2016 International Conference on Power, Energy Engineering and Management (PEEM 2016) ISBN: 978-1-60595-324-3 NARX-Network Based Wind Speed Estimation for Wind Turbines Feng HUO*, Xue-song ZHANG, Guo-rui JI, Zhong-peng LIU, Hai-tao DAI and Jian FENG Guodian United Power Technology Company Ltd, Beijing, P.R. China 100039 *Corresponding author Keywords: NARX neural network, Wind speed, Wind turbine. Abstract. This paper proposes a new wind speed estimation method for the control systems of wind turbines. The nonlinear autoregressive network with exogenous inputs (NARX) is applied to model the wind turbine and estimate the wind speed. The actual data from a wind farm is used to train both the traditional multilayer perceptron neural network (MLP) and the NARX-network. The experimental results show that NARX-network can produce a more accurate estimation than the traditional multilayer perceptron. Introduction Under the pressure of global climate change, most of countries in the world are searching and encouraging the new energy or renewable energy for sustainable development. As an important clean energy, the wind has been considered as a promising choice. With the increasing demand of wind energy conversion, different kinds of wind turbines have been developed from 20kW to 6 MW and even higher capacities in the past two decades. Wind turbines can convert kinetic energy to electrical energy. Greenpeace, an influential NGO, has stated that the electricity contributed by the wind turbines worldwide will go up to 10% by the year of 2020. Then how to convert energy effectively is still the focus of researchers in different countries. Generally, there are two types of wind turbines (WT) which are the fixed speed WT and variable speed WT. The latter is more reliable and widely used. The wind speed is essential to control the wind turbine below the rated wind speed for maximum power extraction. Normally, the wind anemometers are placed some distance away the wind wheel, which means the exact wind speed cannot be measured. The other demerit is that the initial and maintenance cost is quite high. So the wind speed estimation methods are necessary to improve the system performance. There are plenty of papers have been published about this topic vie different methods [1]. Artificial neural networks (ANNs) are popular for their modeling capabilities of the nonlinear time-varing system. There have been many papers published in the wind speed estimation by using ANNs. Some kinds of networks with different structures have been exploited. Li et al. [2] proposed a method to estimate the wind speed using traditional multiplayer perceptron neural network for a permanent-magnet synchronous generator (PMSG). Then the similar structure was applied to a doubly-fed induction generator (DFIG) in [3]. Qiao et al. [4] used Gaussian radial bias function network with three input elements which are power, rotor speed and pitch angle. This method is for varied speed VT and pitch angle is also chosen as an input of the network. The other types of ANN, such like support vector regression [5], echo state network [6], extreme learning machine [7] have also been explored. This paper proposes a new wind speed estimation method based on NARX-network. The idea is from another application of NARX which is presented in [8]. This network has a nature that the past input and output data in time series are used. Compared with traditional multiperceptron, its advantage can help to estimate the wind speed more accurately. The rest of paper is arranged as follows. The second section introduces the model of a wind turbine system. The third section proposes the NARX-network based method to estimate the wind speed. The experimental results are showed in the fourth section. Conclusions follow in the fifth section.