Research On Deep Learning Method For Ultra-Short-term Wind Power Forecast

In September 2020,the Chinese government promised the world to achieve the "double carbon" goal of reaching carbon peak by 2030 and achieving carbon neutrality by 2060.China’s total carbon emissions account for about 30%of the total global carbon emissions,ranking the first in the world,and energy combustion accounts for about 88%of the total carbon emissions.Therefore,energy and electricity are the main battlefields of the dual-carbon goal.Wind power generation is the most important component of clean energy at present.However,wind power generation has the characteristics of randomness,volatility and intermittency,which challenges the stability of the grid.Therefore,it is necessary to accurately predict the power of wind power generation,achieve peak shaving and valley filling,and ensure the stable operation of the grid.In this thesis,for the carbon peaking and carbon neutrality target,we first collected and analyzed the relevant factors of carbon emissions,and based on the scenario analysis method and LSTM network,predicted the annual carbon emissions of China under the three development scenarios of low carbon,baseline and high carbon by 2030,and reached the conclusion that China needs to continue to vigorously implement carbon reduction policies.Therefore,the wind power that China will vigorously develop in the future requires accurate prediction of wind power generation power.Due to the strong randomness and rapid change of wind power time series,it is difficult to predict.At present,the mainstream wind power forecasting methods mainly include physical models,traditional statistical models and machine learning models.The physical model is based on thermodynamics and hydrodynamics theory modeling on a large scale in the region,and it is difficult to accurately predict the particle size of small scale and fine time near the ground because of the complex operation of supercomputers.The traditional statistical model usually requires the time series to have stationarity or differential stationarity,while the wind power generation power series determined by wind speed is a typical nonstationary series with significant randomness and fast spatial and temporal decorrelation,so it is difficult to meet the requirements of power grid for prediction accuracy.The deep learning model can extract the internal time change rule of time series and the correlation relationship between different time series.Using high-quality data samples to train the model can obtain the current optimal performance.At present,the mainstream deep learning model in the precise wind power prediction task,the semantic features extracted are not sufficient,and the prediction result error is large,which can not meet the requirements of engineering practice.In this thesis,firstly,aiming at the problem of insufficient feature extraction,a two-stage spatiotemporal attention model is proposed.On the basis of the traditional temporal attention mechanism to extract the characteristics of the sequence itself,a spatial attention mechanism is added to extract the correlation characteristics between different sequences.The target and non-target mechanisms are introduced into the spatial attention mechanism to further extract the correlation characteristics between the predicted sequence and other sequences.Through prediction error analysis,according to the different characteristics of prediction error distribution in different wind speed intervals,a graph neural network model based on DTW is designed to extract useful information from the prediction error,realize further calibration of the prediction results of the model,and thus improve the accuracy of the model prediction.The performance test is carried out on the real electricity data set of a province in central China,and the results show that the model proposed in this thesis has better prediction accuracy and stability than the baseline model.