policy optimization, to a small scale vertical axis wind turbine and compare their performances to traditional control techniques like maximum power point tracking. Results show that reinforcement learning methods are intrinsically able to build a model representation of the vertical axis wind turbine and its parameters by observing relevant environmental variables like wind speed, rotor speed and load voltage, adjusting to varying wind speeds and changing generator components, thereby obviating the need for re-tuning. It is observed that a reinforcement learning agent with a properly shaped reward function enables the VAWT to extract more energy from the wind, generalizing to wind patterns it had not seen before, and demonstrating the potential of deep reinforcement learning techniques in optimizing the performance of vertical axis wind turbines.