Developing Freeway Travel Time Learning Model under a Logical Traffic Context

摘要

Predicted travel time information is important to successful implementation of many IntelligentTransportation Systems. The artificial neural network (ANN) is one advance approach tofreeway travel time prediction. However, one of the setbacks of ANN models is that many ANNtopologies are developed following a "Black-box" approach, which is difficult to convey theinternal workings of these models to transportation practitioners. Additionally, although inputsettings are one important factor that influences the ANN model performance, there are only ahandful of studies focusing on the impacts of input information. In this study, we first providesome insights on how to logically use knowledge about typical traffic processes to make the"White-box" oriented development of a neural network topology. We then employ a reliableensemble technique to analyze the spatial and temporal effects of input variables on the ANNprediction performances for a study segment on US-290 in Houston. The results have shown thatspeed and occupancy data could be used by themselves or jointly to achieve satisfactoryperformance while traffic volume cannot; better performance can also be achieved by usinginputs from upstream, current and downstream segments, and/or using inputs from current andone or two time steps in the past. At last, we utilize the understandings learned above to developa new ANN topology, the so called time-delayed state-space neural network (TDSSNN). Bycomparing with other popular neural networks, the TDSSNN shows above-average predictionaccuracy and consistency. But more importantly, the model illustrates the possibility of buildinga White-box ANN model.