Ultra-short-term Prediction Of Wind Speed And Wind Power Based On Meta Learning

The intelligent management and control of the wind power system needs to predict the wind speed and power of each wind turbine quickly and accurately as information support.In order to cope with the complicated and changeable environmental wind conditions and unit operating conditions,the model predictive control of wind turbines has higher requirements for the instantaneity and accuracy of the wind speed and power prediction.However,most of the existing wind turbine prediction modeling methods use a large amount of wind power data to train accurate prediction models.For a single wind turbine with limited sample size and computing resources,its ultra-short-term prediction model is difficult to quickly adapt to the real-time working conditions of wind turbine,which limits the prediction accuracy.To solve this problem,an ultrashort-term prediction model of wind speed and power is constructed based on the meta learning algorithm in this paper.The paper mainly includes the following three parts:(1)Aiming at the wind power system,a meta learning algorithm is introduced,and a prediction framework of wind turbine’s key variables including wind speed and power is constructed.Analyzing the operation mechanism of wind turbines,the training sample size and training time of a wind turbine model are limited,which affects the prediction accuracy of the model,so consider introducing meta learning algorithm.Meta learning can extract common knowledge from historical wind power data of different units,which can be used to guide the rapid fitting of target unit samples in new tasks,so that the offline model can quickly adjust the network parameters to realize rapid adaptation of real-time working conditions of target units with few samples,thereby improving the generalization of the model and reducing the sample size and computing cost required for model training.(2)Based on meta learning algorithm,an ultra-short-term wind speed prediction model considering signal decomposition is constructed.The unstable characteristics of the wind speed signal will affect the extraction of wind speed features by the prediction model.Therefore,the Butterworth filtering method is used to decompose the wind speed signal into low frequency trend components and high frequency fluctuation components.At the same time,RNN and LSTM with memory units are considered as the network structure of the wind speed prediction model since the wind speed signal is a time series.The low-frequency trend component and high-frequency fluctuation component of wind speed are used as the input features of LSTM and RNN network layers respectively,to fully exploit the time-frequency characteristics of wind speed and improve the prediction accuracy of the model.Experiments show that the mean absolute percentage error of the ultra-short-term wind speed prediction modeling method based on meta learning and signal decomposition is about 2.6%,and its training samples and training time are much lower than the traditional LSTM prediction model.(3)Based on meta learning algorithm,a cascaded ultra-short-term prediction model of wind turbine rotor speed and power is constructed.A cascaded prediction model of rotor speed and power considering the influence of many factors is constructed to ensure the prediction accuracy of the model,analyzing the operating characteristics of wind turbines and taking into account the influence of unit state quantities,as well as multiple environmental variables including wind speed,wind direction,pitch angle,yaw error and ambient temperature.Among them,wind speed sequence,wind direction sequence,pitch angle,yaw error,ambient temperature and power are the input of the rotor speed prediction network,wind speed sequence,wind direction sequence,pitch angle,yaw error,ambient temperature and predicted rotor speed are the input of the power prediction network,realizing cascaded serial prediction of rotor speed and power.Experiments show that the mean absolute percentage error between the predicted value and the real value of rotor speed and power are about 1.6% and 5.12% respectively.Compared with other methods,the proposed method has better prediction accuracy and requires less sample size and training time.