Study And Application Of Multi-step-ahead Extrapolation Forecast Of Seasonal Time Series

Multi-step-ahead extrapolation forecast of time series is to estimate the values of the target variable in the coming period,which can win a lot of valuable time for decision-making and has great practical significance.In the real world,many time series are presented with seasonal components,while forecasting of seasonal time se-ries has always been the focus and difficulty of academic research.Therefore,this PhD thesis is dedicated to developing effective multi-step-ahead extrapolation forecasting model for seasonal time series in various forecast scenarios.For a seasonal time series with generally clear trend where slight irregular fluctua-tion component is embedded in it,this PhD thesis performs one or both of the following two types of data pre-processing operations in advance before extrapolation forecast to avoid the interference of irregular fluctuation component and seasonal component for forecasting:using the fully data-driven empirical mode decomposition based signal fil-tering to reduce the noise;using seasonal index method to get rid of the seasonal compo-nent.In order to overcome the limitations of the traditional multi-step-ahead extrapola-tion forecast methods,such as the inter-step error amplification,this PhD thesis employs multi-output feedforward neural network(MFNN)to model the preprocessed time se-ries and constructs three multi-step-ahead extrapolation forecasting models:MFE,MFS and MFES.After MFNN has got the forecasts,the latter two models need to recover their seasonal component to obtain the final forecasts since these two models had got rid of the seasonal factor before forecast.When using these three models for empiri-cal analysis,we realize one-week-ahead extrapolation forecast for electric load demand series with double seasonality feature through appropriate data reorganization.In the case that forecasts from multiple seasonal models have been available,this PhD thesis takes all of these individual forecasts into account by forecast combination and makes full use of the details reflected by them.In order to adapt to the changes in the forecast performance of each individual model caused by the time-varying nature of the data,this PhD thesis generalizes a high-order Markov chain model,which was original-ly applied to the categorical data sequence,to update the combining weights.According to this,we propose a time-varying-weight combing method,i.e.HM-TWA.This method regards the combining weight vector at each epoch as a Markov state probability distri-bution vector,and extrapolates them to forecast the out-of-sample combining weights.In addition,this PhD thesis designs a reasonable multi-step-ahead extrapolation fore-cast scheme for combining methods including HM-TWA,which utilizes h-step-ahead individual forecasts to predict the h-step-ahead out-of-sample combined forecast,then HM-TWA realizes one-cycle-ahead extrapolation forecast.For an incremental time series with double seasonality feature,which increases a little cycle(window)of data each time and possibly occurs abrupt changes in its incre-mental process,in order to achieve the dynamic forecast goal that performs a one-little-cycle-ahead extrapolation forecast after each data increment,this PhD thesis establishes a weighted average across windows(WAW)model on the basis of the ensemble idea of machine learning at first.This model estimates one so-called window model for each little cycle without interference between each other,so it has various advantages such as good memory,strong plasticity and small computing cost.The final combined forecast is the weighted average of forecasts of the window models.Then we build the FWAW model which is a WAW model with active selection mechanism for win-dow models based on Fisher optimal division method.With this selection mechanism and the inherent weighting mechanism of WAW,FWAW makes the combined forecastsacross windows reflect the property of the new data as much as possible.Furthermore,the PhD thesis establishes the S-FW model through designing a great-cycle correction mechanism for the window models of FWAW.This mechanism let the window models correctly reflect the double seasonality feature of the time series.Finally,we introduce a regime-switching mechanism that tracks and responds to the abrupt changes for S-FW,and then construct the RS-FW model.In order to verify the effectiveness of the proposed models in the three forecast sce-narios above and expand their real value,this PhD thesis applies them to some forecast fields in electricity system or market such as electric load demand,electricity consump-tion and electricity price.The experimental results show that their forecast performance has great improvement compared with traditional forecasting models.In particular,RS-FW can continuously return effective forecasts in the incremental process of the dou-ble seasonal time series,and can quickly restore the forecast accuracy after an abrupt change.