Battery State Estimation and Diagnosis

Model-Based SOC Estimation & Data-Driven RUL Prediction for Lithium-Ion Batteries

A two-part laboratory project contrasting the two dominant paradigms in battery management. Part I builds a physics-based equivalent-circuit model of a battery pack and estimates its State of Charge (SOC) online with a Kalman filter. Part II trains an LSTM on real cycling data to forecast battery health and predict Remaining Useful Life (RUL).

Authors: Group laboratory project — Franz Wendel and three fellow students (co-authors' names withheld for privacy) Institution: TU Berlin — Chair of Electrical Energy Storage Technology (Fachgebiet Elektrische Energiespeichertechnik)

📄 The full lab reports are in reports/.

---

Overview

Part	Problem	Approach	Tools
I	State of Charge — how full is the cell now?	Equivalent-circuit model + Kalman filter	MATLAB / Simulink
II	Remaining Useful Life — how many cycles are left?	LSTM trained on degradation data	Python / PyTorch

Group Work I — Model-Based SOC Estimation

A 36 V / 10 Ah pack (10s4p, 40 × LG 18650HE4 cells, 360 Wh) is modelled with a first-order Thévenin equivalent circuit — ohmic R0, one R1–C1 branch, and an SOC-dependent OCV(SOC) — cast into state-space form and implemented in Simulink with SOC-driven lookup tables. The model is discretised and used by a discrete-time Kalman filter (predict / update) that estimates SOC from terminal voltage and current. Initialised with a deliberately wrong guess (10 % vs. a true 1 %), the filter converges smoothly to the coulomb-counting reference and tracks it without drift.

📁 01_model_based_soc_estimation/ — Simulink model, parameter script, cell-parameter data.

Group Work II — Data-Driven RUL Prediction

An LSTM forecasts the SOH-vs-cycle degradation curve and, from it, the RUL. Data is the public NASA Li-ion battery dataset (B0006/B0007 train, B0005 validation, B0018 test); SOH = Capacity / 2.0 Ah, with end-of-life at SOH = 0.7. The model is a PyTorch nn.LSTMCell with a sigmoid head (predictions stay in [0, 1]), trained with teacher forcing and run autoregressively at inference. A grid search gave a best configuration of lr 1e-3, 200 epochs, hidden size 128. RUL error shrinks as more history is supplied:

History (cycles)	10	20	30	40	50	60	70	80	90
RUL error (cycles)	12	11	10	8	7	8	7	5	3

📁 02_data_driven_rul_prediction/ — preprocessing.ipynb, lstm.ipynb, raw + preprocessed data, and an example trained checkpoint.

Repository contents

Path	Description
reports/	The two full lab reports (PDF).
01_model_based_soc_estimation/	Simulink model, parameter script and cell data for the ECM + Kalman-filter SOC estimator.
02_data_driven_rul_prediction/	Notebooks, NASA dataset (raw + preprocessed) and a trained model for the LSTM SOH/RUL predictor.

Authors & acknowledgements

A group laboratory project; its results are shared among all members. Only the publishing author is named — the other co-authors' names are omitted for privacy.

Franz Wendel · (three further co-authors — names withheld) · Prof. Dr.-Ing. Julia Kowal (professor)

Carried out at the Chair of Electrical Energy Storage Technology, TU Berlin. The Group Work II data is the publicly available NASA Ames PCoE lithium-ion battery dataset.

License

All rights reserved — published to document an academic project; no reuse, redistribution or modification without the authors' prior written permission. The NASA dataset under 02_data_driven_rul_prediction/data/ remains the property of its original publisher.