A Research On Short-term Load Forecasting Of Distribution Network With Low Redundancy Based On Electricity Behavior Analysis

As the pillar industry of the country,power industry plays an important role in the development of national economy.Electric energy cannot be stored in large quantities,which requires the power department to do a good job in load forecasting.Accurate load forecasting results guarantees the reliable and economical operation of power system.In addition,with numerous of smart meters are installed in households,a large number of users’ electricity consumption data are collected,which also puts forward new requirements and challenges to the power sector in terms of load forecasting.For the problem of the redundancy of the feature set constructed in the process of load forecasting,a feature selection method of short-term load forecasting based on conditional mutual information and Gaussian process regression is proposed.Firstly,the mutual information between features and loads is analyzed quantitatively,and the maximum mutual information value between features and loads is selected.Secondly,the characteristics that can provide the maximum information gain to the load are selected from the remaining features,and the features are ordered by the calculated mutual information value.Then,the optimal feature set is selected under the condition of the predictor of Gaussian process regression and the decision value of the mean absolute percentage error.Finally,a Gaussian process regression model is constructed with the optimal feature subset for short-term load forecasting.The experimental results show that the proposed method reduces the redundancy of feature sets and the complexity of prediction model,and has higher prediction accuracy.The forecast result of the new approach is compared with support vector machine and back propagation neural network combined with Pearson Correlation Coefficient and Mutual information respectively.In order to meet the demand of load forecasting of distribution network based on massive user electricity consumption data,the characteristics of user electricity consumption behavior were analyzed,and the research on clustering and load forecasting of distribution network users was carried out.First,the cluster with GMM is carried out by using 6 features which extracted from the consumption behavior of users.And according to the clustering results,the original feature sets are set up.Benefiting from the conclusion that the accuracy of load forecasting can be effectively improved by using conditional mutual information feature selection method in the process of load forecasting.Then,the feature selection process was carried out by using CMI for choosing the optimal feature subset corresponding to each load of different cluster respectively.Finally,LSTM for load forecasting was trained by each optimal feature subset,separately.The experimental results indicated that proposed method shows significant effect to the promotion of load forecasting precision.