Power Load Forecasting Based On Feature Selection And Ensemble Learning

Electric power industry is an important basic industry of national economy,is also the strategic focus of the development of countries in the world.However,electric energy is a kind of energy which is difficult to store in large quantities.The process of production,distribution and consumption is almost simultaneous.Therefore,to master advanced load forecasting methods and improve the level of load forecasting technology is of great significance to ensure the safe and stable operation of power system and improve social and economic benefits.Based on the analysis of power load characteristics,this paper digs out four factors:historical load data itself,date type,meteorological factors and abnormal or special events.In order to make the observation data more reliable,the threshold value is set to construct the data distribution interval to identify and deal with the outliers from two aspects of the load value at the two time points before and after the longitudinal comparison and the load value at the same time point in the week before and after the transverse comparison.All the original data are normalized to eliminate the influence of different unit dimensions.Aiming at the problem that redundant features will reduce the prediction accuracy and operation efficiency of the model,this paper proposes a feature selection method based on MIC-GPR.This method can fully combine the advantages of the maximum information coefficient method and Gaussian process regression algorithm,eliminate irrelevant features,and select the features that contribute to the power load forecasting as the model input,which greatly reduces the complexity of model training.Aiming at the problems of low prediction accuracy and poor practicability of single power load forecasting model,this paper adopts integrated learning method to build a power load forecasting model of multi model fusion stacking integration.Firstly,by comparing the performance of power load forecasting models based on SVR,LSTM,random forest,GBDT,XGBoost and Light GBM single algorithm,in order to describe the data structure and characteristics from multiple perspectives,SVR model,LSTM model,random forest model and lightgbm model are selected as the base learners in the stacking framework for training and forecasting.Next,through the experimental comparison of the model forecasting effect under different stacking framework construction methods,the best power load forecasting model is built by using LSTM model,random forest model and lightgbm model as the primary learners in the first layer,and SVR model as the meta learners in the second layer.Through the experimental verification and analysis,the results show that compared with the traditional single forecasting model,the multi model integration stacking integrated power load forecasting model significantly improves the forecasting accuracy and the robustness of the model is stronger.