Dynaformer: A Deep Learning Model for Ageing-aware Battery Discharge Prediction

Implementation, data and pretrained models for the paper "Dynaformer: A Deep Learning Model for Ageing-aware Battery Discharge Prediction"

How to get started

• Clone the repository: git clone
• Create a virtual environment: python -m venv venv # Note, we used python 3.6 for this project on Ubuntu 18.04
• Activate the virtual environment: source venv/bin/activate
• Install dependencies: pip install -r requirements.txt
• Install the dynaformer package: pip3 install -e src/
• Download the model and the test data python scripts/download_model_and_test_data.py # This will download the weights of Dynaformer, and the synethetic variable test data

Demo

• Visualize the deme of the model (inference) via streamlit run visualization/visualize_predictions.py

Getting the synthetic data

Generate a synthetic training dataset

In order to generate a synthetic training dataset via python3 scripts/generate_dataset.py. We use Hydra as configuration tool. The configuration file is configs/generate_dataset.yaml. In general, if you want to generate synthetic constant training dataset similar the one in the paragraph of the paper "Performance Evaluation on Constant Load Profiles" you can use the following command:

python3 scripts/generate_dataset.py current.current_type=constant_currents current.N_profiles=1 N_currents=50

Instead if you want to generate a synthetic variable training dataset, similar the one in the paragraph of the paper "Performance Evaluation on Variable Load Profiles" you can use the following command:

python3 scripts/generate_dataset.py current.current_type=variable_currents current.N_profiles=6 N_currents=1000

Please take a look at the configuration file if you want to modify something more in specific.

The generated dataset is saved in data/variable_currents or data/constant_currents depending on the option current.current_type

Getting the real data

Training

How to train the model

• Train the Dynaformer model via the following command:

python3 scripts/train.py method=dynaformer data_dir=data/variable_currents/2022-04-27/14-58-12/data method.batch_size=12

If you want to train the model on a different dataset, you can change the data_dir parameter.

TODO

• [X] Add a demo of the model
• [ ] Add a demo of the model on real data
• [X] Add training dataset
• [ ] Add training/testing dataset generation
• [X] Add training pipeline
• [ ] Add baseline models

Additional information

System Specification

All the experiments were done with Python 3.6 with pytorch 1.9.0+cu111 on Ubuntu 18.04.