Interval Prediction Of Daily Peak Value Of Power Load Series

Nowadays,with the increasing electricity demand,electricity load prediction attracts greater attention due to the fact that it has significant influence on the planning,deployment and management of electric power system.More importantly,the prediction of daily peak of electricity load which is an important indicator of electric power system is a subject worthy of study.Therefore,this paper focused on the prediction of daily peak of electricity load.Electricity load has nonlinear characteristics and are affected by many factors,which complicates the prediction process.The existing methods are hard to achieve high-precision prediction.In the past,the research on electricity load forecasting was mainly about point value forecasting.From the perspective of probabilistic forecasting,this paper made a profound study on interval forecasting to provide volatility range and fluctuation degree to relevant technicians.Most of existing studies on electricity load interval forecasting use the full parameter estimation method to construct the interval.In fact,the error distribution of the electricity load data is hard to be characterized by a known distribution in the practice.In this paper,we consider two heteroscedastic models from location-scale distribution family to describe the heteroscedasticity characteristics of data.Then we consider multiple parametric and semi-parametric estimation methods accordingly in the two model structures.The two heteroscedastic models are ARMA-GARCH model and Linear Double Autoregression(LDAR)model.Due to the fact that semiparametric estimation methods do not make the hypothesis of any known distribution,we introduced them to our application to reflect the real error distribution better in the purpose of higher forecast precision.We used the electricity load data of a County in Jiangsu Province for the empirical analysis.The propose of our research is to predict the confidence interval of the daily peak of electricity load.To do that,we convert interval forecasting to conditional quantile forecasting.We model structures are introduced,ARMA-GARCH model and LDAR model.In the two structures,we apply four conditional quantile estimation methods.They are historical filtering simulation(FHS),skewed-t distribution(for innovation hypothesis)(SSTD),peak over threshold method(POT)and quantile regression(QR).Finally,we use conditional quantile to construct both two-sided intervals and one-sided intervals.Then,we evaluated the interval forecasting performance.The empirical analysis of this paper shows that the model structures and estimation methods in our paper are effective.Moreover,ARMAGARCH model performs better than LDAR models in our data,and semi-parametric estimation methods have advantages in electricity forecasting.We creatively applied semi-parametric estimation methods under two model structures from location-scale distribution family to electricity load interval forecasting,which are widely used in financial risk management.It provides new clues for electricity load prediction problem,enriches the probabilistic prediction method and analyzes the pros and cons of the methods we used.