AI-Driven Battery Management System for Electric Vehicles Using Real-Time State-of-Charge and State-of-Health Estimation

Abstract

Accurate estimation of state of charge (SOC) and state of health (SOH) is fundamental to the safety, range prediction, power management, and lifetime optimization of electric-vehicle batteries. Neither state is directly measurable; both must be inferred from terminal measurements such as voltage, current, and temperature. Conventional coulomb counting is computationally simple but accumulates current-sensor and initialization errors, while model-based observers depend on parameter identification and may lose accuracy as the battery ages. This paper presents an AI-driven battery-management-system architecture that performs coupled, real-time-capable SOC and SOH estimation using two lightweight supervised regression models. The estimators consume causal features derived from voltage, current, temperature, elapsed time, cumulative ampere-hours, and voltage slope. A health-aware supervisory layer uses the resulting states to constrain charge, discharge, thermal, and balancing commands. A proof-of-concept evaluation was conducted on the NASA Ames lithium-ion battery-aging dataset. Cells B0005, B0006, and B0007 were used for training, while B0018 was held out entirely for testing to assess cell-to-cell generalization. On 34,866 held-out samples, the proposed estimator achieved an SOC mean absolute error (MAE) of 2.45 percentage points, SOC root-mean-square error (RMSE) of 3.39 percentage points, and coefficient of determination of 0.988. The corresponding SOH MAE and RMSE were 4.11 and 4.65 percentage points. A fixed-rated-capacity coulomb-counting baseline produced an SOC MAE of 11.19 percentage points on the same held-out cell, illustrating the importance of aging-aware estimation. Measured desktop inference time was approximately 7 microseconds per sample for each estimator, establishing computational feasibility but not replacing an embedded-controller timing certification. The results support lightweight AI as a practical correction and fusion layer for battery-state estimation, while also identifying cross-chemistry validation, uncertainty quantification, and hardware-in-the-loop testing as necessary next steps.