Model-Based Gaussian Process Regression (GPR) for Battery Capacity Forecasting

This repository provides a model-based Gaussian process regression (GPR) framework for forecasting the capacity fade of lithium-ion batteries proposed by Richardson et al. The method combines an empirical model (logistic, exponential, or linear degradation) with a GPR trained on residuals, enabling robust and uncertainty-aware forecasts of battery capacity.

---

Framework Overview

---

Key Features

• Empirical models supported:

  • Reverse Logistic
  • Exponential
  • Linear

• Gaussian Process Regression (with Matern + White kernels) for residual learning

• Normalized and raw capacity forecasting modes

• Built-in error metrics: MAE, RMSE, MAPE, scaled MAPE, and end-of-life MAPE

• Handles different battery chemistries: LFP, NMC

---

Datasets

This repository relies on publicly available datasets:

• NMC (RWTH Aachen University) Dataset: RWTH Aachen

• LFP (MIT / Stanford / Toyota Research Institute) Dataset: MIT Battery Dataset on Matrío

---

Example Usage

Online Training and Forecasting

# Initialize model for LFP dataset
model_LFP = ModelBasedGPR(
    initial_points=300,
    empirical_model_type='reverse_logistic',
    normalize_Q=True,              
    cycle_col='cycle_index',
    capacity_col='discharge_capacity',
    cell_id_col='cellID'
)

# Plot selected cells
model_LFP.plot_all_cells(
    all_datasets['LFP'],
    group_name='test_LFP',
    cell_ids_to_plot=[1,2,3,4,5,6,7,8,9,10]
)