Highway capacity has traditionally been treated as deterministic. In reality, however, capacity can vary from time to time and location to location due to inclement weather, various driving behavior, traffic incident, bottleneck and workzone. This thesis aims to characterize the stochastic variations of highway capacity, and explore its applications.
The stochastic variation of highway capacity is captured through a space-time Autoregressive Integrated Moving Average (STARIMA) model. It is identified following a Seasonal STARIMA model (0,0,2₃ )× (0,1,0)₂, which indicates that the capacities of adjacent locations are spatially-temporally correlated. Hourly capacity patterns further verify the stochastic nature of highway capacity. The goal of this research is to study (1) how to take advantage of second order information, such as capacity variation, and (2) what benefits can be gained from stochastic capacity modeling. The implication of stochastic capacity is investigated through a ramp metering case study. A mean-standard deviation formulation of capacity is proposed to achieve the trade-off between traffic operation efficiency and robustness. Following that, a modified stochastic capacity-constraint ZONE ramp metering scheme embedded cell transmission model (CTM) algorithm is introduced. The numerical experiment suggests that considering second order information would alleviate the bottleneck effect and improve throughput. Monte Carlo simulation further supports this argument. This study helps verify and characterize the stochastic nature of capacity, validates the benefits of using second order information, and thus enhances the necessity of implementing stochastic capacity in traffic operation.
In addition, this thesis presents a multiple days' highway capacity forecasting model and reveals the importance of incorporating capacity randomness separately in short term forecasting. Empirical observations of capacity with variations suggest that the existing deterministic notion of capacity (i.e., a fixed capacity value of 2400 veh/hr/ln) is inadequate to fully describe capacity fluctuations. The goal of this study is to provide insight on forecasting with inherent stochasticity. A hybrid highway capacity forecasting model based on wavelet dynamic neural time series is proposed. Based on 5-day capacity forecasting results, a 3.07% of Mean Absolute Percentage Error (MAPE) demonstrates improved performance against single models. This is achieved by its ability to capture the details and solve a nonlinear time series problem with a dynamic neural network. Following the results, the study concludes with a discussion on why the individual modeling of stochastic information influences the forecasting performance.