Research On Power Load Combination Forecasting Based On Feature Extraction And Bayesian Optimization

Power load forecasting is the basic guarantee for power system planning and development.Accurate forecasting is helpful to reduce cost and optimize decisionmaking.Reasonable load forecasting is very important for power system planning,which is an inevitable requirement for the sustainable,stable and healthy development of power industry,and has significant economic and social benefits.Various factors such as weather and economic conditions will have a certain effect on the fluctuation of power load,resulting in strong instability in load forecasting.Recently,scholars' research on power load forecasting methods has gradually deepened,and many load forecasting methods have been proposed one after another,but the common problem is that using a single model for load forecasting cannot fully show the load change law.The stability of the forecast accuracy cannot be guaranteed for different data sets,so the combining load forecasts model has gradually attracted people's attention.This thesis first introduces the concept and research background of power load forecasting and the combining forecast,as well as the meaning and basic principle of combining forecast.Then the basic principle and algorithm procedure of common machine learning algorithms are discussed including the Random Forest,Quantile Regression Forests,Gradient Boosting Regression Tree,Quantile Regression Gradient Boosting,XGBoost and LightGBM.And the basic principles of combination methods such as Simple Average,Weighted Average based on model performance,Ordinary Least Squares forecast combination,and Least Absolute Deviation forecast combination are presented.In addition,the paper introduces two model evaluation indexes: MAPE and RMSE.Because it is difficult for a single model to obtain the best forecast accuracy for all data sets,in order to deal with this limitation and further improve the accuracy and stability of short-term power load forecast,this paper presents an optimized combination forecast model based on machine learning algorithms: Based on Feature Extraction-Bayesian Optimization(FE-BO)combined forecasting model.The combined forecasting model mainly includes the following four steps: data preprocessing,feature extraction based on the model XGBoost,single models training,and models combination.The data are first normalized and missing values are processed.Then,based on the XGBoost model,the feature importance scores are obtained and sorted.The calculated feature importance scores are used to output a feature subset.This subset removes redundant features.Finally,the hyperparameters of the model are adjusted using Bayesian optimization,and the final combined model is constructed using the selected features and optimized parameters.Based on the actual power load data set and weather data set from January to October 2019 in China's A region,this paper applies the newly proposed FE-BO combination forecasting model for empirical analysis.Firstly the data set is divided into eleven data sets based on units of months,and then six types of single forecasting model: random forest,quantile regression forest,gradient boosted regression tree,quantile gradient boosted regression tree,XGBoost,and LightGBM are built on the eleven power load data sets respectively so as to perform load point forecasting for the last 24 hours.After that,six simple combination models are built respectively based on Simple Average,MAPE-based Weighted Average,RMSE-based Weighted Average,Ordinary Least Squares forecast combination,and Least Absolute Deviation forecast combination.The combined methods perform combination forecast using the error evaluation metrics MAPE and RMSE to select a high-precision reference model,and their forecasting accuracy are compared with the FE-BO combination forecast models which are based on feature extraction and Bayesian hyperparameter optimization.By empirical analysis,compared with the unoptimized original single models and the unoptimized combination forecast models,the new proposed FE-BO combination forecast model proposed in this paper has a great advantage not only in forecasting accuracy but also in forecasting stability.Besides,it has a good performance short-term power load forecasting.