Application Of Pre-treatment For Data In Forecasting Models

This thesis proposed two modeling frameworks. For one thing, forecast and analysis one-dimensional time series taking tendency and period into account firstly. For another thing, forecast and analysis multidimensional time series after reducing the attributes by rough set method. Based on the structural analysis of the data system, these models not only improved the forecasting accuracy but also decreased the time and space costs of computing.For the one-dimensional time series with period but large noise, if we considered the tendency and period and operate certain appropriate pre-treatment and filter tendency and period before employ a model, then forecasted the relatively smooth series, finally we can obtain a higher accuracy compared to the forecasting without pre-treatment. The support vector machine substitutes the traditional principle of minimizing the empirical error by the structural risk minimization principle, whose basic idea is to transform the input space through the non-liner transformation into a high-dimensional space, or even an infinite-dimensional space, and then calculate and obtain the optimal classification surface in this high-dimensional. This thesis considered the tendency and period items combined to theε_ SVM model, finally performed a satisfying result.For the forecasting of multi-factor time series, it is often difficult to ascertain the impact of factors on the target sequence. Taking into account too many factors will easily contain a large number of redundant information, which will not only affects the accuracy but also increase the time and space cost during computing. If the factors we considered are few, it will cause the information loss and change the original system and data structure, also affect accuracy of the model. Rough set theory, good at dealing with imprecise, uncertain and incomplete data, can effectively analyze inaccurate, inconsistent and incomplete information. RS theory are sensitive to noise and weak on generalization ability, but neural network model owns strong adaptability, fault-tolerant ability and generalization ability, therefore can make up for the lack of rough set, then employ particle swarm to optimize the parameters of BP network. In the front of large amounts of data, BP network and GMDH method cannot determine the attribute combination of relative importance, and the constructions of the network are absence of a common method and the reasoning process are not transparent enough, and are lack of explanatory ability. These shortcomings can be improved by rough set method.