Dockless Mobility: Understanding the usage and impact of electric scooters in Austin
This project was built by Aashima Garg, Aneesh Soni, Brian Wilmarth, Vaidehi Duraphe, and Zarif Choudhury. Our data and visualizations can be found here.
Background
Since the Spring of 2018, dockless mobility has dominated coffee shop buzz, tech news, and the streets of many cities world-wide. Not only did independent companies like Lime and Bird set records for start up growth and funding, but also pre-existing ride share companies like Lyft and Uber added the last-mile transportation to their top agendas and entered the dockless mobility market.
But what is dockless mobility? According to the City of Austin, dockless mobility systems consist of devices, such as bicycles or scooters, that do not require fixed docking stations for users to receive or return units. Many companies, some of which are mentioned above, manufacture electric bikes and scooters, deploy them across urban areas, and rely on crowdsourced charging to maintain their vehicles’ lifecycles. This means a combination of operations specialists and everyday consumers collect, charge, and deploy dockless vehicles around cities every day to help these companies maintain operations.
In the past year, the rise of dockless mobility has provided a cheap, fast, and easy way for people to travel a couple miles at a time, without needing to hail a more expensive cab, Uber, or Lyft, but it has also posed a high risk for users who ride vehicles during rush hours, without proper protection, and/or are under the influence. This has forced many city governments to create physical infrastructure as well as many new rules and regulations to ensure the safety of their people. This includes, but is not limited to, restricting vehicle usage and deployment to certain areas, building more bike lanes, creating electric vehicle speed limits, and monitoring users for helmet usage and age restrictions (Los Angeles restricts scooter usage to people of age 18+ by law!).
We found that over the past year, roughly 1,200 electric, dockless bike trips are taken per day and 11,500 electric, dockless scooter trips are taken per day in Austin. Due to the high number of trips per day for scooters and rising scooter usage over time, we decided to focus our analysis of dockless mobility on scooters.
In this particular report, we analyze the impact of electric, dockless scooters on Austin city infrastructure, and we also offer some suggestions for scooter companies to optimize vehicle deployment in Austin for customer usage and safety as well as their own revenue.
Formatting and Cleaning Data
The main dataset we worked with contains dockless vehicle trip data reported to the City of Austin Transportation Department as part of the Dockless Mobility Service operating rules. To clean and format the dataset, we first filtered out any extraneous locations that was from outside Austin city boundaries (we found some locations from China!). Next, we filtered out any erroneous trips by removing trips over 75 miles in length. A dockless scooter’s has a maximum single use range of 20 miles and a dockless bike has a maximum single use range of around 50 miles — All trips over 75 miles were considered invalid.
To visualize the data, we generated a feature correlation chart to find any strongly correlated features that we could begin our data discovery from. However, the correlation factor for most feature pairs was under 0.3, and all factors greater than this were already similarly paired features (i.e. Start Latitude and Start Longitude).
Scooter Usage
In Austin, we found that scooters are used primarily around the downtown area, as you can see on the heat map below.
After implementing a hierarchal clustering algorithm, which we discuss further, we found three primary neighborhoods where scooters are used most.
In these areas, we found that the longest rides are taken from South Congress (SoCo) and Zilker, which makes sense considering these areas of Austin are more isolated and difficult to get to and leave. This shows us that trip taken from SoCo and Zilker are longer, leisure rides and trips from West Campus are shorter, more direct rides, probably to commute to work or class.
Additionally, we discovered that trip distance, in our data set, was collected by taking the difference between starting latitude and longitude. For this reason, we determined trip duration would be more insightful. We found that most trips are between 200 and 250 seconds or only 4 minutes.
To better understand some other generic trends that are not location specific, we used the unsupervised learning technique of hierarchical clustering and mean shifting from the scikit learn library. This works by considering each data point an individual cluster at first and after each iteration, similar clusters merge into one until k clusters are formed. The number of clusters is decided by the algorithm as it sees fit. The “similarity” of a cluster is determined by the proximity of data points to one another.
The first clustering we chose to explore was using the starting and ending latitudes and longitudes in order to get a better idea of where people begin and end their rides outside of the three larger hotspots we displayed earlier. We found there were approximately 23 places in total where rides typically start and end in Austin.
Using the same data (starting latitudes and longitudes), we were also able to create a second visualization that shows how starting positions have changed over time. This change reflects the increase in number of scooters in Austin as the number of trip starts continues to grow over time. However, we wish we had data on the number of scooters in Austin to better estimate demand and each scooter’s deployment position at the start of a given day.
The second clustering we built was a 3D cluster using the Axes3D from the mpl_toolkits library and the features of Month, Hour, and day of the week. The algorithm estimated one main cluster which was in September 2018, particularly on Wednesdays typically around 2–3pm. This made sense to us since students had just gotten back to Austin from Summer vacation and were probably very excited to use lime scooters after getting out of afternoon classes. Another interesting find from the clustering visualization was that people generally just take rides in the Fall and more so in the middle of the day or early in the day as opposed to late night. People generally rode less in the cold Winters and hot Summers.
The last clustering we made was for trip duration and trip distance to identify how people were using the scooters. There were an estimated 14 clusters but after looking at the visualization, it was clear that there were mainly two types of rides. The first type of ride is 5–10 minutes travelling 1/2–1 miles, generally shorter rides with purpose to get from A to B, my guess being rides to class after possibly waking up late, picking up lunch, or running from meeting to meeting in downtown. The other type of ride was 50–90 minutes travelling 4–6 miles with less of a purpose and possibly more for exploration of the city of austin or just enjoyment of using the scooters.
Scooter Usage Trends
We also built some charts to visualize trends in our dataset, as opposed to finding key, distinct insights using clustering.
Demand dips around holidays and in cold weather months. These macro trends can also be explained by fluctuations in the number of scooters deployed as scooter companies enter and exit the marketplace, especially the increase in rides starting in August.
Demand tends to be lowest on Mondays. Demand consistently rises during the week and peaks on Saturdays.
On Mondays through Thursdays, we observed peaks at noon and 6pm, with initial demand rising starting at 6am. This corresponds to the lunch and evening rush hour respectively.
On the weekends, we observe another consistent trend. Saturday consistently has the highest number of rides, which peaks around 6pm. Fridays and Sundays have a more consistent number of rides per hour from around 11am to 7pm.
This regularity of these trends suggest that the day of the week and hour of the day would be very important features in a demand forecasting model.
Weather Influence
2018 was Austin’s 3rd rainiest year in the past decade, with the heavy rain continuing into the first half of 2019. Considering this, our team further investigated the fluctuations in electric scooter demand during periods of precipitation. We combined hourly weather data from the Southern Regional Climate Center with our Austin Dockless Mobility Dataset to confirm assumptions that dockless vehicle trip frequency is inversely correlated with the amount of hourly precipitation, measured in inches.
The chart above compares trip frequency with rainfall over a three week period, roughly. This chart is zoomed in on a subset of our data so that it is easier to see. As suspected, days with a larger amount of hourly rainfall saw dramatically fewer scooter trips, suggesting that predicted rainfall could be very useful as a feature in a short-term demand prediction model.
Next, we wanted to explore the impact of weather on accidents.
There isn’t a clear indication here that precipitation causes more accidents. Accidents occur pretty regularly.
There is an interesting drop in accidents on the week of October 14, which was particularly rainy. It would seem that the drop in accidents should not be attributed to the rain, but to the reduced number of rides because of the rain.
From our data, it is difficult to say confidently that precipitation causes more accidents. Although the increased risk is clear for other reasons. There is a clear example, however, that suggests precipitation reduces the number of rides, which subsequently results in fewer accidents.
Accidents
Additionally, since electric scooters hit the streets of Austin and other large metropolitan areas, there has been a 160 percent spike in Hospital ER visits involving electric scooters for fractures, dislocations and head trauma.
At the end of 2018, the Center for Disease Control announced that it would be conducting a study of the health risks of dockless mobility, specifically scooters, by analyzing the incidents or injuries that occurred in Austin between September 5 to November 4. The study focuses on 37 EMS calls and 68 scooter injuries reported through syndromic surveillance from local hospitals. The time period examined included hundreds of thousands of rides — in October 2018, there were approximately 275,300 scooter rides totaling to more than 264,000 miles. From the study, the CDC found that 45% of incidents relating to electric scooters involved head injuries, determining that the use of helmets would prevent many visits to the ER.
For further discovery on these electric scooter accidents, our team examined 160+ accidents from September 5th to November 29th to find incident trends as they pertain to gender, time of day, and location.
From a first look at the data, we found that 58.8% of patients were male, compared to 41.2% female.
Combining this data in a bar graph with the time the accidents took place, we can compare accidents during the time of day by gender.
From this chart, we gather that the largest discrepancy between genders is during the evening and night time, where there is a difference of almost twice as many males to females involved in electric scooter related incidents. These findings are not surprising, as men have consistently ranked higher than females in terms of car accident and fatality rates — In 2016, the car accident fatality rate for male occupants aged 16–19 was twice the rate of their female counterparts.
Beyond gender, our team suspected that e-scooter incidents are more frequent in certain areas of Austin such as downtown or west campus, where majority of scooters are deployed. Using the incident addresses listed with the CDC accident data, we plotted the locations using Folium, a powerful Python library that helps create several types of Leaflet maps.
KEY
Green = Morning
Red = Afternoon
Light Blue = Evening
Purple = Night
Dark Blue = Late Night
As the map shows, most of the Austin accidents are centered around Downtown and West Campus. These findings are cohesive with our clusterings of electric scooter trip start and end locations, as most electric scooters trips occur around these regions.
Factoring time into our analysis of accidents, one interesting finding is that accidents occurring past in the evening, night, and late-night time periods most frequently occur in the Downtown Area. This could potentially be attributed to the number of bars in the downtown area, suggesting that a combination of drinking and riding scooters is more likely to result in accidents. A solution to the risk that drinking and riding scooters creates could be for scooter companies to incorporate a number of preventative features that ensure a sober state of their riders. These could potentially includes red zones where scooters are locked after a certain time at night, a modified speed limit in accident-prone areas, or a simple sobriety test riders must complete before unlocking a scooter. Additionally, scooter companies could partner with city governments to enforce a higher level of safety on scooters. For example, riders in Los Angeles are required to be above 18 years old and wear a helmet while riding. These requirements are authority-enforced by ticketing and fines.
Conclusion
Going into the project, we would have liked to build more models to predict trends and usage for scooters. However, our limited data restricted the scope of our project immensely. Nevertheless, from our findings, we suggest the following to scooter companies:
- Restrict speed or access to scooters around the downtown area at night (shown below)
- Add tests for users to complete to ensure sobriety before riding scooters around the downtown area at night (shown below)
- Offer discounted rides during holidays to incentivize rides
- Deploy more scooters around the South Congress / Zilker area to incentivize users to take longer rides
And we encourage city governments to:
- Build bike lanes around the three main areas where rides are taken in Austin (Downtown, West Campus, and South Congress / Zilker ) to ensure safer rides
- Enforce safety regulations more (following the example set by Los Angeles) by ticketing riders for riding under the influence, without a helmet, and under certain age restrictions
As far as future work and considerations go, we hope to collect more granular data — specifically regarding supply and deployment of scooters, trip routes, trip cost calculators, and more exact start and end times (our dataset rounded these values to the nearest 15 minutes) — to pursue the following analyses:
- Understanding the affect of supply on rides per day
- Calculating the percentage decrease of rides per hour based on varying levels of precipitation
- Finding exact streets where bike lanes should be built
- Understanding the difference in profit between some long rides to numerous short rides
- Modeling demand forecasts
- Calculating average idle time for scooters and understanding how to reduce it
