A Dataset Construction Method Combined With Compress-filtering And Transfer-expanding For Short-term Load Forecasting

Short-term load forecast has played an important role on the development of the electrical company,in the terms of the programming,planning,scheduling and the consumption of the electric system.In the short term load forecasting,the main steps generally are the construction of dataset,the selection of the feature value and the construction of the forecasting model.A good method of constructing dataset can efficiently improves the prediction accuracy of the forecasting model.However,the existing researches on load forecast mainly focus on the periodicity law of load variation when constructing dataset,taking no consideration of the effects of holidays,economy development and other random factors such as emergencies which have an great impact on load variation.Besides,short term load forecast is usually to forecast a certain city so that the selection of dataset is limited by its city history data,taking no advantage of the history load data of other cities in the same province.For the problems mentioned above,we use Guangdong practical load data and do some research in this paper.Here are the main achievements:1.Considering the double peculiarities of load sequence being effected by both periodic trend and other random factor such as holidays and emergencies,we apply mean compression method of constructing load data2.We propose a compress-filtering based load forecasting method.By mean compression method,we compress month data into week data,which can in line with periodic trend and include randomness of other interference factor,and then screen month data by analyzing compressed data as training data set,whose trend is similar with the load variation pattern of the predicted target.3.On the basis of compress-filtering method of dataset construction,we propose a short term load forecast method which combines the compress-filtering with transfer-expanding.Based on analyzing the similarity of the load data between nearby cities,we apply the method of transfer-learning and introduce load growth rate to reflect differences between load variations caused by every aspect such as climate,economic aggregate and developing speed,and propose a transfer-learning method based on load growth rate,which uses source cities selection method and transfer the data of resource city to the target city data set.The forecasting case of Guangdong Province indicates that,comparing the existing KNN-MI construction based method,the method proposed in this paper,which combines with compress-filtering and transfer-expanding,outperforms in the mean absolute percent error and mean absolute scaled error which is decreased by 27.1%,22.6%and 9.5%,8.7%respectively.,when comparing with the existing mutual information filtering-based predicting method.