Refined Short-Term Load Forecasting Based On Historical Data And Real-Time Influencing Factors

In order to improve the accuracy and speed of load forecasting,this thesis relies on the topic 3 "Research on refined load forecasting method based on data mining" in the national key special project.A refined short-term load forecasting method based on support vector machine(SVM)is proposed,which integrates historical data and real-time influencing factors.The main research contents are as follows:(1)Data collection and pre-processing.The load data and weather data used in this thesis are all from a prefecture-level power grid company in Zhejiang Province.Data quality has a significant impact on forecasting accuracy.Therefore,four aspects of data preprocessing are introduced in this chapter to preprocess the sample data of the region in 2012,mainly including identification and correction of bad load data,interpolation of weather data,quantification of week types and holidays,and normalization of sample data,which lays foundation for analyzing load characteristics and building forecasting models below.(2)Load characteristics analysis and clustering based on data mining techniques.Based on the above data preprocessing,qualitative analysis and quantitative calculation are carried out from three basic tasks of load forecasting.First,data mining techniques are used to qualitatively analyze the regularity of load from three different time dimensions:typical days,weeks and holidays.Secondly,the Pearson correlation coefficient is used to quantitatively calculate the correlation coefficient between load and weather influencing factors from two angles of whole year and summer in 2012.The main weather influencing factors are selected as the input of SVM modeling.Finally,the hierarchical clustering based on weighted average distance is used to cluster the 366-day historical load of the region in 2012,and the forecasting models are established for different clustering results.(3)A refined short-term load forecasting method based on SVM is proposed,which integrates historical data and real-time influencing factors.Based on the statistical theory and regression forecasting theory of SVM,taking the load,weather,week types and holidays of the region in 2012 as sample data,forecasting models are established for six categories of clustering results.The types of kernel functions applicable to different forecasting models are selected,and the cross-validation method is used to perform global search optimization on the core parameters c and g of different forecasting models.The different forecasting models established above are used to predict the daily 96-point load of the same region in 2013,the actual system examples verification are given.And compared with the forecasting results of three traditional forecasting methods,the correctness of the method proposed in this thesis is verified.(4)A method for sensitivity analysis of load to main weather influencing factors under extreme typhoon weather is proposed,which is a special case of the above-mentioned modeling forecasting part.Since Zhejiang Province is located in the southeastern coastal zone and is susceptible to extreme weather such as typhoons,an analysis method combining the change rate of grey correlation coefficient and piecewise regression is proposed in this thesis,and three major typhoons of the region in 2012 are taken as examples for analysis.Based on the change rate of grey correlation coefficient,the main weather factors influencing load are selected.And the multivariate piecewise regression relationship between load and main weather influencing factors is established accurately,the sensitivity of load to main weather influencing factors is analyzed in different sections.And the load during typhoons is forecasted,the validity of the method proposed in this thesis is verified.