Research On Short-term Wind Power Forecasting Based On Gated Recurrent Unit

The randomness and volatility of wind power bring severe challenges to the safety control and stable operation of power system.The uncertainty of wind power often makes the current power system difficult to absorb it,and the effective method to solve this problem is to predict the wind power.With the upgrading of smart grid,deregulation of power market and comprehensive penetration of renewable energy,wind power prediction is becoming more and more important.Based on the important breakthroughs of deep learning technology in speech recognition,natural language processing,computer vision,medical treatment and other fields,this thesis takes the actual short-term wind farm data as the research object,and studies the problems in point prediction and probability prediction of deep learning algorithm-gated recurrent unit.The main research contents include:(1)In order to utilize longer time series characteristics to improve the accuracy of short-term wind power point prediction,a gated recurrent unit is applied to short-term wind power prediction.The time series characteristics of wind power data are analyzed.The network structure and parameters of 3+1 layers deep gated recurrent unit are designed.The performance of the model in iterative prediction,integrated prediction,multi-output prediction and other multi-step prediction methods is systematically compared.It shows that compared with other mainstream machine learning models,the gated recurrent unit has better prediction performance,and the integrated prediction is slightly better than the other two multi-step prediction methods.(2)In order to better quantify wind power uncertainty prediction,a short-term wind power probability prediction based on quantile regression gated recurrent unit is proposed.The proposed model is designed to predict the short-term wind power with different quantiles,and the probability predictions of different confidence levels are constructed.The experimental results showed that the proposed model had higher interval prediction performance than the quantile regression neural network,and the short-term wind power probability density curves with different confidence levels were obtained.(3)For non-stationary short-term wind power data,it is difficult to mine deeper temporal characteristics by single prediction models,a combined algorithm base on ensemble empirical mode decomposition-sample entropy and gated recurrent unit is proposed to improve performance of point prediction,a combined algorithm base on ensemble empirical mode decomposition-sample entropy and quantile regression gated recurrent unit is proposed to improve performance of probability prediction.First,the wind power data is decomposed by the time series decomposition methods,and then the sample entropy is introduced to reconstruct to reduce the number of modeling.In combined point prediction,the gated recurrent unit is used to model components data respectively,and then reconstruct it.In combined probability prediction,the quantile regression gated recurrent unit is used to predict the high frequency components with prediction lag,and the obtained results and components point prediction are reconstructed to obtain the probability prediction of different combination scenarios.Experimental results show that compared with a single model,the proposed combined point prediction model can greatly improve the accuracy of point prediction.The proposed combined probability prediction model can effectively compress the width of prediction interval,and the probability density curves are higher and thinner.