Research On Deep Learning Method For Ultra-short-term Power Prediction Of Renewable Energy

At present,the proportion of clean energy generation is gradually increasing,and its high uncertainty and volatility bring challenges to peak shaving and dispatching of power system.Accurate prediction of renewable energy output is the key to maintain the balance between supply and demand,security and stability of power grid operation.Aiming at the problem of ultra-short-term power forecasting for renewable energy generation,this paper makes an in-depth study on the application of deep learning technology based on data-driven.The specific contents include:1)Aiming at the problems of missing experimental data,abnormal data and noisy data,the feature analysis and preprocessing technology of time series data are studied.The abnormal data detection method based on isolation forest and the missing data filling method based on multiple imputation by chained equations are analyzed,and the correlation analysis and cluster analysis of time series data are carried out.The experimental analysis ameliorates the data quality and improves the generalization and releasability of the prediction model.2)The traditional ultra-short-term generation prediction method is not enough to mine the deep features of input data,and its generalization ability needs to be improved.This paper firstly analyzes the structure and model characteristics of deep learning.Three representative deep learning network models: CNN,LSTM,GRU,and the combination model are selected for comparative experiments.The prediction accuracy,advantages and disadvantages of each method are analyzed,and the power prediction problem in complex environment considering various influencing factors is studied.The effects of different time steps and prediction steps on the proposed method are analyzed experimentally.The results show that the deep neural network model can effectively improve the prediction accuracy and provide guidance for early warning,macro decision-making and power dispatching planning.3)In view of the difference between wind energy and solar power data,and considering the problem that the existing model is easy to lose important information when the time series data is too long,a wind power and photovoltaic power prediction method based on wavelet bidirectional long-short term memory neural network(W-Bi LSTM)with attention is proposed.The wavelet decomposition is used to extract the time domain information and frequency domain information of the input time series.Considering the bi-directional information flow,bidirectional long-short term memory(Bi LSTM)network is used for prediction.Attention mechanism is introduced,and different weights are given to the hidden state of Bi LSTM through mapping weighting and learning parameter matrix.The differences of attention distribution between wind and photovoltaic power at different time points due to the difference of data characteristics are highlighted.Effective information is selectively obtained.Therefore,the time correlation between data is captured and the corresponding driving data for prediction is selected.The experiment is carried out by actual data,and the results show that the proposed model has good prediction performance compared with the comparison model.