Research On Load Feature Extraction And Short-term Load Forecasting Based On Cumulative Contribution Rate

Due to the demand of the electricity trading market and the popularization of smart grids,the reasons for fluctuations in user power consumption collected by the power system have evolved from a single factor to a comprehensive effect of complex and diversified factors In order to ensure that the power system accurately grasps the power consumption of users,short-term load forecasting has become particularly important,especially for the extraction of load characteristics and the study of forecasting methods.As a key link of the power system,short-term load forecasting is of great significance for power companies to formulate reasonable transmission and distribution plans,stimulating social and economic benefits,and ensuring the supply of electricity for residents,businesses and industries.Therefore,short-term load forecasting has always been a key research content of the power industry,and advances in science and technology have also given birth to a variety of research methods in the field of shortterm load forecasting.The specific research content of this article is:(1)In view of the missing data in the original load data set due to the failure of the collection equipment or human input errors,the original load data is first used to repair the missing values by using multiple imputation method;In order to meet the input requirements of constructing short-term load forecasting models,the Z-score method is used for standardized processing.(2)In the process of traditional load feature extraction,the problem of unreasonable load feature extraction caused by setting the threshold of cumulative contribution rate according to general experience.First,the elbow principle is used to optimize the cumulative contribution rate threshold after the PCA algorithm is processed,and the cumulative contribution rate is set according to the optimized threshold and the empirical value,and two different load feature combinations are determined.Secondly,build a feature optimization model based on BP neural network,and verify the output error of the feature optimization model.The cumulative contribution rate threshold after elbow principle optimization is more reasonable than the general experience value,and the error of the feature optimization model is lower.(3)The load characteristics constructed based on the optimized cumulative contribution rate threshold in this paper are used to verify the error between the optimized load characteristics and the real data when the optimized load characteristics are used as the input of the short-term load forecasting model.The simulated annealing algorithm and artificial fish school algorithm with strong global optimization ability are used to construct short-term load forecasting models based on SA-BP and AF-BP respectively.The PSO-GRNN and Attention-LSTM models are compared,and the optimized load characteristics are used as input to compare the error of the output result,the decline process of the error function,and the fitting effect.Experiments have proved that when the load characteristics optimized by the elbow principle are used as model input,the accuracy of short-term load forecasting based on SA-BP and AF-BP is improved to a certain extent,and the effect of short-term load forecasting model based on SA-BP is better.