Ultra-short Term Prediction Of Wind Power Based On Data Decomposition And Reorganization

With the wind power industry booming and gradually replacing traditional fossil energy sources,abundant wind energy resources have become an important source of energy.However,the volatility and randomness of wind energy also bring challenges in terms of stability and security,which limit the further development of wind power industry.Therefore,in order to accomplish grid integration and real-time grid dispatching,it becomes critical to achieve high accuracy ultra-short-term wind power forecasting for wind farms,especially in4-hour and shorter time scales.In order to achieve high-precision wind power prediction,the construction and improvement of prediction models need to be carried out step by step from the following three aspects:identification,cleaning and repair of wind power anomaly data;wind power sequence decomposition;and multivariate prediction combination models.Ultimately,these improvements can help achieve high-precision ultra-short-term wind power forecasting for wind farms,thus solving the stability and security problems brought by the volatility and randomness of wind power to the grid system.The major work of this paper is as follows:(1)In the face of the problem of high dimensionality of raw wind power data input,high proportion of abnormal values,and increased difficulty in model training,measures were taken to preprocess the input data.The source of the data set was explained and input feature selection was carried out to provide data basis for subsequent research.Analyzing the influence of wind power abnormal data on prediction models,multiple analyses were conducted on abnormal data identification and correction,and methods for dual identification of abnormal values based on random forest-quantile regression and abnormal data repair based on BP neural networks were proposed.By comparing wind speed-power graphs and the errors of multiple prediction algorithms before and after data preprocessing,the superiority of the proposed methods was verified.(2)Aiming at the problem of strong volatility and high complexity of wind power series affecting the prediction performance of models,a quadratic decomposition algorithm based~2on ICEEMDAN-VMD was studied.Faced with the problem of too many components generated by one decomposition,a component reorganization based on permutation entropy is proposed to improve computational efficiency.In the face of difficult selection of VMD algorithm parameters,NGO algorithm is introduced to optimize and improve the secondary decomposition results.By comparing the errors of several prediction models before and after the second decomposition,it is verified that the second decomposition is reasonable and effective.(3)In order to break the limitation of single prediction model on multi-component prediction,the multivariate prediction model HKELM-Bi LSTM is proposed.After experimentally comparing the prediction performance of various shallow and deep learning models on each component,the superior performance of this model is verified.And on this basis,for the low-frequency component prediction part,first combine NWP based on K-means for clustering analysis,and then use INFO algorithm for parameter optimization of HKELM to further improve the prediction accuracy of the multivariate prediction model.(4)In order to verify the performance of the prediction model proposed in this paper in all aspects,three sets of comparison experiments are set up from the perspectives of different optimization algorithms,different decomposition algorithms,and different multivariate prediction models.The prediction results show that the multivariate prediction model based on ICEEMDAN-VMD-INFO-HKELM-Bi LSTM proposed in this paper exhibits comparatively higher fit and lower prediction error in the comparison of prediction curve fitting ability and various accuracy indexes.