Research And Application Of Transfer Learning Algorithm For Wind Power Time Series Data

Transfer learning is a method to solve the sparse problem of training samples in related fields by using existing knowledge.By learning the transferable knowledge on the source domain to assist the target domain task training prediction model,it can make up for the shortage of the target domain training samples and improve the predictive effect of the target learner.However,most of the current transfer learning methods focus on the classification of data,and there is no better solution to the problem of time series data regression.The research of time series data is widely used in various fields such as meteorology,energy and economy.Therefore,transfer learning based on time series data is also one of the hot issues in data mining research.It has good theoretical research significance and broad practical application prospects.The current wind power control system is based on the theoretical derivation model to make the control strategy,without considering the deviation from the theoretical model during the actual operation of the wind turbines.At the same time,in actual scenes,the training samples that obtain the changes of key parameters such as motor pitch angle and torque in the real environment are expensive,resulting in sparse of training samples and unable to train accurate power models.Therefore,this thesis takes the related problems of wind turbine system identification model optimization in wind power generation as a case study,combined with SCADA real data and Bladed simulation data,the following research on transfer learning based on time series data:(1)For the optimization of wind turbine system identification model,SCADA real wind power data is taken as the target domain,and Bladed simulation wind power data is introduced as the source domain.Based on the feature mapping idea in transfer learning,principal component analysis(PCA)is used as the feature mapping,looking for a shared low-dimensional subspace between the source domain and the target domain,maintaining the essential information of the data by selecting the first k dimensional principal components of the original data;in order to minimize the difference between the source domain and the target domain,Maximum Mean Discrepancy(MMD)is used as a measure of inter-domain distance to learn meaningful feature mapping between source and target domains by jointly minimizing PCA loss and domain consistency loss between domains.Finally,the converted target domain data and source domain data are merged,and the time series transfer regression model-PCTR is learned by using LSTM in the shared subspace.(2)In the PCA mapping of the PCTR model,only simple mapping relationship between the source domain and the target domain can be exploited.The source domain data is directly mapped to the target domain using the generated confrontation network GAN.In the process of generating network and discriminating network iterative game,the effect of mapping between domains is continuously improved,so as to ensure that the model can learn the complex and non-linear mapping relationship between domains.In addition,for the problem of large network capacity,a cyclic consistency constraint is introduced.A same GAN is inversely coupled on the basis of the original GAN,and the pseudo target domain data is remapped back to the source domain.The hypothesis space of the feature map is further reduced by minimizing the reconstruction error between the original data and the mapped data.Finally,the target domain data and the converted source domain data are merged,and a LSTM is used to learn the time-series deep transfer regression model-CYCTR in the target domain.Finally,the experimental results on a real data set of a wind farm in Jiangsu show that the proposed transfer learning models can effectively improve the prediction accuracy of the wind power model.In addition,by introducing the real data of the wind turbines,the model can reduce the deviation in the actual operation process and gives the optimal fan control parameters.