Research On Short-Term Wind Power Forecasting Method Based On Deep Recurrent Neural Network

Along with the proposal of national sustainable development strategy,wind energy has been more and more widely used.Due to better integrate wind power into the construction of smart grid,it is one of the important means to improve the accuracy level of wind power prediction,which is irreplaceable significance for the completeness of state grid architecture and the strengthening of state grid self-healing ability in China.In this investigation,the wind power intelligent prediction model based on deep recurrent neural network is studied,and the applicability of the model is discussed with the application object of Qidong,Jiangsu wind farm,comparing with the advantages and disadvantages of the model and other forecasting models,the focus is on the following three problems:In order to solve the problem that difficult to identify outliers,based on the local outlier factor algorithm,the data preprocessing is carried out,and the outliers which affect accuracy of wind power prediction model is eliminated.The conclusion that numerical weather prediction data is correctable is obtained due to the statistical analysis of numerical weather prediction data and actual wind speed.In order to solve the problems that slow network training and low prediction accuracy,based on an improved extreme gradient boosting algorithm,the characteristics of the data are analyzed,the lower correlation features are abandoned,the model input dimensions are reduced simultaneously,thus to promote the performance of training speed and forecasting accuracy.In order to solve the problems that slow network training and low prediction accuracy,the paper proposes an improved deep recurrent neural network algorithm,the numerical weather prediction data correction model is established base on it.The output of this model is used as input vector in the power prediction model.Based on the improved deep recurrent neural network algorithm,and then use the model to predict the output active power of wind turbine within 4 hours and 24 hours.The results show that the deep learning algorithm is much more accurate than other algorithm and training faster than original structure,which verifies the effectiveness of the proposed algorithm.