Wind Power Probability Prediction Based On Gaussian Process

Wind power is the most effective green renewable energy,because of the randomness and non-linear characteristics of wind power,the output power of wind farms is often difficult to control,these characteristics also bring challenges and difficulties to the power grid planning and scheduling.Therefore,reliable and accurate wind power prediction is critical to improve the reliability of the power system and optimize the operating costs of the grid,it also becomes one of the very important research direction in new energy field.At present,the commonly used predictive methods as neural network and support vector machine(SVM)have achieved better prediction effects in the wind power prediction.However,due to historical wind speed affected by various weather factors,the wind power error can not be completely avoided.As the neural network and SVM methods can only give the prediction of the predicted value,it to some extent improves the prediction error,if the accuracy of the prediction can be improved at the same time,but also can give the level of the confidence in the prediction,it is more conducive to evaluate the decision-making risk depends on the prediction.Learning from the training data set,Gaussian Processes(GP)can adaptively optimize the hyper parameters of the prior covariance function,and can obtain the predictive variance of the output while giving the mean prediction of the model,so as to explain the confidence level of the model well,and it is also the feature of its modeling.Considering a large-scale training data set,similar to SVM,the GP method can appear covariance function matrix operation difficulties,and the complexity of its calculation increases,these to some extent limit its application.Sparse Gaussian Processes(Sparse-GP)method as a class of probabilistic approximation of GP with fixed “hyperparameters” can be applied to the learning of large-scale data sets.By randomly selecting the data subset of the training data set,three different Sparse-GP methods including Subset of Datapoints(SoD)approximation,Subset of Regressors(SoR)approximation and Projected Process(PP)approximation are given.It maintains the advantages of GP method,and also reduces the computational complexity.The main contents of this thesis are as follows:(1)The corresponding learning algorithm of the GP method is studied,it especially focus on the application of GP method in the regression problem,and the covariance function and the “hyperparameter” used in GP are analyzed.(2)The basic theory of the Sparse-GP model is studied,it mainly focuses on a class of sparse gaussian process with fixed “hyperparameters”,and three methods of them: SoD method,SoR method and PP method have been studied.(3)The three kinds of the Sparse-GP methods are applied to the short-term wind power single-step and multi-step prediction experiments in different regions,under the same conditions,three kinds of the Sparse-GP methods are also comparied with the conventional GP and SVM method.Experiments show that the prediction accuracy of this method is higher than that of the conventional GP method and the SVM method in different single covariance and combined covariance functions,and the prediction accuracy of the same method in the form of the combined covariance functions is higher than that of the single covariance functions.