Research On Wind Power Prediction Method Based On Wind Behavior Analysis

With the urgent need of carbon neutrality,it is an inevitable trend to reduce the large-scale use of fossil energy.Wind energy,which has the most development scale and prospect among clean energy sources,has become a hot research field.The prediction task plays an important role in the efficient and reasonable utilization of wind energy,and is an urgent demand for the location of wind farm and the smooth and safe operation of power system.However,the wind power is influenced by climate,wake effect,topography and other factors,and has great intermittenness and randomness,which increases the difficulty of predicting wind power of wind farms in different environments.Therefore,it is of great theoretical significance and practical value to carry ou t in-depth research on the prediction model based on spatio-temporal coupling behavior characteristics of wind speed and the prediction model incorporating met eorological data.Scholars at home and abroad have proposed many modeling strategies for deep learning models,but there are still some limitations.Single prediction model can not take into account time and space characteristics at the same time,and there is a bottleneck in accuracy improvement.The combined model is difficult to integrate wind speed and meteor ological information into the model.This paper focuses on wind power prediction and puts forward two modeling ideas:(1)For wind farms that cannot obtain meteorological data,the 3D-CNN model with excellent performance in the field of behavior recognitio n is applied to the field of wind power prediction,and a prediction model based on the analysis of wind state behavior characteristics is proposed for the first time.Based on the historical wind speed data,the overall behavior characteristics were extracted to get rid of the over-dependence of temporal prediction on spatial feature extraction.By changing the size of convolution kernel,different spatial sizes and time lengths of data can be learned.In addition,in order to better extract behavioral features,a gradient extraction layer is added to provide more effective behavior change information for 3D-CNN.(2)Adopting the modeling strategy of multi-modal learning,the multi-modal Wind Behavior Characteristics Prediction Model(MBPM)based on meteorological data was proposed to deal with various heterogeneous data in a targeted manner.Under the strategy of behavior feature extraction from historical wind speed data,multiple RNN models are used to train and learn meteorological heterogeneous data respectively.Finally,multiple features are learned together through cross-modal feature fusion layer to ensure the difference and unified treatment of heterogeneous data.(3)For the above two models,the SCADA data of a wind field in Zhangbei County,Hebei Province were calibrated by lidar to verify the validity of the model.The results show that the proposed modeling strategy is effective and has lower error and better performance compared with PR,CNN-LSTM and so on.