Short-Term Load Forecasting Based On Dynamic Similar Day Selection And Improved Stacking Ensemble Learning

Accurate short-term load forecasting of power system is an important basis for guiding power plants to formulate generation plans,coordinating reasonable operation of units,formulating maintenance tasks of units and formulating dispatching plans of power grid system.It plays an extremely important role in safe and efficient operation of power system and improving economic benefits of power grid.With the rapid development of modern power system and the increase of distributed energy,and the more complex factors affecting the short-term load change,the accuracy and practicability of the short-term load prediction of power system are challenged.In the process of electric power system short-term load forecasting,load change is affected by many factors that make its nonlinearity,time sequence and random characteristics such as more and more obvious,the single model prediction precision limited show the generalization ability and the problem of diminishing marginal utility,different prediction models are combined by a certain way,The prediction performance is better than that of single prediction model.Stacking integrated learning is an integrated learning method by combining the advantages of multiple model information to generate a new prediction model.The integrated learning model of multi-model fusion under the framework of Stacking is introduced into the short-term load prediction problem in order to improve the accuracy of the prediction model.And for the defects of the Stacking integrated learning model in the process of cross-validation training and prediction,the Stacking integrated learning model is optimized and improved.In view of the ever-expanding amount of power load data,a better prediction effect can be achieved by using less training data through similar day selection,and problems such as large amount of redundant data input and long training time can be avoided.In the traditional selection of similar days,the difference of the influence degree caused by different impact factors and the poor adaptability under different environments were not considered.Therefore,this paper introduced the ant-lion optimization algorithm to dynamically adjust the weight of each impact factor to improve the practicability and accuracy of the selection method of similar days.In view of the characteristics of time sequence and periodicity of power short-term load,the initial fitting is carried out for the selected time values of similar days and days to be measured,and the fitted values are used as input features to improve the correlation between characteristic indexes and load values.Based on the above methods,the similar day samples selected based on the dynamic similar day selection method are used as the training samples of the improved Stacking integrated learning prediction model,and a complete and effective short-term load prediction model of power system is established.Based on the load data of a region in southern China and a region in New England,the example simulation is carried out.At the same time,they were compared with the STACKING integrated learning prediction model,the improved STACKING integrated learning model-the time series LSTM initial pseudo-cooperation as the input feature,the single prediction model DBN-Nadam model and the traditional DBN model four models.The experimental results show that the proposed method has good generalization ability,robustness and prediction accuracy in the face of complex factors affecting load variation and a large number of historical training samples.This paper has 24 pictures,16 tables,and 62 references...