Accurate estimation of traffic volume is important for a variety of reasons, such as budgeting, traffic planning, speed enforcement, and roadway design. The raw data consists of axle counts, which come from either continuous-count, also known as automated traffic recorders (ATR) sites or weigh-in-motion (WIM) sites, or short-count locations, also known as portable traffic recorder (PTR) sites. There are relatively few ATR locations and most traffic counts come from shortcount locations. In order to convert axle counts from PTR locations into average annual daily traffic (AADT) estimates, MnDOT needs to calculate seasonal adjustment factors (SAFs) and axel correction factors (ACFs) for each PTR location. MnDOT’s current method for estimating SAFs consists of the following steps. MnDOT uses Ward’s clustering algorithm (Ward, 1963) to group ATR data into 12 clusters based on weekday 48-hour count period proportions of AADT (usually 12 in each month) for each ATR/WIM site for the months of year of the typical count season (April through October). Only a few (four or five) of the resulting clusters produce monthly adjustment factors that resemble historical cluster factors. ATR/WIM sites in other clusters are not used to produce adjustment factors except to adjust counts taken on the same road within a reasonable distance from a given site. Each cluster with factors resembling past factors sets and containing a similar roster of ATRs/WIMs is then labeled by its key characteristics, e.g., high weekends and high summer. Independent of this clustering, each PTR site is placed in a group, and each group is attached to either a single ATR or a cluster of ATRs based on professional judgment regarding the similarity of characteristics. Each PTR site in the group inherits the SAFs of the cluster to which this group is attached. Upon knowing SAFs for a PTR site, short-term axle counts are converted to AADT counts. University of Minnesota researchers approached the problem from a different angle and proposed an alternative methodology for categorizing the traffic patterns and calculating the seasonal adjustment factors (SAFs) for portable traffic recorder (PTR) sites. This methodology for estimating SAFs uses seasonal traffic volumes and includes an approach to quantify professional judgment. Researchers prepared the raw data provided by MnDOT using reasonable exclusion of abnormal data. Imputation procedures were then used to compile available data into a dataset with 39 ordered weeks. By analyzing this ATR data, we established a pattern identification method that could be used to group PTRs. A traffic pattern is defined by two components: the weekday traffic volume (referred to as “weekdays”), and the ratio of weekend traffic volume to weekday traffic volume (referred to as “weekend/weekday ratio”). Each component has three attributes. Specifically, for each ATR station, the weekday traffic volume can be categorized as average (A), high (H), or low (L) relative to the average traffic volume across three seasons at that station. Similarly, the weekend to weekday ratio can be categorized as either the same, or high, or low based on pre-determined thresholds. It is possible to categorize traffic into more categories, depending on the degree to which traffic is more or less than the average of three seasons. In this proof-of-concept exercise, we kept the number of categories small for clarity of exposition. According to the above-mentioned scheme, if a station’s traffic pattern is deemed AHA on weekdays, then that means the traffic volume in spring, summer and fall is, respectively, average, high, and average, relative to the average of weekday traffic across all three seasons at that station. Similarly, the SHS weekend traffic means that the weekend daily traffic relative to weekday daily traffic is the same in spring and fall, but high in summer. PTR stations that are located along routes taken by weekend recreational traffic in summer months may exhibit this seasonal pattern. Because data are not collected from PTR sites during winter, we do not consider the winter season in this analysis. That is, the seasonal pattern is categorized only for three seasons: spring, summer, and fall. Based on the results obtained, the most common seasonal traffic pattern is AHA weekday and LLL weekend/weekday, followed closely by AHA weekday and HHH weekend/weekday. The third common pattern is AAL weekday and LLL weekend/weekday. Sixteen to 19 stations each exhibit one of the top three patterns. The forth pattern, AHA weekday and SHS weekend/weekday has a total of five stations. The remaining patterns typically apply to only one or two stations. After identifying these patterns, we calculate the SAFs and the associated confidence intervals (if possible) for each pattern. The results produced by our methodology are not directly comparable to the AADT adjustment factors for short duration weekday traffic volume counts currently in use. In contrast with the existing method, the alternative methodology defines the traffic pattern at the seasonal level. There are over 20 patterns identified in this alternative methodology, whereas there are only five cluster groups used in the existing method. Furthermore, the high and low volumes are defined differently in the two methodologies. Segmenting at the seasonal level may result in more informative and accurate SAF estimates. A critical step in the methodology requires analysts to obtain professional judgment regarding the traffic patterns of PTR sites. Analysts must utilize this information to validate the pattern identification method proposed. To that end, researchers designed a survey tool to automate the process of collecting county engineers’ opinions regarding the traffic patterns of PTR sites. The survey is programmed within Excel and implements an analytic hierarchy process (AHP) methodology to analyze the survey participants’ responses. We emphasize that the survey does not collect opinions about the absolute volume of traffic at any given site. This is the case because we will use the volume information from sampled data (either 48 hours or a week) and extrapolate it using assigned SAFs to obtain the seasonal volumes and overall AADT. Thus, there is no need to ask the respondents about their estimate of total volume in each season. The survey focuses on the top five patterns identified from our ATR data analysis. These patterns cover 80% of the ATR stations. Each remaining pattern only has a small number of ATR stations. For those patterns, using SAFs estimated from ATR data may be too noisy. The survey contains a total of 10 questions. Each question asks the user to compare two traffic patterns at a time and indicate which pattern is more likely to represent the true traffic pattern at the selected PTR site. The survey is also programmed with several options for the user to review the pattern indicated by current answers, manually change the answer to the question that most-likely causes inconsistency in the respondent’s answers, or let the program automatically make the change to meet the consistency requirements. A separate ATR-data-based survey is also created for training and testing purposes. Two MnDOT participants have so far completed the training survey. However, due to the very limited amount of data, it is not possible to identify all of the potential problems that one might encounter if this survey tool were adopted for widespread use. Although there is no reason to believe that this approach will not be successful in the field, more testing will be beneficial. The researchers conjecture that future deployment of this method will benefit from the presentation of background materials and training sessions to county engineers. Finally, researchers carry out a simulation exercise that analyzes the sample-size requirements for desired estimation accuracy at short-count sites. Specifically, researchers propose a simulation methodology that samples and bootstraps continuous-count data to create data records as if they were collected from PTR sites. This approach is illustrated with three sets of attributes: (1) total volume by season, (2) weekend volume by season, and (3) heavy commercial volume by season. Three levels are defined for each attribute, namely, high, average, and low. A traffic pattern is a combination of levels for spring, summer, and fall. The most common pattern is AHA — average attribute level in spring and fall, and high in summer. For each sample size (in terms of weeks), we carry out a simulation with 200 iterations and track the correctness of the identified pattern. The percent of correct classification among the 200 iterations is calculated and treated as classification accuracy for that sample size. We aggregate the station level simulation results obtained and calculate the percentage of stations that show a classification accuracy rate of no less than 50% for each sample size. The results are obtained for all stations individually and subsequently for the three subsets of stations by seasonal traffic pattern. This analysis enables the identification of the minimum sample size that meets two thresholds for each attribute — the thresholds are the sample sizes needed for a minimum 50% of stations to reach a 50% accuracy rate. For most sites, accurate volume estimates can be obtained with three weeks of short-count data — one week for each season. We also carry out additional analysis to demonstrate the robustness of our findings. These findings also indicate that when developing annual short-count plans, MnDOT may choose to collect additional data during seasons that are not represented in the historical data. This does not appear to be the case at the present time. The simulation technique is used to simultaneously validate the professional judgment and to identify the traffic-volume pattern. This technique requires more data than what is currently available or collected. If this approach is adopted, future data collection may be spread over multiple years to avoid excessive effort in any given year of traffic count cycles. The work of this project also sets the stage for identifying which PTR sites are likely to benefit the most from more frequent data collection. These sites are also candidate sites for conversion to ATR systems. (Author/publisher)
Abstract