The Research On Ann Approach In Time Series Forecast

Forecast is necessarily an important link in scientific management and premise before policy-making and layout. It is necessary to forecast and analyze evolution trend of some system. One of the current forecast methods is time series forecast which constructs models according to the historical data before using it to forecast the future. Artificial neural network is an embranchment of artificial intelligent, originated in 1940s and is widely applied to many fields now. Neural network can study and reserve prevenient information and knowledge, which is the theoretical basis when used to forecast the future. As for nonlinear time series forecast, neural network is more efficient and precise than mathematical models.In this paper, definition, significance and several fundamentals of forecast were commented in chapter 1. Five error targets in order to evaluate forecast precision were introduced in chapter 2. Some usual forecast methods was commented in chapter 3. Chapter 4 summarized basic structures, algorithms and some existing problems, such as convergence rate, the global convergence and generalization. In chapter 5, two instances were used to research how to influence precision of forecast when in-out node numbers change, and came to a conclusion that choosing appropriate in-out node numbers according to practical time series is an effective approach to minish forecast errors and improve precision of forecast. In chapter 6, combined artificial network was commented and extended to multiple combined artificial network. Then it was proposed how to classify multiple combined artificial network and a method of numerical value preprocessing, which can effectively improve precision of forecast both in single neural network and in multiple combined artificial network and decrease total numbers of network as well as training time of multiple combined artificial network.