Quantcast
Channel: maps for developers - Medium
Viewing all articles
Browse latest Browse all 2230

Ground-truthing storm weather traffic ETAs in Denver

$
0
0

Improving traffic predictions for logistics and everyday drivers

Typical traffic quality in Denver (left), traffic quality in Denver during snowstorms (right).

By Camilo Sanín

Last winter was one of the snowiest on record, creating challenges in logistics and transportation across the United States. Mail and packages don’t stop on stormy days, but driving conditions do get more unpredictable.

In order to power more accurate ETAs for logistics and everyday drivers during unpredictable weather, the Telemetry Team conducted an analysis of the effect that storm weather had on traffic in Denver, where billions of vehicle-miles were traveled during the winter (estimated 21.7B annual VMT), and thousands of those were trips powered by Mapbox Navigation.

Between November 2018 and April 2019, Mapbox collected about 200M miles of anonymous telemetry that we use to power our traffic data. Of that 200M miles, we set aside 400K anonymous trips (<.05%, totaling 825K miles) to benchmark our traffic products to ground-truth.

With this small subset, we are able to infer very granular traffic dynamics in the city — from heavy-traffic rush hour, Friday nights out, and lazy Sundays. We further singled out moments throughout the winter when traffic was exceptionally slow to evaluate the predictive ability of our traffic models under two dimensions of quality: accuracy and symmetry. ETAs during storm days were consistently optimistic and more inaccurate than typical days. Predictions based on live-traffic data showed a marked improvement over typical-data only ETAs during storm days. This analysis shows the powerful insights that can be derived from a small portion of our telemetry data, and the value of the signal that can be extracted by our traffic models. Here’s a look at how we did it.

Understanding traffic dynamics from anonymous telemetry data

To measure general traffic dynamics in Denver, we calculate the pace (minutes it takes to drive a mile) of each of the trips set aside for evaluation. When we aggregate all those trips, we visualize the variation of pace at which vehicles move hourly though a typical week. It’s worth noting the regularity of traffic: high-pace rush hours during the weekdays as well as faster speeds (low pace) during off-peak times. Other details are also visible during typical weeks; for example, weekend nights are a lot more congested than business days, and traffic on Sundays is in general lighter.

Pace through a typical week in Denver

Assuming a constant VMT (vehicle miles traveled), we translated pace into “number of hours” estimating how much any given time during the winter (in blue in the graph below) deviated from the typical baseline we established (in gray). Positive numbers (in the y-axis) are the number of extra hours in which Denverites were collectively stuck in traffic with respect to the same time in a typical week, while negative numbers are exceptionally low traffic periods. For example, in the second week of December, traffic fell within the baseline and was never particularly fast or slow.

Time spent in traffic through time (in blue) compared to the baseline of typical traffic (in gray)

What happens when it snows

On March 13th a record-breaking storm hit most of the Central US causing major disruptions in Colorado. In Denver in particular, traffic was chaotic and our telemetry data captured a significant slow down in travel speeds.

Time spent in traffic through the storm (in blue) compared to the baseline of typical traffic (in gray)

For both March 13th and 14th, the extra hours driven were well above the typical week. Road conditions were so bad, that on average each person spent 30 minutes longer in traffic than usual (minutes lost per capita) on March 14th. However, this was not the only storm or irregular traffic pattern in the winter.

We repeated the exercise above for every day from November to April calculating the time per capita “lost” in traffic compared to normal conditions. For a typical day, the minutes delayed per capita were ± 5, however, for most days when a storm was reported (and for some holidays) traffic was exceptionally bad. For example, on the Wednesday before thanksgiving, on average 11 minutes were lost per capita, and most winter storms added between 10 and 15 minutes per person.

Mapbox Traffic

Throughout the winter, Mapbox powered the navigation of thousands of trips in Denver through our Navigation SDK and logistics services. ETAs for those services are calculated using traffic data created through two different processes: (1) typical data, which uses historical telemetry, and (2) live data which is created using only data from current telemetry. Because traffic is highly cyclical, typical speeds data maximize the coverage of robust traffic estimates, while live data allow us to navigate in real-time through unexpected conditions.

Mapbox’s turn-by-turn Navigation uses a combination of those two sources of traffic data, however some products like Predictive Navigation and the Directions Matrix API relay only on typical data. In such cases, atypical conditions as those described above reduce the predictive ability of our models. To understand their impact, let’s first zoom out to one of the ways we measure traffic quality in general in the Telemetry team.

ETA Quality and Fan Diagrams

To make sure we understand the strengths and weaknesses of our ETAs, we continually benchmark traffic predictions against ground truth. Daily, we collect about one million records, for which we calculate multiple versions of error with respect to actual travel times.

One such metric is MSE, which for each observation takes a value between -1 and 1. MSEs value for a record is 0 when there is a perfect match to actual travel times. Negative values represent records in which the ETAs were smaller than the ATAs, therefore we had an optimistic estimate. Positive numbers indicate that we estimated a trip to take longer than it actually did.

MSE = {eta-ata}/{max(eta, ata)}

When thousands of records are gathered and grouped together, we can analyze the trend of a population. These groupings can be geographical (e.g. European countries, APAC cities, US States), temporal (e.g. calendar days, business days vs. weekends), any other relevant classification (slow vs. fast trips, types of roads, etc) or even competing models or ETA providers.

For example, below are histograms of the MSE for groups A through D. Notice that there is a difference in (1) the spread of the distribution and (2) how centered around 0 the error is (symmetry).

Simulated distribution of error for four groups (A, B, C, D)

To capture both those dimensions, we can we summarize the above distributions in a fan-like graph, in which the x-axis is the mean error (which captures symmetry), while in the y-axis we plot the mean absolute value of the error (which is the overall error, regardless of its direction). Then we measure traffic quality in two dimensions given the location of a distribution in a polar plot: The magnitude (r) and the angle (θ).

Fan diagram summarizing the distribution of errors simulated above

The magnitude of the vector between the origin and each point (`r`) is related to the overall traffic accuracy — the further from the origin the lower our quality. For example, `A` and `B` in the figure below have similar overall accuracy, whereas `C` is better. The second dimension is the angle of the vector with respect to the dashed line (`θ`). Clockwise estimates are overall optimistic, and counterclockwise ETAs are too high (overly pessimistic). For example `A` and `D` have the same symmetry value (though `A` has greater accuracy), and while ETAs in `C` were overall the closest to the ground truth but when the estimates were off, they were almost exclusively too slow.

In summary — the closer a point is to the tip of the graph, the more accurate our estimates are; and the closer to the line in the center, the more balanced optimistic vs. pessimistic trips are.

Traffic Quality in Snowy days

Mapbox traffic data in Denver is generally very accurate. For more detail on how consistent our ETAs were, we can plot our evaluations grouped by day. In the figure below, each point represents the quality of the typical-only traffic estimates on a day. Similar to the figures above, blue points represent days with significant levels of snow and red ones holidays. “Normal” days (in gray) form a cloud in the high-quality and slightly pessimistic area of the cone, additionally some holidays (which we don’t treat as regular weekdays) have high predictive ability. During storms, as well as days surrounding holidays (e.g. Wednesday before thanksgiving) ETAs based on typical data are less accurate and very optimistic. In other words, for these days our prediction based on baseline-traffic produced business-as-usual ETAs, while travel times on the road were exceptionally long.

Fan-diagram plotting the error of typical-only ETAs for every winter day.

Effect of live traffic

Fortunately, our navigation is also powered with live-data traffic which is especially relevant during atypical days. Anecdotally, we have noticed that the contribution of live-data seems to be most beneficial during extreme conditions, but we also want to understand its impact on ETAs on regular days. For storm days and holidays, with atypical traffic, this seems to be confirmed when plotting the error side-by-side in a fan diagram (small markers below are the errors of typical-only ETAs and larger are ETAs powered by typical+live data).

Fan-diagram contrasting error of typical-and-live ETAs (large points) and typical-only (small points) for holidays and storm days.

While our turn-by-turn ETAs during storm days were still too optimistic, the magnitude of the error is much lower, getting our estimates significantly closer to what drivers were experiencing on the road, but while the improvement is more visible in extreme conditions, live-data overall improves quality every day. To assess the overall contribution of live-data, we quantify the improvement of the typical+live estimates vs. typical-only in the same dimensions described above. In the figure below, the x-axis is the difference in the magnitude of the angle θ regardless it’s direction, thus positive numbers mean that errors of typical+live are more symmetrical than those of typical only. In the y-axis is the difference in r or the over-all error. The vast majority of days fall in the positive-positive quadrant, meaning that live-data improve both the accuracy and symmetry of our ETAs, and this happens for all days, not only those with irregular traffic.

Improvement in traffic quality when live-data is added.

Winter is coming (at least once a year)

Although snow season in the Northern hemisphere is over, it will be back again; and traffic patterns are not only exclusively affected by snow. Other extreme weather phenomena, demonstrations, sports, and other events also change the speed that cities move.

The patterns recovered in this post used only a very small fraction of the total telemetry collected in Denver (less than 0.5%), so harnessing the strength of the millions of miles we collect every day, would enable our live-traffic models to nail these predictions even more accurately in the future. Lastly, there might be some regularity in the traffic during these atypical events. While our logistics customers cannot plan ahead using live-data for a future date, if we know a storm is approaching, providing predictions during a “typical snowy day” might be a significant improvement over our general typical profiles.

Want to incorporate Mapbox traffic into your routing engine, navigation app or embedded system? Access traffic-aware ETAs and routing via the Directions API, Navigation SDK for Android and iOS, or add it to any map including HERE, TomTom, or OSM.

Camilo Sanín - Data Scientist - Mapbox | LinkedIn


Ground-truthing storm weather traffic ETAs in Denver was originally published in Points of interest on Medium, where people are continuing the conversation by highlighting and responding to this story.


Viewing all articles
Browse latest Browse all 2230

Trending Articles