Combination of Neural Networks and Reinforcement Learning for Wind Turbine Pitch Control

Abstract

In this work, a hybrid proposal that combines reinforcement learning control (RLC) and neural networks for wind turbine pitch control is presented. A reward calculator updates a reward or a punishment depending on the value of the power error derivative. These rewards are used to train a neural network, which can learn the rewards expected to be received if an action is carried out at a given state. In this way, the controller executes the action that will receive the greatest reward. The approach is validated in simulation with the mathematical model of a small onshore 7 kW wind turbine. Results show how the error obtained with the RLC is much smaller than the obtained with a conventional PID regulator. Indeed, the RMSE is 30% lower, the MAE is reduced in a 34%, and the STD is 58% smaller.