AI Extends EV Battery Life

By Muhammad OsamaReviewed by Joel ScanlonNov 18 2024

Revolutionizing EV technology: Discover how a machine learning-driven approach unlocks precise battery health insights, ensuring longer lifespans and sustainable energy solutions.

Research: Domain knowledge-guided machine learning framework for state of health estimation in Lithium-ion batteries. Image Credit: IM Imagery / Shutterstock

In a paper recently published in the journal Communications Engineering, researchers presented a novel domain knowledge-guided machine learning framework for accurately estimating the state of health (SOH) of lithium-ion batteries (LIBs) in electric vehicles (EVs). Using readily available data from real-world EV operations, they developed a robust and efficient SOH estimation method that achieves high accuracy even with limited training data. The goal was to address the need for accurate battery health monitoring, which is essential for advancing EV technologies and reducing environmental impact.

Advancements in Electric Mobility Technologies

The global shift toward electrified mobility is largely motivated by the urgent need to mitigate climate change, as transportation contributes approximately 12% of global emissions. Both the Intergovernmental Panel on Climate Change (IPCC) and the International Energy Agency (IEA) emphasize the importance of clean transportation solutions.

LIBs have emerged as the leading energy storage option for EVs due to their high energy density, enabling longer driving ranges than other battery types. However, effective monitoring of LIB health is essential to ensure longevity and efficiency, highlighting the necessity for reliable methods to estimate their SOH and predict their remaining useful life (RUL) to build consumer trust in EV technology. The paper also notes the challenges of using current estimation methods in real-world conditions, emphasizing the need for solutions that perform reliably with partial and noisy data.

A Novel Framework for SOH Estimation

In this paper, the authors focused on designing and evaluating multiple SOH indicators based on empirical data and domain knowledge. They developed a machine learning framework incorporating five significant indicators: power autocorrelation, internal resistance, charging impedance, energy during charging, and energy during discharging. These indicators were derived from experimental datasets that simulate real-world EV driving conditions, enhancing their applicability in practical scenarios.

The researchers employed a comprehensive machine-learning pipeline comprising several vital steps to achieve their objectives. Initially, they extracted SOH indicators from the experimental data, followed by regression analysis to determine the relationship between these indicators and battery capacity. They then trained a linear regression model (LRM) to estimate capacity fade based on the selected indicators. The dataset included Nickel Manganese Cobalt (NMC)/Graphite cells subjected to various cycling conditions over 30 months, with periodic reference performance tests conducted to evaluate battery aging.

The charging impedance indicator was also refined by calculating it within an optimized voltage window, which improved accuracy, particularly in partial charging scenarios. Additionally, the power autocorrelation indicator was noted for its potential in diagnostic applications, even though its real-time implementation may be limited in dynamic driving conditions. A windowed approach was also applied to assess energy consumption during charging, thereby enhancing the effectiveness of energy as a battery health indicator.

Key Findings and Insights

The analysis revealed strong correlations between the SOH indicators and capacity loss, indicating that specific measurable changes occur in battery performance metrics as they age. Notably, the power autocorrelation function emerged as a reliable predictor of capacity fading, exhibiting a strong linear correlation with capacity loss.

The internal resistance of the batteries increased with aging, correlating with reduced power delivery due to higher Joule losses. However, the paper highlights that resistance values must be measured under consistent conditions to ensure accuracy, as factors like temperature and C-rate can influence results. Importantly, these indicators are adaptable across various battery chemistries and do not rely on data like total aging cycles or ampere-hour throughput, which reduces inaccuracies stemming from sensor errors or limited data.

The charging impedance indicator also demonstrated a clear relationship with capacity loss, with values increasing as the cells aged. Energy metrics during both charging and discharging were highly indicative of battery health, reflecting internal cell degradation processes. The models trained with these indicators achieved absolute percentage errors in capacity estimation consistently below 2.5%, with the power autocorrelation feature achieving errors below 1.5%. Furthermore, the study examined the sensitivity of charging energy across different voltage windows, confirming the robustness of energy-based indicators, even with partial charging profiles.

Practical Applications and Real-World Impact

This research has significant implications for advancing the EV industry and energy storage solutions. The proposed framework could be integrated into battery management systems (BMS) to enable real-time battery health monitoring, supporting proactive maintenance strategies. By accurately estimating the SOH and RUL of LIBs, both manufacturers and users can optimize battery performance, enhance safety, and extend battery life.

The findings suggest that integrating these indicators into BMS could improve decision-making regarding battery usage and maintenance schedules. It can also transform energy storage management across various applications beyond EVs, including renewable energy storage and consumer electronics. Nevertheless, the authors caution that further testing with diverse datasets, including those reflecting field conditions, is necessary to validate the framework's adaptability and accuracy.

Conclusion and Future Directions

In summary, the domain knowledge-guided SOH indicators proved effective for accurately estimating capacity fade in LIBs. Specifically, power autocorrelation, energy during charging, and energy during discharging provided highly accurate results, even with limited training data. Integrating these indicators into machine learning models improved the accuracy of battery health assessments and facilitated practical applications in real-world scenarios.

The study acknowledges limitations, including the lack of temperature variation analysis and the use of controlled datasets that may not fully reflect real-world conditions. Future work could explore the addition of more SOH indicators, the use of advanced machine-learning models, and the extension of the framework to different battery chemistries and operational conditions. Overall, this research lays the groundwork for improved BMS, which could enhance EV performance, safety, and sustainability.