Rotor Thrust Estimation Using RNN in TensorFlow - a Preliminary for Wind Turbine Control

EasyChair Preprint 13341

Abstract

In this report, we investigate whether a recurrent neural network (RNN) can be used to act as a virtual sensor and thus provide additional information for the control strategy. Estimators are currently used in control engineering, but they require a very detailed understanding of the model and are therefore difficult to create. In order to realise the thrust model, training data is created using a well-validated simulation model. From the time series created, the thrust force currently acting on the rotor surface is to be recognised on the basis of the last 200 time steps. In our case, the RNN is made up of several LSTM memory cells. The advantage of simulation is that all the required data is available and can be utilised. So the first tests were carried out with ideal initialisation and the past time steps of the thrust force were used to estimate the current thrust force.  As a result, the estimated and original thrust correlated very strongly. Since in reality the thrust force is not measurable and therefore cannot be initialised ideally, the initialisation vector was filled with zeros and is updated with the last estimated thrust at each time step. The results of the models look very promising and show that it is possible to use them as estimators of the thrust force.

Keyphrases: Machine Learnig, RNN, estimation, wind turbine, wind turbine control

@booklet{EasyChair:13341,
  author    = {Jeffrey Stegink and Andreas Klein and Julia Kersten and Reik Bockhahn and Maximilian Basler and Martin Becker and Dirk Abel and Janos Zierath},
  title     = {Rotor Thrust Estimation Using RNN in TensorFlow - a Preliminary for Wind Turbine Control},
  howpublished = {EasyChair Preprint 13341},
  year      = {EasyChair, 2024}}