This paper presents the design of multilayer feedformard neural networks to predict freeway traffic conditions at a loop detector station. The neural networks make use of 30-second volume, occupanc and speed avaraged across all lanes in the past 2 intervals as imputs, and predict the same set of local parameters in the next 1 or 2 time intervals. Networks with various design and training parameters have been trained and evaluated with 2 weeks of morning data collected at I-880 Freeway in the San Francisco Bay Area. The results show that the neural nets have high accuracy in volume, occupancy and speed predictions during low, moderate and perhaps high volume conditions, including recurring congestion and possibly during incidents.
展开▼
