| description abstract | Capacity is an important traffic parameter, extending nonnegligible influence on road network planning, traffic management, and traffic state prediction. The stochasticity of capacity is widely accepted considering the stochastic nature of traffic flow. Previous studies studied stochastic capacity based on long-term observations, lasting for months or even years, at one single site. On the other hand, data-driven methods were applied by researchers to evaluate the impacts of external factors on capacity, in which, however, capacity was always viewed as deterministic. To fully exert the advantages of data-driven methods, this paper proposes a Weibull-distribution-based neural network for capacity estimation on freeways, considering both stochastic nature and external factors. Extremely long-term observation at one single site is no longer essential because this method considers different scenes at the same time and is able to integrate the information automatically. Furthermore, the model has a certain generalization performance. No matter which influencing factor is adjusted, a new distribution can be obtained. The model is verified by open-source data from the California Department of Transportation Performance Measurement System (PeMS) in this paper. Eight easily-fetched explanatory variables are introduced into the model. The mean absolute percentage error between predicted median capacities and observed ones is 0.29 and 70%–80% of observed median capacities are within the prediction band. | |
