Hugging Face's logo

Dharunkumar9
/
battery-capacity-predictor

Time Series Forecasting
English
Model card Files
xet
 Community
battery-capacity-predictor / README.md
Dharunkumar9's picture
Dharunkumar9
Updated readme with transformer model content
1038d5c verified
preview code
|
raw
history blame contribute delete
2.94 kB
metadata 
license: mit
language:
  - en
metrics:
  - mae
pipeline_tag: time-series-forecasting

Battery health prediction using machine learning models

This repository contains transformer model for predicting battery capacity / remaining useful life (RUL).

Project description

Developed deep learning models based on Transformer architectures to predict Li-ion battery capacity degradation and estimate remaining useful life using time-series charge–discharge data. The codebase is implemented using python utilizing pytorch, leveraging voltage, current, and temperature trends with feature engineering and sliding time windows for temporal modeling. The Transformer achieved the best prediction accuracy, while GRU and CNN–LSTM provided efficient alternatives for smaller datasets. The work highlights effective data-driven approaches for battery health monitoring, critical for electric vehicles and energy storage systems. The plot of the capacity of all batteries used is shown below. Figure 1: Capacity degradation trends across cycles for all batteries.

Repository structure

  • transformer.ipynb — Contains code for data processing, model architecture, training and evaluation for transformer based model.
  • artifacts-v1.zip - Contains model artifacts to use for inference.

Requirements

Install core Python packages used by the notebooks:

pip install numpy pandas scipy scikit-learn matplotlib torch transformers

Typical workflow

  • Load raw .mat from dataset/ for transformer model.
  • Extract sequences with utility functions involving extracting capacity.
  • Build sliding windows using build_instances.
  • Train models in the notebooks; Models are saved in saved_models folder.

Results

The table containing performance comparison across different deep learning models is shown below.

Model MAE RMSE MAPE
Basic LSTM 0.0442 0.0522 0.0299
Basic GRU 0.0234 0.0430 0.0154
CNN + LSTM 0.0313 0.0478 0.0206
Transformer 0.0281 0.0349 0.0186

and the plots of predicted vs actual capacity is shown below. Figure 2: Predicted vs Actual capacity of B0018 for LSTM-CNN, GRU, LSTM based models

Figure 3: Predicted vs Actual capacity of B0018 for transformer based based models

Future work

  • Integrate uncertainty quantification to improve capacity prediction reliability.
  • Expand dataset diversity and explore domain adaptation to improve generalization.

References

https://ieee-dataport.org/documents/lithium-ion-battery-data-set

Authors

  • Dharunkumar Senthilkumar
  • Dhruvkumar Patel