This report describes two related projects, the second of which builds on the first. The Iowa Department of Transportation (DOT) and Federal Highway Administration (FHWA) Aurora Transportation Pooled Fund TPF-5(290) partners sponsored both projects, and the Institute for Transportation Center’s Midwest Transportation Center at Iowa State University and U.S. Department of Transportation Office of the Assistant Secretary for Research and Technology co-sponsored Part II. Part I: Improving Estimations of Real-Time Traffic Speeds During Weather for Winter Performance Measurement: Winter weather in Iowa is often unpredictable and can have an adverse impact on traffic flow. The Iowa DOT attempts to lessen the impact of winter weather events on traffic speeds using various proactive maintenance operations. In order to assess the performance of these maintenance operations, a model was developed for estimating expected speed reductions based on weather variables and normal maintenance schedules. Such a model would allow the Iowa DOT to identify situations in which speed reductions are much greater or less than expected for a given set of storm conditions and modify maintenance operations to improve efficiency and effectiveness. For a previous Iowa DOT project, the authors of this report developed a sequential Bayesian dynamic model based on a non-stochastic model proposed in Iowa Highway Research Board Project TR-491. The sequential Bayesian dynamic model was designed to predict speed changes relative to baseline speeds under normal conditions. The model assumes nominal maintenance schedules and incorporates winter weather covariates (snow type, temperature, and wind speed) as measured by roadside weather stations. A limitation of this model is that it is not flexible enough to accommodate temporal heterogeneity. The speed change predictions are inconsistent across different locations and events, and measures of uncertainty in the model do not account for model uncertainty and are therefore overly optimistic. The objective of the work performed for the first part of this project was to improve that model to achieve real-time prediction of traffic speed changes with realistic uncertainty measures. Automated systems that record information on traffic speed and road conditions at particular points on a road system offer a way to improve the information available for building empirical models that relate weather variables to changes in traffic flow at a local level. For this application, we used data from two sources: road weather information systems (RWIS) and automated weather observing systems (AWOS). We further used maintenance crew reports to identify winter weather events throughout the year when reliable information on precipitation, such as snow intensity and type, was not available. Atmospheric variables, along with variables reflecting physical pavement condition, may contain a large number of complex interactions upon which realized changes in traffic speed depend, and these interactions vary with time. To accommodate these types of interactions and temporal dynamics, we developed an empirical adaptive stochastic model. Our approach made use of a Bayesian model formulation in which the effects of weather variables are allowed to adapt over four-hour time segments. Data from prior four-hour time segments provide a prior quantification of the effects that variables such as lane condition, temperature, and wind are expected to have on changes in traffic speed over the next four hours. Additional data in the next time segment are then used to adjust these quantifications to better reflect observed traffic speeds during that period. This model would allow, for example, the effects of wind speed to change somewhat (but not radically) over the course of a longer storm event to reflect the fact that low or high wind speeds have different effects on traffic speeds. The effect of this modeling approach is to circumvent the impossible task of explicitly determining a plethora of temporally dependent interactions. Under this approach, interactions are not explicitly identified and modeled. Rather, the problem is approached by allowing the main effects associated with key factors to undergo small shifts over time. The model also incorporates an autoregressive error structure to reflect temporal dependencies in observations that occur at a reasonably high frequency, such as every few minutes. Part II: Data-Driven Urban Traffic Prediction for Winter Performance Measurements: The model developed in Part I is for data from a single site, which is useful for predicting traffic speed changes in rural areas. In an urban setting, multiple sites collect traffic data from a network of roads, and the data are typically correlated in both time and space. Modeling data from multiple sites jointly can help in detecting abnormal traffic patterns earlier and in more accurately predicting traffic speed changes. The objective of Part II of this report was to use traffic data and limited weather information to develop models for detecting abnormal traffic patterns and predicting traffic speed and volume at any location on a network. We introduced several models for detecting abnormal traffic patterns based on two sources of traffic data, INRIX and Wavetronix, and limited weather information. Because the INRIX data included more locations than the Wavetronix data and the corresponding observations were selfconsistent, we used the INRIX data to detect abnormal traffic areas. We developed a method to estimate the multivariate quantiles for the INRIX observations, and the INRIX data were compared with the estimated quantiles to identify abnormal traffic patterns in both space and time. An online interactive app was developed to visualize the results and help the Iowa DOT make informed decisions about winter weather maintenance. A dynamic Bayesian model was implemented at two Wavetronix sensor locations where weather information was available, with the corresponding median curve as the baseline. The fitting results were satisfactory. Furthermore, we also explored the spatial structure of the traffic data using the INRIX dataset and used curve Kriging to predict traffic speed and volume at any location. The prediction method was tested at the Wavetronix locations and was found to work well. (Author/publisher)
Samenvatting