Forecasting Battery Capacity for Second-Life Applications Using Physics-Informed Recurrent Neural Networks

This repository accompanies our published paper

“Forecasting Battery Capacity for Second-Life Applications Using Physics-Informed Recurrent Neural Networks” (https://doi.org/10.1016/j.etran.2025.100432)

It contains code and processed data for training, evaluating, and visualizing a physics-informed recurrent neural network (PI-RNN) alongside several baselines to forecast battery capacity fade in first- and second-life phases. The raw data is available at: https://digitalcommons.lib.uconn.edu/reil_datasets/3/

---

🚀 Project Structure

.
├── processed_data/               # Preprocessed Cycling & RPT datasets
│   ├── Batch1.pkl                # Processed data from Batch 1
│   └── Batch2.pkl                # Processed data from Batch 2
│                                 # (original filenames: Processed_data_Cycling&RPT_Batch1/2_Capacity_Forecasting_merged_update_Jan2025.pkl)
├── saved_models/                 # Trained model weights
│   ├── b3_scenario3.pth          # Baseline RNN trained on scenario 3
│   ├── pi_rnn_scenario3.pth      # PI-RNN trained on scenario 3
│   └── pi_rnn_s3_mc_dropout.pth  # PI-RNN with MC dropout for uncertainty quantification
├── simulated_PBM_data/           # Physics-based simulations & extracted features (top-performing groups in test set)
│   ├── G2_PBM_Simulated.pkl
│   ├── G3_PBM_Simulated.pkl
│   ├── G4_PBM_Simulated.pkl
│   ├── G16_PBM_Simulated.pkl
│   └── G18_PBM_Simulated.pkl
├── data_utils.py                 # Data loaders and input sequence builders
├── models.py                     # PI-RNN, baseline RNNs, and PBM surrogate models
├── pbm_experiments.py            # PyBaMM-based simulation and feature extraction
├── preprocessing.py              # Raw Excel → merged .pkl conversion + capacity fade plotting
├── RMSE_evaluation.py            # RMSE & MAE computation for single/multi-step forecasts
├── training_strategies.py        # Scenario-based training pipelines for PI-RNN and baselines
├── trajectory_forecast.py        # CLI-based visualization of capacity trajectory forecasts
├── uncertainty_quantification.py # UQ: prediction intervals + calibration curve evaluation
└── requirements.txt              # Required Python dependencies

---

📄 Script Overviews

• data_utils.py

  • Centralizes dataset paths and feature/target definitions
  • load_pbm_surrogate(), load_batch(), load_battery_data(), make_sequences()

• models.py

  • train_pbm_surrogate_for_PI_RNN() — PBM surrogate for capacity-drop injection (random forest)
  • CustomRNNCellWithSurrogate & MultiStepPIRNN — physics-informed RNN
  • BaselineMultiStepRNN — standard RNN baseline
  • GPRBaseline — model-based Gaussian process baseline
  • PBMSurrogate — PBM surrogate model class

• pbm_experiments.py

  • Defines PyBaMM experiments for each group
  • Runs cycling/RPT, extracts features, saves to simulated_PBM_data/

• preprocessing.py

  • Parses “Cycling n” Excel files, computes throughput/time duration features
  • Reads RPT capacities from master Excel, merges, saves to processed_data/
  • Plots capacity fade per group

• RMSE_evaluation.py

  • Trains PI-RNN, Baseline RNN, GPR, PBM surrogate
  • Computes and plots single-step and multi-step RMSE/MAE comparisons

• training_strategies.py

  • Implements three forecasting scenarios (S1: fixed horizon, S2: recursive, S3: maximum horizon)
  • Trains and saves PI-RNN and Baseline RNN for each scenario
  • Provides visualization

• trajectory_forecast.py

  • CLI to load pretrained S3 models (PI-RNN & Baseline)
  • (Optional) fine-tuning on first N points of a selected cell
  • Combines RNN & GPR forecasts, plots trajectories & RMSE bars

• uncertainty_quantification.py

  • Loads or trains the S3 PI-RNN model with MC dropout, saves model state
  • Generates convex-combined prediction of original and recalibrated intervals across life phases
  • Fits isotonic recalibration on “C2” cells (held-out calibration set), plots calibration curves for different number of mc dropout samples

---

🔧 Installation

1. Clone

   git clone https://github.com/your-org/PI-RNN-for-Capacity-Forecasting.git
   cd PI-RNN-for-Capacity-Forecasting

2. Create & activate your Python environment

   python3 -m venv venv
   source venv/bin/activate

3. Install dependencies

   pip install -r requirements.txt

---

▶️ Usage Examples

1. Preprocess data & plot capacity fade

   python preprocessing.py

2. Generate PBM-simulated features

   python pbm_experiments.py

3. Train & evaluate RMSE/MAE

   python RMSE_evaluation.py

4. Train forecasting scenarios

   python training_strategies.py

5. Interactive trajectory forecast

   python trajectory_forecast.py --group G13 --cell C1

6. Uncertainty quantification & calibration curves

   python uncertainty_quantification.py

---

📋 Requirements

See requirements.txt for full dependency list.

---

📖 Citation

Please cite our paper if you use this code:

@article{navidi2025forecasting,
  title={Forecasting battery capacity for second-life applications using physics-informed recurrent neural networks},
  author={Navidi, Sina and Bajarunas, Kristupas and Chao, Manuel Arias and Hu, Chao},
  journal={eTransportation},
  pages={100432},
  year={2025},
  publisher={Elsevier}
}