Research Of Time Series Data Mining And Forecasting

With the development of computer software and hardware, the ability of generating and collecting data by information technology has improved greatly. Millions of databases are employed for commercial managements, government affairs, researches and engineering manufactures which bring us huge storage of data, and new tools and technologies are required urgently to deal with these data and solve the problem of superabundant data and few information, so that datarnining come into being.As the important area of datamining research, time series datamining and forecasting have developed greatly. While as the particularity of data description, Researchers pay much attention to how to apply the traditional datamining technologies to time series datamining and forecasting.After surveying major issues in data mining, problems on time series' representation, searching, modeling and forecasting are analyzed. Some algorithms and solutions of piecewise linear representation of time series, measurement of similarity, time series modeling and forecasting are proposed.The work and contributions are listed as following:1. This paper analyze piecewise linear representation of time series and give a new type representation which are used to build a new measurement rule of similar time series. This new measurement avoids the problem of omitting the similar time series with different length of during time. This paper also discusses the mining algorithms.2. This paper summarizes the statistical characteristics of time series. In complex time series, long-time memory always exists with short-range dependence which needs a lot of parameters to describe. This paper proposed a new method that decomposes the time series into several ones, and uses different sampling frequencies on different new series. Models are then built respectively. This method has a gain on forecasting accuracy with decreasing of computing.3. The presented methods are verified by experiments with stock data and the results are given with figures.Finally, the dissertation is concluded with a summary and prospect of future researches of time series datamining and forecasting.