Ultra-short Term Wind Speed And Wind Power Prediction Based On Wind Turbine Operation Characteristics And Deep Learning

The wind power optimization and the optimal control of wind turbines need high precision wind speed and wind power ultra-short-term prediction for each wind turbine.However,the wind speed and wind power ultra-short-term prediction of one wind turbine is faced with the problem that high prediction accuracy,small sample size and less calculation cannot be satisfied at the same time.The reasons are as follows:(1)The current prediction methods focus on research the characteristic of datasets.But the wind turbines operating characteristics are not considered enough.And then,the prediction accuracy is limited;(2)The current prediction methods emphasis on the data feature extraction from local wind farm data or local wind turbine data.These methods take insufficient consideration of the commonality between multiple wind farms data and multiple wind turbines data,resulting in the consumption of a large number of data samples and high calculation load.In this paper,wind power prediction model considering the wind turbine operating characteristics is established.The prediction model is designed to verify the effect of operating characteristics on the prediction accuracy.Secondly,the commonality between multiple wind farms data and multiple wind turbines data are reused based on transfer learning.The prediction model can be transferred between multiple wind farms and multiple wind turbines,so as to achieve the purpose of saving the data sample size and calculation load.The results show that the prediction accuracy of the wind power prediction model considering the wind turbine operating characteristics is improved obviously.The application of transfer learning not only ensure the model prediction accuracy,but also greatly reduce the amount of data sample size and calculation load,so that the prediction accuracy and the sample size reach the balance.The main research contents of this paper can be divided into the following three parts:(1)A wind power prediction model considering wind turbine operating characteristics based on the long short-term memory neural network is established.The consideration of wind turbine operating characteristics is mainly reflected in the processing and analysis of wind turbine operating data.The wind speed is decomposed to extract the frequency characteristics of wind speed.The decomposition method is used to deal with the problem of the signal abrupt change parts which caused by the fluctuation and uncertainty of wind speed is difficult to predict accurately.There are some other process variables besides wind speed in the process of wind speed convert to wind power.Rotor speed,yaw error,and wind direction,these process variables are introduced as the input of wind power prediction model.Experiments were designed to compare the influence of wind speed decomposition on the prediction accuracy and the influence of the wind turbine operating characteristics on the prediction accuracy.(2)A cascade wind speed and wind power prediction model considering multifactors distribution is established.On the basis of verifying the positive effect of wind turbine operating characteristics on the improvement of prediction accuracy,considering the different effects of multiple factors on the wind speed and power prediction process,a cascaded ultra-short-term prediction model of wind speed and power considering multi-factor distribution was established.Multi-factors distribution was showed as following: The input vectors of wind speed model are composed of wind speed sequence and wind speed decomposition.The input vectors of wind power model are composed of the predicted wind speed sequence,rotor speed,yaw error,wind direction,etc.The cascade model is to forecast the wind speed and power serially,and the wind speed prediction model results are used as the input of the wind power prediction model.(3)Based on the cascading prediction model,a transferable cascading prediction model is established by introducing the transfer learning mechanism.Firstly,a large number of local wind turbine data are used to train a high accuracy cascading prediction model.Then,the local turbine prediction model with parameters is transferred to the target wind turbine,and the target wind turbine data are used for retraining.The transferred target wind turbine model is used for frozen test to frozen part of network layer.Finally,the prediction accuracy,training data sample size and training time are the index of different network layers frozen test.From these comparison,the best transfer scheme are analyzed.The best transfer scheme prediction accuracy is equal to local wind turbine model which used training data samples and training time about1/480 and 1/2000 of the local wind turbine model respectively.These indices were verified the positive effect of the proposed method on the optimization between training precision,data sample size and calculation load.