Study On Feature Selection Based On Learning Theory In Power Systems

The power system is a nonlinear complicated system, which is usually affected by many factors in the same time. The whole system's response will be more intricate and difficult to predict after the deregulation process because the market related stochastic factors will add their influences to the operation manners of the power system. Researchers always hope to clarify which factors act on and how they act on the response of the power system. For this purpose, a method, which is based on the machine learning theory, about feature selection in power systems is discussed in this paper.Based on the review to machine learning and feature selection, several popular feature selection algorithms are compared in details and the process of feature selection and the effect of each algorithm are illustrated by a simple example. The Minimize PRESS method is regarded as the method which can select a feature subset having the most powerful predicting ability. This method is chosen as the major method in the following analysis in power systems in this paper.Firstly, the load forecasting problem is studied. The mathematical description and the solving process of this problem are presented. The different deal strategies of continuous variables and discrete variables are emphasized in this problem. Some methods and skills to handle discrete variables are discussed. Then the chosen features are analyzed and explained according to their physical mechanisms. The accuracy of the chosen subset in prediction is verified by practical data and the results are satisfying.Secondly, the energy loss estimation problem is raised to illustrate how to determine the rank of each selected feature in the model. A practical sample method is applied to sample the training dataset in which the feature selection algorithm works. With the comparison of 1-4 rank models, the conclusion is drawn that the rank of each feature should be determined by its own physical essence if the models' predicting abilities are similar.Finally, the spot price forecasting in electricity market is studied in this paper. Based on the analysis of all possible features, a feature selection method with artificial correction is presented. The accuracy and intricacy of selected models can be highly improved using this method. The numerical tests show that the features selected using this method will be easier to understand and more accurate to predict the future value.The feature selection problem in power systems is quite complicated. A lot of questions such as how to select an appropriate subset from a large number of features, how to find the hidden principle from the volumes of data and build a model to body it and how important for each variable in the model to affect the output must be taken into account and solved carefully. The attempts to solve these questions in this paper will make some contributions to the future research work in this field.