1 of 37

Philadelphia Smart Loading Zones: Analysis and Demand Prediction

Samriddhi Khare, Michael Dunst, Tiffany Luo, Ling Chen, Shengqian Wang

2 of 37

In October 2022, the City of Philadelphia introduced a pilot program that tested 20 paid, curbside loading spaces for delivery drivers in Center City, known as “smart loading zones.”

Delivery companies were able to reserve spaces and times through a smartphone app.

Img Source: https://www.phila.gov/programs/smartcityphl/smart-loading-zones/

3 of 37

The data collected from this pilot program offers interesting insight into reservation patterns that can be used for potential expansion of the program, answering the question:

Where should smart loading zones be located?

Add interesting viz here?

Img Source: https://whyy.org/articles/philly-to-test-new-tech-in-bid-to-stop-delivery-drivers-from-parking-illegally/

4 of 37

Challenge:

Without a strategically located smart loading zones, drivers have no choice but to double park, or spend too much time looking for vacant spots around the city.

Solution:

Our application will serve as a back-end tool that will allow users to predict demand for curbside loading spaces at any location.

5 of 37

Clients

Miriam Cherayil

Smart Infrastructure Project Manager

Akshay Malik

Smart Cities Director

Chris Shelley

Smart Mobility Coordinator

6 of 37

DATA

MODEL

APP

7 of 37

Study Area &

Pilot Site

8 of 37

Study Area &

Pilot Site

9 of 37

Study Area &

Pilot Site

10 of 37

Client Data from the Pilot

  • Booking Data: The number of reservations made through the app, as well as violations (incidences of drivers using the space without booking)
  • Curb Data: Point and linear assets along each curb such as trees and stop signs

11 of 37

Challenge: Differing Hours of Operation

Early morning

Morning

Midday

Evening

Late night

Walnut St

Sansom St

Chestnut St

12 of 37

Challenge: Differing Hours of Operation

Early morning

Morning

Midday

Evening

Late night

Walnut St

Sansom St

Chestnut St

Model Data

13 of 37

Client Data from the Pilot

Average Bookings per Day of the Week

Walnut

Sansom

Chestnut

14 of 37

Understanding Demand for Curb Space

What is the process behind a vehicle choosing to stop alongside a curb?

15 of 37

Understanding Demand for Curb Space

What is the process behind a vehicle choosing to stop alongside a curb?

What might a driver need to stop for?

Distance to the nearest location of a variety of land uses: OpenStreetMap

16 of 37

Understanding Demand for Curb Space

What is the process behind a vehicle choosing to stop alongside a curb?

What might a driver need to stop for?

Distance to the nearest location of a variety of land uses: OpenStreetMap

How many vehicles use the road and why?

Official road classifications

17 of 37

Understanding Demand for Curb Space

What is the process behind a vehicle choosing to stop alongside a curb?

What might a driver need to stop for?

Distance to the nearest location of a variety of land uses: OpenStreetMap

How many vehicles use the road and why?

Official road classifications

What is the time of day?

Separating bookings into times of day

18 of 37

Correlations

Strongest relationships

  • More bookings: closer to retail, civic and institutional buildings, major roads
  • Fewer bookings: closer to clinics and community centers

Strength of Variable Relationships

19 of 37

Initial Data Exploration

External Data

OpenStreetMap Data, Bike and Road Network Data

Client Data

Booking Data, Curb Use Data, Vehicle types

Final Model

Based on correlations, the final variable list included information from a variety of sources: time series information, distance from amenities

20 of 37

DATA

MODEL

APP

21 of 37

1: CREATING NEW VARIABLES

Filtering the data, finding the best variables and creating predictive fixed effect and KNN.

5. CHECKING STATISTICAL ERRORS

Processing the results to see if there are spatial or statistical error clustering,.

4. CROSS VALIDATION

Checking the generalizability of the model using cross validation.

2. IDENTIFYING PREDICTIVE FACTORS

Running further analysis to see booking trends,

MODEL PROCESS

3. MODELLING THE REGRESSION

Running a random forest regression to predict and visualize loading zone demand.

22 of 37

Random Forest Model

Random forests is based on decision trees, where each tree is grown using a random subset of the data set.

mtry=17, ntree = 1000, nodesize = 22, maxnodes = 20

Reduced Overfitting Risk

Improved Model Accuracy

23 of 37

Model Choices and Results

Booking Events by Week and Day

Dependant Variable

Independent Variables (with highest performances)

Civic Center (Nearest Neighbor)

Road Class and Distance to Each Class of the Roads

Time

Top 20 Variables

Major

Arterials

Minor Arterials

24 of 37

Model Accuracy Evaluation: Spatial & Temporal Comparison (MAE)

Spatial

Temporal

25 of 37

Model Results

  • Several weeks where the observed bookings significantly exceed the predictions: potential special events or seasonal peaks

  • Performs better when average bookings are below 1 (good for occupancy check)

Performance of a Random Forest model used for predicting bookings on a weekly basis

26 of 37

Model Evaluation for Occupancy Prediction

The model correctly identifies 79.35% of actual 'occupied' cases

Sensitivity

The model correctly identifies 61.94% of actual 'unoccupied' cases

Specificity

Around 74.52% of the model's predictions are correct.

Accuracy

27 of 37

Generalization

28 of 37

Generalization

29 of 37

Generalization

30 of 37

Generalization

31 of 37

Generalization

32 of 37

DATA

MODEL

APP

33 of 37

Link to Application

Link to Walkthrough

34 of 37

Recommendations for Refining the Pilot Program

Either throughout the entire day or during consistent peak periods, instead of varying hours for different zones

Standardize the operating hours for all curb zones

Distribute curb zones more broadly across the city

Strategically place zones in a variety of neighborhoods, beyond just the center city area

Collect More Data Spatially, Temporally, and Scalably for More Generalizable Results

Problem

Solution

No Consistent Operating Hours

No Data in Diverse Neighborhoods of the City

35 of 37

Recommendations for Refining the Pilot Program

Evaluate the effectiveness across different seasons

Extend the duration of the pilot program

Enhance the visibility of the pilot app

Collaborate with widely-used applications for integration or advertisement.

Avoid the limitations of promoting a standalone new app

Collect More Data Spatially, Temporally, and Scalably for More Generalizable Results

Problem

Solution

Limited temporal data (October to April Only)

Limited App Usage

36 of 37

Conclusion: Takeaways & Next Steps

  • Improve Model Performance
    • Our current model predicts better for lower values, with more time we would like to ensure it predicts well for a range of values
  • Planning vs. Operational use case
    • Thinking about loading zones gross demand per day/week or their occupancy levels at granular levels of time
    • Pay particular attention to weekly patterns or specific events that affect the use case: predicting demand for curbside loading spaces

37 of 37

Thanks!

Samriddhi Khare, Michael Dunst, Tiffany Luo, Ling Chen, Shengqian Wang