A neural networks system for traffic congestion forecasting

Advance Traffic Management Systems (ATMS) must not only control current traffic, but also predict where congestion will occur. Predicting congestion so that preventive actions may be taken in advance will greatly alleviate traffic gridlocks. This paper describes the results achieved utilizing a backpropagation neural network algorithm to predict the traffic flow on surface streets in metropolitan areas. The neural network is trained in two phases. First, an initial learning phase determines the most appropriate connecting weights for data on a typical business day. Second, adaptive learning is employed to learn the special case traffic classes and adapt the weights to the present situation. In the adaptive learning phase, the error function is computed by placing a restriction on the weight changes so that the knowledge learned through the initial learning phase is retained. The prototype system is tested through computer simulations, with results indicating that the application of the neural networks to traffic congestion forecasting is promising.