Research On Ultra-short-term Wind Power Prediction Based On Deep Learning

Facing the increasingly serious environmental problems,reforming the existing energy structure and promoting sustainable economic and social development have become the consensus of all countries in the world.Our government has made it clear that CO₂ emissions will strive to peak in 2030 and strive to achieve carbon neutrality in 2060,and the development and utilization of wind energy is one of the effective measures.Wind power has been widely used because of its mature technology and high achievability,but the volatility and uncontrollability of wind power also bring serious challenges to power system scheduling and safe operation.Therefore,accurate prediction of wind farm ultra-short-term power generation becomes an effective way to reduce the impact of wind power on the grid and improve the stability of the power system.In order to improve the accuracy of ultra-short-term wind power prediction,this paper proposes an ultra-short-term wind power prediction method based on Gated Recurrent Unit（GRU）network.First,based on wind power generation principles and historical monitoring data of wind farms,the impact of meteorological factors on wind power generation is determined through qualitative analysis.Then,the historical data of wind farms are used as the research object,and the three kinds of anomalous data are corrected based on the K-nearest neighbor complementary method and the sample machine correction method,respectively.Then the Spearman correlation coefficient is used to quantitatively analyze the meteorological factors to determine the correlation between each meteorological factor and wind power,and three input data sets are constructed based on the quantitative analysis results.The optimal parameters of the GRU network model are then determined by the experimental method to construct the GRU network model under different input data sets.The effectiveness of the model prediction performance improvement based on the input data sets established by correlation analysis is verified through error analysis.Finally,the GRU network model based on deep learning is verified to have better prediction performance by comparing and analyzing with recurrent neural network,long and short-term memory network and feed forward neural network models.The input of the GRU network model is only the power and meteorological data of the previous moment,but the wind power is influenced by historical power and meteorological factors to a different extent at each moment.The wind power at that moment may be strongly correlated not only with the influencing factors of the previous moment,but also with the influencing factors of multiple historical moments.Therefore,this paper proposes an ultra-short-term wind power prediction method based on Time Division-Gated Recurrent Unit（TD-GRU）network to solve the problem that the GRU network model has unsatisfactory prediction results at abruptly changing moments.Firstly,we use Spearman correlation coefficient to analyze the Time-Division correlation of the strongly correlated factors and construct a time-dependent input data set for 96 moments of the whole day.And the optimal parameters of the network are determined according to the input data set of each moment to build the TD-GRU network prediction model.Finally,the GRU network model and TD-GRU network model are used to predict the power generation of this wind farm at the same time period,respectively,and the error analysis verifies that the TD-GRU network model proposed in this paper has better prediction performance and is more suitable for dealing with the nonlinear prediction problem of wind power,which is extremely time-series and volatile.