Method of Predicting SOH and RUL of Lithium-Ion Battery Based on the Combination of LSTM and GPR

Jiahui Zhao, Yong Zhu, Bin Zhang, Mingyi Liu, Jianxin Wang, Chenghao Liu, Yuanyuan ZhangSustainability 2022 · DOI

TL;DR

This paper introduces a novel hybrid approach combining Long Short-Term Memory (LSTM) networks and Gaussian Process Regression (GPR) to accurately predict the State of Health (SOH) and Remaining Useful Life (RUL) of lithium-ion batteries. By extracting key health factors from charging curves and leveraging LSTM for trend prediction, the method achieves high accuracy in forecasting battery degradation.

What Problem Does This Paper Address?

Ensuring the reliable operation of lithium-ion batteries is crucial across numerous applications, from electric vehicles to grid storage. A key challenge is accurately predicting their State of Health (SOH) and Remaining Useful Life (RUL). These metrics are not directly measurable and are influenced by a complex interplay of factors, making prediction difficult. Existing methods often struggle with the non-linear degradation patterns, capacity regeneration phenomena, and the inherent correlation between SOH and RUL, leading to reduced prediction accuracy, especially in the later stages of a battery's life. This paper aims to address these limitations by proposing a method that jointly estimates SOH and RUL, accounting for their interdependencies and the dynamic nature of battery aging.

Methods

The core of this research lies in a hybrid data-driven approach that combines two powerful machine learning techniques:

  • Health Factor (HF) Extraction and Selection: The method begins by extracting six "health factors" from the charging current and voltage curves of the battery. These factors are designed to capture indicators of battery degradation. To identify the most relevant HFs, the Maximum Information Coefficient (MIC) is employed. MIC is a statistical measure that quantifies the dependency between two variables, capable of detecting non-linear relationships. HFs with an MIC value greater than 0.8 were selected for further analysis, ensuring they strongly correlate with battery capacity.
  • Long Short-Term Memory (LSTM) for Trend Prediction: LSTMs are a type of recurrent neural network particularly adept at learning from sequential data and capturing long-term dependencies. In this study, LSTM is used to predict the future trends of the selected health factors over increasing charge/discharge cycles. This allows for forecasting how these indicators of battery health will evolve.
  • Gaussian Process Regression (GPR) for SOH and RUL Estimation: GPR is a non-parametric, probabilistic approach that provides uncertainty estimates along with predictions. It's well-suited for regression tasks with potentially limited data and complex relationships.
  • SOH Prediction: The predicted health factors from LSTM are used as input to a GPR model to estimate the battery's State of Health (SOH).
  • RUL Prediction: Subsequently, the predicted health factors and the estimated SOH are fed into another GPR model to predict the Remaining Useful Life (RUL). This sequential approach leverages the strengths of both LSTM (trend forecasting) and GPR (probabilistic regression and uncertainty quantification).

Datasets

The study utilized four publicly available lithium-ion battery datasets from NASA [31]. These datasets were collected under controlled experimental conditions and are commonly used for battery aging research. The paper notes that these datasets represent batteries with different depths of charge, allowing for verification of the algorithm's adaptability.

Key Insights and Ideas

  • Synergistic Power of LSTM and GPR: The paper demonstrates that combining LSTM's ability to capture temporal trends with GPR's robust regression and uncertainty estimation capabilities leads to superior SOH and RUL predictions compared to using either method alone.
  • Importance of Health Factor Selection: The use of MIC for selecting the most informative health factors is crucial. It moves beyond simple correlation and identifies features that truly capture the underlying degradation mechanisms, leading to more accurate models.
  • Joint Estimation Advantage: By jointly estimating SOH and RUL, the method implicitly accounts for their strong correlation, which is a significant improvement over methods that estimate them independently. This holistic approach leads to more reliable predictions.
  • Handling Capacity Regeneration: The paper acknowledges capacity regeneration phenomena and suggests that the proposed method, by focusing on health factor trends, can better navigate these fluctuations than direct capacity prediction methods.

Key Results

The proposed hybrid LSTM-GPR method demonstrated high accuracy and reliability in predicting SOH and RUL across four different battery datasets.

  • SOH Prediction: The GPR model, using LSTM-predicted HFs, achieved high-precision SOH predictions. The Mean Absolute Error (MAE) for SOH prediction was consistently low, less than 0.01 across all tested batteries. The relative error for SOH estimation was within 1% when using an 80% training set.
  • RUL Prediction: The RUL prediction results also showed good agreement with experimental data. The relative prediction error for RUL was generally within ±0.5% to ±1% for the tested batteries. The MAE for RUL prediction was less than one cycle for most batteries.

Limitations and Future Work

  • Limitations:
    • The study acknowledges that while the LSTM prediction of HFs followed the general trend, there was still some deviation from experimental values in the late stage of battery recession.
  • The performance of the model is dependent on the quality and quantity of the training data.
  • Future Work:
    • The paper could benefit from exploring a wider range of health factors or more advanced feature extraction techniques.
  • Investigating the impact of varying operating conditions (e.g., temperature, charge/discharge rates) on the extracted health factors and the overall prediction accuracy would be valuable.
  • Further research into optimizing the hyperparameters of both LSTM and GPR models could potentially yield even better results.

Summary updated

Abstract

The state of health and remaining useful life of lithium-ion batteries are important indicators to ensure the reliable operation of these batteries. However, because they cannot be directly measured and are affected by many factors, they are difficult to predict. This paper presents method of jointly predicting state of health and RUL based on the long short-term memory neural network and Gaussian process regression. This method extracts the batteries’ health factors from the charging curve, selects health factors with more relevance than the setting standard as the characteristic of capacity by the maximum information coefficient method, and establishes the battery aging and remaining useful life prediction models with Gaussian process regression. On this basis, the long short-term memory neural network is used to predict the trend of the change in health factors with the increase in cycles, and the results are input into a Gaussian process regression aging model to predict the state of health. Taking the health factors and state of health as the characteristics of remaining useful battery life, a battery remaining useful life model based on Gaussian process regression is established, and the change trend in the remaining useful life can be obtained by inputting the predicted health factors and state of health. In this study, four battery data sets with different depths of charge were used to verify the accuracy and adaptability of the algorithm. The results show that the proposed algorithm has high accuracy and reliability.
BibTeX
@inproceedings{zhao2022method,
  author = {Jiahui Zhao and Yong Zhu and Bin Zhang and Mingyi Liu and Jianxin Wang and Chenghao Liu and Yuanyuan Zhang},
  title = {{Method of Predicting SOH and RUL of Lithium-Ion Battery Based on the Combination of LSTM and GPR}},
  booktitle = {Sustainability},
  year = {2022},
  doi = {10.3390/su141911865},
}
Figure