Incorporating Travel Time Reliability into the Highway Capacity Manual (2014)

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

193 Weather is one of the seven sources of congestion on the transportation network. Weather has a twofold effect on roadways: influencing driver behavior and increasing the likelihood of incidents (another source of congestion). Recent studies have attempted to separate the sources of unreliable travel with varying contributions compared with other sources of congestion, but all point to weather as significantly contrib- uting to unreliable travel. Figure E.1 shows a typical break- down of the sources of unreliable traffic, which shows weather effects at 15%. This appendix describes typical sources of weather data used in transportation studies and the attributes they monitor. Two methods for predicting weather events or probabilities are described, and each is validated against historical weather data. Both methodologies are discussed with respect to their applicability for predicting weather impacts on travel time reliability in the context of SHRP 2 Project L08. Weather Data Sources Various national sources for weather data can be used to obtain historical data and trends for weather patterns for different locations across the United States. Some of the main sources are discussed below in light of their applicability to this project. CLARUS Initiative The CLARUS initiative was established in 2004 by the Federal Highway Administration’s Road Weather Management Pro- gram in conjunction with the Intelligent Transportation System Joint Program Office. The main goal of the CLARUS initiative was to “create a robust data assimilation, quality checking, and data dissemination system that could provide near real-time atmospheric and pavement observations from collective state’s investments in road weather information system, environ- mental sensor stations as well as mobile observations from Automated Vehicle Location equipped trucks.” Weather data can be obtained from the CLARUS system by (1) subscribing to a report that is periodically updated, (2) using the latest quality-checked observations on the map interface, and (3) retrieving an on-demand report for weather observations (Federal Highway Administration 2012). National Climatic Data Center The National Climatic Data Center of the National Oceanic and Atmospheric Administration periodically publishes climatic data summaries from weather stations in each of 284 U.S. cities and territories (National Climatic Data Center 2011). The published document contains 17 statistics related to temperature, wind, cloudiness, humidity, and precipitation. Each statistic is quantified by month of year and based on 10 or more years of data. The National Climatic Data Center also provides storm event data for several thousand locations throughout the United States (including the 284 cities previously mentioned). These data describe the average number of storms, average precipitation depth per storm, average storm duration, and average rainfall rate (i.e., intensity). Each statistic is quantified by month of year. National Weather Service The National Weather Service (NWS) of the National Oceanic and Atmospheric Administration is tasked with monitoring weather across the United States. The NWS provides real- time weather data as well as short-term weather predictions. Real-time reports are primarily from meteorological aviation reports (METARs) from airport weather stations across the United States that contain weather information vital to pilots including wind, visibility, weather type, cloud, temperature, and pressure data. They are typically reported hourly or every 30 min at 50 min past each hour. Special reports (SPECI) are reported when there is a significant change in the weather A p p e n D i x e Weather-Modeling Alternatives and Validation for the Freeway and Urban Street Scenario Generators

194 that occurs between scheduled hourly transmissions. A SPECI will be issued if any of the following occurs: • The ceiling decreases to 1,500 ft or less, or when a cloud layer, previously not reported, appears below 1,000 ft (or below the highest minimum for straight-in instrument flight rules [IFR] landings, or the minimum for IFR departures); • Visibility decreases to below 3 statute miles; • A tornado, waterspout, or funnel cloud is reported; • A thunderstorm begins, intensifies to “heavy,” or ends; • Precipitation begins, changes, or ends; • Winds suddenly increase and exceed 30 knots (speed must double), or when the direction of the winds significantly changes (satisfying the criteria for “wind shift”). Weather Underground (www.wunderground.com) is a weather data service that archives METARs and SPECI for all U.S. NWS weather stations. The service is provided for free and is used by many other services that require historical weather data, such as Wolfram Mathematica. Attributes of Weather Data The following lists indicate the weather attributes that are reported by each of the data sources described above. CLARUS • Air temperature; • Dew point; • Relative humidity; • Surface status; • Surface temperature; • Precipitation intensity; • Precipitation type; • Wind direction; • Wind speed; • Wind gust direction; and • Wind gust speed. National Climatic Data Center Among different weather statistics the following items are of interest to reliability evaluation: • Mean number of days with precipitation of 0.01 in. or more; • Total snowfall; • Normal daily mean temperature; and • Normal precipitation. National Weather Service • Wind direction; • Wind speed; • Visibility; • Weather type or phenomena; • Cloud amount; • Cloud height; • Temperature; • Dew point; • Pressure; and • Precipitation. Weather Effects on Traffic Weather has a significant impact on the operations of both freeway and urban street facilities. Past research has been performed to quantify the effects of weather by category. An extensive review of the literature on weather impacts on traffic operations has been completed by the SHRP 2 L08 proj- ect team that has been summarized in a separate white paper. This section presents a summary of the key findings. Weather affects operations on both freeway and urban street facilities, although the literature is more extensive on weather impacts on freeways. These facility types are discussed separately in the following subsections. Weather Impacts on Freeways Most research on weather effects on freeways has focused on capacity. The 2010 Highway Capacity Manual (HCM2010) (Transportation Research Board of the National Academies Figure E.1. Sources of congestion.

195 2010) includes 15 weather categories with an average and range of capacity effects on freeways. More recent research has also included free-flow effects. In a separate L08 project white paper a synthesis of the literature on capacity and free-flow speed (FFS) effects of weather on freeway facilities has been presented. Table E.1 summarizes the effects for freeways. The factors in Table E.1 can be used to estimate a modi- fied speed–flow relationship for a basic freeway segment. HCM2010 Equation 25-1 uses the capacity adjustment factor (CAF) to fit a speed–flow curve between the FFS and the newly estimated capacity. With the introduction of the speed adjustment factor (SAF) in this project, the revised equation, Table E.2, was adapted from HCM2010. The equation can be used to estimate the speed–flow relationships for different weather events. Figure E.2 shows some illustrative examples. The figure shows the speed–flow relationships for FFS of 75, 65, and 55 mph. The impacts of medium rain and heavy snow are shown for each base FFS in dashed and dotted lines, respectively. The 45 passenger cars per hour per lane (pcphpl) density line is shown to represent the level of service E to F boundary. Figure E.2 illustrates that the SAF results in a downward shift of the speed–flow curve, while the CAF shifts the intercept with the 45 pcphpl density line. The resulting curves in between these points are intuitive, and internally consistent, provided the SAF does not drop the FFS below the speed at capacity (after applying CAF). In that case, the methodology assumes a horizontal speed–flow relationship at a fixed speed equal to speed at capacity (after CAF), thus overriding the SAF input. For nonbasic segments (weaving and merge–diverge sec- tions), the methodology multiplies the FFS by SAF and the segment capacity by CAF in each occurrence in the method. Details on this implementation are provided in a separate working paper. Weather Impacts on Urban Streets The impacts of weather on urban streets are less well defined in the HCM2010, although extensive work has been done in this project to document the state of the practice in the literature. Weather impacts on urban streets primarily affect the saturation flow rate at signalized intersections and the midsegment FFS along an extended urban street facility. predicting Weather probability by Using Historical Averages Weather events can be predicted using Monte Carlo tech- niques. That technique and weather-modeling procedure are described in the section on urban street scenario development in Chapter 5 of the main report. As an alternative approach to the Monte Carlo technique, historical weather averages can be used to estimate the probability of occurrence of weather Table E.2. Estimating Basic Segment Speed from CAF and SAF (adapted from 2010 HCM Equation 25-1) FFS SAF 1 ln FFS SAF 1 CAF 45 CAFp p p p pS e C v C p ( )= + − ( )+ −     Weather Type Capacity Adjustment Factors(CAF) Free-Flow Speed Adjustment Factors (SAF) Free-Flow Speed (mph)* 55 mph 60 mph 65 mph 70 mph 75 mph 55 mph 60 mph 65 mph 70 mph 75 mph Clear Dry Pavement 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 Wet Pavement 0.99 0.98 0.98 0.97 0.97 0.97 0.96 0.96 0.95 0.94 Rain ≤0.10 in/h 0.99 0.98 0.98 0.97 0.97 0.97 0.96 0.96 0.95 0.94 ≤0.25 in/h 0.94 0.93 0.92 0.91 0.90 0.96 0.95 0.94 0.93 0.93 >0.25 in/h 0.89 0.88 0.86 0.84 0.82 0.94 0.93 0.93 0.92 0.91 Snow ≤0.05 in/h 0.97 0.96 0.96 0.95 0.94 0.94 0.92 0.89 0.87 0.84 ≤0.10 in/h 0.95 0.94 0.92 0.90 0.88 0.92 0.90 0.88 0.86 0.83 ≤0.50 in/h 0.93 0.91 0.90 0.88 0.87 0.90 0.88 0.86 0.84 0.82 > 0.50 in/h 0.80 0.78 0.76 0.74 0.72 0.88 0.86 0.85 0.83 0.81 Temp <50 deg F 0.99 0.99 0.99 0.99 0.99 0.99 0.99 0.99 0.98 0.98 <34 deg F 0.99 0.99 0.99 0.98 0.98 0.99 0.98 0.98 0.98 0.97 <-4 deg F 0.93 0.92 0.92 0.91 0.90 0.95 0.95 0.94 0.93 0.92 Wind < 10 mph 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 ≤20 mph 0.99 0.99 0.99 0.99 0.99 0.99 0.98 0.98 0.97 0.96 > 20 mph 0.99 0.99 0.99 0.98 0.98 0.98 0.98 0.97 0.97 0.96 Visibility <1 mi 0.90 0.90 0.90 0.90 0.90 0.96 0.95 0.94 0.94 0.93 ≤0.50 mi 0.88 0.88 0.88 0.88 0.88 0.95 0.94 0.93 0.92 0.91 ≤0.25 mi 0.90 0.90 0.90 0.90 0.90 0.95 0.94 0.93 0.92 0.91 Table E.1. Summary of Weather Capacity and Speed Adjustment Factors for Freeways

196 of a certain category. This approach may seem more data intensive initially, but it eliminates any stochastic element in the application of the method. Each weather category is con- sidered in the scenario generation process, weighted by the probability of occurrence in a particular analysis month and hour based on historical data. Data Source and Processing Historical average probabilities were created from NWS METARs data available from Weather Underground (2012). Historical weather is available in comma separated value (CSV) format for any airport in the Automated Surface Observing System on a daily basis. These CSV files contain all METARs for the airport and day requested. A Python (Version 2.7) script was written to automate the download of all daily CSV weather files for selected airports and years, and a second script compiles all daily files into a single CSV file containing observations from all selected years for a given airport. Once a single CSV file for each airport was created, each file was filtered and analyzed in Microsoft Excel. There were a few issues identified with the hourly reports that had to be fixed before calculating probabilities. First, reports occasionally report “unknown” conditions, for which any field with a number is reported as “-9999.” These values were changed so that they would not be picked up as a weather category with a capacity effect. Additionally, the compiler inserted reports from previous time periods infrequently. This would result in a negative duration between the previous report and the next one. Any repeated reports were removed before analysis. For application in SHRP 2 Project L08, all weather categories outlined in the HCM2010 that reduce capacity by at least 5% are included in the probability calculation. Table E.3, taken from the HCM2010, is shown with weather categories and their associated reduction in capacity; all categories that were modeled are marked with a star. Other weather effects are con- sidered to have minimal capacity effects and are combined with the no weather event probability. To calculate the probability for each weather category, the duration of time between each report is calculated to account for missing reports or SPECI. This time is calculated as the time in hours between the current and previous report. Each report is then classified as one of the 10 weather categories with a capacity effect or a negligible capacity effect. If a report can be classified as having multiple categories, it is assigned to the category with the highest capacity effect. Probabilities for each category are calculated using an Excel pivot table. Each weather category is set as a column header, and the month of year and hour are set as row headers. The sum of durations in each cell is divided by the sum of the durations in each row to calculate the proportion of time for each combination of month, hour, and weather category. Average event duration for each weather category is calculated over the 10-year data set by taking the average of all continuous Figure E.2. Illustrative examples of CAF and SAF application for freeways.

197 event durations. This duration is used in order to model each weather event in the computational engine. Data Characteristics Weather with a significant capacity effect is relatively rare in the United States, even in northern cities with significant winter weather. Figure E.3 shows the proportion of time with negligible capacity effects due to weather for airports in 40 of the largest metro areas in the United States. Las Vegas, Nevada, has the lowest probability (0.18%) of weather with a capacity effect, while Cleveland, Ohio, has the highest probability (9.33%) of the 40 largest metro areas. Figure E.4 shows a summary of the average probability for the 10 weather categories with significant capacity effects for four major metropolitan areas: New York (KLGA); Miami, Florida (KMIA); Chicago, Illinois (KORD); and San Francisco, California (KSFO). If light rain were included in the analysis, it would greatly outweigh other factors, as would the higher temperature categories, which are very common. As shown in Figure E.5, low-intensity snow has a relatively high probability in northern cities; however, it has the lowest effect on capacity of the 10 categories included in the analysis. Figure E.5 shows the average event duration for each of the weather categories for the same metropolitan areas. In general, the durations follow the same trend as the probabilities with one major exception. In Chicago, there was only one severe cold (<4°F) event, which lasted nearly 8 h. Weather-Modeling Procedure The freeway scenario generation approach in Project L08 uses the weather probabilities on a monthly basis for each category. The scenario generator contains a database of weather Note: Stars indicate weather conditions with at least a 5% average reduction in capacity. Source: HCM2010. Table E.3. HCM2010 Weather Categories for Freeway Facilities Figure E.3. Probability of weather with negligible capacity effect in 40 largest U.S. metropolitan areas, 2001–2010.

198 probabilities by month of the year and hour of day for each weather category created from the 10 years of METARs, as described earlier. Once a study period is selected, the scenario generator averages the probabilities across the study period hours (weighted by the fraction of each hour included in the study period in the case of partial hours) to create monthly probabilities by weather category. If months are grouped together (i.e., into seasons), probabilities are averaged across the months to create probabilities for each group of months by weather category (weighted by number of days in each month included in the reliability reporting period). Once the final average probabilities are created, the scenario generation treats each weather type and incident type as independent. Probabilities for each weather type and incident type for each combination of days (or groups of days) and months (or groups of months) are combined independently to generate the probability of scenarios with no weather or incident effects, scenarios with only weather effects, scenarios with only incident effects, and scenarios with both incident and weather effects. Modeling of weather events in the computational engine involves modeling each weather event (and incident event if applicable) once in a single run of the study period. Weather events are assumed to occur at either the start of the study period or the middle of it, and they are always modeled for the average duration, unless the duration does not yield a large Figure E.4. Annual probability of weather types in New York, Miami, Chicago, and San Francisco airports, 2001–2010. Figure E.5. Average weather event durations in New York, Miami, Chicago, and San Francisco airports, 2001–2010.

199 enough probability compared with the remainder of the study period with no weather and incident effects. Weather scenarios modeled in the computational engines are assigned a CAF and free-flow SAF for each weather type for the duration of the modeled weather event taken from the sources mentioned in the introduction. Weather Event Modeling and Probability Validation Comparison of Modeled Weather to Historical Weather Events Four locations in the United States were used to compare the predicted weather events to actual weather events from historical data: Chicago, Miami, New York, New Jersey, and San Francisco. At each location, 10 runs of the Monte Carlo weather event modeling were performed with 10 random seeds. The predicted events were created over 24 h for a full year at 15-min increments using National Climatic Data Center data. The 10-year hourly database (before calculating percentages), in addition to 2011 hourly weather data, was used to identify rain and snow events in historical data. The Monte Carlo model only predicts a maximum of one storm per day, so the number of storms predicted was com- pared with total storms, as well as the number of storm days, from 11 years of historical data. Figure E.6 shows the results of the comparison by month with the average, as well as the range of rain storms or rain storm days from both sources for the Chicago metropolitan area. Each month, the total number of storms was significantly underestimated; the number of storm days was better estimated, although it was slightly lower. Excep- tions are the winter months, when snow storms were predicted frequently in the place of rain storms. A comparison including snow events shows that the model predicts the frequency fairly accurately other than predicting no April or October snow events, while historical data shows that they occur rarely. Weather event characteristics were also compared between the modeled weather and historical weather. Figure E.7 shows that the average rain storm intensity (calculated as the total rainfall in an event divided by the event duration) is under- estimated in the upper probability section of the cumulative density function. Rain event duration was very accurately modeled in Chicago, and the resulting total rainfall per event was also estimated fairly well despite the underestimated intensities. Figure E.6. Chicago airport (KORD) actual versus modeled rain events, 2001–2011 (10 modeled years).

200 Although rain event characteristics were fairly well esti- mated, snow event characteristics had more issues. Average snow event intensities were overestimated, and durations were under estimated. The resulting total precipitation per snow event distribution is shown in Figure E.8, with the modeled events overestimating the total precipitation per event. Modeled weather events are assigned as rain or snow events based on an underlying distribution of temperature, so the underestimation of rain events in winter and lack of snow events in April and October indicate that this type of temperature–event relationship model has room for improvement. Using the same intensity and duration models for rain and snow caused issues with snow event characteristics; snow events would benefit from a separate model with dif- ferent underlying distributions. Probability Confidence Intervals Weather probabilities can vary greatly year to year, so the reported 10-year averages alone are not good indicators of year-to-year variability. Confidence intervals provide upper and lower bounds to the true average probability; 95% con- fidence intervals were calculated and graphed along with 2011 Figure E.7. Chicago airport (KORD) actual versus modeled rain storm intensity, 2001–2011 (10 modeled years). Figure E.8. Chicago airport (KORD) actual versus modeled total precipitation per snow event, 2001–2011 (10 modeled years).

201 probabilities. For the metropolitan area of Chicago, Figure E.9 shows the annual average probabilities by category; Figure E.10 shows average probabilities for January; and Figure E.11 shows average probabilities for April. In Chicago, the lowest-intensity snow category occurs most frequently on an annual basis, but the probability confidence interval is also relatively large, indicating high variability. This trend continues when analyzing January and April separately. While confidence intervals are widest for the most frequent weather categories in Figure E.9 and Figure E.10, Figure E.11 shows that the medium rain and low snow categories have very similar averages but very different confidence intervals. Similar month-to-month differences occur across the four locations analyzed. Estimating Future-Year Probabilities from Historical Averages and Monte Carlo Modeling As shown in the previous section, annual and monthly weather probabilities contain significant variations from year to year. A sensitivity analysis was performed using the hourly weather data from 2001 to 2011 in Chicago to determine what size sample (or look-back period) is appropriate to estimate Figure E.9. Chicago annual average weather probability by type, 10-year average versus 2011. Figure E.10. Chicago average weather probability by type for January, 10-year average versus 2011.

202 future-year weather probabilities. Both 2011 and 2010 were withheld as estimation years, and average probabilities of the 10 weather categories were calculated on a monthly (12 probabilities per weather category) and annual (one probabil- ity per weather category) basis. For 2010, the previous 3-, 5-, 7-, and 9-year average probabilities were compared using root mean square error (RMSE) across all monthly or annual esti- mates. For 2011, the previous 3-, 5-, 7-, and 10-year averages were compared. Both estimation years were compared with a baseline estimate of 0% probability for each category, as shown in Figure E.12 and Figure E.13. These figures show that the averages of all sample sizes tested—across all weather events—were significantly better than estimating 0% for all monthly or annual weather categories, but they indicate a slight increase in estimation error as the sample size increases for monthly probabilities. Further analysis indicates that the monthly weather patterns in 2010 were very similar to the previous 3 years, but overall annual estimation error remains very low as the sample size increases. As shown in Figure E.13, the 2011 estimation error trends downward as the sample size increases for both monthly and annual average probabilities. The Monte Carlo event prediction was also analyzed to create weather probabilities by type for each randomized run of the model with the Chicago weather characteristics. Figure E.14 shows the estimation error for each of the 10 runs when Figure E.11. Chicago average weather probability by type for April, 10-year average versus 2011. Figure E.12. 2010 Chicago annual weather probability sample size sensitivity.

203 Figure E.13. 2011 Chicago annual weather probability sample size sensitivity. Figure E.14. Monte Carlo predicted probability versus monthly average probability error for Chicago.

204 compared with monthly probabilities of each year of historical data, as well as the 10-year average. Figure E.15 shows the same relationship on an annual level, where error is lower. Of all years, only 2006 is better estimated by the Monte Carlo predictions compared with the 10-year average prob- ability. One source of error for all modeled runs is the lim- ited weather types that can be predicted. Rain and snow events make up only five of the 10 weather types used by the freeway scenario generator; however, total error increases when only including those five categories compared with assuming the other categories to have a probability of 0%. Otherwise, error within the categories that are able to be modeled can be attributed to either climatological data informing the model or underlying distributions used in the model. References Federal Highway Administration. ITS Research Success Stories: CLARUS System. http://www.its.dot.gov/CLARUS/. Accessed March 28, 2012. Highway Capacity Manual 2010. Transportation Research Board of the National Academies, Washington, D.C., 2010. National Climatic Data Center. Comparative Climatic Data for the United States Through 2010. National Oceanic and Atmospheric Administration, Asheville, N.C. http://www.ncdc.noaa.gov. Accessed Sept. 21, 2011. Weather Underground. Weather History. http://www.wunderground .com/history/. Accessed April 2012. Figure E.15. Monte Carlo predicted probability versus annual average probability error for Chicago.