Short-term Load Forecasting Methods Basecd On Big Data Analysis Technology

As an important part of energy management system,power system short-term load forecasting is more important for power grid dynamic state estimation,load scheduling and the reduction of generating costs.With the acceleration of the information process of the power grid,a large number of sensors and intelligent devices are deployed.During the long term operation of power grid,not only has accumulated mass load data,but also monitoring data such as temperature and humidity.The data of user side has reached the scale of big data,which greatly increases the difficulty of load analysis and prediction,and the introduction of price competition mechanism also puts forward higher requirements for load forecasting of power grid.How to effectively excavate potential information and knowledge from massive electricity data,and provide decision-making basis for power supply enterprises to expand market and allocate resources rationally,which is one of the main tasks faced by electric load forecasting.In the era of large power data,deep learning theory,data mining technology and Hadoop distributed architecture provide a new thinking mode and means to solve the above problems.The depth model achieves complex function approximation by learning a deep nonlinear network structure,which has strong adaptive perception ability and can discover potential information of big data.In order to further improve the prediction precision of bus load,a deep learning prediction method based on stacked auto-encoder neural network.This method combines the auto-encoder with the logic regression classifier to construct a multi-input single-output prediction model,and then the data about the reconstructed historical load,meteorological elements and others is all input into prediction model,which applies the stacked auto-encoder to the hierarchical learning and the load characteristics extraction.the short-term load prediction is realized by connecting the logical regression model at the top of the network.Finally,the actual load data is compared with other algorithms to verify the feasibility of the method.In order to meet the needs of user level load forecasting,a short-term load forecasting method based on pattern matching is proposed combined with data mining technology.This method applies MapReduce big data computing framework to realize load analysis and prediction process in parallel.Firstly,K-means algorithm is used to cluster the historical daily load to extract the typical load pattern,meanwhile identify the influence factors which have strong correlation with the load variation,the multi time point prediction model is subsequently constructed according to each load pattern.Based on the K-means clustering results,historical meteorology and daily type,the random forest classification model was established.According to the attributes of the forecasting date,the model was quickly matched to the load model,and the corresponding model was selected for prediction.Taking a city grid load data as an example,the simulation analysis under Hadoop platform is carried out,and the effectiveness of the method is verified from two aspects of prediction accuracy and parallel performance.