[Paper Review] State of Health Estimation of Lithium-Ion Batteries in Vehicle-to-Grid Applications Using Recurrent Neural Networks for Learning the Impact of Degradation Stress Factors

Kotub Uddin, J. Edward Schofield+1 more|arXiv (Cornell University)|May 16, 2022

Advanced Battery Technologies Research4 citations

TL;DR

This paper proposes a multivariate long short-term memory (LSTM) recurrent neural network (RNN) to estimate lithium-ion battery state of health (SoH) in vehicle-to-grid applications by learning long-term dependencies on degradation stress factors—temperature, state of charge, depth of discharge, and C-rate. With over 1.15 years of training data, the model achieves <5% mean square error in SoH prediction, demonstrating robust performance for real-time degradation monitoring.

ABSTRACT

This work presents an effective state of health indicator to indicate lithium-ion battery degradation based on a long short-term memory (LSTM) recurrent neural network (RNN) coupled with a sliding-window. The developed LSTM RNN is able to capture the underlying long-term dependencies of degraded cell capacity on battery degradation stress factors. The learning performance was robust when there was sufficient training data, with an error of &lt; 5% if more than 1.15 years worth of data was supplied for training.

Motivation & Objective

• To address the path dependency challenge in lithium-ion battery state of health (SoH) estimation, which is not well captured by traditional model-based methods.
• To improve economic viability of vehicle-to-grid (V2G) services by accurately predicting SoH while accounting for degradation stress factors such as temperature, SoC, DoD, and C-rate.
• To develop a data-driven approach using recurrent neural networks that can learn long-term dependencies in battery degradation data more effectively than kernel methods or classical RNNs.
• To validate the model’s performance across varying training-to-testing data ratios, identifying the minimum data requirement for sub-5% prediction error.
• To explore future enhancements such as transfer learning and structured neural networks to reduce data requirements and improve model interpretability.

Proposed method

• The study employs a long short-term memory (LSTM) RNN with a sliding-window approach to process multivariate time series data including cycle number, current, terminal voltage, and temperature as inputs.
• The model is trained on the NASA Ames Prognostics Center’s Randomized Battery Usage Data Set, excluding datasets with anomalous temperature behavior (RW2, RW3, RW18).
• Adam optimizer is used for training due to its efficiency and suitability for noisy, sparse gradients common in grid-based applications.
• The input features include cell current, terminal voltage, temperature, and cycle number; terminal voltage is used as a proxy for electrochemical overpotentials influencing degradation.
• The model is evaluated across multiple training/test splits (7:3, 5:5, 4:6, 3:7) to assess sensitivity to data quantity and identify minimum data requirements.
• The approach avoids relying on open-circuit voltage (Voc) as a direct SoC proxy, instead using terminal voltage with recognition of its nonlinear relationship to SoC due to iR and polarization effects.

Experimental results

Research questions

• RQ1 Can an LSTM RNN effectively learn long-term dependencies of battery SoH on multiple degradation stress factors (T, SoC, DoD, C-rate) in a multivariate time series setting?
• RQ2 What is the minimum amount of training data required to achieve <5% mean square error in SoH prediction for V2G applications?
• RQ3 How does model performance degrade when training data is limited, and what training-to-test ratio ensures robust generalization?
• RQ4 Can terminal voltage serve as a reliable input for SoH estimation despite its nonlinear relationship to open-circuit voltage and SoC?
• RQ5 To what extent can transfer learning or structured neural networks reduce the data requirements for training accurate SoH prediction models?

Key findings

• The LSTM RNN achieves a prediction error of less than 5% mean square error when trained on more than 1.15 years of data, indicating high accuracy under sufficient data conditions.
• Model performance degrades significantly with less training data, with the 7:3 training-to-testing split showing the best performance among all tested ratios.
• The model's accuracy is highly sensitive to training data length, highlighting a critical data requirement for reliable SoH estimation in real-world V2G systems.
• The use of terminal voltage as an input is justified despite its nonlinear relationship to SoC, as it captures electrochemical overpotential effects influencing degradation.
• The model outperforms classical RNNs and kernel methods by preserving path dependency through internal memory and gradient flow mechanisms.
• Future work should explore using dSoH/dN as the output to better reflect the accelerating nature of degradation over time.