Study On Ultra-short Load Forecasting Based On XGBoost

As an important link in power system,grid load forecasting has been put forward by experts in the industry.Traditional models show weakness when dealing with massive load forecasting data,and cannot achieve high efficiency.With the rapid development of computer science and information technology,machine learning becomesan efficient tool to solve this problem.In this paper,the real grid load data obtainedfromthe 9th Electrical Mathematical Modeling Competition was selected as the sample,and the K fold cross validation and mesh searching method were used to optimize the parameters.The accuracy of the prediction model was predicted by using Mean Absolute PercentageError(MAPE),Root-mean square error(RMSE)and Accuracy(A).Ninety sixtime points(every 15 minutes)of the power load in the Day2015-01-10 was predicted and compared using four kinds of models-RandomForest(RF),the Gradient Boosting Regression Tree(GBRT),Long-Short Term Memory(LSTM),and eXtreme Gradient Boosting(XGBoost).Thesuccess rate defined as the ratio of the samples with relative error<3%to the whole samples and the accuracy rate expressed as error rate of the whole samples were used to assess the predication result using the four different models.The accuracy rate of 94%or higher was considered as successful predictions.The success rates and the accuracy rates when using the RF,GBRT,LSTM,and XGBoost models were 31.25%and 92.42%,57.29%and 96.48%,88.54%and 98.04%,and 85.42%and 85.42%,respectively.The successand accuracy rate of XGBoost model increased by 28.14%and 1.57%compared with the traditional GBRT model,indicating that XGBoost has better fitting capacity for unvolatile single points.In addition,the LSTM model also displayedgood fitting ability,which is a potential development direction for grid load prediction in the future.However,due to the high cost of running equipment and the lower explanatory power of the relative tree model,it is not conducive to the understanding of the business personnel.Compared with that,the XGBoost model belongs to shallow learning,which not only requires low hardware facilities,fast running speed,but also has high prediction accuracy.Thererfore,XGBoost is a method that conforms to the trend of forecasting development of supershort-term load of power grid.This paper also studied the influence of different characteristic dimensions on the grid load prediction,namely,the power system load prediction in the multi-dimensional degree.(1)The XGBoost model only considering the sequence characteristics(Single-XGBoost,S-XGBoost):the success andthe accuracy rate was 44.79%and 96.26%;(2)XGBoost model with single increase of Temperature factor(Temperature-XGBoost,T-XGBoost):the success andaccuracy rate were 95.84%and 98.61%.(3)XGBoost model with single humidity-increasing factor(Humidity-XGBoost,H-XGBoost):the success and accuracy rate were 83.33%and 97.89%.(4)XGBoost model that increases Rainfall alone(Rain-xgboost,R-XGBoost)increase were 64.58%and the accuracy were 96.44%.(5)Multi-dimensional XGBoost model with comprehensive consideration of temperature,humidity and rainfall(Multi-dimensiona-XGBoost,M-XGBoost):the success and accuracy rate were 85.42%and 98.05%.The success rate achiveved was higher that that in H-XGBoost,and the accuracy is 0.72%higher thanin T-XGBoost model,indicating that the temperature displayed a significant impact on the power grid load prediction.Furthermore,the success rate and accuracy obtained in T-XGBoost model were 10.42%and0.56%higher than those of the M-XGBoost model,respectively.Therefore,more characteristic factors in the machine learning is not always better for pridicaiton results,and the feature with highly correlated redundancy factors will reducethe model prediction accuracy.