Wind Speed Prediction Model Based On Data Noise Reduction And Deep Learning

As the proportion of wind energy in the global energy use increases year by year,point prediction,interval prediction and probability prediction technologies in wind speed prediction are increasingly valued by wind power companies and grid companies;these three types of technologies can be applied to power dispatch in deterministic scenarios,wind turbine re-entry control and dynamic economic dispatch based on opportunity constraints.In terms of preprocessing methods for wind speed prediction,the empirical mode decomposition(EMD)based forecasting methods(that is quasi-EMD)cannot be used normally in actual forecasting.There are few researches on wind speed data preprocessing methods for alternative quasi-EMD methods,and wavelet hard threshold noise reduction has problems such as poor noise reduction effect.Moreover,in terms of wind speed prediction methods,the current deep learning algorithm convolutional neural network and long short term memory(LSTM)have many parameters,which are difficult to train and easily cause overfitting.In response to the above problems,this thesis is based on the combined model of preprocessing method wavelet soft threshold denoising(WSTD)and gated recurrent unit(GRU)of deep learning algorithm to study wind speed prediction.Correspondingly,the main research content and contributions of this thesis include:1)In view of the poor noise reduction effect of the wavelet hard threshold noise reduction method,this thesis introduces a WSTD method to preprocess the wind speed sequence.Based on 1000 hour-level historical wind speed values,the wind speed time series is first decomposed into sub-sequences by wavelet transform,and then the noise part of the sub-sequences is eliminated by calculating the threshold,and the noise-reduced wind speed time series is reconstructed.Subsequently,in view of the shortcoming that LSTM is prone to overfitting,this thesis introduces a deep learning method GRU as a wind speed predictor.Combining the WSTD method and the deep learning GRU algorithm,the wind speed point prediction combined model WSTD-GRU proposed in this thesis is obtained.The implementation steps of the GRU algorithm are as follows: Based on the noise-reduced wind speed sequence,first,in the wind speed sequence after noise reduction,the historical wind speed values of the previous 5 hours are input into the GRU neural network.After the loop iteration of the hidden layer,the wind speed value for the first hour in the future is obtained,for example,5 m/s;and then generalize it to multi-step point prediction,that is,the wind speed value for the second or third hour in the future.Finally,the actual wind speed data sets in four different regions are used for prediction simulation,and compared with the other 35 models mentioned in the calculation example section of Section 5.1,it proves that the model proposed in this thesis has advantages such as high accuracy,high prediction speed,low volatility of prediction results and strong adaptability.In addition,this thesis also adjusted the parameters of the GRU through the cross-validation grid search method,and realized the visualization of the tuning process.2)In view of the deficiencies of the heavy calculation burden of the quantile regression and the poor effect of the training process of the bootstrap method in the wind speed interval prediction when the sample size is large(usually no less than 1000samples),this thesis proposes a method based on error statistical interval prediction model.Based on the wind speed point prediction combined model WSTD-GRU,the point prediction model is first trained through the wind speed training set,and the errors in the training process are counted to obtain the root mean square error(RMSE)of the training set.In the prediction set,combined with the characteristics of the model error obeying the normal distribution,the result of the point prediction is taken as the midpoint of the prediction interval,and the confidence interval is constructed by the product of the z value of the standard normal distribution and the RMSE to obtain the wind speed interval in the first hour in the future,for example,4 to 6 m/s;and then generalize it to multi-step interval prediction.Finally,based on the actual wind speed data set,it shows that the interval prediction model has good adaptability,effectiveness and accuracy for multi-step interval wind speed prediction in practical applications.3)Aiming at the problems of model instability and sensitivity to initial parameters in the single predictor probability prediction method,this thesis proposes a wind speed probability prediction model based on error statistics and multiple predictors.Based on the wind speed point prediction combined model WSTD-GRU,train multiple point predictor GRU models through the wind speed training set,and calculate the errors obtained during the training process to obtain the mean and variance of the error statistics during the training process.The sum of the square of the mean and the variance represents the variance of data uncertainty.Subsequently,multiple GRU predictors are used to predict the wind speed prediction set,and the average value of the point prediction results of each predictor is obtained as the mathematical expectation of the wind speed that obeys the normal distribution for the first hour in the future.And calculate the deviation of the results obtained by each predictor,thereby obtaining the variance of the model uncertainty.Add the variance of the above data uncertainty and model uncertainty as the variance of the normal distribution,that is,the variance of the total prediction error,and thus obtain the output 1: the probability density function of the wind speed in the first hour in the future,for example,a probability density function that obeys a normal distribution with an expectation of 5 m/s and a standard deviation of1 m/s.In order to realize the re-entry control of the wind turbine,the result of the probability density function can be transformed into a prediction interval: take the wind speed expectation as the midpoint of the prediction interval,and construct the confidence interval by the product of the z value of the standard normal distribution and the total forecast error,and get output 2: the wind speed value interval for the first hour in the future,for example,4 to 6 m/s;and,because the probability prediction method proposed in this thesis takes into account the more comprehensive uncertainty information,the accuracy of the prediction interval is increased compared with the proposed interval prediction model in Section 3.Finally,the actual wind speed data set verifies the effectiveness and feasibility of the probabilistic prediction model proposed in this thesis.