This is a result of a hands-on seminar during my Master Studies in MIS at University of Bamberg, Germany. The task of this project (roughly): Based on data, predict if a specific electric car goes well with a specific person. Imagine you wanna buy a new electric car and you wanna find the one that fits your needs best. See the summary of this project here: https://medium.com/@mathiasrenner/predicting-the-utility-of-an-electric-vehicle-2091e0823799#.ea5vvb6vb

1. p. 1Mathias Renner | Hands-On Seminar, Predicting Utility of E-cars
Predicting the utility
of an electric vehicle
for a specific mobility behavior
Hands-on Master Seminar
Mathias Renner
2016-07-19
Energy Efficiency Systems Group, University of Bamberg

2. p. 2Mathias Renner | Hands-On Seminar, Predicting Utility of E-cars
Agenda
1. Problem statement and research question
2. Analysis along procedure model for regression
• Data overview
• Data analysis
• Results
3. Lessons learned

3. p. 3Mathias Renner | Hands-On Seminar, Predicting Utility of E-cars
Problem statement of potential buyers:
Does this e-car fit to me?
• Potential buyers of electric vehicles face a high level of uncertainty:
 Can be quantified as
“How much of my goals (in percentage) can I reach?”
• This work’s goal:
Based on existing driving data, predict the percentage of reached goals for
a new person.
 Machine learning, predictive analytics

4. p. 4Mathias Renner | Hands-On Seminar, Predicting Utility of E-cars
Use case: Tesla

5. p. 5Mathias Renner | Hands-On Seminar, Predicting Utility of E-cars
Use case: Tesla

6. p. 6Mathias Renner | Hands-On Seminar, Predicting Utility of E-cars
The classical approach to evaluate utility
230 km
+ =
Source of image on the left:
www.elektroauto-news.net

7. p. 7Mathias Renner | Hands-On Seminar, Predicting Utility of E-cars
The classical approach has drawbacks.
• Result? Max. range Reached goals in percentage
• Practical? Limited due to idealistic Very precise, can be very practical
test conditions and
generalized data
• Data needs? Data only once per car New data for each car and driver
Classic Approach vs. „New“ Approach

8. p. 8Mathias Renner | Hands-On Seminar, Predicting Utility of E-cars
Procedure model:
Predictive data analysis aka. supervised machine
learning with regression models
Source: Adapted from lecture „Business
Intelligence and Analytics“, Sodenkamp 2015

9. p. 9Mathias Renner | Hands-On Seminar, Predicting Utility of E-cars
Data overview
Car ID (1-50) Long./Lat. Highway/Urban Level of battery
… … … …
… … … …
… … … …
… … .. …
500kdatapoints/rows
Ca. 10 features

10. p. 10Mathias Renner | Hands-On Seminar, Predicting Utility of E-cars
25 features have been extracted.
Distance on
Highways
Parking time
per trip
Avg. speed
per trip
Driving time
on weekends
…

11. p. 11Mathias Renner | Hands-On Seminar, Predicting Utility of E-cars
Linear regression model is not applicable.
Linear regression model Logit regression model

12. p. 12Mathias Renner | Hands-On Seminar, Predicting Utility of E-cars
Results: One of univariate models
Reachedgoalsin%
Single trip distance in KM

13. p. 13Mathias Renner | Hands-On Seminar, Predicting Utility of E-cars
Results: Final Model
glm(formula = REACHED_GOALS_SOCBOTH ~ SPEED_AVG_KMH + … )
Estimate Std. Error t value Pr(>|t|)
(Intercept) 5.496e+00 6.778e-01 8.108 2.33e-09 ***
SPEED_AVG_KMH -9.935e-02 1.557e-02 -6.381 3.16e-07 ***
DISTANCE_HIGHWAY_KM -6.441e-05 2.257e-05 -2.854 0.00741 **
ROUND_TRIP_DISTANCE_AVG_KM 8.750e-03 3.466e-03 2.525 0.01656 *
ROUND_TRIP_PARKING_TIME_AVG_H -5.590e-02 1.740e-02 -3.212 0.00294 **
SINGLE_TRIP_PARKING_TIME_AVG_H 2.871e-01 9.314e-02 3.082 0.00413 **
ROUND_NUMBER_TRIPS_AFTERNOON 2.480e-04 1.070e-04 2.319 0.02676 *
Pseudo R2: 0.75
Limitation: Correlations!

14. p. 14Mathias Renner | Hands-On Seminar, Predicting Utility of E-cars
Results: Predicting with model 2
Reachedgoalsin%
Car #
First run
Accuracy:
89,9 %

15. p. 15Mathias Renner | Hands-On Seminar, Predicting Utility of E-cars
Results: Predicting with model 2
Reachedgoalsin%
Car #
First run
Accuracy:
89,9 %

16. p. 16Mathias Renner | Hands-On Seminar, Predicting Utility of E-cars
Results: Predicting with model 2
Car #
5-fold Cross
Validation
Accuracy:
Ø 89,9 %

17. p. 17Mathias Renner | Hands-On Seminar, Predicting Utility of E-cars
Lessons learned
• R has superpowers. It‘s only up to the user to leverage it.
• Lecture Data Analytics in Energy Informatics has been very helpful to
actually understand
• Lecture Business Intelligence and Analytics has been helpful for general
procedures for data analytics
• Performance of R is highly dependent of efficient algorithm description

18. p. 18Mathias Renner | Hands-On Seminar, Predicting Utility of E-cars
Thank you very much for your attention.
Source: Adapted from http://fontawesome.io/

19. p. 19Mathias Renner | Hands-On Seminar, Predicting Utility of E-cars
Backup-Slides

20. p. 20Mathias Renner | Hands-On Seminar, Predicting Utility of E-cars
Distribution of dependent variable

21. p. 21Mathias Renner | Hands-On Seminar, Predicting Utility of E-cars
General procedure model:CRISP-DM

22. p. 22Mathias Renner | Hands-On Seminar, Predicting Utility of E-cars
The classical approach has drawbacks.
• Result? Max. range Reached goals in percentage
• Significant? Limited due to Siginifcant for each driver‘s
generalized data behavior
• Practical? Limited due to idealistic Can be very practical
test conditions
• Data needs? Test data only once New data for each car and driver
Classic Approach vs. „New“ Approach