AV demand forecasting

Approach

Lacking empirical data, we turned to regional travel demand models that are used by cities to predict travel patterns. We modified Miami's regional travel demand model to add a new "AV ride hail service" and derive a realistic set of “ride hail trip requests” in our service area. These trips were then used as inputs for a fleet simulation model capable of simulating the response times of our AV fleet accounting for operational constraints and traffic delays. 

The demand model is agent-based, meaning it has a synthetic population of customers, and probablistic, meaning the agents' mode choices can change in each model run. This allows for sensitivity testing. Its trip data included information about mode, party size, and trip purpose which allowed us to gain insights into customer segments and pent-up demand among households lacking a car. The fleet simulation also allowed us to analyze the impacts of adding vehicles to the AV fleet on customer wait times.

Sensitivity Analysis

User choices are modeled taking the expected cost and travel time of each mode into account. My role was to re-calibrate the user mode choice model to allow for sensitivity testing of various fare levels and wait times, as well as test the potential for ride sharing. I defined parameters for fares, wait times, and detours in the fleet simulator and  then ran a suite of scenarios testing three types of incentives. Each incentive was tested at 25%, 50%, and 75% levels:

• Discounted fares

• Faster in-vehicle travel times

• Reduced wait times

Results

The derived list of “ride hail requests” consisted of approximately 100,000 customers making 170,000 unique trip requests. The majority were for single occupant rides, and about 10% were for shared rides.  Since it was derived from a model, we reality-checked the spatial and temporal distributions against empirical traffic and ride hail data and found them reasonable.

My demand sensitivity analysis revealed that customers are much more sensitive to fares than to wait times. For example, offering 50% discounted rides resulted in 3x sales while cutting the wait time by 50% only increased sales by 50%. Higher income households were the most sensitive to reduced wait times, increasing their ride hail usage the most. The service was increasingly used for work trips as wait times declined. We also found that customers not sensitive to reductions in travel times, that is, if the car arrived after a short wait, they were not concerned about a long journey time.

We also tested how fleet size could reduce customer wait times. Our analysis showed a that ride sales increased as average operating speed increased. This demonstrated that we would need a much larger fleet to achieve the same level of ride sales, if our vehicles were not allowed to use highways. However, we found that as fleet size grows, vehicle utilization efficiency declines, ranging form 2.2 to 4 trips per hour.

Impact

Our findings directly informed strategic business decisions and service planning prior to launch. Ultimately, the AV Demand Forecaster  simulations demonstrated the most efficient fleet for this service area and fare structure was 100 vehicles operating on local streets, due to slower trips with higher per-minute fares. We were able to recommend how to achieve16% improvement in fleet operational efficiency and 24% higher per-mile revenues. I was offered a role on the Corporate Strategy team serving the Board of Directors, focused on autonomous vehicle development strategy.