User-Centric Machine Learning-Assisted Non-Destructive Prediction of Lithium-Ion Battery

The growing demand for lithium-ion batteries has naturally heightened interest in critical issues such as performance and stability. Real-time diagnosis of the battery is crucial to ensure its safe usage. However, managing the exponentially increasing data during operation, without disassembling the battery, presents a significant challenge. To address this issue, predicting the complete the complete charge/discharge curve using machine learning methods has been widely adopted. Accurately predicting the battery’s maximum capacity during its cycle provides valuable insights into its state. It helps to optimize battery lifespan and enables more effective energy management based on the specific usage environment.
Recently, many studies have reported on machine learning-based curve prediction models. These models, developed using machine learning techniques, are capable of handling large amounts of data and variables, learning non-linear patterns within the data, and performing real-time prediction. Especially, efforts have been made to predict the cycle index at SoH (State of Health) 80 using only the initial cycle data, or to predict the charging curve utilizing only a partial section of the corresponding cycle. While making predictions using data from the initial cycle or a fraction of the inputs is practical for real-time prediction, this approach has limitations when applied to data from actual battery users. Battery charging/discharging characteristics vary based on factors such as temperature, c-rate, and component materials, making it difficult to collect sufficient data across a wide range of operating conditions. Since studies often rely on cycled data generated under consistent conditions and protocols, this can lead to poor prediction performance when applied to batteries with different chemical compositions or in varied operating environments.
To ensure reliable predictions for battery users, it is important to train models using datasets that reflects real-world usage patterns. In other words, the dataset should represent various states of battery degradation. Battery materials undergo different degradation modes depending on their exposure to specific voltage. By analyzing dQ/dV plot, it can be inferred that phase transitions occur within certain voltage ranges, which are the primary contributors to degradation. While most battery users tend to fully charge their batteries, the discharging patterns vary significantly. Therefore, this work is focuses on discharging patterns by selecting six different cutoff voltages around the peak from the dQ/dV plot to construct a dataset captures key degradation modes. This study utilizes partial voltage range, voltage shifts, and peak drops at the maximum points of dQ/dV peaks compared to the first cycle as features to predict the entire corresponding discharge curve. The data with different discharging cutoff voltages reflect consistent degradation levels, only 20 cycles with different cycle indices are used up to SoH 80 for each cutoff voltage data. This model demonstrated error less than 2% on the test datasets. To validate the model’s performance in a real-life use environment, a dataset cycled with 50 different discharge voltages, representing randomly cycled data, was generated. The prediction performance was confirmed, showing an error less than 4% on the new dataset. This study demonstrates that entire discharge curve, regardless of the degree of degradation. The model’s robustness and potential for practical application in battery management system under real-world scenarios are highlighted. It suggests that further refinement and validation across a broader range of operating conditions will enable more effective implementation to optimize battery performance and extend battery life.