Research On Few-sample Wind Power Forecasting Method Based On Model-based Transfer Learning And Data Augmentation

Under carbon peaking and carbon neutrality goals,wind power has become the focus of the whole society,but the integration of a large number of new-built wind farms into the power grid will bring great challenges to the safe operation of the power system.Therefore,an accurate wind power prediction for the new-built wind farms is one of the most effective solutions.The combined forecasting methods based on mode decomposition and machine learning is a hot research direction in the field of wind power prediction,however,there are still some bottlenecks in improving the stability of mode decomposition and enhancing the ability of feature extraction.In addition,it is difficult to apply these methods in addressing the few-shot wind power forecasting problem.In response to the above problems,this paper takes the deep learning model as the technical framework,combing the model-based transfer learning and data augmentation methods to carry out research on the relevant theories and methods of few-sample wind power forecasting.The contributions of this paper are summarized as follows:1.For the prediction instability problem caused by high-frequency intrinsic mode functions（IMF）in mode decomposition,variational mode decomposition（VMD）is utilized to further decompose the IMF1 of wind power and wind speed time series obtained by empirical mode decomposition（EMD）,to form a unique EMD-VMD secondary mode decomposition method.In addition,in view of the difficulty of extracting the characteristics,a cascaded CNN-LSTM deep learning framework is designed to extract the coupling relationship and temporal correlation hidden features between wind power subsequences,wind speed subsequences and wind directions.Finally,EMD-VMD-CNN-LSTM wind power forecasting model is established.2.For the problem that the new-built wind farm has few training samples and there are transferable wind farms in the group,a unique serio-parallel CNN-LSTM feature extractor based on transfer learning method and the cascaded deep learning framework proposed above,is designed to extract spatio-temporal coupling relationship between the new-built wind farm（target domain）and transferable wind farms（source domain）so as to make wind power prediction.In order to avoid negative transfer problem,a criteria of selecting source wind farm and a transfer strategy for wind power forecasting are designed,which effectively transfer the information from the source domain to the target domain,and establish a transfer learning approach for wind power prediction based on a serio-parallel deep learning architecture（CLTL）.In addition,for the generalization problem of the prediction model,the fully-connected layer of the CL-TL model is retrained by using the crisscross optimization algorithm with L2 norm regularization term.3.For the problem that the new-built wind farm has few training samples and there are no transferable wind farms in the group,wasserstein generative adversarial network with gradient penalty is firstly applied to generate realistic data with a similar distribution of wind power,wind speed and wind direction of the new-built wind farm to augment the training dataset.In addition,for the instability problem of the augmented dataset,ensemble EMD is utilized to further decompose the augmented dataset so as to reduce the prediction difficulty.To solve the problem of robustness of the prediction model based on the swarm intelligence optimization algorithm,an asexual-reproduction evolutionary neural network is proposed.The prediction model is based on different loss functions under the framework of evolutionary computing,facilitating the network parameters approximating to the global optimum along different error surfaces in the evolutionary process,to greatly improve the adaptive ensemble learning ability.