Soft Actor-Critic based Gust Load Alleviation Control Design for an Experimental Flexible Wing

Abstract

This paper presents a data-driven control framework for active Gust Load Alleviation (GLA), on a flexible wing demonstrator, using the Reinforcement Learning based Soft Actor-Critic (SAC) algorithm. This approach extends RL methods to the domain of aeroelastic control, where traditional model based control methods are limited by the difficulty of obtaining accurate linear Aeroservoelastic (ASE) representations for highly flexible structures. The control law is defined by a neural network architecture that maps sensor inputs to actuator commands. This architecture is trained offline to achieve an optimal control logic using the SAC RL algorithm. The data required for this training process is generated from a high-fidelity linear ASE model of the flexible wing demonstrator. Rather than employing this model explicitly within the control design, it serves as a digital twin to provide the necessary training data. The study details the actor critic neural network structures, the reward model formulation, and the training procedure for controller optimisation. The resulting control law is evaluated in simulation and subsequently validated through wind tunnel experiments. Results demonstrate that the SAC based controller effectively mitigates gust induced loads, highlighting the potential of RL based methods for future active aeroelastic control applications.