CALCE CS2 Battery State of Health (SOH) Prediction

Predicting lithium-ion battery State of Health (SOH) from easily-measured per-cycle signals (voltage, current, internal resistance, cycle number), using the CALCE CS2 battery cycling dataset. Built and validated with Leave-One-Battery-Out (LOBO) cross-validation rather than a single train/test split, since only six battery cells are available.

What this project does

1. Parses 152 raw CALCE cycling files (six battery cells: CS2-33 through CS2-38) into a clean, one-row-per-cycle dataset.
2. Applies several correctness checks during preprocessing:
   • Validates that the assumed discharge step (Step_Index == 7) is actually discharging (negative current) before trusting it.
   • Computes discharge capacity via trapezoidal integration of current over time, rather than trusting the file's own precomputed capacity column.
   • Uses the mean of each battery's first 3 valid cycles as the reference ("100% SOH") capacity, instead of a single, noisier first cycle.
   • Flags (and separately logs, without silently discarding) isolated capacity spikes and near-zero end-of-life cycles for manual review.
3. Trains three models — Linear Regression, Random Forest, Gradient Boosting — to predict SOH from cycle_number, C_rate, average_voltage_V, average_current_A, and internal_resistance_ohm.
4. Evaluates all three with LOBO cross-validation (6 folds, one per battery) as the primary metric, since a single fixed split is a high-variance estimate with only 6 batteries available.

Results summary

Model	LOBO MAE (%)	LOBO RMSE (%)	LOBO R²
Gradient Boosting	3.41	5.06	0.926
Linear Regression	3.51	5.00	0.925
Random Forest	3.74	5.78	0.893

Average discharge voltage is the single most important feature in both tree-based models, consistent with known battery electrochemistry (voltage sag during discharge tracks internal degradation). Full methodology, per-battery breakdowns, and a feature ablation study are in the write-up (report/main.pdf / report/main.tex).

Repository structure

.
├── CALCE_CS2_Battery_SOH_ML_Project_v2.ipynb   # main analysis notebook
├── data/
│   └── CALCE_CS2_raw/                          # raw cycling files (not tracked — see below)
├── outputs/
│   ├── CALCE_CS2_cycle_level_dataset.csv
│   ├── CALCE_CS2_ML_SOH_Dataset.csv
│   ├── CALCE_CS2_LOBO_CV_summary.csv
│   └── CALCE_CS2_Gradient_Boosting_SOH_Model.pkl
├── report/
│   ├── main.tex
│   ├── main.pdf
│   └── figures/
├── requirements.txt
└── README.md

Data

Raw cycling data comes from the CALCE Battery Research Group (University of Maryland), cells CS2-33 through CS2-38. The raw .xlsx files are not included in this repository — they're large (hundreds of MB total) and freely available from CALCE directly. Download them and place the extracted files under data/CALCE_CS2_raw/ before running the notebook (the notebook's EXTRACT_DIR path expects this layout).

Running it

pip install -r requirements.txt
jupyter notebook CALCE_CS2_Battery_SOH_ML_Project_v2.ipynb

Run with Restart Kernel & Run All — later cells depend on state from earlier ones. Expect the discharge-step validation and LOBO CV summary printed near the top and middle of the notebook to be the numbers worth trusting; the single fixed-split results later in the notebook are kept only for the illustrative plots and are explicitly not the headline performance claim.

Requirements

• Python 3.9+
• pandas, numpy, matplotlib, scikit-learn, openpyxl, joblib

License

Analysis code in this repository is provided as-is for academic/coursework purposes. The underlying CALCE CS2 dataset is © CALCE Battery Research Group, University of Maryland — refer to their site for data usage terms.

Author

Ranit Das

M.Tech - Solid Mechanics and Design

IIT Kanpur