Experiences with Model Predictive Control for the W2E-120/3.0fc

Introduction

The dimensioning of a horizontal axes wind turbine mainly depends on the interaction of the mechanical system and the operating controller. In the wind industry classical control schemes based on proportional-integral-derivative (PID) control algorithms are commonly used. Due to increasing market requirements regarding the manufacturing costs of wind turbines, load alleviation should be an additional goal for those control systems. This would allow for a lighter weight design and larger rotor sizes and hence more economic wind turbines. Among many others, one way to achieve these goals is using Model Predictive Control (MPC).

Approach

The approach we followed can be described by the following steps:
- implement a detailed multibody wind turbine model,
- extract fundamental dynamic turbine behaviour to derive reduced order model,
- setup Model Predictive Control (MPC) framework,
- compare controller performance in IEC-61400 load cases to classical controller.


Main body of abstract

Wind Turbine and Multibody Model

This study is based up on the 3 MW wind turbine W2E-120/3.0fc, designed by W2E Wind to Energy. The wind turbine model was implemented in three different types of general purpose multibody programs. An extensive experimental validation of such has shown, that alaska/Wind model is most efficient, when the trade-off between modelling detail and computational time is considered.

Model Predictive Wind Turbine Controller

The Model Predictive Controller comprises a reduced order model and an online optimization. The reduced order model is required to predict the future wind turbine behaviour based upon the current system state. This state is estimated online using an Extended Kalman Filter, which also utilizes the reduced order model.

The reduced order model reproduces the first tower fore-aft eigenfrequency, the first flap-wise eigenfrequency as well as the drive train's first eigenfrequency. The dynamic behaviour of the model is validated against the multibody model simulation results.

The controlled variables of the MPC are the generator speed, electrical output power as well as tower top acceleration. The optimization within the MPC also takes generator torque, rotor speed, pitch angle and speed constraints into account.

Results

For evaluation of the MPC performance it is compared to the classical controller for the 50 year extreme operation gust (EOG50) close to rated wind speed. It can be observed for the electrical power that strong variations occur in the case of classical control while the MPC control remains almost constant. The absolute maximum tower top acceleration can be significantly decreased by the factor of three compared to classical control schemes if MPC is used. The extreme loads in flapwise direction at the blade root can be reduced by 5 % and the tilt moment at tower bottom by 10 %.


Conclusion

The current research work presents Model Predictive Control algorithm for power levelling and load reduction for wind turbines. The results show that even with a significantly reduced control model within the MPC it is possible to reduce occurring loads e.g. in extreme situations. These load cases are design drivers and therefore determining the costs of energy of the wind turbine. So the reduction of such loads is of high importance and can be achieved by the usage of MPC. The load reduction at almost constant power level promises a further load reduction by focusing the optimization of the MPC on the generator speed.


Learning objectives
The learning objectives or lessons learned can be summarized as follows

- Model Predictive Control can be a suitable framework to efficiently combine the contradicting goals power levelling and load reduction
- Modelling depth of the real-time model used within the control system can be chosen relatively low
- In contrast it is strongly recommended to use a significantly more detailed turbine model, e.g. multibody model, for controller tests in order to prove robustness against model deviations