| layout | title | permalink |
|---|---|---|
page |
Divvy Bike Share Project |
/divvy_write_up/ |
Divvy, the city-owned bike share program of Chicago, operated in partnership with Lyft, has seen substantial growth since its inception in 2013. However, a significant challenge that has emerged with this expansion is the imbalance in bike station capacities across the city. Some stations frequently experience shortages, while others have an excess of bikes. To manage this issue, Divvy has employed manual rebalancing strategies, involving paid workers who use vans to redistribute bikes from crowded stations to those with fewer bikes. Despite these efforts, the dynamic nature of station statuses, particularly during peak hours, underscores the critical need for advanced predictive modeling to enhance operational efficiency.
The program's success reflects the city's commitment to sustainable transportation, yet it also poses a challenge: ensuring bike and dock availability are balanced across an extensive network of stations. The problem arises when Divvy rebalancing workers leave for an empty station, but by the time they get there, it is full. Or they leave a general area with stations that will soon be wiped out by morning rush hour Divvy commuters.
This led me to explore predictive modeling, specifically the application of advanced algorithms that can predict potential imbalances 30 and 60 minutes ahead to give rebalancing workers an edge when it comes to route planning.
What makes this model so complex is that it is dynamic in both space and time. When you are predicting station status over time, each station can be thought of as a time series. These could be thought of as independent time series, but it is more complicated by the fact that riders ride from one station to another causing station statuses across stations to be correlated in space. This data would be classified as a joint time series. Joint time series can be tricky and most standard forecasting models such as linear regressions, ARIMA, or Facebook's Prophet model are not directly applicable. All of these could be coerced to fit the problem, but there are more specific models designed for problems like these.
This led me to the StemGNN model, developed by Cao et al. (2020). This model was originally designed for complex network systems like traffic flows in Los Angeles. Each traffic sensor was treated as a node in a network where traffic flowed over time between each sensor. Divvy bike station relationships over time present a very similar setup.
At the heart of this analysis is an expansive dataset sourced from the Chicago Data Portal, offering a granular view of bike availability at Divvy stations, updated every 10 minutes. Covering the years 2013 to 2022, the dataset initially contained an overwhelming 242 million entries. To manage this vast amount of data more efficiently, I focused on a subset from July to September 2022, coinciding with Divvy's peak ridership months. This period was chosen to ensure the data reflected the most dynamic user interactions with the bike-share system.
The study zeroes in on three specific community areas: Lincoln Square, Uptown, and North Center. These areas were strategically selected due to their unique positioning away from the Red Line—Chicago's busiest North-South passenger rail. This separation often compels residents to rely on Divvy bikes for East-West travel, thereby increasing and diversifying ridership within these regions.
Within these neighborhoods, 52 stations were identified (shown below), though only 41 were operational during the selected study period. Each station, or "node" as referred to in our model, plays a critical role in understanding the broader network of urban mobility facilitated by Divvy.
The challenge of forecasting bike availability in Chicago's Divvy program led to the adoption of a specialized model known as StemGNN. This model acts much like a meteorological forecast for bike distribution, predicting where and when bikes will be needed across the city.
The StemGNN model was meticulously fine-tuned to align with Divvy's specific needs, akin to customizing a precision tool for a particular task. Adjustments were made to optimize its learning rate and the amount of historical data it examines, ensuring the predictions were as accurate as possible.
This refined tool now offers a detailed forecast of bike availability, facilitating a smoother operation of the bike-share system. The insights provided by the model are instrumental not just for Divvy but also offer a window into the future of urban mobility management.
For an in-depth technical exploration, my full academic paper detailing the work with Divvy and the StemGNN model can be found here.
Forecasting Time Frames and Field Application
Given that Divvy updates dock data in 10-minute intervals and demand can surge unexpectedly due to rush hours or events, forecasting 30 and 60 minutes ahead—effectively 3 to 6 intervals into the future—offers Divvy's field teams critical foresight. This advanced planning capability is crucial, especially when already stationed at a dock. It allows teams to strategically plan their next moves in the field, enabling them to prioritize which stations to service next and manage stock levels more effectively, thereby enhancing their operational efficiency amidst the city's dynamic demand.
The performance of the model yielded impressive results, as shown below.
The metric used to gauge accuracy was the Root Mean Squared Error (RMSE), expressed as a percentage. This aligns with the data metric for station availability. The model achieved an average accuracy of 4.9% for the 30-minute forecast and 4.8% for the 60-minute forecast. Essentially, this indicates that the model's predictions typically deviate from the actual station usage status by about 5%.
However, RMSE alone doesn't provide a complete picture without a baseline for comparison. That's where the Mean Absolute Scaled Error (MASE) becomes valuable. MASE measures the model's Mean Average Error against that of a basic benchmark model—specifically, a naïve forecasting model that assumes no change over the forecast horizon. For instance, if a station is 70% full at the present moment, the naïve model would predict it to remain 70% full an hour later. This comparative approach effectively positions the StemGNN model against the simplest prediction method available.
The MASE values for our model further confirm its robustness: 0.59 for the 30-minute forecasts and 0.39 for the 60-minute forecasts. This implies that the 30-minute forecast model's error rate is only 60% of the error rate of the naïve model. In contrast, the 60-minute forecast model performs even better, with an error rate of 40% of the naïve model's error. With some arithimtic, you can see that the 60 minute forecasts improve accuracy by 8% compared to having no predictions at all. In layman's terms, for both forecasting horizons, the StemGNN model significantly outperforms the scenario of using no predictive model.
The predictive model demonstrates significant success in forecasting the availability of bikes at Divvy stations. While the Root Mean Squared Error (RMSE) values suggest an average prediction accuracy deviation around 5%, translating this into practical implications reveals its nuanced value. Specifically, an RMSE of approximately 5% equates to a potential forecasting error margin of 1-2 bikes per station. Additionally, the improvement in prediction accuracy of 8% for the 60-minute forecasts, when compared to a baseline of no model, might represent a marginal yet impactful enhancement in operational efficiency, varying by station size.
In practical terms, an 8% increase in forecasting accuracy may have varying degrees of importance depending on the context. For industries like automotive, where inventory costs are substantial, an 8% reduction in surplus could lead to significant savings. However, for bike-sharing operations, the critical question remains whether this level of improved accuracy in station availability forecasts substantively benefits the operational decision-making process. This evaluation ultimately rests with the users and operators of the system.
This study illustrates the capacity of machine learning, specifically through the adaptation of the StemGNN model, to forecast with notable precision within the urban bike-sharing context. The endeavor to refine model parameters such as learning rates and window sizes has yielded a model that predicts dock availability with a relatively small margin of error. This development not only offers Divvy an advanced tool for enhancing its service reliability but also lays the groundwork for broader applications of similar predictive strategies in urban mobility and beyond. Such advancements hold promise for cities globally, proposing a shift towards more data-informed management of transportation infrastructures.