Comparative evaluation of adaptive and neural-network exit demand prediction for freeway control

Reliable on-line predictors that can accurately predict freeway demand in real time are of critical importance in developing optimal control systems for freeway corridors. New freeway exit demand predictors have been developed using two prediction approaches: model-based adaptive-parameter and backpropagation neural network-based prediction. The adaptive-parameter predictor requires prespecified models with parameters determined on line using the Kalman filter. Two such models are formulated. The first model is developed for normal weekdays and requires both historical and current-day measurements. The second model is designed for situations in which no historical information is available. Neural network-based prediction does not require a prespecified functional form that relates traffic measurements to predicted flow. However, an appropriate network structure and training method need to be determined before the network is trained. A three-layer backpropagation neural network was trained with the same data that are used to determine the historical pattern for the adaptive-parameter predictor. The new predictors were tested with real data from the I-35W freeway during a 2-week period and their performance was compared with that of the urban traffic control system (UTCS)-2 predictor. The error indexes from the two new predictors are very close and substantially better than those from UTCS-2 under the same conditions.