Predictive Modeling of Lithium-Ion Battery Degradation Using Supervised Machine Learning Algorithms

Abstract

As the number of EVs and renewable energy storage systems grows, the need for dependable battery management systems (BMS) becomes more critical. Accurately predicting the State of Health (SOH) and Remaining Useful Life (RUL) of Lithium-Ion (Li-ion) batteries is crucial to ensure operational safety, prevent catastrophic failures, and optimize maintenance schedules. The conventional degradation assessment approaches are, however, very complex and highly non-linear, which complicates the assessment of degradation using conventional approaches. Thie research question that requires investigation: What is the best machine learning or deep learning architecture to be used to extract these non-linear patterns and to be able to make accurate, scalable, and interpretable battery health predictions?

In response to this, recent studies have focused on data-driven approaches based on sophisticated computational models that can map battery degradation, avoiding the need to solve complex electrochemical equations. Classical tree-based ensemble models like Random Forest (RF) or Extreme Gradient Boosting (XGBoost) are widely used for their high performance and relatively low computational requirements, whereas sequential deep learning models like Long Short-Term Memory (LSTM) networks have theoretically better potential to capture temporal dynamics among continuous time-series data. The use of explainable artificial intelligence (XAI) systems like Shapley Additive explanations (SHAP) provides a valuable

This study compares and tests the performance of these two different architectural archetypes, based on a thorough toolkit of data-driven metrics developed from open, publicly available data sets from NASA and Stanford. Work on feature engineering involved extracting multi-cycle dynamic characteristics, and the models were rigorously evaluated on Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE). The comparative analysis shows that the sequential models based on LSTM work well for extracting long-term temporal dynamics between continuous-time signals, but XGBoost is able to achieve a high level of accuracy with much less computational load.