Battery Cycle Life Prediction Using Policy-Based Feature Sets and Supervised Machine Learning
Al-Shehri, Wejdan Ahmed, Arbaz Kittaneh, Omar Abdulmajid, Mohammed
Al-Shehri, Wejdan
Ahmed, Arbaz
Kittaneh, Omar
Abdulmajid, Mohammed
Type
Supervisor
Date
2025-12-6
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Accurate prediction of lithium-ion battery cycle life is vital for enhancing performance, reducing testing time, and accelerating deployment in energy storage systems. This study introduces a data-driven framework for battery lifetime forecasting using supervised machine learning techniques. A benchmark dataset containing 124 commercial lithium iron phosphate (LFP)/graphite cells—each cycled under varied charge–discharge protocols—was analyzed. To simulate practical diagnostic scenarios, six policy-specific early-cycle feature sets (F1–F6) were utilized, where each policy includes a different number and combination of charging and performance parameters reflecting incremental diagnostic information. Seven supervised regression models were evaluated using 5-fold cross-validation with the Mean Absolute Percentage Error (MAPE) as the primary performance metric. Among all models, Random Forest demonstrated the most stable and consistent accuracy across all six policy scenarios. Although K-Nearest Neighbors (KNN) achieved a comparable average MAPE (0.2243), Random Forest achieved the lowest average Testing MAPE (0.2201) and maintained higher stability and generalization capability. These results demonstrate that early-cycle data, structured through policy-based feature configurations, can reliably predict battery cycle life, providing a scalable, efficient, and cost-effective alternative to traditional long-term degradation testing. Unlike previous work such as the MIT early-stage forecasting study, this framework extends the analysis through policy-based feature grouping and multi-model comparison to assess generalization performance.
Department
Publisher
Sponsor
NA