The Super-short Fluctuant Load Forecasting Baseing On Combination Forecasting Model

In order to make the super-short load prediction in heavy wave areas more accurateand make a proper96points scheduling plan in this kind of areas,loads are divided intothree types-the basic load,the impact load and the out-put of small hydro-powerplants.The large fluctuation load in this paper is a sudden change of the systemload.The heavy wave area refers to a area where the output of small hydro-power plantsand the impact load account for more than a third of the total load. The super-shortfluctuant load forecasting in this paper indicates the prediction of the output of smallhydro-power plants and the impact load every1hour or every15minutes.Frist,influential factors of the out-put of small hydro-power plants and the impactload will be found.Taking the power load in some area as an example,the out-put ofhydro-power plants is affected by the rainfall,temperature and wind in the basin wherethe hydro-power plant is.Meanwhile impact loads comes from the steel electrical loadmainly in some area,so the impact load in this area is influenced by the "rhythm"ofsteelmaking and the operational state of machines.And corresponding states vector andload data of historical moments should be screened.Two data screening method wasproposed-the improved entropy method and the weighted shape coefficient method.The multiple linear regression model is uesd for the hydropower loadforecasting,and the ARIMA model is used for the Impact load predictionmostly.However,the linear neural network and the the improved Markov model are usedfor the fluctuate load forecasting,and their feasibilities are proved theorily and actuallyin this paper.The weighted shape coefficient method is used to deal with vectors in thispaper.Then the theory of wavelet denoising will be used to process the historical data ofmain influence factors.Then the linear neural network is selected through the linearrelated experiments.Similarly，in order to make vectors easier to reflect the truesituation they are screened in the same way.Then predictions are made through theMarkov chain prediction model.At last, the result baseing on the linear neural networkis combined with the result basing on the improved Markov model through thecombination model based on entropy method.