Research On Short-term Load Forecasting Based On Cluster Analysis And Feature Learning

Accurate short-term load forecasting is of great significance to the dispatching of power generation end and the load distribution of distribution end in power system.With the construction of modern power system,wind power,photovoltaic and other new energy power generation,the application of new technology at the distribution network end makes the power grid structure increasingly complex,single prediction model can not adapt to the regional load prediction under the influence of multiple factors,and puts forward new challenges to short-term load prediction.This thesis focuses on building a combined prediction model.Based on the load data of a certain region in 2020,an accurate and appropriate combined prediction model is established from the selection of training data,the learning of input characteristics and the construction of prediction model.The main research contents of this thesis are as follows:Firstly,a method of screening similar days based on Kmedoids clustering is proposed.Considering the large amount of historical load data and strong diversity,lengthy training data learning value is low.Based on PCC analysis,the factors affecting the load in this region were selected,and the factors really affecting the load change in this region were selected.Then,the clustering model was input together with the historical load data to build the similar day selection method based on cluster analysis.The historical load data of training were optimized to improve the prediction accuracy.Secondly,a feature learning method of high correlation load components is proposed.In the PCC analysis of influencing factors,it is found that there are few influential factors with high correlation of load sequence.Therefore,the EEMD decomposition algorithm is used to mine the rules of load series,generate new load component features as potential inputs of the prediction model,and add load components with high correlation with load series on the basis of original input features to improve the prediction performance of the model.This process is called High correlation load component feature learning(HCFL).Finally,the prediction model is built and improved.Four basic prediction models,BP,GRNN,RNN and LSTM,were established to compare and verify the superiority of LSTM model.The Km-LSTM prediction model was built based on clustering screening similarity days,and the validity of clustering analysis on improving prediction accuracy was verified.The initial predicted value of Km-LSTM was combined with the training set load after clustering,and the combined load sequence was decomposed by EEMD and the highly correlated load component was selected and added into the training set.Considering the increase of dimension of model input,CNN-LSTM was built to enhance the processing ability of prediction model for high-dimension data.Through the simulation of load data in a certain area in MATLAB,it is proved that the Km-HCFL-CNN-LSTM combined prediction model based on cluster analysis and feature learning has better prediction accuracy and robustness for short-term load prediction in a certain area in 2020.The thesis contains 38 figures,24 tables and 80 references.