Research On Wind Power Prediction Based On Deep Learning

The essence of wind energy utilization is to convert the kinetic energy of the atmospheric motion into the form of energy that can be used by human beings.During recent years,the environmental problem and the demand on energy have become increasingly prominent,so countries around the world are aware of the necessity of vigorously promoting the wind power development.However,the randomness and fluctuation of wind power bring challenges to the stable running of the power system.In this context,the high-precision wind power prediction becomes a powerful tool to guarantee safe and stable running of the power system.The ultra-short-term prediction and short-term prediction are taken as the main research content of this thesis and studied from relative aspects,including the pre-processing of wind power data,building of ultra-short-term and short-term prediction models,and example simulation analysis,etc.The principle of wind power generating and performance of wind power are introduced firstly;then,the meteorological factors affecting the wind power are analyzed: the Spearman correlation analysis method is used to conduct quantitative analysis on factors that might affect the historical wind power.On this basis,for problems like low prediction accuracy caused by abnormal and missed data in data pre-processing aspect,the K-means Clustering(K-means)algorithm is applied to detect abnormal data while the Multiple Imputation(MI)method is used to restructure the missed data,so to pre-process the original data collected from the wind farm.Finally,the original data is intercepted through the sliding window and compose the feature data that are required to be input in the prediction model,thus providing premium data for the study of the following prediction models.Targeting at the problems of ultra-short-term prediction models of wind power,including the complex characteristic extraction and low prediction accuracy,the Convolutional Neural Networks(CNN)is applied to extract in-depth wind power data characteristics firstly.Then,the Long Short-Term Memory Networks(LSTM)is used to learn the time series data.After that,in order to further improve the wind power prediction accuracy,the Bi-directional LSTM(Bi LSTM)is built combining with the forward and reverse information of the input sequences.Finally,the Grey Wolf Optimizer(GWO)algorithm is adopted to search optimal from hyper parameters of two modules in the model,so to obtain the optimal hyper parameters and build the complete combined ultra-short-term prediction model for wind power.The experiment results show that,the proposed combined model can further improve the wind power prediction accuracy.For the problems of large time span,unstable prediction,and lack of spatio-temporal characteristic extraction encountered during the process of building the short-term wind power prediction models,a series of methods as listed below are adopted to tackle them.First of all,in order to fully mine the spatio-temporal characteristic information of long time data,a parallel convolutional network composing of causal convolution and dilated convolution is established to capture and restructure the spatio-temporal characteristics of wind power data.Then,to maximize the fusion of spatio-temporal characteristics of the output layer and the data characteristics of the bottom layer,the spatio-temporal characteristics extracted by the parallel convolutional network are input into the improved LSTM network,that is: introducing the residual connection method in the improved model,so to make the model able to cross the circulation layers and transfer the dependence information of bottom layer,thus further improving the prediction accuracy of short-term wind power.Finally,the models are used to predict the actual wind power data and the results are then compared,thus verifying the practicability and effectiveness of the improved model proposed in this thesis.