Spatio-temporal Features And Error Processing In Wind Power Prediction

Wind energy is a clean energy with great development potential.high-precision wind power prediction plays a significant role in the development and utilization of wind energy.In recent years,machine learning has been applied to the field of wind power prediction and has drawn much attention due to its good performance in shortterm prediction,but there are still some problems,and the accuracy of the prediction system needs to be further improved.In this thesis,two problems of machine learning in short-term wind power prediction are discussed: 1)spatio-temporal features;2)prediction error processing.The thesis analyzes and summarizes the previous work,and on this basis,puts forward the corresponding improvement methods.On the one hand,The spatio-temporal features,including wind power historical information and spatial information,can effectively improve the accuracy of wind power prediction,but the advantages of spatio-temporal features have not yet been fully explored.This thesis focuses on the variance of spatio-temporal features and designs a hybrid wind power prediction method based on the research results.Firstly,the value of each dimension is regarded as the observation value of wind power to calculate the variance.According to the variance,the training datasets are divided into several groups averagely.Secondly,each group of data is used to train machine learning models separately.Finally,each model is evaluated and assigned a weight.The weighted average of the results obtained from each model will be used as the final result when making prediction.The method of grouping training set based on the variance and the method of calculate models' weights are two main innovations of this part.On the other hand,Experiments show that the predicted errors exhibit some rules when machine learning algorithms are used to predict wind power,such as hysteresis.The error data of existing models are collected in this paper,and an additional prediction model is adopted to estimate the error of the existing model prediction.Compared with the wind energy data,the randomness and uncertainty of error data are higher,although wind energy data itself has high randomness and uncertainty.based on the above reasons,when predicting the error data,the model is prone to over-fitting problem,to alleviate the above problem,a k-nearest neighbor data smoothing algorithm is designed for the data pretreatment to ease the problem.The feature extraction method and data smoothing method used in error prediction are the two main innovations of this part.Experiments on NREL datasets prove that the proposed method outperforms current advanced methods in predicting accuracy.Compared with Support vector machine regression,K-nearest neighbor regression,Decision tree regression and MLP neural network,the mean absolute error(MS E)are reduced 4.644%,12.088%,17.176%and 5.629% respectively.