Fast online reinforcement learning control of small lift-driven vertical axis wind turbines with an active programmable four bar linkage mechanism

Maintaining high power generation for small lift-driven vertical axis wind turbines in a changing wind environment has not been well studied yet, due to the challenges inherited from the unpredictable turbulent flow-blade interaction and complex blade interferences. Herein, a fast online reinforcement learning pitch control using an active programmable four bar linkage mechanism is proposed, making it possible for turbines to quickly adapt to wind changes and maintain high power output in operation. We formulate the pitching mechanism using a drag-link configuration with a variable frame link length into an optimization problem and further solve it by the interior point algorithm under a wide range of tip speed ratios. Then, a parameter explorative policy gradient reinforcement learning method is designed for the turbine to adaptively tune the frame link length. Since the design significantly reduces the number of parameters needed to depict a whole pitch trajectory, the proposed online learning process can converge quickly, making it capable of handling complex wind conditions in an urban environment. The transient behavior overlooked in much of the literature is also studied. Comparisons to two benchmarks have demonstrated that our proposed system has a superior performance.