Research Of Freight Volume Forecast Model Based On RBF Neural Network

Freight Volume Forecast (FVF) plays an important role in national and regional development plan, as well as a base of investment decision making for transport infrastructure. Research on FVF aims at how to mine useful information from transportation system and use forecast model to process it, in order to provide accurate and effective FVF for traffic management department and transportation enterprise to help them arrange transportation reasonably and control transportation process conveniently.FVF has been studied for a long history. Quantitative prediction was commonly used in practice. Both time series analysis and regression analysis are based on classical statistics. Construction of them is mature technology and has been used in some FVF application. Recently grey prediction method and neural network, which have still great research potential, are the hot topic in the field of FVF research. Especially Radial Basis Function (RBF) neural network, with the character of approximation and global optimal, is superior to other methods.Based on research at home and abroad, in this paper, RBF neural network is used for FVF. Result of traditional RBF network may not be the best, if network structure is defined by empirical formula or the tester before training. As for it, by improving learning algorithm of traditional RBF neural network, a new dynamic cluster-based self-generated method for hidden layer nodes is proposed. It is a two-phase learning algorithm, in which the first phase is to gain center and spread constants by cluster and the second phase is to get weight matrix by least squares method. Network structure is adjusted with networks parameter learning, which could reduce error. Lastly, the learning algorithm is realized by programming on the MATLAB7.0 platform. Two groups of functions are defined to prove learning efficiency and extrapolation approximation ability of Dynamic Cluster Learning Algorithm (DCLA) validly.As for some applications, twice forecast method is used. When DCLA is applied to FVF, an extended matrix, combined with time domain information, is defined. The extended matrix is used to improve storage structure of input and output data, so that a simplified forecast model is built, as well as the learning data of neural network can be used well. As a result, learning ability to RBF neural network is promoted effectively and cumulated error is eliminated. Eventually, the DCLA based on extended matrix is used for road freight volume forecast and total freight volume forecast. Experimental result shows that the method is effective.