Reducing Plant-Model Mismatch for Economic Model Predictive Control of Wind Turbine Fatigue by a Data-Driven Approach

This paper considers the inclusion of an adaptive element in the model-predictive control of a wind turbine. In fact, an adaptive internal model can reduce the plantmodel mismatch, in turn potentially leading to an improved performance. A Reduced Order Model (ROM) is augmented by training a Neural Network (NN) offline. The improvement in state predictions due to model augmentation is assessed and compared with the non-augmented ROM. The augmented ROM is then used as the internal model in an Economic Nonlinear Model Predictive Controller (ENMPC), which maximizes profit by optimally balancing tower fatigue damage costs with revenue due to power generation. The tower cyclic fatigue costs are formulated directly within the controller using the Parametric Online Rainflow Counting (PORFC) approach. The designed ENMPC is implemented using the state-of-the-art ACADOS framework. The performance of the controller and the impact of a reduced plant model mismatch is assessed in closed loop with the NREL 5MW onshore wind turbine, simulated using OpenFAST. Results show that the ENMPC utilizing the augmented ROM yields higher economic profit, slightly higher torque travel, and significantly lower pitch travel, compared to the ENMPC utilizing only the baseline ROM.