About

This is only a basic prototype of an application that tries to predict the range that the EV can travel depending upon certain parameters. The application uses different tensorflow models to continuosly capture necessary data from the sensors in the EV and make predictions for SOC, SOE, and finally the range of the EV depending upon the current scenario/ environment of the vehicle.

The application uses three tensorflow models


1. SOC Predictor

This model focuses on predicting the SOC (State of Charge) of the battery from the battey's voltage (V), current (A), and temperature (degree Celcius) data.

For training the model, we have used the drive cycle data from the Panasonic 18650PF Li-ion Battery dataset. Various charging and discharging cycle tests were performed on the lithium-ion battery, at different temperatures and rates. We are interested in the drive cycle data.The drive cycle power profile is calculated for an electric Ford F150 truck with a 35kWh battery pack scaled for a single 18650PF cell.

For training the model for SOC estimation, we have used the drive cycle data available at 25℃, taking into consideration, the real life temperature scenario of driving an EV. The required data was extracted out, it was preprocessed, outliers were removed, and it was normalized before using for training the tensorflow model. The neural network architecture used in the model is as follows:


2. SOE Predictor

This model builds upon the SOC model. It focuses on predicting the SOE (State of Energy) of the battery from the voltage (V), current (A), temperature (degree Celcius), and the predicted value of SOC.

For training the model, we have used the same dataset as we did for SOC model. The preprocessed data along with the new predicted value for SOC, is used as an input to the neural network. We need the SOE for calculating the amount of energy (kWh) remaining in the battery for range prediction.The neural network architecture used in the model is as follows:
Hyperparameters used for both models Value/ Type
Activation function for hidden layers RELU
Activation function for output layer Linear
Technique used to prevent overfitting Early Stopping
Optimizer Adam
Loss Function Mean Squared Error
Epochs 700
Batch Size 512

For more information click here




3. Range Predictor

This model predicts the range that the EV can travel depending upon the features like amount of battery energy remaining (kWh), type of route, status of A/C or heater in the car, type of tyres used, the driving style of the driver, and average speed of the car.

For training the model, we have a used a dataset that contains real-life driving data of two Volkswagen e-Golf cars, with year of manufacture as 2014 and 2016 respectively. The data is available at the Spritmonitor website.

1. Volkswagen e-Golf, year 2014, 85 kW (116 PS)
2. Volkswagen e-Golf, year 2016, 85kW (116 PS)
The data was scrapped using a python crawler (vehicle_crawler.py) available here .

The file includes data about 3615 trips with a total travel distance of around 152167 kilometers. The data was preprocessed, necessary features were one-hot encoded and standardized before being used to train the model.

Hyperparameters used for both models Value/ Type
Number of hidden layers 2
Number of nodes in each hidden layer 32
Activation function for hidden layers RELU
Activation function for output layer Linear
Technique used to prevent overfitting Early Stopping
Optimizer RMSprop
Loss Function Mean Absolute Error
Epochs 1000
Batch Size 16

For more information click here


Note: Due to the unavailability of a physical testing and data collection environment, the application will take in values of the required features from a predefined csv file that contains data for SOC & SOE estimation. A few features required for range prediction, needs to be provided by the user on the next page.