Wind turbines condition monitoring and fault warning have important practical value for wind farms to reduce maintenance costs and improve operation levels. Due to the increase in the number of wind farms and turbines, the amount of data of wind turbines have increased dramatically. This problem has caused a need for efficiency and accuracy in monitoring the operating condition of the turbine. In this paper, the idea of deep learning is introduced into wind turbine condition monitoring. After selecting the variables by the method of the adaptive elastic network, the convolutional neural network (CNN) and the long and short term memory network (LSTM) are combined to establish the logical relationship between observed variables. Based on training data and hardware facilities, the method is used to process the temperature data of gearbox bearing. The purpose of artificial intelligence monitoring and over-temperature fault warning of the high-speed side of bearing is realized efficiently and conveniently. The example analysis experiments verify the high practicability and generalization of the proposed method. INDEX TERMS Adaptive elastic network, condition monitoring, deep learning, wind turbines. I. INTRODUCTION

Fluctuations are a key characteristic of the wind resource. It is important to quantitatively analyze wind direction fluctuation due to its influence on the optimization of wind turbine yaw control. Based on wind resource data available from SCADA systems, a method is proposed to describe wind direction fluctuations in terms of fluctuation amplitude A and fluctuation duration T. A Weibull distribution is employed to fit the marginal probability density of both these two measures of wind direction fluctuations, and a mixed Copula used to connect the marginal distributions, establishing the joint probability density function. This representation has been verified through comparison with the real operating SCADA data. A set of indicators are extracted from the probability distribution which can accurately quantify the local wind direction fluctuation characteristics of a wind turbine. These indicators can be helpful in the optimization of the yaw control system parameters, facilitating an improvement in the power generating performance of the wind turbine.

With increasing size and flexibility of modern grid-connected wind turbines, advanced control algorithms are urgently needed, especially for multi-degree-of-freedom control of blade pitches and sizable rotor. However, complex dynamics of wind turbines are difficult to be modeled in a simplified state-space form for advanced control design considering stability. In this paper, grey-box parameter identification of critical mechanical models is systematically studied without excitation experiment, and applicabilities of different methods are compared from views of control design. Firstly, through mechanism analysis, the Hammerstein structure is adopted for mechanical-side modeling of wind turbines. Under closed-loop control across the whole wind speed range, structural identifiability of the drive-train model is analyzed in qualitation. Then, mutual information calculation among identified variables is used to quantitatively reveal the relationship between identification accuracy and variables’ relevance. Then, the methods such as subspace identification, recursive least square identification and optimal identification are compared for a two-mass model and tower model. At last, through the high-fidelity simulation demo of a 2 MW wind turbine in the GH Bladed software, multivariable datasets are produced for studying. The results show that the Hammerstein structure is effective for simplify the modeling process where closed-loop identification of a two-mass model without excitation experiment is feasible. Meanwhile, it is found that variables’ relevance has obvious influence on identification accuracy where mutual information is a good indicator. Higher mutual information often yields better accuracy. Additionally, three identification methods have diverse performance levels, showing their application potentials for different control design algorithms. In contrast, grey-box optimal parameter identification is the most promising for advanced control design considering stability, although its simplified representation of complex mechanical dynamics needs additional dynamic compensation which will be studied in future.

This paper analyzed the input and output data of wind farm based on deep neural network, developed intelligent model, and realized the predictive modeling of important parameter variables and control of wind turbine. By establishing the Deep Extreme Learning Machine(DELM), the higher-order nonlinear model is simplified. In this structure, unsupervised hierarchical ELM is conducted for feature extraction, and the features of the lower layer are transferred to the higher layer through layer by layer coding to form a relatively complete feature representation. Finally, the Extreme Learning Machine (ELM) is used to complete the mapping of feature representation to target output to minimize the loss of information in the transmission process. The target output is used as reference data for Pitch control of wind turbine, which is proposed by using a radial basis function (REF) neutral network. Simulation results from GH-Bladed show that proposed control algorithm can mitigate the loads effectively. The algorithm provides a practical reference for the design of wind turbine controller.

In this paper, a receding-horizon optimization strategy is introduced to optimize the wind farm active power distribution with power tracking error and transmission loss. Based on the wind farm transmission connections, a wind farm can be divided into clusters, in which the wind turbine generator systems connected to one booster station can be taken as one cluster, and different clusters connected from booster stations to the farm-level main transformer output the electric power to the grid. Thus, in the proposed strategy, the power tracking characteristic of the wind turbine generator system is modeled as a first-order system to quantify the power tracking error during the power tracking dynamic process, where the power transmission losses from wind turbine generator systems to booster stations are also modeled and considered in the optimization. The proposed strategy is applied to distribute the active power set-point within a cluster while the clusters' set-point still follows the conventional strategy of the wind farm. Simulation results show significant improvement for both optimization targets of the proposed strategy.