Variable Pitch Control Of Wind Power Generation Based On Reinforcement Learning

Earth’s resources are depleting, however, due to the socialdevelopment and population growth and other factors, these resourceshave been over-development and utilization. Ensuing greenhouse effectand atmospheric pollution makes humans had to seek a sustainable greenenergy, wind energy as a clean and efficient non-polluting energy isattracting more and more attention.Wind power has become an industry, countries have made manyachievements in the wind power theory and fan design. In this paper,variable pitch control as the key technology of wind power is the focus ofthe study, it is combined with the current advanced and sophisticatedintelligent control method to achieve high utilization of wind energy andrated output of wind turbines.Based on the analysis of aerodynamics and variable pitch workingprinciple we get the variation of wind power utilization factor, and how toobtain the best wind power utilization factor due to the pitch angleadjustment. According to the characteristics of wind turbine, this paperestablished the model of wind system, transmission system, actuator andgenerator. Through a combination of whole models we get aMathematical model that can reflect the dynamic behavior and is suitablefor control requirement. The model laid a foundation of control algorithmfor the next study.Without the mathematical model of controlled object is the biggestadvantage of reinforcement learning method to solve control problems ofcomplex system, and reinforcement learning method has self adaptabilitycan be very good to overcome the effect of perturbed motion and thedisturbance. This paper presents an adaptive heuristic evaluationalgorithm which is based on RBF neural network. The method uses theglobal approximation of RBF neural network to online approximateevaluation function and movement function. It also can handle continuousstate space, and solves the generalization problem to a certain extent. Ifthe method is applied on the pitch control, it only need to set thereinforcement signal, and update the network parameters by gradient descent method. Pitch control system can converge to better controlstrategy after traversing a certain amount of state action space.This paper introduces the sliding mode control theory to reduce thelearning time and accelerate the speed of convergence. Sliding modecontrol has the advantage of simple design, easy realization and gooddynamic response, it has a complete robustness in the sliding mode. Weuse the approximate analysis of wind turbine to get the sliding mode pitchcontrol method. In order to eliminate the adverse effects of chattering, weuse the saturation function to replace sign function, then we get the quasisliding mode control. Quasi sliding mode control lost robustness becauseof the continuous control in the boundary layer. In order to overcomingthe effect of disturbance on system control, this paper uses the slidingmode function as input vector and dynamically compensates the lostrobustness of reaching phase and sliding mode phase. Matlab simulationproves the validity and advance of the algorithm.