ADAPTIVE OPTIMAL CONTROLLER FOR WIND TURBINE USING REINFORCEMENT LEARNING

Abstract

Currently, numerous researchers are prioritizing renewable energy sources as the cost of traditional fuels continues to escalate in daily operations, while also aiming to mitigate environmental impacts. Wind energy stands out as one of the swiftest expanding energy sources, renowned for its cost efficiency and widespread availability. The adaptive optimal controller using reinforcement learning for optimizing wind power extraction from available wind speed is the focus of this thesis. It has been difficult to harness energy from the wind because wind speed is so unpredictable. The proposed control system which is based on reinforcement learning (Q-learning) is used for an adaptive control to adapt the system and adequately achieve high performance compare to dynamic programming because of its model-free nature and its learning rate capability from the environment that makes this algorithm best for adaptive optimal control system. The addition of Kalman filter estimator collaboration with q learning makes the control system more robust and accurate in estimation and prediction of the control state which is again compute by q learning agents. In this thesis Q learning algorithm is employed for adaptive optimal control and dynamic programming algorithm is used for optimal control in place of traditional torque control to optimize the wind energy capture under fluctuating wind scenario in the wind turbine systems of region 2 operations. The MATLAB software is employed to analyze the effectiveness of this control system. The result indicates the tip speed ratio is 2.0 to 7.95 and 6.173 to 9.1787 respectively using q learning and dynamic programming with an aerodynamic efficiency of 0.411. Furthermore, the adaptive controller using q learning captured 2.08% and 10.17% more power than the optimal dynamic programming controller and STC respectively using wind speed from the input of piecewise step. In the sinusoidal function-based wind speed, the adaptive optimal controller using q learning collected 3.59% and 30.61% more power than the optimal controller using dynamic programming and STC respectively. As the given result depicted that the proposed control system which is adaptive control using reinforcement (Q learning) perform better compare to dynamic programming (DP) and standard torque control. The smaller percentage rise for the given condition is because of narrow scope of available wind speed in the simulation.