Research On Short-term Load Probability Forecast Based On Copula And Nonlinear Quantile Regression

Electric load forecasting is very crucial to plan,operation and regulation of the grid system.Experts are studing how to strengthen the accuracy of forecasting,plan and dispatch resources more effectively,and reduce energy waste.But,which is affected by abundant uncontrollable components.The point forecasting can only provide a single output and cannot measure the uncertainty of electric load forecasting.But,interval prediction and probability density prediction can respectively provide the fluctuation range and probability information of future load,and provide more valuable prediction information than point prediction.Therefore,this article is dedicated to studying the short-term load probability prediction of nonlinear quantile regression,and considering the importance of variable selection for the accuracy of model prediction,a new variable selection method is proposed which is a combined approach of Copula method and XGBoost algorithm.The model can not only solve the complex nonlinear problems of power load and influencing factors,but also solve the uncertainty problem of load forecasting.This article first introduces the Victorian power load data set for the whole year of 2019.This article compares the pros and cons of the two methods of screening variables.The specific operation that the two methods of screening variables select the same number of independent variables,respectively use the selected variables,and put them into three forecasting models to predict the future electric load demand.The specific forecasting model which are Quantile Regression Long and Short-term Memory,Quantile Regression Forest and Quantile Regression Gradient Boosting.The results show that the method of the Copula method and the XGBoost algorithm has a greater impact on the dependent variable than the independent variables selected by the single XGBoost algorithm,and the final prediction precision of the three models has been raise.This article is based on the Victorian electric load data set for the whole year of 2019,using the newly proposed variable screening method and three nonlinear quantile regression models for substantial analysis.Since the scatter plot of temperature and power load is roughly presented in the shape of the letter "V",in order to predict the demand of electric load more accurately,point prediction,interval prediction and probabilities of electric load demand are made from two data sets in summer and winter.The results show that the three prediction models have high prediction accuracy in point prediction,the largest MAPE is only 1.2%,and the predictions are all at a high level,and the QRGB model shows good predicted performance regardless of the summer or winter data sets,its MAPE can be controlled at about 0.77%.In the interval prediction,use the PICP,PINAW and Corrected Predictive Interval Accuracy(CPIA)to evaluate the prediction precision.The prediction effects of the three prediction models are better,and the interval prediction of the QRLSTM model performs the best.In the probability density prediction,the probability density is plotted the prediction map can show the probability density prediction distribution of the three models at different times,and use the average pinball loss(APL)indicator to evaluate the probability density prediction effect.This article is also based on the full-year electric load data set of a city in China in 2019,using the newly proposed variable screening method and three nonlinear quantile regression models to predict the electric load demand for the next three days.From the prediction performance on the summer and winter data sets,it can be known that the point prediction effect of the QRGB model is the best,the interval prediction performance of the QRLSTM model is the best.In summary,combined with the forecast performance on the two data sets of Victoria and a city in China,we can know that based on the method of Copula and XGBoost to select variables,using QRLSTM,QRF and QRGB three models for electric power load forecast is better.This method has universal applicability.For different data sets,the prediction performance results are relatively consistent,and the prediction ability is strong,and the prediction effect is excellent.