Electric car driving range prediction
- 0 Collaborators
Range anxiety is a major issue that discourages drivers from considering electric cars as an alternative mode of transportation to the standard combustion engine vehicle. This project aims to provide such drivers with an upper bound prediction of travel time if they were to use an electric vehicle based on their own driving habits. The goal is to help such drivers estimate whether it would be reasonable for them to consider this more environmentally-friendly alternative without significantly impacting their habits. ...learn more
Project status: Under Development
Overview / Usage
In this project, a battery storage system in electric cars has been simulated under a very basic model of energy transfers. (This is a very common system consisting of two storage devices, and can be found in many cars today). AI is used in the form of Reinforcement Learning, to determine the most optimal robotic control of the energy transfers for the given driving route. By learning from the user's historical driving patterns, the software agent is able to both minimize energy losses and increase the battery lifetime, to allow the driver to travel further in the future. As far as we know, such techniques are currently only at the research stage today, but this simulator will approximate the time with some accuracy provided that the driver can input his/her driving speed data (which could be easily measured from an odometer during the trip).
Methodology / Approach
The model-based RL framework used to solve the problem is called approximate dynamic programming (ADP). The simulator has been built in Matlab using a basic physical model and is currently running on four CPU cores. The program is using the historical data to predict the best way internally control the energy transfers for driving. Although we are ignoring many losses and making some assumptions, the time for fully depleting the stored energy under optimal conditions gives a rough estimate of the driving range. (The state of charge over time is what is shown in current visualizations of the trip, currently when fed with random driving data.)
We are porting the code to Python to leverage Tensorflow. This is necessary as the RL "loss" (value) function must be approximated in order to reduce compute, as it is currently running for 50 "epochs" with 10,000+ "training examples" during the learning process and cannot be parallelized. No common standards or toolkits were used in developing this agent, as this is primarily an academic project with not much other work done using the ADP optimizer.
Technologies Used
Python (TF-Eager, TensorBoard), Matlab.
Repository
https://github.com/AlokD123/Hybrid-Storage_Project