Research On Short-term Wind Power Prediction Based On Auto-encoder And Recurrent Neural Network

Since the energy shortage and environmental pollution becomes more serious,more attention is paid to the development of new energy technology.Among the new energy sources,wind power is in a high-speed development stage due to its unique advantages and characteristics.However,due to the strong intermittence,fluctuation and uncertainty of wind power,it brings great difficulty to power grid dispatch,which seriously affects the security and stability of power grid operation.Short-term prediction of wind power is an effective solution to this problem.Based on the current research status of short-term wind power prediction,combined with the actual historical data of a wind farm A along the southeast coast of China,this paper studies the point prediction and interval prediction of short-term wind power,and proposes several methods to improve the prediction accuracy.The main contents of this paper are as follows:(1)An algorithm based on Local Outlier Factor(LOF)and weighted K nearest neighbor(KNN)for anomaly data correction is proposed.The processed data will help to improve the prediction accuracy of the model and lay a foundation for future research.(2)The advantages and disadvantages of Denoising Auto-encoder(DAE)and Stacked Denoising Auto-encoder(SDAE)compared with traditional feature extraction methods(Principal Components Analysis,PCA)are compared through experiments.The results show that DAE and SDAE have better performance than PCA without considering the time cost.On the basis of choosing the appropriate descent dimension and combining with the prediction model,it can help the model to reduce the prediction error.(3)Because of the cumulative error of rolling prediction,the experimental comparison proves that the Recurrent Neaural Network(RNN)and the Long-ShortTerm Memory(LSTM)neural networks can effectively mitigate the cumulative error of rolling prediction to a certain extent.At the same time,a DAE-LSTM model based on Denoising Auto-encoder and LSTM neural network is proposed,which can improve the one-step prediction accuracy of LSTM model.(4)Because of the inherent defects of the model and the noise of the data,the point prediction results inevitably have errors.Based on the point prediction results of DAELSTM model,the k-means clustering algorithm is proposed to divide the power interval,analyze the error characteristics of each sub-interval,and then use the nonparametric kernel density estimation confidence interval prediction method to obtain the confidence interval of wind power based on the statistical error characteristics.The experimental results show that the proposed method is effective.