E.V.A.C

Electric Vehicle Assisted Charging

​

Where should the next electric vehicle charging station be placed?

UC Berkeley: Data-X (Team 17)

Avinash Jain, Jason Ma, Nimalen Sivapalan,

Perry Trinh, Shoumik Jamil

Background

What’s going on in the EV space?

Background

• 1.1 Million Electric Vehicles are currently on the road in the U.S.
• Currently have 2.1% market share → expected to rapidly grow
• Cumulatively 5 million electric vehicles around the world

​

• Volkswagen aiming for 22 million EV vehicles by 2030
• Every single automobile company focused on producing EVs

​

• California Governor Jerry Brown signed Executive Order B-48-18
• 5 million EVs in California by 2030, and 250,000 EV chargers in the ground by 2025.

​

Stats reported from: CNBC, CNN, Wikipedia

As the number of electric vehicles exponentially grows, there becomes an immediate need for more charging stations around the country

“

Solution

What’s the solution to this problem?

E.V.A.C

​

​

A Machine Learning model that can accurately predict the best locations to implement electric vehicle charging stations in Columbus, Ohio

​

​

In partnership with:asdasdasdasd

Approach & Results

Process and discovery

Approach

• Built a machine learning model that recommends the best places in Columbus, Ohio to implement electric charging stations
• Classified a random set of points within Columbus as good or bad locations for electric vehicle charging stations
• Tested various models (K-Nearest Neighbors, Neural Networks, SVMs) before settling on Logistic Regression

​

• Factors that we took into account:
• Population Density
• Points of Interest
• Police Cameras and Security
• Natural Disasters

​

Population Density

• Our model uses Census Data to calculate what the population density is of a given coordinate
• This photo is a simple visualization of the granularity we had for our data (using Census Block Groups)

Points of Interest

• Our model considered all kinds of points of interest - retail, offices, education, public places, etc.
• We classified the categories into three groups - 0, 1, and 2, in terms of relevancy to having an EV charging station

Police Cameras and Security

• Utilized Police Camera dataset to figure out safe zones throughout Columbus, OH

​

• Calculated distance between latitude/longitude of each camera with our locations

​

• Utilized 1’s and 0’s to classify whether E.V. Charging Stations were in safe locations or not

​

Natural Disasters

• Columbus has suffered mainly from tornadoes over the last half-century
• Our model uses tornado distance and magnitude as a proxy for tornado risk

Locations of 40 Tornadoes in Columbus, 1950 - 2010

Architecture

Choosing, building, and refining our model’s architecture

Architecture of Solution

• Model is given a dataset of random coordinates that it can classify as either good or bad for an EV charging station
• Model is also given coordinates of existing charging locations as reference of good charging locations
• Factors like population density and closest points of interest are automatically determined for each coordinate
• If a coordinate is missing a data type, we assign it to be the same value as the nearest available coordinate that is at most 1-2 miles away (distance varies for each data type)

Data Model

Location Data

Points of Interest

Population Density

Security Cameras

Tornado

Data

Clean up Data

Combine Datasets

Data Processing

Modeling

Create train/ validate/test sets

Train our proposed models on training data

Validate different models and parameters

Test our final selected model

Classifying locations for new chargers

Outputs

Generating visualization/UI of chargers

Machine Learning Model

• We tried different approaches to this classification problem:
• Logistic regression
• K-nearest neighbors
• Neural nets
• Soft-margin SVM
• Kept 10% of our data to test our final model
• Kept 10% of our data to validate our models and hyperparameters
• Normalized the population density and tornado distance columns
• Ensured that all columns were equally weighted

Model Tuning

• Logistic Regression and SVM:
• Tuned regularization parameter C using our validation set
• Found the radial basis function in SVM to be the best kernel
• K-nearest neighbors:
• Experimented with different K-values → Selected 3 based on validation accuracy
• Neural network:
• Experimented with different architectures with varying numbers of neurons and layers
• Final neural net consisted of → 4 hidden layers that had 150, 100, 50, 25 neurons respectively

86.67%

86.67 %

80%

80%

Logistic Regression Validation

Neural Net Validation

3 - Nearest Neighbors Validation

SVM Validation

Results

• Training and Test set:
• 73% of existing E.V. charging stations classified correctly
• 92% of locations without charging stations classified correctly
• Prediction:
• Classified 100 random points without charging station
• Good or Bad for building E.V. Charging Station
• 8 of these points were recommended

81.10%

82.28%

87.50%

Training accuracy

Overall accuracy of logistic regression model

Test accuracy

Good Location:

Fairgrounds, Columbus, 43211

• Coordinates: 39.9994337, -82.9887561
• Points of Interest:
• Public Place
• Police training academy
• State expo center
• Hotel
• Population Density: �1195 people / square mile
• Security camera present

User Interface

Design choices

User Interface

• The intended user interface is a Google MyMap with the list of points that the model predicted on
• The UI has three main components:
• General boundary of the area of Columbus
• Markers for points that are existing E.V. charging stations in Columbus, and our model accuracy for those points
• Markers for points that our model classified as good locations to implement new E.V. charging stations

Process for user interface

Train/test/valid. on 158 locations

79 existing &

79 random (no E.V.) locations

Predict on 100 random locations in Columbus, OH

Find 100 random locations in Columbus, OH

User Interface

• ​

Learning Path

From start to finish

Initial Design Decisions

• Brainstormed a variety of topics useful for Honda before ultimately deciding on Electric Vehicle Charging Stations�
• Highway safety with vehicle to vehicle communication�
• Managing collisions based on road conditions
• Decided that population density, points of interest, and safety (using security camera and natural disaster proximity) were relevant factors to use to determine best charging locations
• Concluded that a classification model (yes or no to good charging spots) was best for our project, as opposed to assigning each location a score or rank

Ideas We Tried

• Intended Extensions to EVAC Project (potential future upgrades)
• Incorporate more factors like public transportation, prevalence of crime, and neighboring gas stations
• Look into large cities like San Francisco and Austin
• Additional feature on map to filter by various metrics
• Real time recommendations on interactive map given a user’s custom coordinates

Reaching the Solution

• Created datasets based on open source metrics data from SmartColumbus, US Census, and ChargeHub
• Selected Logistic Regression for our model due to high accuracy and simplicity
• Used Google MyMaps as our user interface to visualize the model’s predictions

THANK YOU!

Any questions?

Github repository: https://github.com/data-x-sp19/EVAC

​

More info on how to use this template at www.slidescarnival.com/help-use-presentation-template

This template is free to use under Creative Commons Attribution license. You can keep the Credits slide or mention SlidesCarnival and other resources used in a slide footer.

​

​

UC Berkeley: Data-X (Team 17)

Avinash Jain, Jason Ma, Nimalen Sivapalan, Perry Trinh, Shoumik Jamil