| Short term load forecasting plays an extremely important role in the dispatch of power system.Obtaining accurate short-term load can greatly improve the safety and economy of a power grid operation.A single forecasting model cannot meet the characteristics of large fluctuations and high frequency of the grid load.Hybrid prediction can combine the advantages of each algorithm,but the existing hybrid model combined with decomposition technology still has the problem of irregular mutations and noise interference in the input data,resulting in inaccurate component information and low prediction accuracy.Aiming at the deficiencies of existing prediction methods,this paper proposes a hybrid intelligent system prediction model of "clustering first,decomposition and prediction",namely DPC-EMD-LSTM-(FA)WNN,to improve the accuracy and robustness of prediction Sex and universality.The main tasks are as follows:(1)From the aspect of the composition and characteristics of power load,two important characteristics of periodicity and randomness of power system load are analyzed.The steps of power system forecasting are introduced,and the objectives,main tasks and key research contents of this paper are determined.Perform abnormal value correction and normalization processing on historical load data to improve data quality and reduce data gaps.Introduced the basic principles of prediction accuracy evaluation criteria and various learning algorithms.(2)Use the DPC clustering algorithm to divide the data set into four categories with strong similarity in trend to reduce noise interference.Use EMD to decompose the four types of data sets into different eigenmode functions with frequency from high to low,calculate the complexity of each component through sample entropy,and reconstruct the components with complexity greater than 0.5 into high-frequency components,The rest are low frequency components;(3)Established a load forecasting model based on a single model of shallow learning and deep learning.Through comparison,it is found that LSTM and WNN have better prediction accuracy,but WNN has the shortcoming of unstable output.To this end,a model based on swarm intelligence algorithm to optimize WNN is established,and firefly algorithm is used to optimize WNN weights,scaling factors and translation factors.Improve the robustness of the output.However,the single model FA-WNN and LSTM still has the problem of weak generalization ability.Therefore,LSTM is used to predict the high-frequency components and capture the information of the unsteady sequence;FA-WNN predicts the low-frequency components to improve the cognitive ability of the model.The two results are superimposed to output the final result.(4)Using data from a northwestern region from 2016 to 2018 to verify the"DPC-EMD-LSTM-(FA)WNN" hybrid prediction model proposed in this article.By comparing with LSTM,FA-WNN,EMD-LSTM,EMD-FA-WNN,EMD-LSTM-(FA)WNN algorithms,the simulation experiments on different date types are carried out.The results show that the prediction method proposed in this paper is at least 20%higher than the prediction accuracy of a single model,and at least 10%higher than the model using decomposition technology,which proves that the built model has good prediction accuracy and universal applicability. |
