Machine Learning Approaches for Lithium-Ion Battery Health Parameters Estimation

Lithium-ion batteries (LIBs) have become a focus of research interest for electric vehicles (EVs) due to their high volumetric and gravimetric energy storage capability, lower self-discharge rate, and excellent rechargeability coupled with high operational voltage as compared with the lead-acid batteries. This paper presents different machine learning approaches to predict health indicators & usable cycle life of LIBs. Here, we focus on two important battery health indicators i.e., battery discharge capacity and Internal resistance (IR). We used publicly available multi-cycled data of the Lithium Iron Phosphate (LFP), Lithium-Nickel-Manganese-Cobalt-Oxide (NMC) and Lithium Cobalt Oxide (LCO) cells. The approach proposed for predicting health indicators involves using a time-series model in the areas where the actual data i.e., from the Beginning of life (BOL) to the End of life (EOL) is not available. This methodology includes dynamically training a time-series based regression models with the last 100 cycles of information. It includes formulating the equations for individual C-Rates with discharge capacity, and internal resistance of the last ‘T’ cycles as an input to estimate the future discharge capacity and internal resistance after ‘X’ cycles. The accurate results for predicting battery health indicators have been achieved using the concept of dynamic training and timeseries model. This approach helps for quick estimation of battery State of Health (SOH). In the second approach, we have suggested a method for useful cycle life estimation using early cycle data. Here, we extracted battery voltage, current and temperature values of initial 100 cycles for training the model. This method has helped us achieve a minimum RMSE of 8.9 %, showcasing a noteworthy accuracy.