Short-Term Freeway Traffic Volume Forecasting Using Radial Basis Function Neural Network

Abstract

A radial basis function (RBF) neural network has recently been applied to time-series forecasting. The test results of an RBF neural network in forecasting short-term freeway traffic volumes are provided. Real observations of freeway traffic volumes from the San Antonio TransGuide System have been used in these experiments. For comparison of forecasting performances, Taylor series, exponential smoothing method (ESM), double exponential smoothing method, and backpropagation neural network were also designed and tested. The RBF neural network model provided the best performance and required less computational time than BPN. It seems that RBF and ESM can be a viable forecasting routine for advanced traffic management systems. There are some tradeoffs between RBF and ESM. Although the performance of ESM is inferior to RBF, the former does not need a complicated training process or historic database, and vice versa. However, even in the best performance case, 35 percent of the forecast traffic volumes showed 10 percent or more percentage errors. This means that we cannot heavily depend on the forecast traffic volumes as long as we are utilizing the models tested. Further work is needed to provide a more reliable traffic forecasting model.