The problem proposed to us was to forecast Electric Vehicle adoption rate from now (2018) through 2050 and its effect on oil demand. The problem was given by Pioneer Natural Resources company. The expectation of the problem sponsor company (Pioneer) was to see a model which is flexible enough to add/change inputs that affect the EV adoption rate and oil demand in future and re-forecast the numbers. This would allow them to tailor their business strategy on the future breakthroughs in EV technology.
[We were one of the 9 finalists and we won in the Pioneer sponsor theme]
We approached this as a machine learning problem.
EV Adoption Rate ML Model: After collecting data on factors that affect EV adoption rate, I developed and trained the a nonlinear gradient descent model on the data from 2010 through 2016 and used it to predict the rate for 2017. Then I used this predicted rate (of yr 2017) along with the published “predicted” input features as training inputs to the model and re-trained the model from 2010 to now 2017 – and then predicted the rate for 2018. We executed this process repeatedly to predict the rates till 2050.
EV adoption rate = function (initial cost, battery infrastructure, charging infrastructure, govt subsidies and crude oil price).
Oil Demand Model: We used the same model architecture with different input features to predict the oil demand from now through 2050. The EV adoption rate was one of the inputs to this model
Oil Demand = function(EV adoption rate, total supply of oil, total demand for oil, demand for renewables, policy changes)
To validate the accuracy, we trained the data from 2010 to 2016 and then predicted the outputs for 2017. We then compared these to the actual published numbers of 2017. The error between the predicted and published numbers is shown below
PyQT GUI design:
To make the model more flexible, I designed a GUI based off python which allows the user to select a CSV file for his inputs. He can disable/enable various inputs as needed to isolate the effects of each feature on the outputs.
Flexibility of the Model:
Incorporate tech breakthroughs:
e.g If tomorrow there is a breakthrough in battery technology which brings the $/KWhr price of the pack down by 20%. The user can then input only this data and disable all other factors in the model to predict how the EV adoption rate (and hence oil demand) would get affected due to this breakthrough.
Influence business decisions:
e.g. This model was designed to give stakeholders enough information to influence their business strategy. Say if Ford has decided to sell x% EVs by 2020, and want to evaluate how many charging stations they need to install in order to reduce the range anxiety of their customers and hence increase the EV adoption rate, they could now do this using this model and inputting variety of data for charging infrastructure to see what numbers helps them achieve their targets. This would then help Ford to allot the needed budget for the project.
Annual Energy Outlook 2017 (with projections to 2050) : https://www.eia.gov/outlooks/aeo/
Good article on EV trends: https://www.iea.org/publications/freepublications/publication/GlobalEVOutlook2017.pdf
International Energy Outlook 2017:
Charging infra projections
EV number prediction article
Predicting about big crash
The Electric Car Revolution is Accelerating
Dynamics in the Global Electric-Vehicle Market