CBI Paper for ISEHP 2016
- 1. doi:10.1016/j.trpro.2016.06.041 Proposed Data-Driven Performance Measures for Comparing and Ranking Traffic Bottlenecks David K. Hale1 , Ali Hajbabaie2 , Jiaqi Ma3 , Jia Hu4 , Hyoshin Park5 , and Joe Bared6 1 Leidos Inc., Reston, U.S. 2 Washington State University, Pullman, U.S. 3 Leidos Inc., Reston, U.S. 4 Federal Highway Administration, McLean, U.S. 5 University of Maryland, College Park, U.S. 6 Federal Highway Administration, McLean, U.S. [email protected], [email protected], [email protected], [email protected], [email protected], [email protected] Abstract To justify investments towards improved traffic operations, engineers and policy-makers need scientific and accurate methods of congestion measurement. However, status-quo methods are limited and/or outdated. Peak-hour analyses are becoming outdated as a sole source of traffic assessment, because they fail to account for changing conditions throughout the year. There has been a movement towards “reliability” modeling, which attempts to capture these annual effects. But due to significant input data and calibration requirements, the reliability models suffer from practicality issues. Next, there have been recent improvements in data-driven ITS technologies, which identify congestion in real time. However, there is room for improvement in the robustness of performance measures derived from these technologies. Finally, some engineers have compared and ranked congested locations (i.e., bottlenecks) on the basis of experience and judgment. Despite their cost-effectiveness, judgment-based qualitative assessments will lack credibility unless backed by quantitative results. In a recent Federal Highway Administration study, congestion measurement was a primary area of emphasis. This paper discusses project-specific software development, which produced new and innovative performance measures for congestion measurement. It will present concepts and evidence to imply superiority of the proposed new measures. This paper is intended to serve as a preview of a future full journal paper; which will rank ten or more real-world bottlenecks according to new and old performance measures, to demonstrate impacts of the new measures. It is hoped that the new performance measures will be adopted by states and/or commercial products, for a new level of robustness in congestion measurement. Keywords: performance measures, traffic bottlenecks, INRIX, reliability Transportation Research Procedia Volume 15, 2016, Pages 483–494 ISEHP 2016. International Symposium on Enhancing Highway Performance Selection and peer-review under responsibility of the Scientific Programme Committee of ISEHP 2016 c The Authors. Published by Elsevier B.V. 483
- 2. 1 Introduction Precise congestion measurement is one of the best ways that traffic engineers can demonstrate a solid return on investment, in a time of tight budgets. It can also be an important first step towards identifying and ranking congested locations (i.e., bottlenecks) (Elhenawy, Chen, & Rakha, 2015). However improved methods of congestion measurement are needed, because the status-quo methods are either limited or outdated. Congestion measurement should be more cost-effective, precise, and scientific. When considering the status-quo methods, it is possible to associate each one of them with one or more fundamental disadvantages. Back in the 1990s, most traffic engineers were modeling the peak 15-minute period, and using the peak hour factor. However this paradigm did not take into account the significant factors that come into play throughout the year; such as seasonal demand volume fluctuation, weather, and incidents. These days, improved procedures are available for analyzing the entire year; but they are extremely data-intensive, which makes them difficult to use. Improved measurement technologies deployed by companies such as INRIX are now available. However performance measures derived from these new data sources have not adequately accounted for variability, reliability, or throughput. This paper will describe some recent improvements in congestion identification, which leverage the latest advances in both research and ITS technology. Some drawbacks of the status-quo methods are briefly summarized below. In the 1990s, peak 15-minute modeling was the norm, and was recommended by the Highway Capacity Manual (Transportation Research Board, 2010). The peak hour factor, used to convert hourly flows into peak 15-minute flows, was a required concept for transportation engineers and software products. In micro-simulation analysis, peak 60-minute analyses were also popular. However regardless of the time period duration being used, peak period modeling did not take into account significant factors that come into play throughout the year; such as seasonal demand volume fluctuation, weather, and incidents. Instead, it was necessary to analyze and design for the peak 15- minute period, and hope that design would work well throughout the year. Without the data-driven technologies available today, it was difficult for the 1990s engineer to independently verify how well their design was working throughout the year. Modeling frameworks are now available for analyzing the entire year, and reflecting some of the elements that were not recognized by peak period modeling. The drawbacks of today’s reliability models include a relatively extensive input data requirement, which increases their risk for engineers who would use them. The time, money, and expertise needed to fulfill the input data requirement essentially increase the probability of late deliverables, funding shortages, and human errors. Moreover, when micro-simulation is used to analyze scenarios occurring throughout the year, excessive computer run times become a problem, further reducing the practicality of reliability modeling. Finally, popular reliability model performance measures (e.g., Travel Time Index) assume inflexible comparisons to the free-flow speed, and do not directly account for traffic volume levels. In contrast to the 1990s, it is now common for traffic measurement data sources (e.g. INRIX, HERE, PeMS, TomTom) to be constantly collecting data in the field. Some organizations have used these rich sources of data to develop bottleneck rankings based on traffic measurements, as opposed to traffic models. However, this paper will attempt to show that some bottleneck rankings in the industry are relatively insensitive to traffic volume levels, congestion variability, and travel time reliability. Through extensive personal observation, traffic engineers can develop a sense for specific areas containing the worst traffic problems. However without precise annual measurements, it is more difficult to communicate the duration and extent of traffic problems to decision-makers. When operational improvements are made, it is more difficult to verify the beneficial impact of these improvements. This paper will propose and describe new, data-driven performance measures, designed to overcome limitations of the status-quo methods described above. Data-Driven Performance Measures for Comparing and Ranking Traffic Bottlenecks D. Hale et al. 484
- 3. 2 Background Data-driven methods now facilitate a more precise quantification of congestion and bottlenecks. The spatiotemporal traffic state matrix (STM) (Elhenawy, Chen, & Rakha, 2015) shown in Figure 1 can be considered a step in the direction of “big data”, compared to traditional traffic engineering analyses. The STM facilitates monthly and annual analyses, instead of just peak-hour analyses, providing a more comprehensive picture of traffic problems. The STM makes it possible to put a more precise price tag on transportation investments. Bottlenecks can now be quantified according to D.I.V.E. – Duration, Intensity, Variability, and Extent. The STM is a fundamental pre-requisite to generating these measures for congestion identification. Various types of measured data (e.g. INRIX probe data, loop detector data) could be used to generate STMs. STMs can also be generated by traffic models like micro-simulation, or the Highway Capacity Manual (HCM) procedures (Transportation Research Board, 2010). In the case of HCM modeled data, traditional capacity analysis generates results for a single STM cell. However a recent HCM advancement is to generate STM results throughout the analysis year(s). The first step of the HCM procedure is to intelligently define a “reliability analysis box”, to capture the congestion both spatially and temporally. Figure 1 shows that two-dimensional heat maps can be expanded into a third “reliability dimension”, by modeling a significant number of days. Red cells represent congested conditions, yellow boxes denote “at- capacity” conditions, and green cells indicate the uncongested regime. Regardless of whether derived from measurements (like INRIX) or models (like the HCM), the STM is an important first step towards a more precise identification of congestion and bottlenecks. Figure 1: HCM Reliability Modeling Concepts (Transportation Research Board, 2010) The upcoming Methodology section will describe recent improvements in congestion identification; which leverage the latest advances in transportation research, and in ITS technology. Recent research produced an INRIX-based software tool for congestion and bottleneck identification (CBI), which displays a typical STM. One of the main performance measures from the CBI tool is called “bottleneck intensity”. For example, if 15% of the Figure 1 analysis box were colored red, intensity would be 15% for than particular day. The red area could represent vehicles traveling below a “cutoff speed” (e.g., 45 mi/h). Alternatively, other performance measures (e.g. travel time, percentage of stopped vehicles, density) could be used to identify the congested sections. When comparing and ranking bottlenecks, it is more practical to compare speeds, instead of travel times or other measures. To compare travel times, it would be necessary to define cutoff travel times for each bottleneck, and these travel times would be affected by numerous factors. At any rate, obtaining the bottleneck intensity is a good first step in comparing and ranking bottlenecks. However it is not sufficient, because it does not take into account variability or reliability throughout the year. The Data-Driven Performance Measures for Comparing and Ranking Traffic Bottlenecks D. Hale et al. 485
- 4. Regional Integrated Transportation Information System (RITIS) website* has rich functionality for congestion identification. It is capable of generating STMs for a number of different performance measures. It has a “slider” feature to easily adjust thresholds for different colors in the “heat map”, and a second slider for specifying which hours of the day will be plotted. In summary, recent advancements in data-driven technology and reliability-based research are now making it possible to achieve a more precise quantification of congestion and bottlenecks. However, these technologies and concepts are relatively new to the transportation industry. Today’s reliability models have extensive input data requirements. In some cases, excessive computer run times are a problem. Popular reliability measures like the Travel Time Index do not allow custom cutoff speeds, and are insufficiently sensitive to throughput. The upcoming Methodology and Case Study sections will give a better understanding of these limitations, and will demonstrate new ideas for overcoming them. 3 Methodology Before the innovative performance measures were developed, the CBI tool was intended to rank traffic bottlenecks according to conventional industry performance measures. The RITIS website currently provides bottleneck rankings based on an “impact factor”. The RITIS formula for computing impact factors is shown below as Equation 1. Assuming consistent performance measures and cutoff thresholds (e.g., 45 mi/h) for differentiating between congested and uncongested conditions, the impact factor computed by this formula is essentially equivalent to the bottleneck intensity discussed in the prior section, aggregated for all days of the analysis period (e.g., one year). The RITIS impact factor is a reasonable method for comparing and ranking traffic bottleneck locations, because it simultaneously accounts for the average duration and extent of congestion throughout the year. However, it does not account for the instability of congestion throughout the year. In statistics, instability is often quantified through the variance and/or standard deviation. In reliability modeling, which is a relatively new concept in transportation operations, instability is often quantified through the Travel Time Index (Transportation Research Board, 2010). A problematic trend in many countries is the decreasing reliability of surface transportation. As stated by Texas Transportation Institute researcher Bill Eisele, when our transportation facilities become less reliable, “we have to allow increasingly more time to ensure an on-time arrival” † . (1) Where: IF = impact factor for ranking bottlenecks (dimensionless) D = duration of congestion (minutes) L = length of congestion (miles) N = number of days in the analysis period (e.g, one year) 3.1 Innovation #1: Annual Reliability Matrix In order to integrate the reliability concept into the comparison and ranking of traffic bottlenecks, the annual reliability matrix (ARM) was developed. The ability to compute bottleneck intensity for every relevant day of the year, as illustrated earlier in Figure 1, is a pre-requisite to the ARM. Suppose one wished to compare and rank a group of known bottleneck locations throughout the year 2014, excluding weekends and Fridays. In this case, there would be approximately 210 relevant days to consider. Likewise, there would be approximately 210 bottleneck intensity values to consider. These * https://www.ritis.org/ † http://mobility.tamu.edu/ums/media-information/press-release/ Data-Driven Performance Measures for Comparing and Ranking Traffic Bottlenecks D. Hale et al. 486
- 5. intensity values may range between 0% and 100%. Thus imagine a deck of 210 cards, where each card would contain a value between 0 and 100. These 210 cards could then be sorted in a lowest-to-highest sequence, in which each new card would contain a higher value than the previous card. Now imagine all of the 210 card values being plotted left-to-right onto a two-dimensional, X-Y graph. The Y-axis could represent the bottleneck intensity value, which ranges between 0% and 100%. The X-axis could then represent 210 relevant days of the year; with the far left containing the best day of the year (i.e., lowest intensity), and the far right containing the worst day of the year (i.e., highest intensity). Once the annual values have been plotted in this manner, it becomes possible to visualize traffic bottlenecks in terms of both intensity and reliability. The upper half of Figure 2 compares the STM to the ARM. Whereas the STM displays a bottleneck intensity for one particular day, the ARM shown below can display dozens or hundreds of bottleneck intensities throughout the year, or across multiple years. By displaying and storing a distribution of bottleneck intensities in this manner, the ARM facilitates visualization and/or computation of useful performance measures. For example, the amount of red area “under the curve” in an ARM is essentially equivalent to the impact factor from RITIS. However the 85th percentile day of the year can now be visualized fairly easily as having an intensity of approximately 50%. The 85th percentile intensity would be a reasonable way to compare and rank bottlenecks, because it reflects some of the most congested days of the year. By contrast the 100th percentile intensity might reflect a fluke accident, and the 50th percentile intensity might not reveal how congested the worst days are. As stated earlier, the ARM was developed to integrate reliability into the comparison and ranking of traffic bottlenecks. To demonstrate this, the lower half of Figure 2 illustrates two ARM diagrams, for two known bottleneck locations. Each ARM diagram has approximately the same amount of red area. According to the RITIS bottleneck ranking formula, both bottlenecks would produce approximately the same impact factor, and would thus be considered approximately equal priority. However the ARM is intended to show that for bottleneck #2 on the right, roughly one-third of its days exhibit more delay than will ever be experienced at bottleneck #1. This means that a driver of bottleneck #2 will need to allow more time to ensure an on-time arrival, in comparison to bottleneck #1. Bottleneck #2 is thus less reliable, and should be considered a higher priority for mitigation efforts. Figure 2: Illustrations of ARM and STM Concepts Data-Driven Performance Measures for Comparing and Ranking Traffic Bottlenecks D. Hale et al. 487
- 6. The ability to generate both STMs and ARMs was implemented within the CBI software tool. Because Intensity is the product of Duration and Extent, the ARM simultaneously accounts for all four standard D.I.V.E. performance measures (i.e., Duration Intensity Variability Extent). One limitation that was noticed in the ARM concept was the inability to quantify “speed drops”. For example if two bottlenecks exhibit a similar ARM shape, it is quite possible that one of the two bottlenecks may in fact be much more serious than the other one, because their vehicle speeds in the congested red region might be much lower. The inability to quantify speed drop is apparently one of the fundamental disadvantages of bottleneck intensity, as a performance measure. The first step taken to address this disadvantage was to compute daily and annual average speed drops in the CBI tool, and to display these speed drop values as numeric performance measures. In this manner, the speed drop values could conceivably be used as a “tiebreaker” when comparing two bottlenecks with similar ARMs. 3.2 Innovation #2: Integration of the California Method In the year 2015, a series of workshops was conducted across the U.S., to share FHWA research on congestion and bottlenecks. At the Oakland California workshop, local participants shared a method for evaluating bottlenecks on the basis of vehicle-hours of delay. One of the advantages of this method was the ability for bottleneck rankings to explicitly reflect traffic volume demands. For example, a four-lane freeway would tend to exhibit twice as much delay as a two-lane freeway, and would become a doubly higher priority for mitigation. Indeed, the RITIS impact factor and intensity-based ARM would give similar priority ratings to a four-lane freeway and a two-lane freeway, as long as they had similar traffic density levels. This provided an incentive for integrating this “California method” into the CBI tool. Vehicle delay is computed according to the 2013 Most Congested Freeways Report and Methodology (Iteris, 2014). Equations (2) and (3) illustrate California’s use of a 35 mi/h cutoff speed. When integrating this method into the CBI tool, instead of assuming a fixed value of 35 mi/h, cutoff speeds were obtained from a flexible set of CBI software features. (2) (3) The California method also required specification of a bottleneck volume. According to the Iteris report, bottleneck volumes should be measured immediately downstream of the downstream end of congestion. The CBI tool was updated to allow users to enter a bottleneck volume. In addition, it was upgraded to display vehicle-hours of delay as a numeric performance measure. Vehicle delay is believed to be an important performance measure for comparing and ranking bottlenecks, because it explicitly captures the effects of both speed drops and demand volumes. Figure 3 illustrates an ARM based on vehicle-hours of delay. This is in contrast to the previous intensity-based ARMs, in Figure 2. Figure 3: Example ARM Based on Vehicle-Hours of Delay, as Shown in the CBI Tool Data-Driven Performance Measures for Comparing and Ranking Traffic Bottlenecks D. Hale et al. 488
- 7. 3.3 Innovation #3: Bottleneck Intensity Index Although the shape and size of delay-based ARMs are useful for simultaneously visualizing bottleneck intensity and variability, respectively, it was believed that a numerical performance measure was needed, to quantify the ARM into a single number. This is because matrices and graphs allow room for human interpretation, or might be confusing to some engineers. By contrast, a single number could conceivably avoid interpretation errors. In other words, the highest congestion number could be much more easily interpreted as the worst number. In response, an ARM-based performance measure called the Bottleneck Intensity Index (BII) was conceived for implementation within the CBI tool. This value represents a delay level below which 85% of the ARM’s red area exists. The BII is an attempt to capture size and shape of the ARM into a single number. As shown in the previous Figure 3, the BII can be superimposed as a horizontal black line across the ARM. If the “Hotspots” checkbox is turned off, this horizontal line will be hidden. Also in Figure 3, the BII is computed as 824 vehicle- hours, because 85% of the annual red area falls below the 824 vehicle-hour mark. Other numeric performance measures on the screen are daily values as opposed to annual values. The 85th percentile BII is believed to be more effective than 85th percentile delay, for comparing bottlenecks, because 85th percentile delay conveys less information about delays that occur throughout the year. This relationship of 85th percentile BII to 85th percentile delay is similar to the relationship of mean to median. The 85th percentile BII explicitly reflects 85% of the red area. The 85th percentile delay simply says that 85% of the days had a lower delay. The 85th percentile BII is believed to be more effective than the 100th percentile BII, because the 100th percentile BII only reveals which bottleneck experiences the highest-delay day of the year, which may have been caused by a fluke accident or event. The 85th percentile BII is believed to be more effective than the 50th percentile BII, because the 50th percentile BII focuses on 50% of the best days of the year. The 85th percentile level provides a reasonable number for focusing on some of the worst days of the year, without over-emphasizing a small number of absolute worst days. Because of this, the BII is always an 85th percentile value, regardless of what is selected in the CBI tool’s “Centile” control. By contrast, standard D.I.V.E. performance measures all use the chosen percentile from the “Centile” control. An alternative concept to the BII is the “reliability ratio” (Li, Hensher, & Rose, 2010), defined as marginal rate of substitution between expected travel time and travel time reliability. The reliability ratio appears to be another promising method of simultaneously accounting for intensity and variability. Because the BII method was developed before learning of the reliability ratio, a comprehensive comparison between the two has not been performed yet. This could be a topic of future research, as the BII and reliability ratio could both be made available as performance measures from the same tool. Finally, the CBI tool encourages intelligent specification of cutoff speeds, to consider many factors. Popular reliability performance measures like the Travel Time Index assume inflexible comparisons to the free-flow speed, but the perception of congested conditions depends on more than free-flow speed. As pointed out by (Elhenawy, Chen, & Rakha, 2015), drivers affected by severe weather and/or limited visibility conditions may not consider themselves delayed at speeds well below the posted limit. In some areas, factors like lane width, lateral clearance, and pavement quality may not be properly reflected in the posted speed limit. In some areas, cutoff speeds should vary significantly on different segments within the same corridor. The important point is that engineering judgment must be used in properly defining cutoff speeds on a case-by-case basis. Without this judgment, a roadway exhibiting low travel speeds due to severe weather could mistakenly be identified as having the same congestion as a corridor having nearly identical speeds, but mostly unaffected by weather. Once the cutoff speeds have been properly chosen for each corridor, a fair comparison of known bottleneck locations can take place. As a pre-requisite to a fair comparison, a consistent analysis box should be chosen for each bottleneck, in terms of both time and distance. For example, each bottleneck could be analyzed for the same number of miles, months of the year, days of the week, hours of the day, and interval duration. Most testing of the CBI tool was performed under 5- Data-Driven Performance Measures for Comparing and Ranking Traffic Bottlenecks D. Hale et al. 489
- 8. minute interval durations. Larger interval durations of 15 minutes or 30 minutes could be used, but this might compromise accuracy of the results. Smaller interval durations of 1 minute could be used, but this might cause the software to become slow in performance. Once a consistent analysis box and interval duration have been chosen, direct comparisons can occur. Figure 4 compares three real-world bottleneck locations along three different freeway corridors. In this example, the I-695 bottleneck exhibits the worst score, with a BII of 3840 vehicle-hours. It is possible for multiple bottlenecks to be compared along the same corridor, to see which part of the corridor should be mitigated first. It is also possible to evaluate known bottleneck locations in this manner, soon after mitigation strategies have been implemented, to assess possible improvements in a precise and scientific manner. Figure 4: Comparison of Bottlenecks Based on Bottleneck Intensity Index (BII) 3.4 Innovation #4: Wavelet Filtering Method It is also possible to evaluate surface arterial corridors in this manner. Some researchers have expressed concern over the application of this type of congestion identification, along signalized arterials. To be sure, the software tools can easily apply the same models in exactly the same way on both arterials and freeways. However, along surface arterials, traffic signals consistently produce speed drops and delays, even under light congestion. Although this is necessary for safety reasons, it makes congestion and bottleneck measurement more difficult. Referring to the STMs in Figure 2, repetitive traffic signal delays cause “spotty patterns” instead of contiguous red shapes, making it more difficult to identify or measure existing bottlenecks. Another concern is that speeds and travel times are measured less accurately on signalized arterials (compared to freeways), although the ITS technologies used on arterials are expected to improve over time. In response to concerns about the spotty congestion patterns caused by traffic signals, FHWA researchers recommended the wavelet concept (Zheng, Ahn, Chen, & Laval, 2011) for filtering out portions of delay that appear to be unrelated to congestion. Statistical implementations of the wavelet filtering method are available through the Matlab computing environment. Because the CBI tool is not easily integrated with Matlab, the first attempt at incorporating wavelet functionality was a simplified version (i.e., “pseudo-wavelet”). More recently the explicit, statistical wavelet method was implemented in the CBI tool without requiring Matlab. This paper will describe the logic of the pseudo-wavelet method, whereas follow-on studies will hopefully demonstrate impacts of the “pure” method. The pseudo-wavelet method assumed that delays on the 0th percentile day were not caused by congestion. Indeed, when performing an annual reliability analysis, the 0th percentile day is often a holiday with no work trips. On signalized arterials, low speeds observed on the 0th percentile day are believed to be caused by signal timings and intersection safety requirements. The CBI tool allows the user to easily activate and deactivate (pseudo-)wavelet filtering. Bottleneck comparisons and rankings can thus be made with and without the unavoidable delays. It is certainly conceivable that bottleneck rankings, in some cases, might be unaffected by the wavelet filtering. However in cases where wavelet filtering actually changes the bottleneck rankings, an engineer at this point would need to make an assessment: What is causing the so-called unavoidable delay? Is it truly unavoidable? If the signal timings and intersection safety requirements could be Data-Driven Performance Measures for Comparing and Ranking Traffic Bottlenecks D. Hale et al. 490
- 9. modified in such a way as to reduce the unavoidable delay, then perhaps the non-wavelet bottleneck rankings should be used. Conversely if wavelet-filtered delay is deemed unavoidable, then presumably the filtered rankings should be used. Figure 5 illustrates the simple methodology. The example involves PM peak analysis of a signalized arterial with six TMC segments. Although the CBI tool is capable of providing unique cutoff speeds on each TMC segment, a flat 25 mi/h cutoff speed is applied to all segments in this example. As discussed earlier, the first step is to examine the 0th percentile day, on which traffic congestion is assumed to not exist. Any low speeds on this day are assumed to be caused by signal timings, intersection safety requirements, or other unavoidable factors. On segment 1, the highest observed speed is 30 mi/h. Because vehicles are clearly capable of traveling above 25 mi/h, no adjustment to the cutoff speed is required on this segment. On segment 2 the highest observed speed is 27 mi/h, so no cutoff speed adjustment is required on this segment. However on segment 3, the highest observed speed is 24 mi/h. Because this is the 0th percentile day, this implies that vehicles (on average) are never capable of traveling above 24 mi/h on this segment. The cutoff speed is thus lowered to 24 mi/h on segment 3, and the 24 mi/h cutoff speed is then applied on segment 3 for each day of the year. Similar logic is applied to other segments, as shown in Figure 5. Figure 5: Sample Illustration of Pseudo-Wavelet Cutoff Speed Logic Figure 6 illustrates a pair of ARMs for the US-13 signalized arterial in Pennsylvania. The diagram on the left exhibits 14% lower delays throughout the year. The diagram on the left illustrates what would happen when unavoidable delays are filtered out by the pseudo-wavelet method. As stated earlier, bottleneck comparisons and rankings could be made with and without wavelet filtering. If wavelet filtering actually changed the rankings, the engineer would need to assess whether or not the filtered delays were truly unavoidable, before deciding on which set of rankings to use. In the HCM context, wavelet filtering may offer a more pure assessment of Level of Service (LOS) thresholds. For example, without wavelet filtering, certain urban street segments may never achieve LOS better than ‘C’, because so much of the delay is unavoidable (without significant infrastructure changes). In such a case, the typical HCM threshold for LOS C may be re-classified as an LOS A threshold for certain analyses, in terms of what is truly possible from low-cost operational improvements. Regardless of LOS thresholds, conducting an analysis with and without wavelet filtering would presumably provide the engineer with more information than they would have otherwise had. Figure 6: Sample ARM Diagrams With and Without Pseudo-Wavelet Filtering Data-Driven Performance Measures for Comparing and Ranking Traffic Bottlenecks D. Hale et al. 491
- 10. 4 Case Study In a future, full journal paper, the tentative plan is to rank ten or more real-world bottlenecks according to both new and old (conventional HCM) performance measures. It is expected that the two sets of measures and methods will produce unequal rankings, and that these discrepancies will make it possible to demonstrate improved insights provided by the new measures. For now, this paper will provide a preliminary case study analysis, of eastbound Interstate 40 near Raleigh NC. More specifically, this case study will analyze and quantify congestion for the times and locations listed below. To obtain and demonstrate the “new” performance measures, the CBI tool was used. An INRIX dataset for I-40 near Raleigh was downloaded from the RITIS website, and subsequently loaded into the CBI tool. As shown in Figure 7, most congestion on a typical day occurs downstream, near mile marker 291. I-40 freeway between mile markers 278 and 291 between 2-8 PM each day on weekdays only (no weekend days included) throughout the year 2010 Figure 7: CBI Performance Measures for the 85th Percentile Worst Day (I-40 Case Study) In terms of vehicle delay, the 85th percentile worst day of the year was April 16th, which experienced 3776 vehicle-hours of delay. To obtain this delay estimate, it was first necessary to specify a cutoff speed and a bottleneck volume. According to the model developed by (Elhenawy, Chen, & Rakha, 2015), the cutoff speed is 72% of the free-flow speed under clear weather conditions. Therefore, this case study applies cutoff speeds equal to 72% of segment-specific INRIX free-flow speeds throughout the corridor. If this stretch of roadway were consistently affected by weather and/or visibility issues, it would be appropriate to further reduce cutoff speeds at this stage. Regarding the assumption of bottleneck volume, the number of mainline lanes near mile marker 291 (three) was multiplied by a typical freeway capacity of 2100 veh/ln/h (Transportation Research Board, 2010). This produced a bottleneck volume of 6300 veh/h. According to the proposed method of ranking traffic bottlenecks, this stretch of roadway would have a bottleneck intensity index (BII) of 3599 vehicle- hours, because 85% of red area on the annual reliability matrix (ARM) falls below this line. This ARM for I-40 is illustrated in Figure 8. Until a significant number of roadways can be ranked, the BII cannot be easily compared to the HCM’s Travel Time Indices. However compared to the TTI, BII is more explicitly sensitive to cutoff speeds, throughput (and number of lanes), and speed drop. On signalized arterials, BII is also capable of filtering out unavoidable delay, via the wavelet method. To obtain and demonstrate the “old” (i.e., conventional HCM) performance measures, current HCM reliability analysis methods were used. These methods are currently implemented in a Java version of the Freeval software. Freeval results for I-40 are illustrated in Figure 9. Although the CBI and Freeval tools both provide several numeric performance measures, the measures themselves are Data-Driven Performance Measures for Comparing and Ranking Traffic Bottlenecks D. Hale et al. 492
- 11. fundamentally different. Whereas the CBI reports on bottleneck intensity, variability, extent, duration, speed drop, and delay, Freeval reports the reliability rating, travel time index, misery index, and planning time index. A case study of one corridor cannot establish the superiority of new performance measures versus old performance measures. However, it does set the stage for a future comparison, involving rankings of a significant number of corridors. Given the different ranking orders produced by both methods, it should become possible to demonstrate improved robustness under the new measures, with specific numeric examples. Figure 8: CBI Annual Reliability Matrix (I-40 Case Study) Figure 9: HCM-Based Reliability Results from the Freeval Software (I-40 Case Study) 5 Conclusions In a recent Federal Highway Administration study, congestion measurement was a primary area of emphasis. The software development effort produced a congestion and bottleneck identification (CBI) tool containing both numeric and graphical performance measures. The CBI tool was designed to compare and rank traffic bottlenecks in more detail than existing methods. The initial version of the CBI tool emphasized annual bottleneck intensities, similar to the RITIS website. When the ARM concept was implemented, this integrated reliability (variability) into the comparison and ranking of traffic bottlenecks. However, it was noticed that the ARM concept was unable to quantify “speed drops”, so the overall bottleneck rankings could not be fully trusted. Next, this disadvantage was addressed by computing daily and annual average speed drops in the CBI tool, and displaying these values as numeric performance measures. In this manner, the speed drop values could conceivably be Data-Driven Performance Measures for Comparing and Ranking Traffic Bottlenecks D. Hale et al. 493
- 12. used as a “tiebreaker” when comparing two bottlenecks with similar ARMs. However at the Oakland workshop, local participants recommended vehicle-hours of delay to explicitly reflect traffic volume demands, such that four-lane freeways would nominally become doubly higher priorities for mitigation than two-lane freeways. When the California method was integrated into the CBI tool, this allowed bottleneck rankings to explicitly reflect throughput impacts. Finally, along signalized arterials, some portions of delay might be completely unrelated to congestion. In response, the wavelet concept was incorporated, for filtering out unavoidable portions of delay. Table 1 summarizes the evolution of the CBI tool, in terms of the new elements that can now affect bottleneck rankings. Unaddressed Issue Innovation Expected Benefit Cutoff Speed CBI Interface bottleneck measurement Variability ARM illustrates reliability Numeric Index BII quantifies the ARM Throughput California prioritizes congested roads Signal Delay Wavelet reveals unavoidable delay Table 1: Summary of Performance Measure Innovations for Annual Reliability Modeling These proposed methods provide a new level of robustness in identifying congestion and bottlenecks. They provide an important first step towards prioritizing problem areas, and selecting countermeasures. Unlike subjective assessments of congestion that rely on experience and judgment, these automated methods offer scientific justifications for transportation investments. Unlike prior research, these methods reflect impacts of elements not reflected by posted speed limits, such as weather and visibility. Once the bottleneck locations have been precisely identified and prioritized, an assessment of mitigation strategies can then begin. In cases where mitigation strategies have recently been implemented, the proposed methods would have the ability to verify mitigation strategy impacts with more precision and robustness, compared to status-quo methods. Future research should rank real-world bottlenecks according to new and old performance measures, to demonstrate impacts of the new measures. It is hoped that the new performance measures will be adopted by states and/or commercial products, for a new level of robustness in congestion identification. In addition, the proposed methods could be incorporated into future editions of the HCM. References Elhenawy, M., Chen, H., & Rakha, H. (2015). Traffic Congestion Identification Considering Weather and Visibility Conditions Using Mixture Linear Regression. Transportation Research Board 94th Annual Meeting. Iteris. (2014). 2013 Most Congested Freeways Report and Methodology. San Francisco Bay Area: Metropolitan Transportation Commission. Li, Z., Hensher, D. A., & Rose, J. M. (2010). Willingness to pay for travel time reliability in passenger transport. Transportation research part E: logistics and transportation review, 384-403. Transportation Research Board. (2010). Highway Capacity Manual. Washington, DC. Zheng, Z., Ahn, S., Chen, D., & Laval, J. (2011). Applications of wavelet transform for analysis of freeway traffic: Bottlenecks, transient traffic, and traffic oscillations. Transportation Research Part B, 372-384. Data-Driven Performance Measures for Comparing and Ranking Traffic Bottlenecks D. Hale et al. 494