This is the Pytorch implementation of MAGNet in the paper: Forecasting battery degradation trajectory under domain shift with domain generalization.

@article{TAN2024103725,
title = {Forecasting battery degradation trajectory under domain shift with domain generalization},
journal = {Energy Storage Materials},
pages = {103725},
year = {2024},
issn = {2405-8297},
doi = {https://doi.org/10.1016/j.ensm.2024.103725},
url = {https://www.sciencedirect.com/science/article/pii/S2405829724005518},
author = {Ruifeng Tan and Xibin Lu and Minhao Cheng and Jia Li and Jiaqiang Huang and Tong-Yi Zhang}
}

Key dependencies:

• Ubuntu 20.04

• Python 3.8.13

• CUDA 11.1

• torch == 1.9.0

• matplotlib == 3.5.3

• jupyter == 1.0.0

• numpy == 1.23.4

• pandas == 1.4.4

• Wasserstein == 1.1.0

Dependencies can be installed using the following command:

pip install -r reqruiements.txt

To help the reproducibility, we have provided the trained model weights and other necessary files in the Google drive, thus you can easily see the main results in this paper. First of all, you need to download the checkpoints, dataset, detailed_results, fig5 and results folders from the Google drive and then put them under the project root folder. After that, you can reproduce the results using the following scripts:

• evaluate_model.py: You can evaluate the model performance using this Python script.

• plot_fig3_MAE_and_RMSE.ipynb: This notebook gives the draft fig. 3 of this paper.

• visualize_enc_new_conditionLevel.py: You can visualize the embeddings on the NCA dataset using this Python script.

• visualize_enc_new_conditionLevel_NE.py: You can visualize the embeddings on the LFP dataset using this Python script.

• Compare_PCA_and_mechanism_NCA.py: Compute the superiority matrix.

The datasets and pretrained model weights are released at Google Drive: https://drive.google.com/drive/folders/1iHu32xrkbNOzqbQM9qH-RXMdxxzj3TZm?usp=sharing

Note that the directory names for pretrained MAGNet weights are as follows:

• NCA dataset: NCA_meta_Informer_Batteries_cycle_SLMove_lr1e-06_metalr0.005_mavg15_ftM_sl10_ll10_pl150_dm8_nh4_el1_dl2_df2_fc4_fc21_ebCycle_dtFalse_valratio0.33_test_lossawmse_vallossnw_dp0.0_bs32_wd0_mb0.25_agamma0.25_lradjtype4_0

• LFP dataset: NE_meta_Informer_Batteries_cycle_SLMove_lr1e-06_metalr0.0075_mavg15_ftM_sl20_ll20_pl500_dm12_nh4_el2_dl2_df4_fc5_fc21_ebCycle_dtFalse_valratio0.5_test_lossawmse_vallossnw_dp0.0_bs128_wd0_mb2_agamma0.2_lradjtype4_0

• NCM dataset: NCM_meta_Informer_Batteries_cycle_SLMove_lr1e-06_metalr0.0075_mavg15_ftM_sl10_ll10_pl150_dm12_nh4_el2_dl2_df2_fc4_fc21_ebCycle_dtFalse_valratio0.33_test_lossawmse_vallossnw_dp0.0_bs32_wd0_mb2_agamma0.2_lradjtype4_0

To reproduce the results, the same dependencies and core computing devices should be used. You can install the dependencies according to the Requirements. Additionally, you need a Ubuntu 20.04 machine with GTX 3090 GPUs.

Data processing Initially, you need to process the raw data to get the processed dataset. For the NCA dataset, you need to download the NCA_data_process from the Google Drive. After installing the dependencies shown in the requirements.txt, you can run the following scripts in order to get the processed dataset used in this work:

1. abnormal_utils.py
2. 2022NC_cycle_provider.py

For the LFP dataset, you need to download the LFP_data_process from the Google Drive. After installing the dependencies shown in the requirements.txt, you can run the following scripts in order to get the processed dataset used in this work:

1. BuildPkl_Batch1.ipynb; BuildPkl_Batch2.ipynb; BuildPkl_Batch3.ipynb
2. Load Data.ipynb
3. utils.py
4. process_data_Autoformer.ipynb

For the NCM dataset, you can download the raw data from this link. After that, you can process the dataset using process_NCM_dataset.ipynb notebook.

Then, you can put the processed datasets under the dataset folder.

Train the model

You can reproduce the results in this paper using the same hyperparameters listed in the supplementary information. The code for training models on the NCA, NCM and LFP datasets are run_meta_weighted_NCA.py, run_meta_weighted_NCM.py and run_meta_weighted_NE.py respectively. The training might cost 20 hours to finish in the LFP dataset.

# Example
# The following instruction runs the code in background and saves the ouput in the log.out file.
nohup python -u run_meta_weighted_NCA.py &> ./log.out&