Application Of Granular Computing Based On Quotient Space Theory In Time Series Forecasting

Forecasting is a guidance to decision-making Time Series Forecasting isa method of forecasting the future based upon historical data and itsresearching scope involves multiple subjects. But the time series forecastingbased on the traditional statistics methods can't satisfy the demands ofnonlinear and uncertain multi-variable system in complex reality usually. Inthe process of time series forecasting, Artificial Neural Networks can refinevaluable information from complex patterns easily for its self-learning andself-adaptive ability. Traditional neural network can not satisfy therequirement because of its low speed of processing, the difficulty in definingthe structure and estimating the parameters. So it is very necessary tointroduce new ideas and methods into time series forecasting.In this dissertation, different time series have been discussed on the basisof quotient topology. By this means, complicated problem can be solved byanalyzing the complex time series with different grain-size. Also, it isvaluable to introduce the constructive learning methods to this model.The author's work is concentrated on the following areas:1. Granularity selecting from time series has been discussed. Differentgranularity can be obtained by granulating and synthesizing under the theoriesof granularity model. A suitable granularity can be selected according to thenature of the problem to provide corresponding knowledge. By this approach,a higher success ratio and a lower computational cost can be obtained.2. Further discussion about constituting and forecasting of time serieshave been given on the basis of quotient topology.In consideration of time interval, different-size granules selectingcorresponds with different time interval analyzing in time series. In QuotientSpace Model, relative series are combined synthetically for the problem of time-series forecasting since it take full advantage of granular informationfrom different layers based on relativity of information.In consideration of data overlapping, time series can be constituted byoverlapping. Since the information in it has the characteristic of nonlinearityand imperfection. In order to reduce the impact of the incomplete informationand obtain more accurate information and rules, the time series have beenoverlapped on the basis of quotient topology. By researching these time-serieswhich are interdependency as an integration namely multi-variable time-seriesanalysis, the properties of these time-series will be realized better. Theexperiment results of four-day-overlapped series 2 and nine-day-overlappedseries 3 in atmosphere quality forecasting both show that it is a reasonable andeffective method.3,It is more precise and readable with its simple and clear networkstructure and high testing speed by introducing the constructive coveringalgorithm into time series forecasting. The experimental results in yieldforecasting Of winter wheat shows that the method based on coveringalgorithm is more practical, flexible and steady. Kernel covering algorithm notonly overcomes limitation of traditional approach to prediction but alsoovercomes the weakness of both SVM and constructive machine learningmethod with characteristic of low computing and high accuracy. Theexperimental results also prove it is more effective in coal price seriesforecasting.It's only a primary discussion and attempt to apply quotient space theoryto the research of time series forecasting. With the deeply research incomputational intelligence field, more methods will be applied to timing seriesanalyse and prediction. It can make contribution to the development of theeconomy and the society.