Towards the development of a wake meandering model based on neural networks

Abstract

In this work, we develop a neural network model for predicting the instantaneous wake position, which is crucial for a wake meandering model. The data used for training are from the large-eddy simulation of a utility-scale wind turbine. A neural network of four hidden layers with 128 units for each layer is found to be effective when training the model. Effects of different input features on the accuracy of the trained model are systematically tested. It is found that the input features including the downwind and crosswind velocities at two locations upwind of the turbine and the thrust and torque acting on the turbine are enough to guarantee the accuracy of the trained model. Without using the thrust and torque as the input features, the accuracy of the model is significantly worse.