Application Of Semiparametric Estimation Methods To High Frequency Power Data

With widespread use of smart grid and high cost of electrical energy storage,it is vital to predict electrical load accurately for certain area in grid management and power supply decision strategies.Therefore,scholars have tried various kinds of solutions to enhance the accuracy of load forecasting.However,the forecasting performance of electrical load is affected by various uncertain aspects such as weather,calendar effects,economic and so on.With the fast spread of information,certain new and policies also have certain impact on real-time electrical load.In this case,only referring conditional mean forecasting of electrical load to decision making will bring the risk of power shortage.In this paper,we made some research on the conditional quantile forecasting of load forecasting to reduce the risk.The existing research mainly focused on non-parametric methods,while these methods did not take intraday periodicity and heteroscedasticity of intraday electrical load into consideration.So we learn conditional quantile estimation methods from other domains to predict conditional quantile of electrical load in this paper.For the purpose of getting more valuable information from electrical load and obtaining more reliable conditional quantile prediction,we extracted the intraday period from electrical load.After removing the intraday period,the characteristics of leptokurtosis,fat-tail and volatility clustering are showed through descriptive statistical analysis.Because of the similarity of all these characteristics between our data and financial time series,we use three kinds of methods classified by statistic perspective from financial time series forecasting to predict conditional quantile of electrical load.They are parameter estimation method,non-parametric estimation method and semiparametric estimation method.For the parameter estimation method,we used Risk Metrics which is popular in the financial field.For the non-parametric estimation method,we used Historical Simulation(HS).For the semi-parametric estimation method,we tried out Conditional Autoregressive Value at Risk by Quantile Regression(CAVia R),Filtered Historical Simulation(FHS)and Auto-Regression with Generalized Autoregressive Conditional Heteroscedastic error(AR-GARCH).For comparison,we chose Support Vector Quantile Regression model.We used the electrical load data in Lianshui County,Jiangsu Province,China from January 1st,2017 to January 4th,2018 with a frequency of five minutes for empirical analysis in this paper.We carried out various methods to predict the upper 1% conditional quantile,then evaluated the performance of all the methods by comparing the forecasting result with the actual value and the backtest result of conditional quantile.The empirical analysis shows that the performance of two semi-parametric estimation methods,FHS and AR-GARCH are much better than the others.So if we use this two methods’ results as reference for the power supply of smart grid,lots of energy will be saved and power shortage will rarely happen.The research has highly practical use.In summary,after removing the intraday period,FHS and AR-GARCH perform better in the conditional quantile estimation of electrical load and can provide a reliable reference for the grid staff.