Behavioral cloning for fatigue-oriented wind turbine optimal predictive control

Abstract

As wind energy production keeps on increasing, it is necessary to optimize the fatigue/performance trade-off of operating wind turbines. However, considering fatigue directly in optimal control needs to be carefully done because its faithful model does not necessarily fit standard forms commonly required by general purpose solvers. It was recently shown that a signal variance and its induced fatigue can be statistically related. This suggests that a fatigue-related cost function can be expressed as a non-quadratic cost function, and used in an open-loop optimal control problem. However, it was also shown that such cost function becomes very efficient only for relatively long prediction horizons. This makes the on-line solution of the underlying open-loop optimal control problem computationally demanding and undermines the real-time applicability of associated MPC schemes. In this paper, a behavioral learning solution is proposed to imitate the optimal controller obtained through open-loop, long prediction horizon-based optimization. The behavioral cloning controller is compared to a finely tuned quadratic MPC regarding its ability to reduce fatigue. Preliminary results show that the proposed solution enables significant fatigue reduction on a broad range of realistic disturbances while being realtime implementable.