INTEGRATING AN AUTOMATED BOTTLENECK DETECTION TOOL ...

INTEGRATING AN AUTOMATED BOTTLENECK DETECTION TOOL INTO AN ONLINE FREEWAY DATA ARCHIVE

Jerzy Wieczorek Department of Mathematics and Statistics Portland State University P.O. Box 751 Portland, OR 97207-0751 Phone: 503-725-4285 Fax: 503-725-5950 Email: [email protected] Huan Li Department of Civil and Environmental Engineering Portland State University P.O. Box 751 Portland, OR 97207-0751 Phone: 503-725-4285 Fax: 503-725-5950 Email: [email protected] Rafael J. Fernández-Moctezuma Department of Computer Science Portland State University P.O. Box 751 Portland, OR 97207-0751 Phone: 503-725-4036 Fax: 503-725-3211 Email: [email protected] Robert L. Bertini Department of Civil & Environmental Engineering Portland State University P.O. Box 751 Portland, OR 97207-0751 Phone: 503-725-4249 Fax: 503-725-2880 Email: [email protected]

Submitted for presentation and publication to the 88th Annual Meeting of the Transportation Research Board January 11–15, 2009

November 2008

Word Count: 7442

TRB 2009 Annual Meeting CD-ROM

Paper revised from original submittal.

Wieczorek, Li, Fernández-Moctezuma and Bertini

2

Abstract. Bottlenecks are key features of any freeway system. Their impacts are of increasing importance as congestion worsens in major urban areas. In the U.S., the FHWA has been working to identify and monitor key bottlenecks in each state. In Oregon, a freeway data archive known as PORTAL records count, occupancy, and speed measurements from over 600 freeway locations. This archive has enabled development of online freeway performance and reliability analysis tools. This paper describes the results of a project aimed at developing an automated tool for identifying recurrent freeway bottlenecks using historical data within the framework of the data archive. Efforts have focused on identification and display of active bottleneck features using graphical tools and the selection of optimal variables that enable the careful identification of active bottlenecks. This research aims to detect bottleneck activation in real time and to expand the use of reliability techniques. Ultimately the results of this research will improve the prioritization of improvements and implementation of operational strategies on the freeway network. INTRODUCTION Understanding traffic dynamics at freeway bottlenecks is a foundation for understanding how the freeway system operates. A bottleneck is defined as a point upstream of which one finds a queue and downstream of which one finds freely flowing traffic. Bottlenecks can be static (e.g. a tunnel entrance) or dynamic (e.g., an incident or a slow moving vehicle). Bottlenecks can be activated or deactivated due to a decrease in demand or spillover from a downstream bottleneck (1). Congestion imposes great costs on the movement of people and freight, motivating the development of new freeway performance measures and reporting systems. In Oregon, the Portland Oregon Regional Transportation Archive Listing (PORTAL—see http://portal.its.pdx.edu/) has been established to collect count, occupancy, and speed data from over 600 locations at 20-second intervals. In addition to archived data, performance measurement tools and analytical tools are available. The objective of this paper is to describe the development of an automated system to identify freeway bottlenecks using historical data within the PORTAL environment. Using ground truth knowledge of when and where bottlenecks occurred during a sample period, a working prototype was developed and tested with the intent to accurately identify, track, and display active bottleneck features using graphical tools. Previous research, conducted in California by Chen et al. (2), was validated for use in PORTAL, and suggestions for further research are provided. BACKGROUND Past research has sought a greater understanding of where freeway bottlenecks formed and how and when they were activated. One method uses oblique curves of cumulative vehicle arrival number versus time and cumulative occupancy (or speed) versus time constructed from data measured at neighboring freeway loop detectors (3, 4, 5, 6, 7). Using this method it is possible to observe transitions between freely-flowing and queued conditions and to identify important features. This method has been applied here to systematically define bottleneck activations and deactivations as ground truth for testing and implementing an automated bottleneck detection algorithm. Previous research has also developed techniques for automatic bottleneck activation and deactivation identification by comparing measured traffic parameters and applying thresholds for those values. Notably, Chen et al. (2) developed an algorithm to identify bottleneck locations and their activation and deactivation times using 5minute aggregations of freeway loop detector data from San Diego, California, focusing on speed differences between adjacent detectors. Zhang and Levinson (8) implemented a similar system to identify bottlenecks using occupancy differentials. Banks (9) used 30-second speed drop thresholds to identify bottleneck activation in San Diego, and Hall and Agyemang-Duah (10) developed a threshold using the ratio of 30-second occupancy divided by flow on a Canadian freeway. Bertini (11) tested other signals including the variance of 30-second count as characterizing bottleneck activation at several sites. Building on the past research, the objective of this paper is to test the Chen method using a sensitivity analysis approach with data from a freeway in Portland, Oregon. EXTENDING PORTAL’S FUNCTIONALITY In addition to providing loop detector data at different aggregation levels, the PORTAL archive includes automatic data analysis tools such as oblique plots, travel time analysis, congestion reports, and information layers in timespace surface plots with incident and variable message sign information (12). In response to the FHWA’s initiative of identifying bottleneck locations, a new PORTAL module is being developed to provide bottleneck identification and analysis capabilities. One of PORTAL’s main features is the use of surface plots to represent the time-space

TRB 2009 Annual Meeting CD-ROM

Paper revised from original submittal.

Wieczorek, Li, Fernández-Moctezuma and Bertini

3

evolution of a particular measurement. One goal for the bottleneck identification module is to visually represent the bottleneck activation and deactivation times, as shown hypothetically in Figure 1. Bottleneck activation is marked as well as the average propagation speed of the shockwave. This tool will provide the most relevant information to the user: activation time, duration, deactivation time, location, and shock propagation speed.

B

A Bottleneck Activation Deactivation Estimated Propagation Speed A – 25 mph

B – 21 mph

FIGURE 1 Desired output of the bottleneck analysis on historical data module for PORTAL. One objective of this research is to incorporate historical data into live data displays. Figure 2 illustrates a notional example of how previously known bottleneck locations can be displayed on the time-space diagram in a real time environment. Bottlenecks A, B, and C were detected and mapped on the plot, and the bounding boxes illustrate durations and locations that occurred (for example) 90% of the time over a previous period of time. These durations and locations can be learned by analyzing data from previous archived days. Bottleneck detection thresholds may lead to false positives (or negatives), as with bottleneck A. In this case, displaying the additional layer of historical bottleneck locations provides the user with current analysis output and historical knowledge. Historical knowledge of when and where to expect bottlenecks may be used to improve forecasting of propagation speeds as new data arrives. Bottleneck information may also be displayed in combination with incident information, providing a rich tool for users who may want to distinguish recurrent congestion from incident related congestion. ANALYTICAL FRAMEWORK Many Portland freeway corridors contain recurrent bottlenecks that form with front stationary shockwaves coupled with backward forming queues. Therefore, this analysis begins with identifying bottleneck locations, and activation and deactivation times, followed by tracking of queue propagation and mapping the congested conditions in time and space. To identify a bottleneck, the Chen et al. algorithm (2) focuses on longitudinal detector pairs and uses as its parameters the maximum speed measured by the upstream detector and the minimum speed differential between each detector pair. Using San Diego freeway speed data aggregated every 5 minutes, Chen et al. selected a 40 mph maximum upstream speed and a minimum 20 mph speed differential as their thresholds. The optimal values of these two parameters may be different in other cities, perhaps due to different freeway network configurations, geometry, weather, pavement, traffic conditions and the nature of the available data. As an example, the Portland freeway network contains several locations where one freeway ―tees‖ into another, resulting in recurrent congestion. In this

TRB 2009 Annual Meeting CD-ROM

Paper revised from original submittal.

Wieczorek, Li, Fernández-Moctezuma and Bertini

4

paper, the maximum upstream speed and minimum speed differential speed thresholds and data aggregation levels comprising the Chen method were tested for bottleneck identification in Portland using PORTAL data. In addition, ground truth bottleneck locations and activation/deactivation times identified using the method described in Cassidy and Bertini (4, 5) were used as baseline reference points for evaluating the outcomes of the Chen method using different data sets and a range of parameters.

C

A

Bottleneck Activation Deactivation

B

90% percentile of historical bottlenecks Estimated Propagation Speed A – 25 mph B – 22 mph C – 21 mph

FIGURE 2 Desired output of incoming data analysis in combination with historically learned parameters, such as expected bottleneck locations and activations.

SITE DESCRIPTION AND DATA PREPARATION The implementation of the Chen method was tested with archived data from the northbound Interstate 5 (I-5) corridor in the Portland, Oregon metropolitan region, shown in Figure 3. There are 22 detector stations along this 24-mile corridor. The average detector spacing is approximately 1 mile. Detectors are located just upstream of metered freeway on-ramps. In order to evaluate the Chen method against ground truth on Portland data, a judgment sample of five representative mid-week days from 2006 to 2008 was chosen (February 7, 2008; December 5, 2007; October 3, 2007; March 28, 2007; and August 8, 2006). Data were extracted from PORTAL at the lowest available resolution of 20 seconds between 5:00 am–midnight for each day, including count, occupancy and time mean speed in each lane. Minor data cleaning was performed during the aggregation into 1-minute data. This facilitated further aggregation into the 3-, 5- and 15-minute data sets. For each station, a volume weighted mean speed for 1 minute aggregation was compared to arithmetic averaging, and the results showed that the difference was not significant. Therefore, mean speeds by station were calculated at further levels of aggregation (1 minute, 3 minutes, 5 minutes, and 15 minutes) using arithmetic averages. Thus, for each station, measurements from multiple lanes were aggregated into a single quantity using simple arithmetic averaging for simplified calculation. The obtained value is extrapolated across a region of influence of a sensor station, i.e., the segment of the corridor between the current sensor station location and the next upstream sensor location.

TRB 2009 Annual Meeting CD-ROM

Paper revised from original submittal.

Wieczorek, Li, Fernández-Moctezuma and Bertini

5

BASELINE BOTTLENECK ANALYSIS In order to establish a baseline to act as ―ground truth‖ for this experiment, a detailed analysis of bottleneck activations was conducted. Figure 4 shows a speed plot from PORTAL for October 3, 2007. The x-axis is time, the y-axis is the milepost, and the color is the speed at that location and time. Black triangles indicate automated algorithmic detection of bottleneck activation, while squares show the true activation and deactivation times. The first active bottleneck (A) occurs at MP 299.7 between 7:45 and 10:00 am. Location (B) consists of a bottleneck at MP 290.54 between 7:15 and 9:15 am. The third bottleneck (C) is detected at MP 307.9 between 2:30 and 7:30 pm. The activation and deactivation times of each bottleneck were carefully diagnosed and verified using oblique curves of cumulative vehicle arrival number versus time and cumulative occupancy (or speed) versus time constructed from data measured at neighboring freeway loop detectors. Previous research describes the procedures in detail, particularly those described in Horowitz and Bertini (7) for the detector configuration in Portland (only upstream of on-ramps) and other sources (3, 4, 5, 6, 11). The value of this visual tool is that the prevailing flows and speeds measured at the particular station are seen as slopes of the oblique cumulative plot. This method filters out the noise while not altering the times at which flow and speed changes occurred. This manual procedure was used to carefully confirm queue formation and propagation, such that active bottlenecks separated freely flowing traffic from queued traffic. A detailed example of this diagnosis for this project is available in (14). The procedure described above was performed for all candidate bottleneck activations. Knowing the times at which a transition from unqueued conditions to queued conditions occurred at a station also allowed for the estimation of a shock speed between stations, and is also shown on Figure 4 for each activation. This method provides an excellent baseline for approximating the ground truth. However, it requires a great deal of user intervention and is heuristic; this is part of the motivation for the search for an automated algorithm that is close to this method in accuracy. EXPERIMENTAL DESIGN Data Collection The judgement sample of five days was selected by searching for representative days with high data quality. In total, 51 bottlenecks were found to be active for more than 60 hours on the days selected for testing the model. The bottleneck activation and deactivation times identified by the oblique curve procedure serve as the baseline ―ground truth‖ for experiments aimed at developing and testing the automated bottleneck detection algorithm. These activations occurred at 8 distinct locations on northbound I-5, and had a mean duration of 1 hour 11 minutes. Having this baseline facilitates a sensitivity analysis approach that reveals the best combination of data aggregation level and speed thresholds to be used within PORTAL’s bottleneck identification system. In order to test the Chen method on the Portland dataset from PORTAL, we implemented the method for identifying the locations of bottlenecks in the MATLAB programming environment. The Chen method compares each pair of longitudinally adjacent detectors at each 5-minute time point and declares that there is a bottleneck between them when: the speed at the upstream detector is below the maximum speed threshold vi, and the difference in the speeds at the upstream and downstream detectors is above the speed differential threshold Δv. Given that a bottleneck separates unqueued (downstream) from queued (upstream) traffic states, Chen et al. used values of vi = 40 mph, Δv = 20 mph, with data aggregated at 5 minute intervals. Thus, if the upstream speed is less than 40 mph and the downstream speed is more than 20 mph greater than the upstream speed, the Chen method identifies a bottleneck between those detectors during that interval. However, Chen et al. recognized that each of the three parameters (maximum upstream speed, minimum speed differential, and aggregation level) may need to be adjusted for the algorithm to work optimally in new situations. In this research, we have tested ranges of values of vi, Δv, and the aggregation level, as shown in Table 1, with a total of 125 possible combinations.

TRB 2009 Annual Meeting CD-ROM

Paper revised from original submittal.

Wieczorek, Li, Fernández-Moctezuma and Bertini

6

Vancouver MP 307.46 Marine Dr.

MP 307.9 Jantzen Beach

MP 306.51 Delta Park MP 305.12 Portland Blvd. MP 304.4 Alberta St. MP 303.88 Going St. MP 302.5 Broadway MP 301.09 Morrison Bridge

Portland

MP 299.7 Macadam Ave.

MP 297.33 Terwilliger Blvd. MP 295.18 Capitol Hwy.

MP 296.6 Multnomah Blvd. MP 296.26 Spring Garden

MP 293.74 Pacific Hwy. MP 293.18 Haines St. MP 292.18 OR 217/Kruse W.y MP 291.38 Upper Boones MP 290.54 Lower Boones Ferry MP 289.63 Nyberg Rd. WB MP 289.4 Nyberg Rd. EB MP 286.3 Stafford Rd. WB MP 286.1 Stafford Rd. EB MP 283.93 Wilsonville

FIGURE 3 Northbound Interstate 5 corridor in the Portland region.

TRB 2009 Annual Meeting CD-ROM

Paper revised from original submittal.

Wieczorek, Li, Fernández-Moctezuma and Bertini

7

Timeseries speed surface plot for I-5 NORTH on October 03, 2007 (Units in mph) Interstate Timeseries speed surface plot forfor I-5 NORTH onon October 2007 in in mph) TimeseriesTimeseries speed surface plot for I-5 NORTH on October 03, 200703, (Units in(Units mph) speed surface plot October 03, 2007 (Units mph) Timeseries speed surface plot I-5 forNORTH I-5 NORTH on October 03, 2007 (Units in mph) Bridge MP 308

CC C C C C C

C I-405 MP 304

D

I-405 MP 300

C

C

I-84 MP 302

DD

DAA

ADA A AA A

OR-217 MP 292

B

B

BB B BB

A

A

B

B

Bottlenec Bottleneck Bottleneck Bottleneck Bottleneck Activation Activation Activation Activation Activation Deactiva Deactivation Deactivation Deactivation Deactivation Algorithm Algorithm Algorithm Algorithm detection Algorithm detection detection detection detection Estimate Estimated Estimated Estimated Propagation Estimated Propagation Propagation Speed Speed Propagation Speed Propagation Speed A – 20 mph A –A20 mph A – 20 mph –A 20 mph – 20 mph B – 17 mph B – 17 mph B – 17 mphB – 17 mph Bmph – 17 mph C – 15 mph C –C15 C – 15 mph –C 15–mph 15 mph D – 15 mph D –D15 mph D – 15 mph –D 15 –mph 15 mph

B

I-205 MP 288

FIGURE 4 Baseline analysis of bottleneck locations and algorithm detection on northbound Interstate 5 corridor on October 3, 2007. Also following Chen et al., a ―sustained bottlenecks‖ filter was added that smoothes the results of the base algorithm. This filter discards transient false positives that are isolated in the time dimension from other bottleneck detection instances at the same detector. Similarly, it fills in small gaps in bottleneck detection at a given detector. For instance, if several points in time at a given detector are classified as bottlenecks, but one point in the middle of these others fails to register as a bottleneck, the sustained-bottlenecks filter fills in this gap, assuming that it is unlikely for a bottleneck to deactivate only for a few minutes in the midst of a time period of heavy congestion. After the original bottleneck detections are performed, this filter runs by scanning each detector one at a time and checking to see whether each detection is part of a sustained, active bottleneck. If, within the span of 7 consecutive time periods, fewer than 5 of those periods include bottleneck detections, the individual detections in that period are assumed to be false positives and are deleted. However, if 5 or more of the 7 time periods registered as active bottlenecks, then any gaps in between them are also registered as bottleneck detections. This filter smoothes the data as described, improving overall accuracy. TABLE 1 Input Parameters Tested for Bottleneck Identification Input Parameters Values Tested Max speed detected by upstream detector (mph) vi 30, 35, 40, 45, 50 mph Speed differential (mph) Δv 10, 15, 20, 25, 30 mph Data aggregation level for measured speed data 20 sec, 1 min, 3 min, 5 min, 15 min

TRB 2009 Annual Meeting CD-ROM

Paper revised from original submittal.

Wieczorek, Li, Fernández-Moctezuma and Bertini

8

RESULTS Initial Graphical Results The result of our tests is an assessment of how many bottleneck instances are captured by the Chen method with the different threshold combinations (Δv, vi) and using different aggregation levels for the measured speed data. Taking one sample day’s speed map as an example, Figure 5 visually shows the outcomes for all 25 threshold combinations using 5 minute speed aggregation. Each time-space plot shows a combination of speed differential and maximum speed measured at the upstream detector, labelled as (Δv, vi). Previous research has developed a quality index for the accuracy of a congestion reporting system as compared with ground truth data (15). Using a similar model, the success rate and false alarm rate for the given thresholds are considered as a function of the input values. The success rate is the proportion of speed-pairs which are real bottlenecks that are successfully captured by the Chen method. These correct detections are labelled as light triangles in Figure 5. The false alarm rate is the proportion of speed-pairs captured by the Chen method even though they are not real bottlenecks; these false alarms are labelled as dark triangles in Figure 5. From Figure 5, the correct detections and false ones are easy to recognize visually. As can be seen in this figure, lowering Δv and raising vi relaxes the algorithm’s definition of a bottleneck and increases the overall number of detections (both correct and false ones). The same procedure was repeated for the other four aggregation levels shown in Table 1. Figure 6 provides a visual comparison of the effects of different parameter choices, shown on speed plots using data from December 5, 2007 on the I-5 corridor. The selected (Δv, vi) pairs are shown on the right side of the figure. Due to the expected fluctuations in the data, the 20-second aggregation level data shows an unacceptably large number of false alarm detections. Meanwhile, the 15 minute aggregation level data misses many of the true bottlenecks at some threshold combinations. The data aggregation level directly affects the total number of data points available to be tested for bottleneck activation. Thus, the raw numbers of successful bottleneck detections and false alarms are not comparable from one aggregation level to another. Instead, following the lead of (15), these numbers were converted into proportions for the comparison: the success rate, true_found/all_true, and the false alarm rate, false_found/all_found. The algorithm’s output is compared against the ground truth to find the numbers that compose these ratios. The variable true_found is the number of the algorithm’s bottleneck detections that were indeed bottlenecks according to the ground truth; the ―detections‖ that were not real bottlenecks are added up into false_found. Note that the denominators of these two ratios are different: all_true is the total number of true bottleneck instances according to the ground truth, while all_found is the total number of the algorithm’s bottleneck detections regardless of whether they are true bottlenecks or not. These are standard measures for evaluation of a new test, but because of their different ratios they are not directly comparable to one another: one cannot simply add or subtract them to obtain a single score. Figure 7 A and B show the success rate and false alarm rate (vertical axis), respectively, calculated across all 5 sample days. These figures show how the success and false alarm rates vary across different thresholds (horizontal axes) and different speed data aggregation levels (the different surfaces). For instance, the success rate is maximized when the maximum upstream speed is set to 50 mph, the minimum speed differential between a pair of upstream-downstream detectors is 10 mph, and the data aggregation occurs at 1 minute intervals. It is clear that, at each data aggregation level, the success rate increases as the maximum upstream speed increases and the minimum speed differential decreases, i.e., when the standards for defining a bottleneck are relaxed. Unfortunately, the false alarm rate also experiences an increase under the same conditions. Statistical Analysis of Results Since the success and false alarm ratios are not directly comparable, the optimal choice of parameters depends on the user’s choice of a composite score function that takes into account the relative costs of missed bottlenecks and false alarms. Since there is no inherently ―right‖ score to use, three different score functions were proposed for the evaluation of Chen’s method on the Portland dataset. First, a simple sum can be maximized: sum = success_rate + (1 – false_alarm_rate), assuming that an increase in successful detections is equally desirable to a decrease in false alarms and that they are comparable in an additive manner. Second, one may choose to maximize the product of these two terms, rather than their sum: product = success_rate (1 – false_alarm_rate). Finally, thinking in terms of costs rather than benefits and drawing from (16), one might decide to put a heavier penalty on missing a real

TRB 2009 Annual Meeting CD-ROM

Paper revised from original submittal.

Wieczorek, Li, Fernández-Moctezuma and Bertini

9

bottleneck than on false alarms, and consequently wish to minimize the resulting score: cost = false_alarm_rate + 3 (1 – success_rate), where the penalty weight of 3 is assumed here for the sake of argument. milepost

305 295 285 (10, 30)

(10, 35)

(10, 40)

(10, 45)

(10, 50)

(15, 30)

(15, 35)

(15, 40)

(15, 45)

(15, 50)

(20, 30)

(20, 35)

(20, 40)

(20, 45)

(20, 50)

(25, 30)

(25, 35)

(25, 40)

(25, 45)

(25, 50)

(30, 30)

(30, 35)

(30, 40)

(30, 45)

(30, 50)

milepost

305 295 285

milepost

305 295 285

milepost

305 295

milepost

285 305 295 285 6

12 18 24 time

6

12 18 24 time

6

12 18 24 time

6

12 18 24 time

6

12 18 24 time

FIGURE 5 Outcome of bottleneck identification with different input value combinations (Δv, vi) using 3 minute aggregation of speed data (I-5 North, December 5, 2007).

milepost

20 second aggregation

1 minute aggregation

3 minute aggregation

5 minute aggregation

15 minute aggregation

305 (10, 50)

295

milepost

285 305 (15, 35)

295

milepost

285 305 (30, 30)

295 285 6

12 18 24 time

6

12 18 24 time

6

12 18 24 time

6

12 18 24 time

6

12 18 24 time

FIGURE 6 Comparison of outcomes with different aggregations of speed data (I-5 North, December 5, 2007).

TRB 2009 Annual Meeting CD-ROM

Paper revised from original submittal.

Wieczorek, Li, Fernández-Moctezuma and Bertini

10

A statistical analysis of variance was performed independently for each of the three score functions, treating our results as the output of a 3-factor, 5-level full factorial experiment with no replications. The three factors and their levels were described in Table 1. The terms of the ANOVA model included all 3 main effects and all 3 two-way interactions. Since there were no replications, the three-way interaction was used to estimate the error term. For example, the ANOVA table for the values of the cost score function is reproduced in Table 2. TABLE 2 Analysis of Variance for cost Using Adjusted SS for Tests Source DF Adj SS Adj MS F P AggLevel 4 0.7006 0.1752 216.2 0.000 MinDiff 4 10.5544 2.6386 3257.7 0.000 MaxUpstream 4 0.4651 0.1163 143.6 0.000 16 0.204741 0.0128 15.8 0.000 AggLevel MinDiff 16 0.1380 0.0086 10.7 0.000 AggLevel MaxUpstream 16 1.4299 0.0894 110.3 0.000 MinDiff MaxUpstream Error 64 0.0518 0.0008 Total 124 13.5445 S = 0.0285 R2 = 99.6% R2(adj) = 99.3% The variation in the scores was partitioned differently for each score function, but in each case the ANOVA results showed that each of the main effects and each of the 2-way interactions were highly significant. In other words, changing two factors at once does not necessarily have the same effect as changing each factor separately one at a time. Thus, it is not enough to select the factor values that independently have the most desirable main effects; instead, it is necessary to compare all possible sets of parameter settings. It turned out that each score function was optimized at the same set of the three parameters: 3 minute data aggregation level, 15 mph minimum speed differential Δv, and 35 mph maximum upstream speed vi. A visual cross-check with Figure 7 confirms the outcome of the statistical analysis. The point at (Δv, vi) = (15, 35) on the 3 minute aggregation surface does appear to be a reasonable point to use on both figures. In Figure 7A, we see that the success rates level off near this point, so its relatively high success rate could be increased only slightly. Similarly, in Figure 7B, this point has a relatively low false alarm rate, and the surfaces level off near this point so the false alarm rate cannot be lowered much further. Thus, it makes sense that this point is as close as we can get to maximizing the success rate and minimizing the false alarm rate simultaneously. Hence, the original Chen et al. settings used for the San Diego data—20 mph differential, 40 mph maximum upstream speed, and 5-minute aggregation—are close to, but not the same as, the optimal settings for this Portland freeway. Such a result is to be expected, since this is a parametric method. If researchers and transportation operations analysts in other cities wish to implement a system using the Chen et al. algorithm, they can perform a similar analysis to configure the optimal parameter settings for their own network. CONGESTION TRACKING The analysis so far has focused on setting the optimal speed thresholds for automatically identifying the location and activation times of bottlenecks, i.e., front stationary shockwaves. In the next step, the functionality was extended to examine the duration and location of the congestion that builds up upstream of the bottlenecks. To locate this congested area on the speed contour plot, the expanded algorithm analyzes each active bottleneck location and assesses the conditions at the next detector upstream to determine whether its measured speed is also below the maximum upstream speed threshold vi. If this is true, the algorithm continues stepwise further upstream until a detector is reached with a speed greater than the threshold vi. At this point in time and space, the detector is considered to be free of congestion, and all the detectors downstream toward the bottleneck are labeled as part of the congested regime. It is assumed that such a technique will only find congestion that is directly caused by an active bottleneck. Thus the total congested region on the time-space plane for the corridor under scrutiny can be combined with measured flow information to estimate the total delay caused by a given bottleneck. With historical analysis of a recurrent bottleneck, its consequent delay in vehicle-hours could be translated into the financial cost of wasted time, environmental cost of pollutant and carbon emissions, etc. caused by this bottleneck, which could help transportation planners to prioritize their spending intelligently.

TRB 2009 Annual Meeting CD-ROM

Paper revised from original submittal.

11

Success rate

Wieczorek, Li, Fernández-Moctezuma and Bertini

False alarm rate

A.

B. FIGURE 7 (A) Proportion of true bottleneck instances successfully captured. (B) Proportion of false bottleneck detections.

TRB 2009 Annual Meeting CD-ROM

Paper revised from original submittal.

Wieczorek, Li, Fernández-Moctezuma and Bertini

12

Incorporating Historical Information For traveller information purposes, the onset and duration of congestion around a certain time and location can be predicted by using the historical data to find the percentage of days that congestion has occurred at a particular time and location. The historical percentile methodology (perhaps coupled with incident information) also makes it possible to distinguish between recurring and nonrecurring congestion. In order to test and demonstrate this procedure, one month’s worth of weekday data from northbound I-5 in February 2008, a total of 21 days, is analyzed. The goal is to map various percentiles to illustrate times and locations most prone to bottleneck conditions. To illustrate the results of applying this process, Figure 8A shows a time-space diagram for northbound I-5, with the highlighted area outlining the congested region (upstream of a bottleneck activation) that occurred at least 90% of the time during February 2008. For each cell in the time-space plane, this means that the highlighted region fell upstream of a bottleneck and met the congestion criteria on about 19 of the 21 days analyzed. This illustrates that between 4:00 and 6:00 pm, traffic approaching the Interstate Bridge (just beyond the furthest-downstream detector) experiences congestion nearly every weekday, and the queue propagates upstream as far as three miles almost every day. In addition to the 90th percentile, for the month of February 2008, the recurrent congestion that occurred on at least 75% of the days is highlighted in Figure 8B. On at least three-quarters of weekdays, traffic approaching the interstate bridge experienced congestion as early as 2:30pm. Figure 8C contains the 50th percentile analysis, illustrating the regions (in time and space) in which bottleneck-activated congestion occurred on at least 50% of the days analyzed. It can be seen that during at least half of all weekdays, there was also congestion in the morning upstream of milepost 300, near the exit for Interstate 405. It is clear that this is a useful tool for performance measurement and possibly also for predicting conditions in real time. Ongoing work is aiming toward processing all PORTAL data since July 2004 for all freeway corridors so that similar historical analysis can be conducted. Tracking the Edges of the Congested Regime Once the active bottleneck has been detected and the upstream congested regime has been delineated, it is possible to analyze the backward-forming shockwave edge and compute its slope, which translates into the speed of the queue propagation. The bottleneck detection tool could combine historical information with a measured real-time queue propagation speed to estimate when the bottleneck-related congestion will affect traffic at a given point upstream. To illustrate this, a comparison of the algorithmically-derived shockwave speeds was made against those found using the ground truth data. The shock speeds for the 51 activations described earlier were calculated manually and plotted against the automated computations, shown in Figure 9. This indicates that the algorithm’s results tend to be close to the ground truth but contain a few outlying cases in which we overestimate the propagation speed. This is arguably better than the alternative of underestimation, since travelers are likely to be more frustrated by a queue appearing sooner than expected than they would be by a queue that they never meet. Even so, further work to improve the accuracy of this tool is ongoing. CONCLUSIONS This paper performed a detailed graphical and statistical analysis to test the best combination of parameters (concluding 35 mph maximum upstream speed, 15 mph speed differential, and 3 minute aggregation level) toward implementing an automated bottleneck detection procedure in Portland, Oregon. The paper also presented a useful implementation that detects bottleneck activations and deactivations; considers their persistence over time; traces and maps resulting congestion upstream; analyzes historical congestion patterns; and measures shock propagation velocities. Based on a rigorous and detailed comparison with ground truth data, it appears that the procedures can be extended to the remaining freeway corridors in Portland. In terms of traffic information, appropriate quality assessment criteria should characterize the timeliness of traffic messages and their consistency with the traffic situation that would be experienced by the driver on a given route. This paper presents information that will be useful in the planning and operations environment but also to travellers, by mapping recurrent congestion in time and space. It is clear that each user’s optimal choice of thresholds and data aggregation level will depend on variable factors in terms of geography, traffic pattern and driver behaviours in a certain region. It can be said that the definition of congestion could be varied from region to region.

TRB 2009 Annual Meeting CD-ROM

Paper revised from original submittal.

Wieczorek, Li, Fernández-Moctezuma and Bertini

13

A.

B.

C. FIGURE 8 Congestion tracking. (A) 90% of days. (B) 75% of days. (C) 50% of days.

TRB 2009 Annual Meeting CD-ROM

Paper revised from original submittal.

Wieczorek, Li, Fernández-Moctezuma and Bertini

14

Shockwave Propagation Speed (mph)

Ground Truth Shockwave Speed (mph)

30

25

20

15

10

5

0 0

5

10

15

20

25

30

Automatic Detection Shockwave Speed FIGURE 9 Comparison of shockwave propagation speeds from algorithm and ground truth. FUTURE WORK The results described here are promising and are leading to additional research toward automating the process of identifying bottlenecks on Portland freeways. One next step consists of comparing the percentiles of congestion area by day of the week and by weather or seasonal conditions, using additional historical data. A further step is to incorporate volume data from PORTAL for quantifying total delay caused by such recurrent bottlenecks which can be translated into the cost of the time wasted and externalities such as fuel consumption and emissions unnecessarily produced by these bottlenecks. When this bottleneck detection tool is fully implemented in PORTAL, users will be able to make such comparisons and explore what might be causing this congestion. This will facilitate some simple forecasting that can be shown on the time-space speed plots as the loop detector live feed proceeds. Finally, the reliability of travel time predictions on a given corridor may be more important than the travel times themselves for travelers, shippers, and transportation managers. In addition to identifying recurrent bottlenecks using measures of delay, we plan to test several reliability measures including travel time, 95th percentile travel time, travel time index, buffer index planning time index, and congestion frequency. ACKNOWLEDGMENTS The authors acknowledge the support of the Oregon Department of Transportation, the Federal Highway Administration, TriMet and the City of Portland for their ongoing support and guidance in the development of the Portland ADUS. This work is supported by the National Science Foundation under Grant No. 0236567, and the

TRB 2009 Annual Meeting CD-ROM

Paper revised from original submittal.

Wieczorek, Li, Fernández-Moctezuma and Bertini

15

Mexican Consejo Nacional de Ciencia y Tecnología (CONACYT) under Scholarship 178258. The Portland metropolitan area’s TransPort ITS committee has also encouraged development of this project. REFERENCES 1. 2.

3.

4. 5. 6.

7.

8. 9.

10.

11. 12.

13.

14.

15.

Daganzo, C. F. Fundamentals of Transportation and Traffic Operations. Elsevier Science, New York, 1997. Chen, C., A. Skabardonis, and P. Varaiya. Systematic Identification of Freeway Bottlenecks. In Transportation Research Record: Journal of the Transportation Research Board, No. 1867, Transportation Research Board of the National Academies, Washington, D.C., 2004, pp. 46–52. Cassidy, M. J. and J. R. Windover. Methodology for Assessing Dynamics of Freeway Traffic Flow. In Transportation Research Record: Journal of the Transportation Research Board, No. 1484, Transportation Research Board of the National Academies, Washington, D.C., 1995, pp. 73–79. Cassidy, M. J. and R. L. Bertini, R.L. Some Traffic Features at Freeway Bottlenecks. Transportation Research, Part B, Vol. 33, No. 1, 1999, pp. 25–42. Cassidy, M. J. and R. L. Bertini. Observations at a Freeway Bottleneck. Proceedings of the Fourteenth International Symposium on Transportation and Traffic Theory, Jerusalem, Israel, 1999, pp. 107–124. Bertini, R. L. and A. Myton. Using PeMS Data to Empirically Diagnose Freeway Bottleneck Locations in Orange County, California. In Transportation Research Record: Journal of the Transportation Research Board, No. 1925, Transportation Research Board of the National Academies, Washington, D.C., 2005, pp. 48–57. Horowitz, Z. and R. L. Bertini. Using PORTAL Data to Empirically Diagnose Freeway Bottlenecks Located on Oregon Highway 217. Presented at the 86th Annual Meeting of the Transportation Research Board. Washington, D.C., 2007. Zhang, L. and D. Levinson. Ramp Metering and Freeway Bottleneck Capacity. Transportation Research Part A, 2004, forthcoming . Banks, J. H. Two-capacity Phenomenon at Freeway Bottlenecks: a Basis for Ramp Metering? In Transportation Research Record: Journal of the Transportation Research Board, No.1320, Transportation Research Board of the National Academies, Washington, D.C., 1991, pp. 83–90. Hall, F.L. and K. Agyemang-Duah. Freeway Capacity Drop and the Definition of Capacity. In Transportation Research Record: Journal of the Transportation Research Board, No. 1320, Transportation Research Board of the National Academies, Washington, D.C., 1991, pp. 91–98. Bertini, R. Toward the Systematic Diagnosis of Freeway Bottleneck Activation. Proceedings of the IEEE 6th International Conference on Intelligent Transportation Systems, Shanghai, China, 2003. Bertini, R.L., S. Hansen, A. Byrd, and T. Yin. PORTAL: Experience Implementing the ITS Archived Data User Service in Portland, Oregon. In Transportation Research Record: Journal of the Transportation Research Board, No. 1917, Transportation Research Board of the National Academies, Washington, D.C., 2005, pp. 90– 99. Bertini, R.L., H. Li, J. Wieczorek, and R. Fernández-Moctezuma. ―Using Archived Data to Systematically Identify and Prioritize Freeway Bottlenecks.‖ 10th International Conference on Application of Advanced Technologies in Transportation, Athens, Greece, May 27– 31, 2008. Bogenberger, K., N. Henkel, E. Göbel, and R. Kates. How to Keep Your Traffic Information Customers Satisfied. Presented at the 10th World Congress on Intelligent Transport Systems, Madrid, Spain, November 1620, 2003. Bertini, R.L., and D. Lovell. Developing an Optimal Sensing Strategy for Accurate Freeway Travel Time Estimation and Traveler Information. Presented at the 87th Annual Meeting of the Transportation Research Board, Washington, D.C., 2008.

TRB 2009 Annual Meeting CD-ROM

Paper revised from original submittal.