Lithium-ion battery state of health estimation using machine learning algorithms
Ivanov, Yury (2025)
Diplomityö
2025
School of Energy Systems, Sähkötekniikka
Kaikki oikeudet pidätetään.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi-fe2025062473104
https://urn.fi/URN:NBN:fi-fe2025062473104
Tiivistelmä
Efficient and accurate State of Health (SOH) estimation is essential for optimal lithium-ion battery usage and it has to be performed continuously. The specific focus of this study is on a fast SOH estimation technique within a narrow voltage interval. A fully controlled charging phase is selected for taking measurements. This allows taking the measurements quickly and accurately during almost every charging event.
In this research, data is estimated from eight datasets provided by Joint Research Centre (JRC) and Aalto University. Measurements are conducted across four different current rates at four different temperatures The voltage and current signals are monitored for the whole period from a brand new cell to the end-of-life state. Voltage curves are analysed across various conditions. Patterns are examined across a narrow voltage interval and compared against the total SOH. Factors such as duration of measurement and initial State of Charge (SOC) of the cell are considered. The voltage interval is chosen to suit the widest possible range of charging protocols and ambient conditions.
Incremental capacity analysis is applied to determine the optimal voltage interval position and width, which represents the overall SOH of the battery with the strongest correlation. After noise filtering and formatting, the voltage-time curves are converted into uniform one-dimensional time vectors and combined with the current and temperature for automated processing. Among various machine learning algorithms, a Long-Short Term Memory (LSTM) neural network is found the most suitable due to its ability to handle short time sequences of adjacent time curves. The model achieves a decent level of accuracy, with the worst-case RMSE <1% .
In this research, data is estimated from eight datasets provided by Joint Research Centre (JRC) and Aalto University. Measurements are conducted across four different current rates at four different temperatures The voltage and current signals are monitored for the whole period from a brand new cell to the end-of-life state. Voltage curves are analysed across various conditions. Patterns are examined across a narrow voltage interval and compared against the total SOH. Factors such as duration of measurement and initial State of Charge (SOC) of the cell are considered. The voltage interval is chosen to suit the widest possible range of charging protocols and ambient conditions.
Incremental capacity analysis is applied to determine the optimal voltage interval position and width, which represents the overall SOH of the battery with the strongest correlation. After noise filtering and formatting, the voltage-time curves are converted into uniform one-dimensional time vectors and combined with the current and temperature for automated processing. Among various machine learning algorithms, a Long-Short Term Memory (LSTM) neural network is found the most suitable due to its ability to handle short time sequences of adjacent time curves. The model achieves a decent level of accuracy, with the worst-case RMSE <1% .