Short-term Power Forecasting Of Wind Power Based On Improved Backpropagation Neural Networks

In order to deal with the energy crisis and environmental problems caused by the longterm and large-scale use of fossil energy,vigorously developing renewable new energy such as wind energy has gradually become the consensus of countries around the world to develop the power industry.In recent years,the wind power industry in the world,especially in China,has a large scale and rapid growth.However,due to the randomness and volatility of wind power generation,large-scale grid connected operation of wind power not only poses new challenges to the safe and stable operation of the power system and the high-quality guarantee of power quality,but also increases the cost of wind power generation in disguised form.As one of the key technologies to solve the above problems,the power prediction of wind power generation can ensure the security of power system and improve the consumption level of wind power.Based on the error back propagation algorithm,back propagation neural network has the characteristics of fast execution,strong stability and strong nonlinear expression ability.In this thesis,an improved back propagation neural network model will be established based on multivariate wind farm data to predict the output power.The main work of this thesis is divided into the following three parts:First,analyze the weather forecast data,wind measurement data and power data in the original data,and test the integrity and rationality of the data respectively.Then,according to the industry standards,the inspected data are processed accordingly.The validation results of the data set show the effectiveness of the test and processing.After that,this thesis selects various numerical indicators to evaluate the results of wind power prediction from multiple perspectives,and sets up comprehensive indicators for overall evaluation.Second,in order to further improve the quality of data sets and the accuracy of prediction,this thesis proposes a hierarchical clustering method based on similarity measure to clean abnormal data and reconstruct.This method uses similarity measures such as Euclidean distance,longest common subsequence and dynamic time warping distance to accurately capture data features and define the feature distance between time series.Then,taking this distance as the clustering index,we carry out condensed hierarchical clustering,and clear the abnormal data and reconstruct the missing data according to the clustering results.Then the Stacked Denoising Autoencoders(SDAE)based on sliding window processing is used to extract the most relevant features of power output,which effectively improves the operation speed.Through the validation of this data set,the above methods can effectively extract data features and clean abnormal data.Thirdly,on the basis of high-quality data sets,an improved back-propagation neural network model is established,and this model is used to predict the short-term power of wind power.In order to solve the problems of over fitting,slow convergence and falling into local optimal value of traditional back-propagation neural network,L2 regularization,momentum gradient descent,exponential decay learning rate and Softplus activation function are used to improve it.Finally,the effectiveness and superiority of this method are verified by single prediction and multiple prediction average results.According to the two-year data set verification,the monthly average accuracy rate of this method is 91.76%,and the monthly average qualified rate is 99.25%.