| In order to make full use of wind energy,wind farms are usually located at high latitudes and high altitudes.Wind turbines operating in these areas often face various harsh climatic conditions,which brings a high risk of blade-icing.The blade-icing of a wind turbine will change its aerodynamic efficiency and torque,resulting in a reduction in power generation,aggravating its fatigue load,and affecting the economic benefits and stable operation of the wind farm.Traditional wind turbine blade-icing detection methods are usually based on manual detection or external sensors/tools.These technologies are limited by human expertise and additional costs.The mathematical based methods rely on domain knowledge and is prone to misjudgment.With the widespread application of sensor technology in wind turbines,a large amount of data has been collected,which makes the data-driven detection of wind turbine blade icing possible.Compared with other methods,data-driven methods do not require prior knowledge of domain knowledge,special materials or mechanical tools,which are gaining more and more attention.The data-driven blade-icing detection is considered as the problem of multivariate time series classification in this thesis.The main works and contributions can be summarized as follows:1.Aiming at the multivariate time series classification,a novel attention model is proposed.The proposed model is composed of a channel attention module and a temporal attention module,which aims to enhance the feature learning ability of convolutional neural network from the spatial and temporal dimensions of multivariate time series.2.Aiming to construct the data-driven models based on the fully labeled sensor data of blade-icing for wind turbines,a temporal attention conventional neural network(TACNN)is proposed.The proposed TACNN can learn the importance of each sensor in each time step,overcome the limitation of traditional CNN treating each sensor equally,and enhance its feature learning ability.Based on the proposed TACNN,an end-to-end framework for blade icing detection of wind turbine is proposed.The framework can deal with the highly imbalanced sensor data well and can carry out automatic feature learning and effective recognition of icing state.3.Aiming to construct data-driven model based on paritially labeled sensor data,a class-imbalanced semi-supervised model is proposed.The main motivation of this work is that the data-driven model is highly dependent on the labeled data,but it takes a lot of manpower and resources to label the data.At the same time,the data of normal state will be much more than the data of ice state in the collected sensor data,which causes the serious class imbalance between normal state and ice state.The proposed model integrates semi-supervised learning and class imbalanced learning into a unified framework,which can use the unlabeled data and overcome the class imbalance.4.Aiming to construct distributed data-driven model and enhance data privacy,a fedearated learning model for blade-icing detection of wind turbine is proposed.The model can utilize the data from different wind farms in a distributed manner for training,avoiding data leakage.The model combines the idea of transfer learning to improve the generalization and robustness of the model.The model introduces a privacy preserving class imbalance algorithm to overcome the class imbalance of data. |
