Icing condition prediction of wind turbine blade by using artificial neural network based on modal frequency

Abstract

Wind turbines are increasingly installed in cold regions because of better wind resources. In these regions, high humidity and low temperature in winter increase the risk of ice accumulation on wind turbine blades. Different icing location and icing mass will lead to different natural frequency variation. In order to obtain an explicit relationship between the natural frequency of the iced blade and icing mass as well as icing location, the natural environment icing experiments and iced-blade modal experiments are carried out aiming at a 2 kW wind turbine respectively. In these experiments, the ice accumulates on the blade tip, blade middle, blade root or the whole blade respectively. The relationships between the first four-order natural frequencies of the iced blade and icing mass are constructed under four icing cases. Two artificial neural networks are used to predict the ice location and ice mass of the iced blade based on the established relationships in this paper. 335 data sets are generated to train the Back Propagation (BP) neural network and the Radial Basis Function (RBF) neural network. 12 groups of random mass samples are used to test the prediction ability of two neural networks. The prediction results show that the relative percentage error of local region (Elocal) icing is higher when a small amount of ice accretes on the blade. However, it decreases when more ice accretes on the blade. The mean relative percentage errors of the whole icing blade (Ewhole) predicted by BP and RBF neural network are all less than 7.5% when all three zones of blade are covered by ice. The mean Ewhole using the BP neural network is 2.83%, which is lower than that of the RBF neural network. The proposed method is also demonstrated on the 2 MW wind turbine blade data by using Finite Element Method (FEM), and the mean Ewhole using the BP neural network is 13.21%.