Development of Neural Network Based Traffic Flow Predictors Using Pre-processed Data

Abstract

Neural networks have commonly been applied for traffic flow predictions. Generally, the past traffic flow data captured by on-road detector stations, is used to train the neural networks. However, recently research mostly focuses on development of innovative neural networks, while it lacks development of mechanisms on pre-processing traffic flow data priors on tr aining in order to obtain more accurate neural networks. In this chapter, a simple but effective training method is proposed by incorporating the mechanisms of back-propagation algorithm and the exponential smoothing method, which is proposed to pre-process traffic flow data before training purposes. The pre-processing approach intends to aid the back-propagation algorithm to develop more accurate neural networks, as the pre-processed traffic flow data is more smooth and continuous than the original unprocessed traffic flow data. This approach was evaluated based on some sets of traffic flow data captured on a section of the freeway in Western Australia. Experimental results indicate that the neural networks developed based on this pre-processed data outperform those that are developed based on either original data or data which is preprocessed by the other pre-processing approaches.