Ultra-short-term And Short-term Forecasting Techniques For Wind Farm Power Outputs Using Recurrent Neural Networks

Wind energy is an environment-friendly,pollution-free renewable energy.It is also believed as one of the main forms of future energies.However,wind has a strong random and intermittent nature,leading to unstable wind power outputs and difficulties of wind power outputs controlling.Therefore,the increasing of wind power penetration has brought big challenges to the optimization of operation,scheduling and many other aspects of power systems.It has become an urgent need to improve the wind power output forecasting accuracy.The study of wind power forecasting methods could offer references to the optimal operation of wind power integrated power system.Therefore,this thesis mainly focuses on wind farm power forecasting methods.The main contributions of this thesis include these following:Due to the data transmission error or some other factors,abnormal power data might exist within wind farm historical operation data.It may impair the accuracy of wind farm power forecasting.Therefore,based on the weighted k-Nearest Neighbor(kNN)distance,this thesis has proposed wind farm abnormal power data identification model and abnormal data repair model.The proposed identification model adopts the weighted kNN distance to define the degree of outliers of the data points in the historical power data,and then identify the anomalous data according to the degree of outliers.The abnormal data repair model use the average of k nearest points' power value as the repaired value of abnormal data.The integrity and chronological of the raw data is maintained in such model.Based on the case study of a wind farm in Spain,4.2% anomaly data is identified and fixed using the proposed model.The correctness and validity of the abnormal power data identification and repair model are thus verified.Since the wind farm output power is influenced by many factors such as wind speed,wind direction,air pressure and so forth,reducing the input dimension of forecasting model and minimizing the timing and correlation between the forecasting outputs could improve the forecasting accuracy of the wind farm output power forecasting model.Therefore,based on condition mutual information,the input feature dimension reduction method is proposed,and the recurrent neural networks model of wind farm output power ultra-short-term forecasting is established.First,the historical power data and meteorological information are adopted as the original input feature.Based on the conditional mutual information method the features can be selected so that to provide useful information for wind power forecasting.The redundant information have been removed and input feature dimension of the forecasting model has also been reduced.Then,based on the recurrent neural networks,the wind farm output power ultra-short-term forecasting model is established as to forecast the next 2 hours output power value.Simulated on the wind farm data in Michigan,the results have shown that 72 features are selected from 252 input features using the proposed model,and the forecasting accuracy of the recurrent neural networks model has been improved by 9.33% when compared to the BP neural network model.In order to expand the forecasting period of the forecasting model,to further improve the forecast precision and to speed up the model training,a spectral clustering and recurrent neural networks based wind farm output power classification forecasting model is proposed,with the NWP data adopted as the forecasting model input.First,according to the trend of the next 24 hours' predicted wind speed,the similarity data of NWP data is classified by spectral clustering method.Then,for each kind of similar data,the forecasting model is established based on the recurrent neural networks,respectively.Finally,when conducting new forecasting,the class of the forecasting day's NWP data need to be judged first,and then the forecasting can be conducted using the corresponding forecasting model.Simulated on the wind farm data and NWP data in Michigan,the results have shown the forecast accuracy has been improved by 6.01% when the classification forecasting model is used,and the model training time is reduced by 85%.