Affiliations: [a] Computational Intelligence Group of the University of the Basque Country (UPV/EHU), Basque, Spain | [b] Polytechnical School, UPV/EHU, Spain | [c] ENGINE Centre, Wroclaw University of Technology, Wroclaw, Poland
Correspondence:
[*]
Corresponding author: Manuel Graña, Department of CCIA, Facultad de Informatica, University of the Basque Country (UPV/ EHU) Paseo Manuel Lardizabal, Donostia-San Sebastian, Spain. Tel.: +34 943018000; E-mail:[email protected]
Abstract: The control of Variable Speed Wind Turbines (VSWT) to
achieve optimal balance of power generation stability and rotor angular
speed is impeded by the non-linear dynamics of the turbine-wind interaction
and sudden changes of wind direction and speed. Conventional approaches to
design VSWT controllers are not adaptive. However, the wind shear phenomenon
introduces a strongly non-stationary environment that requires adaptive
control approaches with minimal human intervention, i.e. very little
supervision of the adaptation process. Reinforcement Learning (RL) allows
minimally supervised learning. Specifically, Actor-Critic is designed to
deal with continuous valued state and action spaces. In this paper we apply
an Actor-Critic RL architecture to improve the adaptation of the
conventional VSWT controllers to changing wind conditions. Simulation
results on a benchmark VSWT model under strongly changing wind conditions
show that Actor Critic RL approach with functional approximation provide
great enhancement over state-of-the-art VSWT controllers.
